BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Drupal iCal API//EN X-WR-CALNAME:Events iCal Start May 2024 X-WR-TIMEZONE:America/New_York BEGIN:VTIMEZONE TZID:America/New_York X-LIC-LOCATION:America/New_York BEGIN:DAYLIGHT TZNAME:EDT TZOFFSETFROM:-0500 TZOFFSETTO:-0400 DTSTART:20260308T070000 END:DAYLIGHT BEGIN:DAYLIGHT TZNAME:EDT TZOFFSETFROM:-0500 TZOFFSETTO:-0400 DTSTART:20250309T070000 END:DAYLIGHT BEGIN:DAYLIGHT TZNAME:EDT TZOFFSETFROM:-0500 TZOFFSETTO:-0400 DTSTART:20240310T070000 END:DAYLIGHT BEGIN:STANDARD TZNAME:EST TZOFFSETFROM:-0400 TZOFFSETTO:-0500 DTSTART:20251102T060000 END:STANDARD BEGIN:STANDARD TZNAME:EST TZOFFSETFROM:-0400 TZOFFSETTO:-0500 DTSTART:20241103T060000 END:STANDARD BEGIN:STANDARD TZNAME:EST TZOFFSETFROM:-0400 TZOFFSETTO:-0500 DTSTART:20231105T060000 END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:6a09ef91119df DTSTART;TZID=America/New_York:20260515T143000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260515T153000 LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Pulkit Agrawal CLASS:PUBLIC DESCRIPTION:Speaker: PULKIT AGRAWAL\, Associate Professor\, Department of E lectrical\nEngineering and Computer Science\, Massachusetts Institute of\n Technology\, and Co-founder\, Eka Robotics\n\nTalk Title: What’s Missing in Embodied Agents: Force Intelligence\nand Lifelong Learning\n\nModern r obots can plan sophisticated motions\, yet they remain slow\,\nbrittle\, a nd unreliable on tasks humans find effortless. The missing\npiece is not b etter planning\, but better force reasoning: knowing\nwhen\, where\, and h ow much force to apply under uncertainty and across\ndiverse tasks. Force intelligence\, I argue\, is a unifying principle\nfor scalable robotics— bridging dexterous manipulation and whole-body\ncontrol. However\, even a force-aware robot that cannot learn from its\nown experience will remain b rittle. Today’s systems are effectively\nfrozen after training\, unable to adapt once deployed. Real-world\nautonomy instead demands learning in d eployment: the ability to\nimprove continuously from interactions\, failur es\, and successes. In\nthis talk\, I will present our lab’s recent work on lifelong learning\nand outline a future path for how combining it with force-centric\ndesign could enable reliable\, useful robots in the real w orld.\n\n—\n\nPulkit Agrawal is an Associate Professor in the Department of\nElectrical Engineering and Computer Science at MIT and a co-founder o f\nEka Robotics. He earned his Ph.D. from UC Berkeley and a bachelor’s\n from IIT Kanpur\, where he was awarded the Director’s Gold Medal. His\nw ork has received multiple Best Paper Awards\, the IEEE Early Career\nAward in Robotics and Automation\, the IROS Toshio Fukuda Young\nProfessional A ward\, the IIT Kanpur Young Alumnus Award\, the Sony\,\nSalesforce\, and A mazon Research Awards\, the Signatures Fellow Award\,\nand the Fulbright S cience and Technology Award. He previously\nco-founded SafelyYou Inc. \n\ n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9112d45 DTSTART;TZID=America/New_York:20260515T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260515T143000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Thesis Proposal - Timothy Kim CLASS:PUBLIC DESCRIPTION:Speaker: TIMOTHY KIM\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Efficient Data Storage Pr ovisioning\, Placement\, and\nTransitions at Scale\n\nExascale storage sys tems are under increasing pressure to store more\ndata while providing gre ater performance per byte. Historically\,\nhyperscale storage systems have relied on a two-tiered hierarchy:\nhard-disk drives store most bytes\, wh ile smaller flash tiers absorb\nrequests for hot and performance-critical data. This design is\nbecoming increasingly difficult to sustain. Datacent er data is getting\nwarmer with the proliferation of AI/ML and analytics-h eavy workloads\,\nwhile storage devices are becoming denser without propor tional\nimprovements in per-byte performance or endurance. As a result\,\n enabling denser storage devices at exascale requires improving both\nthe s oftware storage system and the hardware provisioning strategies\nfor moder n datacenter workloads.\n\nThis thesis shows that exascale storage systems can enable denser\nstorage options by jointly reducing internal IO and im proving data\nplacement/hardware provisioning decisions. The first part of this\nthesis\, Morph\, reduces IO associated with lifetime redundancy\ntr ansitions. Morph introduces a novel hybrid redundancy scheme for\nearly-li fe data and a system designed around a new erasure-code for\nlate-life\, r educing ingest and transcode overheads. The second part of\nthis thesis de velops a total-cost-of-ownership (TCO) model and\noptimizer for exascale s torage provisioning. This model determines the\nminimum-cost set of hardwa re necessary to serve datacenter workloads\nand shows how heterogeneous co nfigurations can cost-effectively enable\ndense devices in modern datacent ers.\n\nWe propose work that connects these two directions by modeling sto rage\nworkloads from the bottom up\, using fine-grained lifetime and\ntemp erature transition behavior to reason about provisioning and\nplacement de cisions. By understanding how data cools and survives\nthroughout its life time\, this work aims to produce a more precise\nframework that co-optimiz es the heterogeneous storage mixture and the\nplacement of data across the storage tiers. Together\, these\ncontributions show how storage systems c an cost-effectively service\ncontemporary storage workloads with denser me dia and enable the\nmassive growth in data demand.\n\nThesis Committee\n\n Greg Ganger (Co-Chair)\n\nRashmi Vinayak (Co-Chair)\n\nGeorge Amvrosiadis\ n\nSaurabh Kadekodi (Google)\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9113312 DTSTART;TZID=America/New_York:20260514T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260514T120000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:PoP Seminar - Liron Cohen CLASS:PUBLIC DESCRIPTION:Speaker: LIRON COHEN\, Associate Professor\, Institute for the Theory of\nComputing\, Faculty of Computer and Information Science\, Ben-G urion\nUniversity\n\nTalk Title: First-Class Effects: A Unified Effectful Type Theory for\nContinuity\, Choice\, and SearchThursday\, May 14\, 2026\ , 11am – 12pm\nProgramming languages usually treat effects with cautio n: state\,\nfailure\, nondeterminism\, choice\, and interaction complicate reasoning\nand threaten metatheoretic properties. In this talk\, I will a rgue that\neffects are not merely impurities to be controlled\, but first- class\nsemantic capabilities that can reveal and shape the logic of type\n theory. I will present boxTT\, a unified effectful type-theoretic\nframewo rk in which modalities and choice operators make effects part\nof the stru cture of the theory itself. This lets us ask a foundational\nquestion: whi ch logical principles become valid when a type theory has\ncertain computa tional effects and a certain discipline for controlling\nthem? As case stu dies\, I will show how stateful computations can\ninternalize continuity b y computing moduli of continuity for\nhigher-order programs over infinite streams\, and how effectful models\ncan separate variants of Markov’s pr inciple that correspond to\ndifferent strengths of search and witness extr action. Ultimately\, the\naim is to move beyond isolated effectful models toward foundational\ntype theories with native effects\, where constructiv e principles can\nbe systematically validated\, separated\, and transporte d according to\nthe computational capabilities available.—Liron Cohen is an\nassociate professor in the Faculty of Computer and Information Scienc e\nat Ben-Gurion University and is currently visiting Cornell University.\ nHer research lies at the intersection of programming languages\, type\nth eory\, constructive mathematics\, and proof assistant formalization\,\nwhe re she develops formal frameworks for reasoning about programs and\nproofs with the broader aim of building foundations for reliable\nsoftware and f ormalized mathematics. Her recent work focuses on\nbuilding semantic and t ype-theoretic foundations that broaden the\nnotion of computation\, especi ally by making computational effects such\nas state\, choice\, nondetermin ism\, and control part of the foundations\nof type theory. Her work has ap peared in leading venues including\nJACM\, LICS\, and POPL\, has been reco gnized by awards such as the BSF\nPazy Memorial Research Award\, and inclu des recent verification work\nthat received TACAS 2025 Distinguished Paper and Distinguished\nArtifact Awards. She was a Fulbright postdoctoral rese archer at\nCornell University and received her PhD and MSc from Tel Aviv\n University.Faculty Host:  Stephanie Balzer Event Type: Seminars Room\nNu mber: In Person Building: Traffic21 Classroom\, Gates Hillman 6501\nSpea ker's Name: LIRON COHEN Speaker Website: www.lironcohenlab.com\n[http:// www.lironcohenlab.com] Speaker's Professional\nTitle: Associate Professor \, Institute for the Theory of Computing\nFaculty of Computer and Informat ion Science\, Ben-Gurion University\nTalk Title: First-Class Effects: A U nified Effectful Type Theory for\nContinuity\, Choice\, and Search For Mor e\nInformation: mstanle2@andrew.cmu.edu Affiliations: Computer Science\n Department (CSD) Organization(s): School of Computer Science Event\nWebsi te Title: Series Website Event Website URL: www.cs.cmu.edu\n[http://www. cs.cmu.edu]…\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911395b DTSTART;TZID=America/New_York:20260513T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260513T113000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Thesis Proposal - Sophia Roshal CLASS:PUBLIC DESCRIPTION:Speaker: SOPHIA ROSHAL\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Adjoint Types for Funct ional Programming\n\nSubstructural type systems offer a principled foundat ion for reasoning\nabout resource usage by controlling the structural rule s of weakening\,\ncontraction\, and exchange. Linear\, affine\, strict\, a nd ordered types\neach enable distinct optimizations and semantic guarante es. However\,\nsystems restricted to a single structural discipline are of ten too\nrigid for practical programming\, limiting expressiveness and\npr eventing programs from simultaneously exploiting multiple structural\nprop erties.\n\nThis thesis develops Adjoint Types\, a logically founded substr uctural\ntype system based on Adjoint Logic\, that supports the principled \ncombination of multiple substructural modes within a single language.\nT his design preserves the practical expressiveness of an unrestricted\nlang uage\, while still permitting mode based optimizations on the\npieces of t he program that allow them.\n\nThe proposal proceeds in four parts. First\ , we present Adjoint Natural\nDeduction\, including both a declarative and algorithmic type system\,\nand establish soundness and completeness with respect to a sequent\ncalculus formulation. Second\, we develop a mode inf erence procedure\nthat supports mode polymorphism and eliminates the need for manual\nmode annotations. Third\, we extend the framework with ordered modes\,\nestablishing decidability for a corresponding type system\, whil e\noutlining ongoing work toward practical type-checking algorithms.\nFour th\, we propose generalized pattern matching in the adjoint\nsetting\, dev eloping a logically sound approach to nested patterns via\nfocusing and ou tlining extensions to support wildcards and ordered\nmodes.\n\nTogether\, these contributions establish ordered adjoint types as a\nlogical foundati on for substructural functional programming languages\nand logical framewo rks\, enabling structural properties to drive\ncompiler optimizations and intensional program behavior without\nsacrificing practical usability. \n \nThesis Committee:\n\nFrank Pfenning (Chair) \n\nJan Hoffmann\n\nJonatha n Aldrich\n\nBrigitte Pientka (McGill University)\n\nChris Martens (Northe astern University) \n\nAdditional Information\n\nIn-person and Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9113e37 DTSTART;TZID=America/New_York:20260508T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260508T170000 URL:https://csd.cmu.edu/academics/doctoral-resources/dsr-schedule LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Student Review (DSR) General Meeting CLASS:PUBLIC DESCRIPTION:Talk Title: Doctoral Student Review (DSR) General Meeting - Yea rs 4-n\n\nStudents in years 4-n\, special cases\, LOAs\, ghosts\, and case s referred\nfrom the Thursday area meetings will be reviewed at the Genera l\nMeeting on Friday.\n\nSee email announcement(s).\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911413f DTSTART;TZID=America/New_York:20260513T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260513T150000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:PoP Seminar - Brian Milnes CLASS:PUBLIC DESCRIPTION:Speaker: BRIAN G. MILNESTalk Title: APAS-VERUS: AI Paired Proof \nEngineering Techniques and Experience\n\nThis talk focuses on my experie nces using LLM AIs to generate Rust\ncode for Acar and Blelloch's 'Algorit hms Parallel and Sequential'\ntextbook. I implemented all of its 121 algor ithm variants\, 81 of which\nare parallel\, and then verified them in the Verus formal verification\nsystem.   This work puts to rest the question of: can current AI\nLLMS be used to prove practical Industrial-grade\, pe rformant\nalgorithms. \n\nI'll discuss my experiences\, the time\, the co sts\, and the software\nengineering discipline developed using programmati c techniques to\ndrive down the agentic coding and proof costs. I'll revie w Rust's\nstrengths and weaknesses\, discuss Verus's strengths and weaknes ses and\nreview AI agents and their interfaces. \n\nFrom this work I am a ble to provide three metrics\, which I call the\nCPR$: - C\, a reasonable costs per lines estimate of code in agentic\nprogramming a textbook\,  - P\, and a reasonable costs per lines of\nproof while proving that code and learning Verus.\n\nThe bottleneck in agentic programming and proving is n ow the time to\nreview the code and proofs. I derive an exact measure\, R\ , of how\nproven code reduces the cost of programmer review. \n\n—\n\nI have done two years of industrial AI programming languages\ndevelopment\, followed by seven years of AI research at Carnegie Mellon\nSchool of Comp uter Science (SOAR). Then two years of programming\nresearch in programmin g languages focusing on their use in systems\nprogramming and networking ( Fox Project).\n\nI was then drafted into the Internet boom and directed gr oups working\nin systems\, performance\, operations and architecture for L ycos\, Amazon\nand Zillow. (M.Sc. Computer Science\, University of Washing ton\; B.Sc.\nMathematics (Computer Science)\, Carnegie Mellon )\n\nIn Per son and Zoom Participation. See announcement.\n\nPasscode: 1TRrEL\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91145da DTSTART;TZID=America/New_York:20260512T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260512T183000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Thesis Proposal - Gabriele Oliaro CLASS:PUBLIC DESCRIPTION:Speaker: GABRIELE OLIARO\, Ph.D. Student\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Rethinking the AI Sta ck: Systems for Agents and Agents for\nSystems\n\nLarge language models ha ve become central infrastructure for modern AI\napplications\, but running them efficiently remains a major systems\nchallenge. Larger and more capa ble models require more GPU\ncomputation\, GPUs remain expensive\, and sop histicated applications\nsuch as agents require low latency and predictabl e service-level\nobjectives to be practically usable. At the same time\, i nference\noptimization increasingly depends on deep specialized expertise in\nscheduling\, memory management\, and GPU kernel engineering. This\nexp ertise is difficult to scale because model architectures and\naccelerator platforms evolve rapidly\, introducing new operators\,\nprecision formats\ , parallelization patterns\, and hardware-specific\noptimization requireme nts.\n\nThis thesis develops systems that use model-driven and agentic\nte chniques to automate parts of this optimization process. SpecInfer\nuses s maller language models to speculate on the outputs of larger\nmodels\, con verting otherwise serial autoregressive decoding into\nparallel verificati on. SuffixDecoding extends this idea to agentic\nworkloads by caching and reusing prior generation patterns to\nspeculate with minimal GPU overhead. FlexLLM automates fine-grained\nresource allocation between latency-criti cal inference and\nthroughput-oriented finetuning\, allowing both services to share GPUs\nwhile preserving inference service-level objectives. The r emaining\nwork moves from optimizing inference systems with AI techniques to\nusing AI agents to optimize the systems themselves. FastKernels\nprovi des a production-faithful benchmark for LLM-based GPU kernel\nagents\, and the proposed kernel agent will use compiler feedback\,\ncorrectness tests \, runtime measurements\, and hardware feedback to\ngenerate optimized ker nels for rapidly evolving operators such as\nlinear attention\, state-spac e models\, mixture-of-experts routing\,\nquantized inference\, and multimo dal fusion.\n\nThesis Committee\n\nZhihao Jia (Chair)\n\nTianqi Chen\n\nPh illip Gibbons\n\nHao Zhang (University of California\, San Diego)\n\nAddit ional Information\n\nIn-person & Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9114a9b DTSTART;TZID=America/New_York:20260513T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260513T163000 LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Varun Sundar CLASS:PUBLIC DESCRIPTION:Speaker: VARUN SUNDAR\, Ph.D. Student\, Computer Sciences Depar tment\,\nSchool of Computer\, Data &amp\; Information SciencesUniversity o f\nWisconsin-Madison\n\nTalk Title: Quanta Perception as Probabilistic Eve nts\n\nAutonomous systems ultimately rely on extracting information from\n light\, yet remain brittle in extreme environments\, from nighttime\nnavig ation to high-speed robotics. This limitation stems from a\nclassical imag ing abstraction: conventional sensors integrate photon\nflux over fixed ex posure windows\, imposing trade-offs between\nsensitivity\, dynamic range\ , and temporal resolution that degrade\nperception when photons are scarce or dynamics are rapid. Emerging\nquanta (single-photon) image sensors ove rcome these limits by\ndetecting individual photons\, but they generate ph oton streams that\nexceed the compute and latency budgets of real-time sys tems by orders\nof magnitude.\n\nHere we introduce probabilistic events\, a computational primitive for\nreal-time quanta perception at the limit o f individual photons. By\ncomputing the posterior distribution over the ti me since the last\nabrupt intensity change\, we represent photon streams a s recursively\ncomputed belief states. Rather than the binary\, fixed-thre shold\ntriggers of event cameras\, this recursive Bayesian formulation yie lds\nthree simultaneous\, low-latency signals: motion-adaptive scene flux\ ,\nhigh-fidelity activity maps\, and an entropy measure quantifying\nperce ptual uncertainty. This representation enables perception in\nextreme cond itions\, including detecting and estimating the pose of a\nrunning person at ~0.05 lux illumination—without retraining standard\nvision models. Ou r approach sustains input throughputs exceeding\n50\,000 quanta frames per second on commodity GPU hardware—four to\nfive orders of magnitude fast er than state-of-the-art quanta\nreconstruction baselines—yielding kiloh ertz-scale outputs even for\nmegapixel arrays. By replacing frame reconstr uction with direct\nprobabilistic inference over photon streams\, this wor k enables\nreal-time perception at the photon limit and bridges photon-cou nting\nquanta sensing with practical robotic vision.\n\n—\n\nVarun Sunda r is a graduate student at the University of\nWisconsin–Madison\, pursui ng a Ph.D. in computer science. At\nUW–Madison\, he is advised by Prof. Mohit Gupta\, where he focuses on\nsingle-photon imaging techniques. His w ork has been published at\nvenues such as CVPR\, ICCV\, and SIGGRAPH\, and has included live demos\nat ICCP 2023\, CVPR 2024 and SIGGRAPH 2024 (whic h won the best-in-show\naward in the Emerging Technologies track). In 202 6\, he was awarded\nthe Ivanisevic Award at UW–Madison\, which recognize s outstanding\ncomputer science dissertators. He previously received a ba chelor’s\ndegree in electrical engineering from the Indian Institute of\ nTechnology\, Madras in 2020.\n\nThe VASC seminar is generously sponsored by HeyGen \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911504b DTSTART;TZID=America/New_York:20260430T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260430T133000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:MSCS Thesis Presentation - Hunter Rhoades CLASS:PUBLIC DESCRIPTION:Speaker: HUNTER RHOADES\, Master's Student\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Virtual Reality for Drones: Assessing Physical Actuation\nvia Idealized Sensor Feedback\n\nCo mputer vision-enabled autonomous drone flight creates challenges\nwith rep roducibility and behavior assessment due to inherent\nvariability in the m odels that provide cognitive function to the\nvehicle. These challenges ar e most visible when attempting to validate\ncontrol logic in the automatio n pipeline and when benchmarking for the\npurpose of making objective asse ssments on agility and\nmaneuverability. In this work we begin by explorin g the current state\nof simulation technology and its particular applicati ons in the\ndevelopment and fielding of autonomous drone systems\, exposin g a\nsudden jump from simulation-heavy prototyping to end-to-end full\nsys tem testing involving physical flight that increases the difficulty\nof va lidating correctness and incurs unnecessary risk. We present the\nconcept of a SensorTwin\, a novel application of simulation technology\nthat seeks to reduce the identified gap between early phase software\nand hardware-i n-the-loop style simulation and late phase digital twin\nstyle simulation by isolating the visual feed and cognitive engine\ncomponents of a physica l vehicle-automation pipeline system. We\nprovide a concrete implementatio n of a SensorTwin built over the\nSteelEagle Autonomous Drone system and d emonstrate proof of concept by\nevaluating the performance of the SensorTw in-augmented system on\nexisting cognitive task evaluations. We also intro duce a novel set of\nagility evaluation tests\, referred to as the Autonom ous Agility\nBenchmark\, with the intent of enabling future relative compa rison\nacross vehicle platforms\, autonomy pipelines\, and cognitive funct ion\nimplementations. An initial dataset for the benchmark suite is\nprodu ced through physical flight augmented by the SensorTwin\nimplementation an d analyzed to provide initial objective and\nsubjective insights on agilit y assessment specific to the automated\ndrone setting.\n\nThesis Committee \n\nMahadev Satyanarayanan (Chair)\n\nBabu Pillai\n\nMihir Bala\n\nAdditio nal Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9115597 DTSTART;TZID=America/New_York:20260504T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260504T140000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Speaking Skills Talk - Matias Scharager CLASS:PUBLIC DESCRIPTION:Speaker: MATIAS SCHARAGER\, Ph.D. Student\, Computer Science De partment\,\nCarnegie Mellon University\n\nTalk Title: Types of States in S eparation Logic\n\nThe store passing translation allows us to provide func tional types to\nimpredicative operations. We build an intuitive understan ding of the\ntype of stores and references in this setting. We discuss\nge neralizations to these type definitions and their relevance to\nseparation logic.\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRe quirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911592a DTSTART;TZID=America/New_York:20260504T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260504T123000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Speaking Skills Talk - Siddharth Jayashankar CLASS:PUBLIC DESCRIPTION:Speaker: SIDDHARTH JAYASHANKAR\, Ph.D. Student\, Computer Scien ce\nDepartment\, Carnegie Mellon University\n\nTalk Title: A Multi-GPU Fra mework For TB-Scale Encrypted Inference\n\nFully Homomorphic Encryption (F HE) is a transformative cryptographic\nprimitive that enables direct compu tation on encrypted data\, offering\nrobust security guarantees for sensit ive workloads. By ensuring user\ndata remains encrypted throughout the ent ire execution pipeline\, FHE\nfacilitates secure inference even in the pre sence of untrusted model\nproviders or malicious attackers. However\, desp ite this potential\,\npractical adoption has been hindered by massive comp utational\noverheads. While ASICs have been proposed to mitigate these cos ts\,\nthey rely on advanced semiconductor manufacturing and high-end\npack aging\, rendering them currently impractical for widespread\nproduction.\n \nIn this talk\, I will present my work on Cardamom\, an FHE framework\nth at achieves ASIC-competitive performance using commodity datacenter\nGPUs. Cardamom introduces a holistic approach to optimization across\nthe compu te\, memory\, and communication stacks to automatically\ngenerate high-per formance implementations for FHE programs—ranging\nfrom small-scale CNNs to Large Language Models (LLMs) like Llama3-8B.\nOur evaluation demonstra tes that Cardamom outperforms expert-optimized\nlibraries by up to 2.25 ti mes and achieves a milestone in FHE\nperformance: the first sub-10ms CKKS bootstrapping at 7.5ms.\nFurthermore\, Cardamom enables end-to-end encrypt ed inference for\nBERT-Base in 8 seconds and Llama3-8B in 43 seconds\, rep resenting a\nsignificant leap toward the practical deployment of secure\,\ nlarge-scale AI.\n\nPresented in Partial Fulfillment of the CSD Speaking S kills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9115e0c DTSTART;TZID=America/New_York:20260429T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260429T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Meghal Gupta CLASS:PUBLIC DESCRIPTION:Speaker: MEGHAL GUPTA\, Ph.D. Student\, Department of Electrica l\nEngineering and Computer Sciences\, University of California\, Berkeley \n\nTalk Title: Optimal Quantile Estimation for Streams\n\nEstimating qua ntiles is one of the most basic problems in data\nsketching. In this prob lem\, a stream x1\, x2\, x3\, …\, xn of elements\nfrom some universe of size U\, a rank r\, and an accuracy ε are given.\nThe goal is to give a s pace-efficient algorithm that outputs an\nelement with rank between r-εn and r+εn. For example\, this captures\nmedian estimation and 99th percen tile estimation.\n\nIt has long been known that a quantile sketch can be made more\nspace-efficient than storing every element individually (which would\ntake nlogU memory). The previous best algorithms all improved\nsu bstantially on nlogU but did not meet the lower bound of Ω(ε-1·\nlogεn +logεU)) . In this talk\, I’ll describe a\ndeterministic quantile s ketch that uses the optimal\nO(ε-1·(logεn+logεU)) bits of memory.\n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91161b3 DTSTART;TZID=America/New_York:20260505T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260505T113000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Thesis Proposal - Carlos G. Martin CLASS:PUBLIC DESCRIPTION:Speaker: CARLOS G. MARTIN\, Ph.D. Student\, Computer Science De partment\,\nCarnegie Mellon University\n\nTalk Title: Solving infinite gam es with deep multiagent reinforcement\nlearning\n\nIn this thesis\, we stu dy the problem of solving infinite games. Such\ngames can have infinitely many states\, actions\, players\, and steps.\nUnlike mean‑field games\, our players need not be symmetric or\nexchangeable. Furthermore\, we allow such games to have partial\nobservability\, hidden information\, imperfec t recall\, stochastic state\ntransitions\, discontinuous utility functions \, and interdependent\nsocial preferences (i.e.\, matrix-valued discount f actors). Together\,\nthese properties can model a wide range of highly com plex\, real-world\nscenarios that defy traditional game-theoretic solvers. \n\nTo tackle this problem\, we propose a unified framework grounded in\nd eep multiagent reinforcement learning. It includes five core\ncomponents.\ n\nFirst\, it introduces randomized policy networks (RPNs) to model\nobser vation-dependent mixed strategies over infinite action\nspaces.Second\, it represents complex strategy profiles across an\ninfinite continuum of pla yers using player-to-strategy networks\n(P2SNs).Third\, it evolves these r epresentations through a\nshared-parameter simultaneous gradient (SPSG)\, which extends the\nstandard simultaneous gradient to this shared-parameter regime.Fourth\,\nto ensure computational efficiency\, it estimates this g radient using\nrandomized parameter perturbations via a joint-perturbation \nsimultaneous pseudo-gradient (JPSPG).Fifth\, it employs approximate\nexp loitability descent (ApproxED) with learned best-response functions\n(BRFs ).\n\nWe propose to benchmark our approach on a diverse suite of real-worl d\ndomains. These include financial markets\, traffic flow\,\nepidemiologi cal contagion\, energy grids\, and evolutionary ecology.\n\nThesis Committ ee:\n\nTuomas Sandholm (Chair)\n\nVincent Conitzer\n\nFei Fang\n\nIan Gemp (Google)\n\nAdditional Information\n\nIn-person and Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91166c4 DTSTART;TZID=America/New_York:20260504T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260504T110000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Thesis Oral Defense - Nicole Feng CLASS:PUBLIC DESCRIPTION:Speaker: NICOLE FENG\, Ph.D. Candidate\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Algorithms for Generali zed Signed Distance and Winding\nNumbers\n\nIn this talk\, I'll discuss al gorithms for generalized inside/outside\nand signed distance computation. By \"generalized\"\, I mean that these\nalgorithms make geometric inferenc es from imperfect data comprising\nincomplete\, inaccurate\, or ambiguous observations or representations\nof shapes. In other words\, these algorit hms generalize from imperfect\ndata and implicitly approximate the true un derlying curve or surface.\nA theme is that generalization can often be ac hieved by processing\nglobally-defined functions encoding the geometry of interest\, rather\nthan the original\, defective curve or surface. For bot h inside/outside\nand signed distance computation we can unlock further co ntrol over\ngeometry and topology by processing higher-order derivatives o f these\nfunctions. Another theme is that inside/outside and signed distan ce\ncomputation are closely related problems\; towards this end\, I'll\npr ovide a formalization of their relationship that justifies the\ndesign of our algorithms.\n\nThesis Committee: \n\nKeenan Crane (Chair)\n\nNancy Po llard\n\nIoannis Gkioulekas\n\nChris Wojtan (Institute of Science and Tech nology Austria)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9116b1c DTSTART;TZID=America/New_York:20260504T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260504T140000 LOCATION:Newell-Simon Hall SUMMARY:Doctoral Thesis Oral Defense - Arjun Lakshmipathy CLASS:PUBLIC DESCRIPTION:Speaker: ARJUN LAKSHMIPATHY \, Ph.D. Candidate\, Computer Scien ce\nDepartment\, Carnegie Mellon University\n\nTalk Title: Contact Areas f or Dexterous Manipulation and Beyond\n\nHumans use their hands to effortle ssly manipulate objects of\narbitrarily complex geometries and physical pr operties every day\;\nhowever\, adapting these behaviors to dexterous robo ts and virtual\ncharacters is a difficult task. Understanding how humans e xploit\ncontact to perform these manipulations has the potential to greatl y\nadvance progress towards this goal.\n\nUnsurprisingly\, research effort s have analyzed contact in the context\nof dexterous manipulation for deca des. We now have numerous metrics\nfor evaluating grasp quality in terms o f contacts\, sophisticated\nmodels of contact states\, efficient means of computing physically\nsimulated contacts\, and strategies that exploit con tact\ncorrespondences between hands and objects to synthesize grasps and\n manipulations. But the majority of existing works fundamentally\ncharacter ize contact the same way: as points\, lines\, or planes of\ninteraction.\n \nBut contact in the real world is much more complicated. Real bodies\nins tead interface with one another via areas of contact which greatly\nvary w ith the geometries of the contacting surfaces. If we wish to\nmodel the co mplexities of manipulations as they actually occur\, then\nwe must progres s beyond such simplifying assumptions and deal with the\nmessy nature of r eality.\n\nThis thesis aims to do so by presenting foundational frameworks and\nalgorithms for the modeling\, capture\, mutation\, and exploitation of\ncontact areas. Our intention is to establish the foundations necessary \nto elevate contact regions to first-class primitives and demonstrate\nth eir inherent value across a range of practical applications in\ndexterous manipulation and adjacent domains.\n\nFirst\, we introduce three novel con tact area models alongside\noperations supported by each model designed to run on real geometries\nrather than primitive shapes. Next\, using area-b ased primitives\, we\nintroduce: a set of intuitive artist tools for digit ally drafting high\nquality grasps\, a kinematic motion retargeting pipeli ne for dexterous\nmanipulations\, a contact-driven control framework for m ulti-fingered\nhands in physical simulation\, and two practical extensions to\ndifferent domains. We then shift our focus to the real world by\nintr oducing approaches for capturing and reconstructing contact areas\nduring human-object and human-human interactions. Finally\, we present\nan end-to -end system architecture framework for constructing fully\nfunctional robo t systems from contact-rich human demonstrations.\n\nThesis Committee:\n\n Nancy S. Pollard (Chair)\n\nJessica K. Hodgins\n\nKeenan Crane\n\nZackory Erickson\n\nC. Karen Liu (Stanford University)\n\nIn-person and Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911712e DTSTART;TZID=America/New_York:20260428T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260428T120000 LOCATION:Wean Hall 8220 SUMMARY:Mathematics Colloquium CLASS:PUBLIC DESCRIPTION:Speaker: JIANGENG LIU\, J.B. Duke Distinguished Professor\, Dep artment\nof Mathematics\, Duke University\n\nTalk Title: State-space model s through the lens of ensemble control\n\nThe interactions between the fie lds of optimal control and deep\nlearning have been quite fruitful in rece nt years. In this talk\, we\nconsider the state-space models (SSM)\, recen tly emerged effective\narchitectures for sequential modeling\, through the lens of ensemble\ncontrol\, where a shared control law governs a populat ion of\ninput-dependent dynamical systems.\n\nWe derive Pontryagin's maxim um principle (PMP) for this ensemble\ncontrol formulation\, providing nece ssary conditions for optimality.\nMotivated by these conditions\, we intro duce an algorithm based on the\nmethod of successive approximations. We pr ove convergence of this\niterative scheme and establish sufficient conditi ons for global\noptimality. The resulting framework provides a control-th eoretic\nperspective on SSM training.  Joint work with Ye Feng.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91174af DTSTART;TZID=America/New_York:20260428T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260428T163000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:5th Year Master's Thesis Presentation - Gary Gao CLASS:PUBLIC DESCRIPTION:Speaker: GARY GAO\, Master's Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: From Scores to Workflows: An Interactive Framework for\nData Attribution in Foundation Models\n\nTr aining data attribution seeks to identify which training examples\nare mos t related to a model prediction or user query\, yet applying\nattribution methods to modern foundation models practice often brings\ndifficulties to users as such methods could be computationally\nexpensive\, method-specif ic\, and hard to inspect within a single\nanalysis process. This thesis pr esents an interactive framework for\ntraining data attribution that combin es live query computation with\nprecomputed training-side reference repres entations\, enabling users to\nretrieve\, compare\, and interpret ranked t raining examples for live\nqueries or selected validation examples within a unified workflow. \nThe framework centers on gradient-based attribution methods\, including\ngradient similarity\, DataInfstyle influence approxi mation over\nprojected gradient features\, LESS-style low-rank gradient ma tching\,\nand LoGra-style features. It also supports multi-query analysis\ nthrough score aggregation and comparative inspection across methods\,\nal lowing attribution to be analyzed as a workflow rather than as an\nisolate d score. Demonstrations are conducted on public\nquestion-answering datase ts with Pythia-family models\, on a multimodal\ndriving-video setting with Qwen3-VL\, and on a medical-domain setting\nthat extends the framework to private and domain-specific data. This\nthesis shows that training data a ttribution becomes substantially more\nuseful when it is treated not only as a scoring method\, but as a\npractical workflow for querying\, comparin g\, and interpreting training\nexamples across models\, methods\, and data sets.\n\nThesis Committee\n\nChenyan Xiong (Chair)\n\nAlexander Rudnicky\n \nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9117930 DTSTART;TZID=America/New_York:20260430T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260430T163000 LOCATION:Newell-Simon Hall SUMMARY:Doctoral Thesis Proposal - Noah G. Singer CLASS:PUBLIC DESCRIPTION:Speaker: NOAH G. SINGER\, Ph.D. Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Expansion of coset com plexes and applications\n\nIn computer science and discrete mathematics\, a hypergraph is a\ncollection of subsets of a set of vertices\; in particu lar\, a graph is\na collection of (unordered) pairs of vertices. Graphs ar e used to\nencode pairwise relationships between objects\, and hypergraphs to\nencode more general multi-way relationships.\n\nHigh-dimensional expa nders are\, informally\, hypergraphs which are\n\"well-connected\" in a qu antitative sense. While expansion in the\n\"low-dimensional\" case of grap hs has been studied intensively since\nthe 1970s\, high-dimensional expand ers have recently gained visibility\nfor their role in a number of theoret ical breakthroughs\, such as the\nconstruction of efficient sampling algor ithms for spanning trees\n(Anari–Liu–Oveis Gharan–Vinzant\, STOC 201 9\, Ann. Math. 2024)\,\noptimal locally testable codes\n(Dinur–Evra–Li vne–Lubotzky–Mozes\, STOC 2022)\, and efficient\nlow-soundness PCPs (B afna–Minzer–Vyas\, STOC 2025).\n\nKaufman and Oppenheim (STOC 2018\, E ur. J. Comb. 2023) gave a\nremarkably simple construction of high-dimensio nal expanders by using\ncarefully chosen groups of matrices to instantiate a classical\nconstruction known as a \"coset complex\". In this thesis\, we further\ninvestigate the expansion of these complexes and various\ngene ralizations\, as well as applications of this expansion in computer\nscien ce and mathematics.\n\nIn two preliminary works (joint with Ryan O’Donne ll)\, we:\n\nInvestigated a particular notion of high-dimensional expansio n\n(\"1-cosystolic expansion\") on certain generalized KO complexes by\nco mbining intricate group-theoretic arguments with computer-assisted\nmatrix rank calculations.Showed that KO complexes have a\n\"low-soundness agreem ent testing\" property which allows them to be\nused inside the PCP of Baf na et al.\, replacing a much more complicated\nconstruction.\n\nIn this pr oposal\, we propose several additional directions for the\nPh.D. thesis. T hese include:\n\nUsing the KO complex to improve the complexity of the Baf na et al. PCP\n(e.g.\, the verifier or prover runtimes)\,systematically ch aracterizing\nthe coboundary and cosystolic expansion of generalizations o f KO\ncomplexes\; andimproving existing constructions of agreement testers by\nusing alternative kinds of tests.\n\nThesis Committee:\n\nRyan O'Donn ell (Chair)\n\nAayush Jain\n\nYang P. Liu\n\nIrit Dinur (Institute for Adv anced Study)\n\nMadhur Tulsiani (Toyota Technological Institute at Chicago &\nUniversity of Chicago)\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9117e5f DTSTART;TZID=America/New_York:20260430T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260430T120000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Speaking Skills Talk - Yash Savani CLASS:PUBLIC DESCRIPTION:Speaker: YASH SAVANI\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Antidistillation Sampling : Protecting Reasoning Models\nfrom Capability Theft\n\nReasoning models g enerate extended chain-of-thought traces that\nsimultaneously serve users and expose valuable training signals to\npotential competitors. A competit or's model can be fine-tuned on these\ntraces to replicate frontier capabi lities at a fraction of the\noriginal training cost\, while stripping away the safety features that\ntook significant effort to build. Current defen ses\, such as\nsuppressing reasoning traces or blocking suspected users\, impose costs\non legitimate users and undermine safety monitoring. In this talk\, I\nwill present Antidistillation Sampling\, a decoding-time method that\nmodifies the teacher model's token sampling distribution to make it s\noutputs less useful for distillation while preserving their utility\nfo r end users.\n\nPresented in Partial Fulfillment of the Speaking Skills Re quirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91181cf DTSTART;TZID=America/New_York:20260424T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260424T180000 LOCATION:Newell-Simon 4305 SUMMARY:5th Year Master's Thesis Presentation - Allen Zheng CLASS:PUBLIC DESCRIPTION:Speaker: ALLEN ZHENG\, Master's Student\, Computer Science Depa rtment.\nCarnegie Mellon University\n\nTalk Title: Attention Over The Past for Data Efficiency in RL\n\nWe introduce a framework that incorporates a ttention over past states\ninto reinforcement learning (RL) in two complem entary ways. First\, we\nuse an attention mechanism over a trajectory buff er of previously\nvisited states to construct a history-aware critic\, rep lacing the\nstandard neural network critic with an estimate computed as an \nattention-weighted average over stored values. Second\, we replace GAE\, \nwhich is computed through a single trajectory\, with an \nattention-wei ghted advantage  that\, in addition to stepping forward\nthrough time\, a lso steps according to similar states. States more\nsimilar to the current one contribute more to the return estimate\,\nproviding a smooth\, simila rity-weighted alternative to the sequential\nrollout. Together\, these two mechanisms reduce variance by pooling\nsignal across similar states rathe r than relying on a single\ntrajectory.\n\nThesis Committee\n\nGeoff Gordo n (Chair)\n\nJeff Schneider\n\nAdditional Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911858b DTSTART;TZID=America/New_York:20260424T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260424T143000 LOCATION:Tepper Building 4242 SUMMARY:Operations Research Seminar - Hyung-Chan An CLASS:PUBLIC DESCRIPTION:Speaker: HYUNG-CHAN AN\, Associate ProfessorDepartment of Compu ter\nScience and EngineeringYonsei University\n\nTalk Title: Handling LP-R ounding for Hierarchical Clustering and\nFitting Distances by Ultrametrics \n\nIn this talk\, we present an improved approximation algorithm for\nhie rarchical correlation clustering. Given L layers of complete graphs\non a common vertex set V\, where each edge is labeled either + or -\,\nthe prob lem is to compute a clustering of V for each layer so that\nlower-layer cl usterings refine upper-layer ones and the total weighted\nnumber of disagr eements over all layers is minimized. Here\, a + edge\nis counted as a dis agreement if its endpoints are separated\, and a -\nedge if its endpoints are placed in the same cluster. This problem is\nboth a natural generaliza tion of correlation clustering (the case L=1)\nand a formulation of the L1 ultrametric fitting problem arising in\nnumerical taxonomy. Unlike the si ngle-layer setting\, for which\nsubstantial algorithmic progress has been made in recent years\, the\nhierarchical case has seen little progress sin ce the first\nconstant-factor approximation algorithm of Cohen-Addad\, Das \,\nKipouridis\, Parotsidis\, and Thorup. We give a new LP-rounding\nalgor ithm achieving an approximation ratio of 25.8\, significantly\nimproving t he previous guarantee.\n\nJoint work with Mong-Jen Kao\, Changyeol Lee\, a nd Mu-Ting Lee.\n\n—\n\nHyung-Chan An is an Associate Professor in the D epartment of Computer\nScience and Engineering at Yonsei University. His r esearch interests\ninclude approximation algorithms\, online algorithms\, combinatorial\noptimization\, and their applications to engineering proble ms. Before\njoining Yonsei University\, he was a Postdoctoral Researcher a t École\nPolytechnique Fédérale de Lausanne (EPFL). He received his Ph. D. in\nComputer Science from Cornell University and his B.S. in Computer\n Science and Engineering from Seoul National University. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911899b DTSTART;TZID=America/New_York:20260427T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260427T163000 LOCATION:Newell-Simon 4305 SUMMARY:VASC Seminar - Guha Balakrishnan CLASS:PUBLIC DESCRIPTION:Speaker: GUHA BALAKRISHNAN\, Assistant Professor\, Electrical a nd\nComputer Engineering Department\, Rice University\n\nTalk Title: Learn ing Through Fitting: Advancing Non-Pixel\nRepresentations for Visual Infer ence\n\nGridded pixel and voxel representations form the backbone of visua l\ncomputing\, but they struggle to scale efficiently to large\,\nhigh-dim ensional data\, such as volumetric medical scans and complex\nscientific s imulations. Consequently\, continuous\, non gridded models\nsuch as implic it neural representations (INRs) and Gaussian splatting\nhave gained signi ficant research traction over the past five years.\nHowever\, their use ha s largely been confined to signal reconstruction\nrather than acting as fo undational data types for downstream analysis.\nIn this talk\, I will pres ent our recent work on elevating continuous\nmodels beyond mere signal rep resentation. First\, I will discuss how\ninjecting learned priors into INR s via strategic parameter\ninitialization enables powerful new capabilitie s\, including rapid\,\namortized fitting to novel signals and even semanti c segmentation.\nSecond\, I will briefly outline our recent efforts in per forming visual\nrecognition tasks directly on 2D Gaussian image representa tions.\nFinally\, I will highlight interesting future directions in this\n “learning through fitting” paradigm of visual computing.\n\n—\n\nGuh a Balakrishnan is an Assistant Professor in the Electrical and\nComputer E ngineering Department at Rice University. His research group\ntackles a di verse range of problems across computer vision and\nimaging\, with a prima ry focus on developing efficient neural\nrepresentations for complex visua l signals and advancing responsible\nAI through uncertainty estimation and interpretability techniques. He\nfrequently grounds these methods in real -world applications by\ncollaborating with domain experts in scientific di sciplines such as\nmedicine and the geosciences. His scientific contributi ons have been\nrecognized with several honors\, including the NSF CAREER A ward and the\nMICCAI Best Paper Award. Before joining Rice\, he completed his Ph.D.\nat MIT’s Computer Science and Artificial Intelligence Laborat ory\n(CSAIL)\, and earned his undergraduate degrees in Computer Science an d\nComputer Engineering from the University of Michigan\, Ann Arbor.\n\nTh e VASC seminar is generously sponsored by HeyGen \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9118e1e DTSTART;TZID=America/New_York:20260429T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260429T123000 LOCATION:CIC SUMMARY:Doctoral Speaking Skills Talk - Kevin Kuo CLASS:PUBLIC DESCRIPTION:Speaker: KEVIN KUO\, Ph.D. Student\, Computer Science Departmen t\,\nCarnegie Mellon University\n\nTalk Title: Exact Unlearning of Finetun ing Data via Model Merging at\nScale\n\nTo improve trust and adoption of c ollaborative learning systems\, it is\nimportant to provide opt-out guaran tees which enable removal of\nunwanted (e.g. harmful or private) informati on after it has been used\nto train a model. A popular solution to this pr oblem is approximate\nunlearning\, which efficiently updates an LLM so tha t it behaves\n(roughly) as if it was not trained on a subset of data to be gin with.\nHowever\, existing methods are brittle in practice and can easi ly be\nattacked to reveal supposedly unlearned information. To alleviate\n issues with approximate unlearning\, we instead propose SIFT-Masks\n(SIgn- Fixed Tuning-Masks)\, an exact unlearning method based on model\nmerging.\ n\nSIFT-Masks addresses two key limitations of standard model merging:\n(1 ) merging a large number of tasks can severely harm utility\; and (2)\nmet hods that boost utility by sharing extra information across tasks\nmake ex act unlearning prohibitively expensive. SIFT-Masks solves these\nissues by (1) applying local masks to recover task-specific\nperformance\; and (2) constraining finetuning to align with a global\nsign vector as a lightweig ht approach to determine masks independently\nbefore merging. Across four settings where we merge up to 500 models\,\nSIFT-Masks improves accuracy b y 5-80% over naive merging and uses up\nto 250x less compute for exact unl earning compared to other merging\nbaselines.\n\nPresented as part of the CyLab Student Seminar\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9119200 DTSTART;TZID=America/New_York:20260429T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260429T153000 LOCATION:Newell-Simon 1305 & Zoom SUMMARY:Doctoral Thesis Proposal - Zhengyao Lin CLASS:PUBLIC DESCRIPTION:Speaker: ZHENGYAO LIN\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Verifying Asynchronous D ataflow Compilers\n\nThe paradigm of asynchronous dataflow circuits\, in w hich parallel\noperators are dynamically scheduled and data-driven\, unloc ks\nsubstantial energy efficiency and performance gains in reconfigurable\ ndataflow architectures (RDAs) and dynamic high-level synthesis (HLS)\ntoo lchains. A key challenge hindering their mainstream adoption is the\ndiffi culty of correct\, efficient\, and general-purpose compilation\ntowards as ynchronous dataflow.\n\nIn this proposal\, I present my efforts to apply f ormal verification to\nasynchronous dataflow\, with the goal of developing an end-to-end\,\nprovably correct dataflow compilation pipeline.\n\nThesi s Committee:\n\nBryan Parno (Chair)\n\nStephanie Balzer\n\nRuben Martins\n \nBrandon Lucia\n\nMilijana Surbatovich (University of Maryland\, College Park)\n\nAdditional Information\n\nIn-person and Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91195b8 DTSTART;TZID=America/New_York:20260428T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260428T133000 LOCATION:Newell-Simon Hall 3002 & Zoom SUMMARY:Doctoral Thesis Proposal - Lingjing Kong CLASS:PUBLIC DESCRIPTION:Speaker: LINGJING KONG\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Causal AI for Transfera ble\, Interpretable\, and\nControllable Machine Learning\n\nFoundation mod els are rapidly becoming capable assistants for\nknowledge work\, but thei r deployment in real settings is limited by\nthree gaps: they do not trans fer reliably across environments\, their\ninternal reasoning is opaque\, a nd their behavior is hard to control\nprecisely. In this talk\, I argue th at these limitations are not only\nabout model size — they are fundament ally about whether learning\ncaptures and leverages the underlying structu re of the data-generating\nprocess. I use causal thinking as a practical l ens to model what is\ninvariant\, what changes\, and what can be intervene d on\, and I further\nshow how this leads to learning principles that impr ove\ntrustworthiness. I will first present methods for learning unifying\n mechanisms from heterogeneous data\, across domains and modalities\, to\ne nable reliable transfer and controllable generation. Next\, I will\nshow h ow structured concepts can be recovered even from seemingly\nunstructured data\, by analyzing and improving self-supervised\nobjectives (such as mas king and diffusion) through hierarchical\nlatent-variable models. These co ncept structures can then be used to\ninterpret generative models and supp ort targeted\, multi-level edits.\nFinally\, I connect these two threads t o generalization beyond the\ntraining distribution. I will discuss natural conditions for\nextrapolation and a compositional generation framework th at improves\nprompt following for novel concept combinations. I will concl ude with\na brief outlook on self-improving world models.\n\nThesis Commit tee:\n\nKun Zhang (Co-chair)\n\nYuejie Chi (Co-chair)\n\nEric Xing (Co-cha ir)\n\nTom Mitchell\n\nKevin Murphy (Google DeepMind)\n\nAdditional Inform ation \n\nIn-person and Zoom \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9119a1d DTSTART;TZID=America/New_York:20260427T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260427T133000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Thesis Proposal - Kevin Kuo CLASS:PUBLIC DESCRIPTION:Speaker: KEVIN KUO\, Ph.D. Student\, Computer Science Departmen t\,\nCarnegie Mellon University\n\nTalk Title: Algorithms for Efficient an d Safe Collaborative Learning\nSystems\n\nData for AI systems is becoming increasingly sparse. Within a decade\,\nLLMs are projected to be trained o n datasets the size of the total\nstock of public human data\, while indiv iduals and organizations are\nincreasingly restricting access to their dat a due to economic and\nprivacy concerns. Collaborative learning has the po tential to fuel\ndata-hungry AI systems by enabling access to restricted s ources of\ndata---but only if it can provide meaningful guarantees regardi ng data\nprivacy\, quality of service\, and computational cost. This thesi s\nstudies three unique ML system architectures that offer data\nprotectio n by design: (1) multi-round federated learning\, (2) model\nmerging-and-l ocalization\, and (3) proxy tuning. We leverage principles\nfrom model com pression and transfer learning to improve the utility\,\nefficiency\, and privacy guarantees of these frameworks.\n\nThesis Committee:\n\nVirginia S mith (Chair)\n\nAditi Raghunathan\n\nGauri Joshi\n\nHolger Roth (NVIDIA)\n \nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9119e2a DTSTART;TZID=America/New_York:20260428T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260428T163000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Speaking Skills Talk - Hongyi Jin CLASS:PUBLIC DESCRIPTION:Speaker: HONGYI JIN\, Ph.D. Student\, Computer Science Departme nt\,\nCarnegie Mellon University\n\nTalk Title: Event Tensor: A Unified Ab straction for Compiling Dynamic\nMegakernel\n\nModern GPU workloads\, espe cially large language model (LLM) inference\,\nsuffer from kernel launch o verheads and coarse synchronization that\nlimit inter-kernel parallelism. Recent megakernel techniques fuse\nmultiple operators into a single persis tent kernel to eliminate launch\ngaps and expose inter-kernel parallelism\ , but struggle to handle\ndynamic shapes and data-dependent computation in real workloads. We\npresent Event Tensor\, a unified compiler abstraction for dynamic\nmegakernels. Event Tensor encodes dependencies between tiled tasks\,\nand enables first-class support for both shape and data-dependen t\ndynamism. Built atop this abstraction\, our Event Tensor Compiler (ETC) \napplies static and dynamic scheduling transformations to generate\nhigh- performance persistent kernels. Evaluations show that ETC\nachieves state- of-the-art LLM serving latency while significantly\nreducing system warmup overhead.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911a1a3 DTSTART;TZID=America/New_York:20260506T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260506T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting CLASS:PUBLIC DESCRIPTION:See Email Announcement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911a443 DTSTART;TZID=America/New_York:20260424T143000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260424T160000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Thesis Proposal - Sagar Bharadwaj Kalasibail Seetharam CLASS:PUBLIC DESCRIPTION:Speaker: SAGAR BHARADWAJ KALASIBAIL SEETHARAM\, Ph.D. Student\, Computer\nScience Department\, Carnegie Mellon University\n\nTalk Title: OpenFLAME: Federated Spatial Intelligence\n\nSpatial applications\, i.e.\, applications that tie digital information\nwith the physical world\, have improved many of our daily activities\,\nsuch as navigation and ride-shar ing. This class of applications also\nholds significant promise of enablin g new industries such as augmented\nreality and robotics. The development of these applications is enabled\nby spatial intelligence systems\, or sy stems that can provide\nphysically-grounded services such as routing\, sp atial search\, and\nlocalization. Today\, mapping platforms provided by or ganizations like\nGoogle and Apple serve as spatial intelligence systems. These maps are\ncentralized and primarily cover outdoor spaces. We envisio n that\nfuture spatial applications\, such as persistent world-scale augme nted\nreality and robotics\, would require detailed and precise spatial\n intelligence across indoor and outdoor spaces. The scale of\ncartography e fforts required to survey indoor spaces and their privacy\nneeds inhibit e xisting centralized maps from incorporating such spaces\ninto their platfo rm.\n\nThis thesis proposes the design and implementation of OpenFLAME\, a \nfederated spatial intelligence platform built on two core pillars.\nFirs t\, it introduces a federated mapping infrastructure where\nindependent pa rties can manage and serve their own maps of physical\nregions. This unloc ks scalability of map management\, isolation\, and\nprivacy of maps. This is implemented on top of the existing Domain\nName System (DNS)\, which e nables us to leverage its existing\ninfrastructure. Second\, it explores t he development of spatial\nservices\, such as search\, routing\, and local ization\, by transforming\nraw spatial data (e.g.\, images and videos) of a space into structured\,\nqueryable maps that support these capabilities. \n\nThesis Committee:\n\nSrinivasan Seshan (Co-Chair)\n\nAnthony Rowe (Co- Chair)\n\nJustine Sherry\n\nHari Balakrishnan (Massachusetts Institute of Technology)\n\nAdditional Information\n\nIn-person & Zoom \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911a8b7 DTSTART;TZID=America/New_York:20260424T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260424T163000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year Master's Thesis Presentation - Taekseung Kim CLASS:PUBLIC DESCRIPTION:Speaker: TAEKSEUNG KIM\, Master's Student\nComputer Science Dep artment\nCarnegie Mellon University\n\nTalk Title: Towards Faster Parallel Algorithms for Tree Decompositions\n\nThe tree decomposition problem\, al so known as the treewidth problem\,\nis a central topic in graph algorithm s that has been studied\nextensively over the past 30 years. Its importanc e comes from the fact\nthat treewidth serves as a parameter under which ma ny otherwise\nintractable problems become algorithmically tractable. As a result\,\nthe study of treewidth has played a major role in the developmen t of\nfixed-parameter tractability and has also influenced related areas\n such as database theory and constraint satisfaction.\n\nIn this talk\, we survey previous results on treewidth computation\,\nranging from work by B odlaender in the 1990s to more recent\ndevelopments by Korhonen. We begin by defining the problem and\nreviewing the separator-based algorithms that were prominent until the\nearly 1980s\, developed by Robertson and Seymou r. This line of research\nled to Reed’s work\, which established an nlog n bound for computing\ntreewidth. We then turn to algorithms from the 1990 s\, mainly due to\nBodlaender and colleagues\, who used compression and im provement\ntechniques to obtain better bounds. With further development of these\nideas\, Bodlaender proved the existence of a linear-time algorithm for\ncomputing treewidth\, and also showed that a parallel version of thi s\nalgorithm exists. We then briefly discuss algorithms developed after\nt he 2010s\, including approximation algorithms obtained by Bodlaender\nand later improved by Korhonen in 2021. We also briefly cover the\ndynamic alg orithm proposed by Korhonen in 2023.\n\nIn addition\, we consider these re sults from the perspective of\nparallel algorithms\, where some approaches appear parallelizable while\nothers remain difficult to parallelize. We d iscuss the intuition and\noverall paradigms underlying treewidth algorithm s\, and examine them\nthrough a parallel lens\, asking why some techniques can be\nparallelized and others can not. Although there has been earlier\ nresearch in this direction\, the parallel study of treewidth has been\nli mited since the 1990s\, especially after the paper by Bodlaender and\nHage rup. At the same time\, several modern techniques have been\ndeveloped sin ce then. While further progress is needed on parallel\nstatic algorithms f or computing treewidth\, parallel batch dynamic\ncomputation for treewidth also remains open. While Korhonen’s 2023\npaper suggests an amortized d ynamic algorithm for treewidth\ncomputation\, this method is not suitable for parallelization since it\nuses splay tree style rotation\, and paralle l batch dynamic computation\nfor treewidth has not yet been solved.\n\nFin ally\, we suggest future directions in which this line of work could\nlead to parallel static and batch dynamic algorithms\, which we hope to\nestab lish. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911ad99 DTSTART;TZID=America/New_York:20260420T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260420T173000 LOCATION:Remote Access- Zoom SUMMARY:Database Seminar - Steve Schirripa CLASS:PUBLIC DESCRIPTION:Speaker: STEVE SCHIRRIPA\, Co-founder and Chief Technology Offi cer\,\nVillageSQL\n\nTalk Title: The Extensibility Tax: Decisions\, Princi ples\, and Lessons\nfrom a Year of Teaching MySQL New Tricks\n\nWhat does it take to add custom data types and indexes to a database\nthat was never designed for them? Over the past year\, we built an\nextensibility framew ork for MySQL — and discovered\, one by one\,\nevery place the codebase assumes it knows all the types that will ever\nexist.\n\nThis talk walks t hrough the design decisions that survived and the\nones that didn't. We'll cover why extending MySQL's type system\nthrough the type system is impos sible\, why injecting metadata into\nexisting classes beat every attempt a t wrapping or subclassing\, and\nwhy eight separate comparison code paths all needed interception.\nWe'll look at the surprises — CREATE TABLE and ALTER TABLE being\nessentially different codebases\, SQL features like LO AD DATA and\ndefault values that each bypass the layers you'd expect\, and \nnon-deterministic query plans exposing latent bugs that only appear on\n the second execution. And we'll share the principles that emerged:\nfollow the existing patterns exactly\, treat backward compatibility as\na first- class constraint\, and make the Data Dictionary sacred ground.\n\nWhether you're building database extensions\, working with large legacy\ncodebases \, or just curious about what's hiding inside MySQL\, you'll\nleave with p ractical lessons about making architectural decisions when\nthe codebase i s fighting you at every turn.\n\n—\n\nSteve Schirripa is alumni of Carne gie Mellon University (MSE'00) and\nthe technical co-founder/CTO of Villag eSQL. Prior to that he was a\nDistinguished Software Engineer at Google fo r 21 years where he was\nthe Technical Lead for Bigtable and the Colossus File System.\n\nThis talk is part of the PostgreSQL vs. The World Seminar Series\n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911b22a DTSTART;TZID=America/New_York:20260416T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260416T173000 LOCATION:Gates Hillman 7501 and Zoom SUMMARY:Crypto Seminar - Saachi Mutreja CLASS:PUBLIC DESCRIPTION:Speaker: SAACHI MUTREJA\, Ph.D. Student\, Computer ScienceColum bia\nUniversity\n\nTalk Title: Classical Obfuscation of Quantum Circuits v ia Publicly\nVerifiable QFHE\n\nA classical obfuscator for quantum circuit s is a classical program\nthat\, given the classical description of a quan tum circuit Q\, outputs\nthe classical description of a functionally equiv alent quantum circuit\nÕ that hides as much as possible about Q. Previous ly\, the only known\nfeasibility result for classical obfuscation of quant um circuits\n(Bartusek and Malavolta\, ITCS 2022) was limited to \"null\" security\,\nwhich is only meaningful for circuits that always reject. On t he other\nhand\, if the obfuscator is allowed to compile the quantum circu it Q\ninto a quantum state Õ. there exist feasibility results for\nobfus cating much more expressive classes of circuits: all\npseudo-deterministic quantum circuits (Bartusek\, Kitagawa\, Nishimaki\nand Yamakawa\, STOC 20 23\, Bartusek\, Brakerski and Vaikuntanathan\, STOC\n2024)\, all unitaries (Huang and Tang\, FOCS 2025)\, and even arbitrary\nquantum circuits (Huan g and Tang)\n\nWe show that (relative to a classical oracle) there exists a\nclassical obfuscator for all pseudo-deterministic quantum circuits.\nA s our main technical step\, we give the first construction of a\ncompact q uantum fully-homomorphic encryption (QFHE) scheme that\nsupports public ve rification of (pseudo-deterministic) quantum\nevaluation\, relative to a c lassical oracle. To construct our QFHE\nscheme\, we improve on an approach introduced by Bartusek\, Kitagawa\,\nNishimaki and Yamakawa (STOC 2023)\, which previously required\nciphertexts that are both quantum and non-comp act due to their heavy\nuse of quantum coset states for their publicly-ver ifiable properties.\nAs part of our core technical contribution\, we intro duce new\ntechniques for analyzing coset states that can be generated on t he\nfly\, and prove a new cryptographic property—the \"delayed\ncollapsi ng'' property—of the one-shot signature scheme of Shmueli\nand Zhandry ( CRYPTO 2025). Our techniques allow us to produce QFHE\nciphertexts that ar e purely classical\, compact\, and\npublicly-verifiable. This additionally yields the first classical\nverification of quantum computation protocol for BQP that\nsimultaneously satisfies blindness and public-verifiability. \n\nIn Person and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911b6a3 DTSTART;TZID=America/New_York:20260422T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260422T133000 LOCATION:Panther Hollow Conference Room 4105\, Mehrabian Collaborative\nInn ovation Center SUMMARY:Special SDI Seminar - Ethan Miller CLASS:PUBLIC DESCRIPTION:Speaker: ETHAN L. MILLER\, Engineer\, Office of the Chief Techn ology\nOfficer\, Everpure (formerly Pure Storage)\n\nTalk Title: Towards A rchival Storage Using QLC Flash\n\nThis talk describes early design ideas for an alternate approach to\narchival storage: low-cost\, high-capacity f lash storage. Long-term\narchival storage is typically implemented using d isk or nearline media\nsuch as tape and optical media. While this approach is relatively\ninexpensive\, it limits access to archival data behind a s low interface\nwith limited bandwidth. In contrast\, high-density QLC flas h storage\nshipping today can fit half an exabyte into a single rack\, wit h\nincreases in density forecast to exceed an exabyte per rack in less\nth an 18 months. QLC flash has low power consumption\, much higher read\nband width\, and longer durability than existing archival storage. Using\nflash for archival storage also greatly reduces the cost of media\nmigration\, currently a major issue for traditional archival media.\nMore excitingly\, archival QLC flash enables new ways of interacting\nwith archival storage via \"background\" data accesses that piggyback on\nperiodic data scrubbi ng to leverage the massive bandwidth available at\nlow cost and power from flash storage. Using storage under this model\,\nresearchers can run many more experiments on their archived data at\nrelatively low cost compared to today's approach of fetching archival\ndata to \"fast\" storage and ana lyzing it.\n\n—\n\nEthan L. Miller is an engineer in the Office of the C TO at Everpure\n(formerly Pure Storage) and a Professor Emeritus in the Co mputer\nScience and Engineering Department at the University of California \,\nSanta Cruz\, where he held the Veritas Presidential Chair in Storage.\ nHe helped co-found Pure Storage\, and has worked with the company since\n 2009 to develop reliable high-performance flash-based storage systems.\nAt Everpure\, his focus is on reliability\, efficient data structures\,\ndat a security\, and the design of future data storage systems. His work\nat E verpure has resulted in over 200 granted patents. He is a Fellow\nof the I EEE and an ACM Distinguished Scientist.\n\nAt UC Santa Cruz\, Prof. Miller was an active member of the faculty\nfrom 2000-2023 and the Director of t he NSF IUCRC Center for Research\nin Storage Systems (CRSS) from 2013-2020 . He was also a founding\nmember of the Storage Systems Research Center (S SRC) at UC Santa Cruz.\nHe has co-authored over 170 papers in a range of t opics in file and\nstorage systems\, operating systems\, parallel and dist ributed systems\,\ninformation retrieval\, and computer security. He was a member of the\nteam that developed Ceph\, a scalable high-performance dis tributed file\nsystem for scientific computing that is now being adopted b y several\nhigh-end computing organizations. His work on reliability and s ecurity\nfor distributed storage is also widely recognized\, as is his wor k on\nsecure\, efficient long-term archival storage and scalable metadata\ nsystems. \n\nFaculty Host: George Amvrosiadis \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911bb9d DTSTART;TZID=America/New_York:20260420T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260420T103000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:5th Year Master's Thesis Presentation - James Kim CLASS:PUBLIC DESCRIPTION:Speaker: JAMES KIM\, Master's Student\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: Dissecting Reinforcement Le arning: Mechanisms Behind\nCompositional Reasoning in LLMs\n\nRecently\, L arge Reasoning Models (LRMs) have achieved impressive\nperformance on a va riety of reasoning tasks\, including mathematics and\ncode generation. The ir long chains of reasoning enable scaling of\ninference-time computation\ , allowing them to solve increasingly\ncomplex problems. LRMs are typicall y post-trained from base models\nusing supervised fine-tuning (SFT)\, rein forcement learning (RL)\, or a\ncombination of both. RL is often hypothesi zed to be a key driver of\nreasoning ability\, as it enables models to exp lore and discover new\nsolutions. However\, recent work suggests that RL m ay instead\nconcentrate probability mass on existing solutions. The mechan isms by\nwhich RL leads to reasoning ability remain poorly understood. In this\nthesis\, we study RL training mechanisms through the lens of\ncompos itional generalization—a key sub-skill of reasoning that\ninvolves combi ning atomic skills to solve more complex problems. We\nfind that RL-traine d models substantially outperform those trained\nwith standard SFT on this task. To isolate the effects of RL\, we\ndecompose it into three componen ts: on-policy data\, the use of\nnegative samples\, and objective design. By ablating each component\, we\nestablish a progression from SFT to RL an d identify their respective\ncontributions. Empirically\, we find that bot h on-policy data and\nnegative samples are critical for the emergence of c ompositional\ngeneralization\, while objective design choices (e.g.\, grou p\nnormalization) have a relatively small impact. Our findings suggest\nth at effective post-training of LLMs requires understanding and\ncarefully d esigning the individual components of the training\npipeline\, rather than treating RL as a monolithic improvement.\n\nThesis Committee\n\n Chenyan Xiong (Chair)\n\nAditi Raghunathan\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911c050 DTSTART;TZID=America/New_York:20260421T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260421T133000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:5th Year Master's Thesis Presentation - Hao Kang CLASS:PUBLIC DESCRIPTION:Speaker: HAO KANG\, Master's Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Training Mixture-of-Exper ts Language Models\n\nMixture-of-Experts (MoE) has become the dominant arc hitecture for\nfrontier language models\, offering a favorable trade-off b etween model\ncapacity and per-token computation. This thesis studies the training\nof MoE language models from two perspectives: modeling and syste ms.\n\nOn the modeling side\, we present FLAME-MoE\, a transparent researc h\nplatform providing a suite of MoE models across seven scales\, with all \ncode\, data pipelines\, intermediate checkpoints\, and routing logs\nrel eased publicly. We establish MoE scaling laws and show that the\nresulting models outperform dense baselines at matched compute. Using\nthe full tra ining trace\, we conduct empirical analyses of expert\nbehavior\, finding that expert specialization emerges gradually\,\nco-activation remains spar se but intensifies in deeper layers\, and\nrouting decisions converge earl y in training.\n\nOn the systems side\, we present PithTrain\, a Python-na tive MoE\ntraining framework that delivers production-grade throughput in\ nroughly 10K lines of code. PithTrain supports 4D parallelism\, a\nDualPip eV pipeline scheduler that overlaps computation with\ncommunication\, FP8 training via DeepGEMM\, and fused Triton kernels for\nexpert dispatch. At this scale\, the entire codebase fits within the\ncontext window of modern AI coding tools\, making it end-to-end\nreadable by both humans and agent s. We also explore the implications\nof this minimal design for agentic de velopment.\n\nThesis Committee\n\n Chenyan Xiong (Chair)\n\nTianqi Chen\n \nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911c45b DTSTART;TZID=America/New_York:20260422T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260422T130000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Speaking Skills Talk - Andrew Brady CLASS:PUBLIC DESCRIPTION:Speaker: ANDREW BRADY\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Faster Parallel Batch-Dy namic Algorithms for Low\nOut-Degree Orientation\nwindow.MathJax = {\ntex: {\ninlineMath: [['$'\, '$']\, ['\\\\('\, '\\\\)']]\n}\n}\;\n\nA low out-d egree orientation directs each edge of an undirected graph\nwith the goal of minimizing the maximum out-degree of a vertex. In the\nparallel batch-d ynamic setting\, one can insert or delete batches of\nedges\, and the goal is to process the entire batch in parallel with\nwork per edge similar to that of a single sequential update and with\nspan (or depth) for the enti re batch that is polylogarithmic.\n\nIn this paper we present faster paral lel batch-dynamic algorithms for\nmaintaining a low out-degree orientation of an undirected graph. All\nresults herein achieve polylogarithmic depth \, with high probability\n(whp)\; the focus of this paper is on minimizing the work\, which varies\nacross results.\n\nOur first result is the first parallel batch-dynamic algorithm to\nmaintain an asymptotically optimal o rientation with asymptotically\noptimal expected work bounds\, in an amort ized sense\, improving over\nthe prior best work bounds of Liu et al.~[SPA A~'22] by a logarithmic\nfactor. Our second result is a $O(c \\log n)$ ori entation algorithm\nwith expected worst-case $O(\\sqrt{\\log n})$ work per edge update\,\nwhere $c$ is a known upper-bound on the arboricity of the graph. This\nmatches the best-known sequential worst-case $O(c \\log n)$ o rientation\nalgorithm given by Berglin and Brodal ~[Algorithmica~'18]\, al beit in\nexpectation.\n\nOur final result is a $O(c + \\log n)$-orientatio n algorithm with\n$O(\\log^2 n)$ expected worst-case work per edge update. This algorithm\nsignificantly improves upon the recent result of Ghaffari and\nKoo~[SPAA~'25]\, which maintains a $O(c)$-orientation with $O(\\log^ 9\nn)$ worst-case work per edge whp.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911c8bc DTSTART;TZID=America/New_York:20260421T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260421T143000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Aditi Nandkishor Kabra CLASS:PUBLIC DESCRIPTION:Speaker: ADITI NANDKISHOR KABRA\, Ph.D. Candidate\nComputer Sci ence Department\nCarnegie Mellon University\n\nTalk Title: Verified Contro l Envelope Synthesis\n\nMany cyber-physical systems\, such as trains\, pla nes\, and self-driving\ncars\, are safety-critical but difficult to reason about. The task of\ndesigning controllers for such systems is complex (th e subject of an\nentire field\, control theory)\, made even more challengi ng by the need\nto ensure correctness over all of the infinitely many poss ible\nscenarios that the system may face. This thesis develops techniques\ nthat let computers automatically synthesize the conditions that define\nc orrect control solutions\, with mathematical guarantees that these\ncondit ions are correct.\n\nSymbolic control envelopes are our representation of the control\nconditions that characterize sets of safe control solutions. They are\nrepresented parametrically in symbols that can be instantiated w ith\nany real-valued input (e.g.\, for a train control envelope\, train\nw eight w). Control envelopes provide a path to designing complex\ncontrolle rs that still have mathematical correctness guarantees by\nallowing separa tion of concerns during controller design. A verified\n(i.e.\, mathematica lly correct) safe control envelope can first\nidentify the set of control solutions that ensure the safety-critical\ncontrol objectives\, and then n on-formal techniques\, e.g.\, machine\nlearning\, can optimize within that envelope for secondary objectives.\n\nThe thesis automates the process of designing symbolic control\nenvelopes by creating the first framework for symbolic control\nenvelope synthesis. The framework takes as input the sh ape of a\ncontrol system\, which indicates what control and environment be haviors\nare physically possible and what the desired control behavior is\ ,\nmaking the synthesis question well-defined. The framework\nautomaticall y identifies the symbolic control conditions indicating\nwhen a given cont rol action is correct\, which is shown to correspond\nto the nondeterminis tic control policies of players in hybrid games\n(games with both continuo us and discrete dynamics).\n\nThis thesis tackles the hybrid game control envelope synthesis problem\nin its full generality\, developing the theory to solve for all of\ndifferential game logic. It introduces a specialized procedure for an\ninteresting subset of problems (time-triggered\, where the controller\nloops with some maximum time latency) that is total comput able under\nsome reasonable assumptions. By strategically using large lang uage\nmodels along with verification\, it provides a general approach to\n sound\, scalable synthesis.\n\nThesis Committee\n\nAndré Platzer (Co-Chai r)\n\nStefan Mitsch (Co-Chair)\n\nEunsuk Kang\n\nArmando Solar-Lezama (Mas sachusetts Institute of Technology)\n\nIn Person and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911cdf8 DTSTART;TZID=America/New_York:20260421T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260421T103000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Speaking Skills Talk - Sophia Roshal CLASS:PUBLIC DESCRIPTION:Speaker: SOPHIA ROSHAL\, Ph.D Student\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: Adjoint Types for Efficient Functional Programming\n\nThere has been growing interest in programming languages in which\ntypes control resource usage. In most languages\, type s only track the\ntypes of expressions\, but they do not capture how the u nderlying\nresources are used. Substructural type systems on the other han d track\nusage explicitly. Linear types require that each resource is used \nexactly once\, affine types require that it is used at most once\, and\n strict types require that it is used at least once. The “standard”\nty pes which do not restrict resource usage we call unrestricted in\nthis set ting.\n\nThese distinctions have practical use cases. As just one example\ ,\nlanguages such as Rust use an affine type system to enforce thread\nsaf ety by ensuring that resources cannot be accessed more than once.\nThe sam e restriction of single use can be used to implement\nmemory-reuse compile r optimizations that can immediately re-use the\nmemory of a value that we know can never be read again.\n\nIn practice\, most programs are not pure ly linear\, affine\, strict\, or\nunrestricted\, but instead have pieces o f programs at different modes.\nThis talk introduces Adjoint Types\, a typ e system for a functional\nprogramming language that allows these modes to coexist within the\nlanguage. This allows us to exploit the benefits of l inear\, affine\,\nstrict\, and unrestricted typing while retaining the fle xibility of a\nprogramming language with only unrestricted types.\n\nPrese nted in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911d1ff DTSTART;TZID=America/New_York:20260420T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260420T173000 LOCATION:Gates Hillman 7501 SUMMARY:5th Year Master's Thesis Presentation - Jerick Shi CLASS:PUBLIC DESCRIPTION:Speaker: JERICK SHI\, Master's Student\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: The Structure of Deception : How LLM Agents Lie\, Break\nPromises\, and Exploit Trust in Multi-Agent Settings\n\nLarge language models are increasingly deployed not just as ch atbots\nbut as autonomous agents that negotiate\, trade\, and act on behal f of\nusers in multi-agent systems. When these agents are able to\ncommuni cate intentions and then privately deviate from them\, deception\nbecomes a concrete safety problem\, distinct from the single-model\nreliability is sues that most current evaluations address. This thesis\nstudies when and how LLM agents deceive across multiple scales of\ninteraction. We first un ify the fragmented literature on LLM\ndeception\, spanning hallucination\, sycophancy\, alignment faking\, and\nstrategic scheming\, into a single t axonomy organized by\ngoal-directedness\, object of deception\, and mechan ism\, and apply it to\n50 existing benchmarks to reveal systematic gaps in evaluation\ncoverage. We then place frontier LLMs in one-shot game-theore tic\nsettings with public commitments\, finding that agents break promises \nin over half of all scenarios\, that most deviations serve\nself-interes t\, and that the dominant failure mode is unreflective\npayoff optimizatio n rather than deliberate deception. Extending to\nrepeated games with endo genous promises and mixed-model groups\, we\nshow that deception is predom inantly premeditated yet not a fixed\nmodel trait\, that different models interpret communication through\nincompatible frameworks producing persist ent exploitation\, and that\nself-reported trust is decoupled from actual outcomes. Across all\nsettings\, deception is shaped more by the structure of the environment\nthan by model identity\, aggregate metrics obscure qu alitatively\ndistinct failure modes\, and the monitoring tools we currentl y rely on\nmiss the most common patterns. We conclude by discussing implic ations\nfor the deployment of multi-agent AI systems in domains where\ncoo rdination\, trust\, and accountability matter\, and outline directions\nfo r studying deception in richer economic environments where\nadversarial in centives are not explicitly assigned.\n\nThesis Committee\n\nVincent Conit zer (Chair)\n\nAditi Raghunathan\n\nAdditional Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911d689 DTSTART;TZID=America/New_York:20260420T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260420T143000 LOCATION:Gates Hillman 5117 SUMMARY:Doctoral Thesis Proposal - Zhibo Chen CLASS:PUBLIC DESCRIPTION:Speaker: ZHIBO CHEN\, Ph.D. Student\nComputer Science Departmen t\nCarnegie Mellon University\n\nTalk Title: CoLF - A Coinductive Logical Framework\n\nThe LF logical framework represents judgments as types\, and objects\nand derivations as finitary terms. Twelf\, an implementation of L F\, is\na metalogical framework that incorporates metatheorem checking and a\nlogic programming engine. There is no direct encoding of infinitary\no bjects and derivations in LF and Twelf. In this thesis\, we propose a\nnew logical framework CoLF\, building on top of LF\, for encoding\ninfinitary objects and derivations as infinitary terms. We develop\nthe type theory and implementation of CoLF. We also propose to\ninvestigate the metatheor etic reasoning principles and a new logic\nprogramming semantics in the pr esence of infinitary terms.\n\nThesis Committee\n\nFrank Pfenning (Chair)\ n\nKarl Crary\n\nIliano Cervesato\n\nAlberto Momigliano (University of Mi lan)\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911dc30 DTSTART;TZID=America/New_York:20260415T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260415T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Bernardo Subercaseaux CLASS:PUBLIC DESCRIPTION:Speaker: BERNARDO SUBERCASEAUX\, Ph.D. Student\, Computer Scien ce\nDepartment\, Carnegie Mellon University\n\nTalk Title: Breaking down g raphs and hypergraphs into structured\npieces\, optimally and efficiently\ n\nWe will consider the problem of writing an arbitrary graph as an\nedge- disjoint union of complete bipartite graphs\, and its natural\ngeneralizat ion to hypergraphs. At a very high level\, the question at\nhand is to wha t extent we can summarize complicated structures by\ndecomposing them into very structured pieces. I will present several\noptimal asymptotic bounds and algorithms\, with applications to graph\ncompression\, SAT solving\, cryptographic secret sharing\, and\napproximations for the densest subgrap h problem. More concretely\, our\nmain result is that every n-vertex d-uni form hypergraph H can be\nwritten as the union of a family F of complete d -partite hypergraphs\nsuch that every vertex of H belongs to at most (n ch oose d)/(n lg n)\ngraphs in F. This improves on results of Csirmaz\, Liget i\, and Tardos\n(2014)\, gives the best upper bound for some secret sharin g questions\,\nand answers several 40-year-old questions of Chung\, Erdős \, and\nSpencer (1983). The heart of our proof is a simple idea from word\ ncombinatorics\, which allows us to balance the number of d-cliques each\n vertex belongs to.\n\nThis talk is based on joint work with Andrew Krapivi n\, Benjamin\nPrzybocki\, and Nicolás Sanhueza-Matamala.\n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911dfef DTSTART;TZID=America/New_York:20260416T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260416T160000 LOCATION:Gates Hillman 8102 SUMMARY:Doctoral Speaking Skills Talk - Zhouzi Li CLASS:PUBLIC DESCRIPTION:Speaker: ZHOUZI LI\, Ph.D. Student\nComputer Science Department \nCarnegie Mellon University\n\nTalk Title: BOA Constrictor: Squeezing Per formance out of GPUs in the\ncloud via Budget Optimal Allocation\n\nThe pa st decade has seen a dramatic increase in demand for GPUs to\ntrain Machin e Learning (ML) models. Because it is prohibitively\nexpensive for most or ganizations to build and maintain a large GPU\ncluster\, organizations ins tead choose to rent GPUs from cloud\nproviders.  The customer is responsi ble for devising a policy for (i)\ndeciding how many GPUs to rent at every moment in time to process a\nstream of ML training jobs and (ii) allocati ng the rented GPUs among\nthe currently active jobs in the system.  Becau se ML training jobs\ncan be parallelized across different numbers of GPUs\ , the customer\ngenerally has many options for how many GPUs to use for ea ch job.\nAllocating more GPUs to a single training job will cause the job to\ncomplete more quickly. However\, the customer pays for each GPU-hour\n they use\, and a training job receives a diminishing marginal benefit\nfro m running on additional GPUs. Hence\, allocating too many GPUs to a\nsingl e training job can dramatically increase the overall cost that\nthe custom er pays to the cloud provider.  This gives rise to a\ncost-performance tr adeoff that customers must balance when running\ntraining jobs in the clou d.\n\nTo balance the cost-performance tradeoff\, we develop BOA Constricto r\,\na new scheduler for ML training jobs which uses a Budget-Optimal\nAll ocation (BOA) policy to squeeze the highest level of performance\nout of a cloud-deployed GPU cluster given a fixed budget constraint.\nFor a given budget level\, we demonstrate that BOA Constrictor can\nreduce average JCT by 1.6 times in small-scale implementation\nexperiments and by 2 times in detailed\, large-scale simulations\ncompared to state-of-the-art heuristi c based schedulers.\n\nPresented as Part of the Catalyst Group Meeting\n\n Presented in Partial Fulfillment of the CSD Speaking Skills\nRequirement  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911e40f DTSTART;TZID=America/New_York:20260420T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260420T123000 LOCATION:Newell-Simon 1505 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Jonathan Laurent CLASS:PUBLIC DESCRIPTION:Speaker: JONATHAN LAURENT\, Ph.D. Candidate\nComputer Science D epartment\nCarnegie Mellon University\n\nTalk Title: Oracular Programming: A Modular Foundation for Building\nLLM-Enabled Software\n\nLarge Language Models (LLMs) can solve previously intractable tasks\ngiven only natural- language instructions and a few examples\, but they\nremain difficult to s teer and lack a key capability for building\nreliable software at scale: t he modular composition of computations\nunder enforceable contracts. As a result\, they are often embedded in\nlarger software pipelines that use do main knowledge to decompose tasks\nand improve reliability through validat ion and search. Yet the\ncomplexity of writing and maintaining such pipeli nes has so far\nlimited their sophistication.\n\nWe propose oracular prog ramming: a foundational paradigm for\nintegrating traditional\, explicit c omputations with inductive oracles\nsuch as LLMs. It rests on two directin g principles: the full\nseparation of core and search logic (allowing the latter to freely\nevolve without breaking the former)\, and the treatm ent of few-shot\nexamples as grounded and evolvable program components (allowing\ntheir consistency with the rest of the program to be enforced through\nits evolution\, with breakages easily identifiable and repairable ).\nWithin this paradigm\, programmers express high-level problem-solving\ nstrategies as programs with unresolved choice points. These choice\npoint s are resolved at runtime by LLMs\, which generalize from\nuser-provided e xamples of correct and incorrect decisions.\nAn oracular program is comp osed of three orthogonal components:\na strategy that consists of a nond eterministic program with choice\npoints that can be reified into a search tree\, a policy that\nspecifies how to navigate this tree with the help of LLM oracles\, and\na set of demonstrations that describe successful and unsuccessful\ntree navigation scenarios across diverse problem instanc es. Each\ncomponent is expressed in a dedicated language.\n\nWe address th e key programming language design challenges of modularly\ncomposing oracu lar programs and enforcing consistency between their\ncomponents as they e volve. We also demonstrate universal\nself-improvement mechanisms for orac ular programs\, in which training\nand tuning data is automatically extrac ted from successful and\nunsuccessful runs. Finally\, we present Delphyne \, an open-source\nframework for oracular programming based on Python\, an d empirically\nevaluate our approach through several case studies.\n\nThes is Committee\n\nAndré Platzer (Chair)\n\nMarijn Heule\n\nZico Kolter\n\nA rmando Solar-Lezama (Massachusetts Institute of Technology\n\nIn Person an d Zoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911e933 DTSTART;TZID=America/New_York:20260415T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260415T120000 LOCATION:Newell-Simon 4201 SUMMARY:Doctoral Speaking Skills Talk - Lingjing Kong CLASS:PUBLIC DESCRIPTION:Speaker: LINGJING KONG\, Ph.D. Student\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: Causal AI for Transferable \, Interpretable\, and\nControllable Machine Learning\n\nFoundation models are rapidly becoming capable assistants for\nknowledge work\, but their d eployment in real settings is limited by\nthree gaps: they do not transfer reliably across environments\, their\ninternal reasoning is opaque\, and their behavior is hard to control\nprecisely. In this talk\, I argue that these limitations are not only\nabout model size — they are fundamentall y about whether learning\ncaptures and leverages the underlying structure of the data-generating\nprocess. I use causal thinking as a practical lens to model what is\ninvariant\, what changes\, and what can be intervened o n\, and I further\nshow how this leads to learning principles that improve \ntrustworthiness.\n\nI will first present methods for learning unifying m echanisms from\nheterogeneous data\, across domains and modalities\, to en able reliable\ntransfer and controllable generation. Next\, I will show ho w structured\nconcepts can be recovered even from seemingly unstructured d ata\, by\nanalyzing and improving self-supervised objectives (such as mask ing\nand diffusion) through hierarchical latent-variable models. These\nco ncept structures can then be used to interpret generative models and\nsupp ort targeted\, multi-level edits. Finally\, I connect these two\nthreads t o generalization beyond the training distribution. I will\ndiscuss natural conditions for extrapolation and a compositional\ngeneration framework th at improves prompt following for novel concept\ncombinations. I will concl ude with a brief outlook on self-improving\nworld models and AI-assisted s cientific discovery.\n\nPresented in Partial Fulfillment of the CSD Speaki ng Skills\nRequirement\n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911ed6e DTSTART;TZID=America/New_York:20260414T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260414T135000 LOCATION:Hamburg Hall 1002 and Zoom SUMMARY:Privacy Seminar - Shreya Kochar CLASS:PUBLIC DESCRIPTION:Speaker: SHREYA KOCHAR\, Pre-doctoral Research Fellow\, Softwar e and\nSocietal Systems Department\, Carnegie Mellon University\n\nTalk Ti tle: Predictive Privacy: Quantitatively Modeling Privacy Harms\n\nPrivacy has long been recognized as important\, yet its harms remain\ndifficult to define\, with courts and regulators often dismissing\nviolations as too a bstract to warrant action. This challenge has\nintensified in the digital age\, where data-driven applications\nincreasingly collect/share personal information\, and data brokers\naggregate information across sources. Addi tionally\, machine learning\nsystems use this data to infer sensitive attr ibutes about people that\nare not directly observed. These inferences may not be necessarily\naccurate\, but they still have the potential to cause harm (ex.\nreputational damage\, anxiety\, etc). \n\nHowever\, courts an d regulators often fail to recognize privacy harms\ndue to their intangibl e and difficult-to-define nature\, as seen in\ncases like Spokeo v. Robin s and TransUnion LLC v. Ramirez. Filing\nsuit in Federal court requires standing\, but as Justice Kavanaugh\nremarked in TransUnion\, “No concr ete harm\, no standing.” How\,\nthen\, are we to understand and address the effects of privacy harms?\n\nTo address this gap\, we propose Predict ive Privacy\, a framework for\nmodeling and quantifying privacy harm in mo dern data ecosystems. Our\napproach introduces a “harm function” that accounts for attribute\nvalues\, inference probabilities\, social norms\, and the identity of the\nobserver\, allowing harm to be structured and app roximated. Using\nsynthetic data representative of the U.S. population\, s urvey responses\nfrom human evaluators\, and machine learning clustering t echniques\, we\nempirically estimate perceived harm across scenarios.\n\nO ur findings show that individuals consistently perceive harm from the\nexp osure of sensitive data\, even absent tangible consequences. This\nwork es tablishes a quantitative model of privacy harm and offers a\npractical met hodology for assessing injury in legal and regulatory\ncontexts.\n\n—\n\ nShreya Kochar is currently a Pre-doctoral Research Fellow at Carnegie\nMe llon\, working with Professor Norman Sadeh. Her research sits at the\ninte rsection of algorithmic auditing\, policy\, and security\, with a\nfocus o n mitigating privacy risks in data aggregation and\ncommercialization. She completed her masters degree at Columbia\nUniversity and her undergraduat e degree at Wellesley College.\nPreviously\, she worked at Microsoft as a software engineer for 3\nyears. She will be starting her PhD in the fall.\ n\nSpring 2026 Privacy Seminars are sponsored by the Masters in Privacy\nE ngineering Program and the Carnegie Bosch Institute \n\nIn Person and Zoo m Participation. See announcement.\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911f241 DTSTART;TZID=America/New_York:20260415T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260415T103000 LOCATION:Rescheduled SUMMARY:Doctoral Thesis Oral Defense - Nicole Feng - Talk Rescheduled CLASS:PUBLIC DESCRIPTION:Speaker: NICOLE FENG\, Ph.D. CandidateComputer Science\nDepartm entCarnegie Mellon University\n\nTalk Title: Algorithms for Generalized Si gned Distance and Winding\nNumbers\n\nIn this talk\, I'll discuss algorith ms for generalized inside/outside\ncomputation (via winding numbers) and signed distance computation. By\n\"generalized\"\, I mean that these algor ithms make geometric inferences\nfrom imperfect data comprising incomplete \, inaccurate\, or ambiguous\nobservations or representations of shapes. I n other words\, these\nalgorithms generalize from imperfect data and impli citly approximate\nthe true underlying curve or surface. A theme is that g eneralization\ncan often be achieved by processing globally-defined functi ons\nencoding the geometry of interest\, rather than the original\, defect ive\ncurve or surface. For both inside/outside and signed distance\ncomput ation we can unlock further control over geometry and topology\nby process ing higher-order derivatives of these functions. Another\ntheme is that in side/outside and signed distance computation are\nclosely related problems \; towards this end\, we provide a formalization\nof their relationship th at justifies the design of our algorithms.\n\nThesis Committee\n\nKeenan C rane (Chair)\n\nNancy Pollard\n\nIoannis Gkioulekas\n\nChris Wojtan (Insti tute of Science and Technology Austria)\n\nIn Person and Zoom Participatio n.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911f61f DTSTART;TZID=America/New_York:20260422T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260422T130000 LOCATION:Newell Simon Hall 3305 and Zoom SUMMARY:AI-SDM Seminar - Aaditya Ramdas CLASS:PUBLIC DESCRIPTION:Speaker: AADITYA RAMDAS\, Associate Professor\, Department of S tatistics\nand Data Science and Machine Learning Department\, Carnegie Mel lon\nUniversity\n\nTalk Title: Conformal Changepoint Localization\n\nWe re visit a classical problem of identifying the location of a\nchangepoint\, but from a modern distribution-free lens. Our algorithm\,\nCONformal CHang epoint localization (CONCH) produces a confidence set\nfor the index at wh ich a change occurs without making any assumptions\nabout the distribution s before or after the change. Existing methods\noften rely on parametric a ssumptions\, tail conditions\, or asymptotic\napproximations\, or only pro duce point estimates. By proving a\nconformal Neyman-Pearson lemma\, we de rive principled score functions\nthat yield informative (small) sets\, who se normalized length shrinks\nto zero under weak assumptions. We also esta blish a universality\nresult showing that any distribution-free changepoin t localization\nmethod must be an instance of CONCH. Experiments suggest t hat CONCH\ndelivers precise confidence sets even in challenging settings\n involving images or text.\n\n—\n\nAaditya Ramdas is an Associate Profess or (with tenure) at Carnegie\nMellon University in the Department of Stati stics and Data Science and\nthe Machine Learning Department. He was a post doc at UC Berkeley\n(2015–2018) mentored by Michael Jordan and Martin Wa inwright\, and\nobtained his PhD at CMU (2010–2015) under Aarti Singh an d Larry\nWasserman\, receiving the Umesh K. Gavaskar Memorial Thesis Award . His\nundergraduate degree was in Computer Science from IIT Bombay (2005- 09\,\nAll India Rank 47). His work has been recognized by the Presidential \nEarly Career Award (PECASE)\, the highest distinction bestowed by the\nU S government to young scientists. He has also received a Kavli\nfellowship from the National Academy of Sciences\, a Sloan fellowship\nin Mathematic s\, the CAREER award from the National Science Foundation\,\nthe Emerging Leader Award from COPSS (Committee of Presidents of\nStatistical Societies )\, early career awards from the Bernoulli Society\nand the Institute of M athematical Statistics\, and faculty research\nawards from Adobe and Googl e. He was recently elected Fellow of the\nIMS\, was awarded Statistician o f the Year 2025 by the ASA’s\nPittsburgh Chapter\, and is the program ch air of AISTATS 2026. He has\npublished over 150 peer-reviewed papers\, abo ut half at top journals\nlike The Annals of Statistics\, Biometrika\, IEEE Transactions on\nInformation Theory and PNAS\, including prestigious disc ussion papers\nat the Journal of the Royal Statistical Society and Journal of the\nAmerican Statistical Association\, and about half at the top AI\n conferences like NeurIPS\, ICML\, ICLR\, UAI and AISTATS\, including over\ na dozen orals/spotlights. He has given several keynote talks\,\nincluding at Lunteren\, AISTATS and VCMF\, and invited tutorials at\nCUSO\, KDD and ICML.\n\nAaditya’s research in mathematical statistics and learning has an\neye towards designing algorithms that both have strong theoretical\ng uarantees and also work well in practice. His main interests include\npost -selection inference (multiple testing\, simultaneous inference)\,\ngame-t heoretic statistics (e-values\, confidence sequences) and\npredictive unce rtainty quantification (conformal prediction\,\ncalibration). His areas of applied interest include privacy\,\nneuroscience\, genetics and auditing (elections\, real-estate\, finance\,\nfairness). He co-organizes of the St atML Group at CMU. He loves to\ntalk about backpacking adventures through over 70 countries\,\ntrash-free living\, completing the Ironman triathlon\ , long-distance\nbicycle rides\, books and parenthood.\n\n REGISTER → r egister to attend in-person or on Zoom\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911fb88 DTSTART;TZID=America/New_York:20260409T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260409T173000 LOCATION:Gates Hillman 7501 and Zoom SUMMARY:Crypto Seminar - Zachary DeStefano CLASS:PUBLIC DESCRIPTION:Speaker: ZACHARY DeSTEFANO\, Ph.D. Candidate\, Computer Science \nDepartment\, New York University\n\nTalk Title: Sum-check protocol for a pproximate computations\n\nMotivated by the mismatch between floating-poin t arithmetic\, which is\nintrinsically approximate\, and verifiable comput ing protocols for\nexact computations\, we develop a generalization of the classical\nsum-check protocol. Our generalization proves claims of the fo rm ∑x\n∊g(x) ≈ H\, where g is a low-degree v-variate polynomial ov er an\nintegral domain 𝕌. The verifier performs its check in each round of\nthe protocol using a tunable error parameter δ. If Δ is the error i n\nthe prover's initial claim\, then the soundness error of our protocols\ ndegrades gracefully with δ/Δ .\n\nUnlike the classical sum-check protoc ol\, which is fundamentally\nalgebraic\, our generalization exploits the m etric structure of\nlow-degree polynomials. The protocol can be instantiat ed over various\ndomains\, but is most natural over the complex numbers\, where the\nanalysis draws on the behavior of polynomials over the unit cir cle. We\nalso analyze the protocol under the Fiat-Shamir transform\, revea ling a\nnew \"intermediate security\" phenomenon that appears intrinsic to \napproximation.\n\nPrior work on verifiable computing for numerical tasks typically\nverifies that a prover exactly executed a computation that onl y\napproximates the desired function. In contrast\, our protocols treat\na pproximation as a first-class citizen: the verifier's checks are\nrelaxed to accept prover messages that are only approximately\nconsistent with the claimed result. This establishes the first\nblack-box feasibility result for approximate arithmetic proof systems:\nthe protocol compiler is indepe ndent of how arithmetic operations are\nimplemented\, requiring only that they satisfy error bounds. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef911ffb3 DTSTART;TZID=America/New_York:20260408T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260408T123000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY: 5th Year Master of Science Thesis Defense - Michael Cui CLASS:PUBLIC DESCRIPTION:Speaker: MICHAEL CUI\, Master's Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: A Unified Framework fo r Paper Assignment\n\nAssigning submitted papers to appropriate reviewers is a fundamental\ncomponent of the peer-review process in large academic c onferences. In\nmodern conference settings\, this task has become increasi ngly\nchallenging due to the scale of submissions and the need to satisfy\ nmultiple competing objectives simultaneously. In particular\, program\nc hairs must balance reviewer expertise\, diversity considerations\, and\nro bustness to strategic behavior\, while also ensuring that the\nassignment process remains practical at scale. This thesis aims to\nimprove the paper -assignment process by making it more effective\, more\nrobust\, and more practical for real-world peer review.\n\nThis thesis studies the problem o f large-scale paper assignment from\nan optimization perspective. It exami nes the limitations of existing\nassignment methods\, which often optimize only a subset of the relevant\nobjectives or become computationally impra ctical in realistic\nconference settings. To address these limitations\,  the thesis\npresents Robust Assignment via Marginal Perturbation (RAMP)\, a unified\nframework for scalable\, robust\, and diversity-aware reviewer\ nassignment. The proposed framework combines a linearized\nperturbed-maxim ization objective with soft constraints that\nincorporate multiple practic al desiderata into a single optimization\nprocedure\, together with an att ribute-aware sampling method for\nconverting fractional assignments into i ntegral ones.\n\nIn addition to presenting the algorithmic framework\, thi s thesis\ndiscusses the practical challenges and lessons that arose in dep loying\nthe method for major AI conferences\, including AAAI 2026\, AAMAS 2026\,\nand EC 2026. It also describes an interface that enables future\nc onference organizers to run the matching process directly\, helping\nbridg e the gap between optimization research and real\nconference workflows.\n \nThesis Committee\n\nFei Fang (Chair)\n\nNihar Shah\n\nAdditional Informa tion \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91203f7 DTSTART;TZID=America/New_York:20260413T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260413T130000 LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar - Nihar Shah CLASS:PUBLIC DESCRIPTION:Speaker: NIHAR SHAH\, Associate Professor\, Computer Science De partment\nand Machine Learning Department\, Carnegie Mellon University\n\n Talk Title: Security\, Privacy\, and Integrity in Peer Review\n\nScientifi c peer review is a high-stakes distributed system\, and like\nany such sys tem\, it presents a large attack surface for exploitation.\nAs a result\, adversarial actors constantly attempt to compromise the\nsystem\, through tactics including identity theft\, collusion rings\,\nreviewer snitching\, and other forms of manipulation. We will discuss\ninvestigations into the se attacks\, present existing defenses\, and\noutline a roadmap of high-im pact open problems where security research\ncan substantially strengthen t he integrity of science.\n\n—\n\nNihar B. Shah is an Associate Professor in the Machine Learning and\nComputer Science departments at CMU. His res earch is on the Science of\nEvaluation and the Evaluation of Science. His group develops\ncomputational tools\, with rigorous mathematical guarantee s\, and\ndesigns and conducts empirical experiments for evidence-based pol icy\ndesign. His work has been used in the evaluation of several hundred\n thousand papers and thousands of proposals in well over 200 venues\, as\nw ell as in applications such as competitions and admissions. He is a\nrecip ient of the David J. Sakrison memorial prize from UC Berkeley for\na \"tru ly outstanding and innovative PhD thesis\,\" the Young Alumnus\nMedal from the Indian Institute of Science\, a JP Morgan faculty\nresearch award\, G oogle Research Scholar Award\, an NSF CAREER Award\n2020-25\, the Microsof t Research PhD Fellowship 2014-16\, the Berkeley\nFellowship 2011-13\, and several Best Paper Awards.\n\nIn Person and Zoom Participation.  See ann ouncement.\n\n→ Cylab Seminars are only open to CyLab Partners and curre nt CMU\nfaculty\, staff and students. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9120cdc DTSTART;TZID=America/New_York:20260408T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260408T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Shyamal Patel CLASS:PUBLIC DESCRIPTION:Speaker: SHYAMAL PATEL\, Ph.D. StudentTheory of Computation\nGr oupColumbia University\n\nTalk Title: Faster Algorithms for Learning k-Ter m DNFs with Queries\n\nA fundamental task in learning theory is to unders tand the\ncomputational complexity of learning k-term DNFs in the\ndistrib ution-free PAC learning model. Unfortunately\, given only random\nexamples \, a variety of lower bounds suggest that learning DNFs with a\nsuperconst ant number of terms requires superpolynomial time. That\nsaid\, work by Bl um and Rudich showed that such a bound can be\ncircumvented if we have que ry access to the function\, proving k-term\nDNFs can be learned in time n · 2k in this setting.\n\nIn this talk\, we’ll describe an algorithm to learn k-term DNFs in\ntime n · 2√k. To do so\, we construct an adapt ive set of features\nthat allow us to represent the function as the sign o f a low-degree\npolynomial over these features. We then use tools from jun ta testing\nand attribute-efficient learning to effectively reduce the num ber of\nvariables and achieve our result.\n\nBased on joint work with Josh Alman\, Shivam Nadimpalli\, and Rocco\nServedio\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91210c4 DTSTART;TZID=America/New_York:20260409T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260409T160000 LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Victor Reiner CLASS:PUBLIC DESCRIPTION:Speaker: VICTOR REINER\, Professor\, School of Mathematics\, Un iversity\nof Minnesota\n\nTalk Title: Ehrhart theory and a q-analogue\n\nC lassical Ehrhart theory begins with this fact: for a convex polytope\nP wh ose vertices lie in the integer lattice Zn\, the number of lattice\npoints in the positive integer dilates mP grows as a polynomial\nfunction of m. We will review highlights of the classical theory\, and\nexplain a new \"q -analogue\": it replaces the number of lattice points\nin mP by a polynomi al in q that specializes to the lattice point count\nat q=1. There are q-a nalogues for many classical Ehrhart theory\nresults\, some proven\, others conjectural. In particular\, a certain new\ncommutative ring plays an int riguing role.\n\n(Based on with Brendon Rhoades.)\n\n4:00 pm → Jane Stre et-sponsored tea and cookies in the Math Lounge\n(bring your mug!). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9121421 DTSTART;TZID=America/New_York:20260410T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260410T140000 LOCATION:Tepper 2700 SUMMARY:Operations Research Seminar - Stephen Arndt CLASS:PUBLIC DESCRIPTION:Speaker: STEPHEN ARNDT\, Ph.D. Student\, Ph.D. Program in Algor ithms\,\nCombinatorics and Optimization\, Tepper School of Business\n\nTal k Title: Approximation Algorithms for Matroid-Intersection Coloring\nwith Applications to Rota's Basis Conjecture\n\nWe study algorithmic matroid in tersection coloring. We give the first\npolynomial-time O(1)-approximation algorithm to color O(1) general\nmatroids. Notably\, for two general matr oids we achieve a\n2-approximation. Furthermore\, we give a fully polynomi al randomized\napproximation scheme (FPRAS) for coloring the intersection of two\nmatroids when the maximum chromatic number is large. This yields t he\nfirst polynomial-time algorithm for an asymptotic variant of Rota's\nB asis Conjecture. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9121749 DTSTART;TZID=America/New_York:20260407T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260407T120000 LOCATION:Gates Hillman 6121 SUMMARY:Carnegie Mellon AI Safety Initiative Seminar - Nikola Jurković CLASS:PUBLIC DESCRIPTION:Speaker: NIKOLA JURKOVIĆ\, Member of Technical Staff\, Model E valuation\nand Threat Research (METR)\n\nTalk Title: How to check if an AI model is safe? Lessons from\nevaluating frontier AI\n\nLLMs are quickly l earning capabilities needed to increase catastrophic\nrisks. When a model pushes the frontier of capabilities\, how can we\nknow whether it's safe o r not? In this talk\, I will go into METR's\nrecent work on evaluating fro ntier AI models. I will talk about\nmeasuring AI capabilities over time\, assessing risks from frontier AI\nmodels\, and what AI evaluations and saf ety assessments might look like\nin the future as we approach AGI.\n\n—\ n\nNikola Jurković is a researcher at METR (Model Evaluations & Threat\nR esearch) and works on evaluating frontier models and assessing model\nrisk s. Recently\, he led METR's reviews of Anthropic's Sabotage Risk\nReports and  contributed to METR's pre-deployment evals of GPT-5 and\nGPT-5.1-Cod ex-Max.\n\nLearn More about CASI\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9121abd DTSTART;TZID=America/New_York:20260409T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260409T133000 LOCATION:Newell-Simon 3305 and Zoom SUMMARY:AI-SDM Seminar - Emma Brunskill CLASS:PUBLIC DESCRIPTION:Speaker: EMMA BRUNSKILL\, Associate Professor\, Computer Scienc e\nDepartment\, Stanford University\n\nTalk Title: Estimating the Value of Personalization\n\nFrom medicine to marketing to social sciences\, the pr omise of\ntailoring interventions to individual characteristics is undenia ble.\nHowever\, personalization often comes with costs— from logistical\ nchallenges to lack of shared context to concerns about fairness. In\naddi tion\, personalized decision policies can be more fragile\, because\nthey typically require more data to learn accurately compared to\nidentifying a single best intervention for all. In this talk I’ll\nintroduce a new st atistical estimator that quantifies\, given\nhistorical data\, if there is evidence that a personalized intervention\npolicy provides significantly superior expected outcomes compared to\ndeploying the best single overall intervention. We present results\nacross four diverse datasets— from job training to medical\ntreatments— to highlight the wide range of setting s where\nquantifying the impact of personalization can be helpful\, and th e\nstrength of our proposed estimator over prior related approaches.\n\nJo int work with Zhaoqi Li.\n\n—\n\nEmma Brunskill is an American computer scientist. Her research\ncombines machine learning with human–compute r interaction by\nstudying the effects of AI systems in human-centered ap plications\nincluding educational software and healthcare\, and the the ory\nof reinforcement learning in situations where mistakes impose high\ nrisks or costs. She is an associate professor of computer science\nat St anford University\, where she also holds a courtesy appointment\nin the S tanford Graduate School of Education and is an affiliate of\nthe King Cen ter on Global Development.\n\nREGISTER → register to attend in-person o r on Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9121ef4 DTSTART;TZID=America/New_York:20260408T084500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260408T101500 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Thesis Proposal - Bernardo Subercaseaux CLASS:PUBLIC DESCRIPTION:Speaker: BERNARDO SUBERCASEAUX\, Ph.D. Student\nComputer Scienc e Department\nCarnegie Mellon University\n\nTalk Title: SAT Encodings: Fro m Art to Science\n\nAutomated reasoning engines\, and SAT solvers in parti cular\, have\nbecome a powerful tool for tackling hard combinatorial probl ems. While\nSAT solvers have improved dramatically\, their effectiveness s till\ndepends critically on how problems are encoded into conjunctive norm al\nform (CNF). Encoding choices routinely account for runtime differences \nof several orders of magnitude\, yet their design remains more art than\ nscience—guided by intuition and hard-won experience rather than a\nprin cipled theory. This thesis aims to shed light on the design of\neffective SAT encodings\, from both practical and theoretical\nperspectives.\n\nThe first part is dedicated to the art of encodings. We show how\nclever probl em-specific encodings\, together with other automated\nreasoning technique s\, have allowed us to solve a variety of open\nproblems in discrete mathe matics ranging from graph coloring to\ndiscrete geometry. These encodings benefit from mathematical insights\nincorporated via additional constraint s or auxiliary variables that\nrepresent semantically important aspects of the problem. Notably\, some\nof these insights are themselves inspired by computational\nexperiments\, establishing a symbiotic relationship betwee n automated\nreasoning and discrete mathematics. We also apply this art to \ncomputational problems in explainable AI—uncovering patterns in\ndecis ion trees\, nearest-neighbor classifiers\, and other symbolic\nmodels—de monstrating the transferability of our techniques beyond\nmathematics.\n\n The second part is dedicated to the foundations of a science of\nencodings . We present a landscape of theoretical results regarding the\nnumber of c lauses and auxiliary variables required to encode several\nbuilding blocks of propositional encodings\, from cardinality\nconstraints to arbitrary k -CNF functions. We posit that encoding\nboolean functions into CNF with as few clauses as possible offers a\nfascinating yet largely unexplored terr itory for circuit complexity.\nBy leveraging auxiliary variables and wide clauses\, this\nclause-minimization model permits rich combinatorial struc tures that\nsurpass known lower bounds for circuits. Furthermore\, while\n clause-minimization does not always correlate with solver performance\,\nt heoretical developments in this model have led us to novel encodings\nthat run faster on practical problems. More broadly\, this thesis aims\nto est ablish that the clause-minimization model is not only an elegant\ntheory\, but also a realistic path towards empirical speedups.\n\nThesis Committee \n\nMarijn Heule (Chair)\n\nJeremy Avigad\n\nRuben Martins\n\nStefan Szeid er (Technische Universität Wien)\n\nRyan Williams (Massachusetts Institu te of Technology)\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9122407 DTSTART;TZID=America/New_York:20260406T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260406T110000 LOCATION:Newell-Simon 3305 SUMMARY:Doctoral Speaking Skills Talk - Ava Pun CLASS:PUBLIC DESCRIPTION:Speaker: AVA PUN\, Ph.D. Student\nComputer Science Department\n Carnegie Mellon University\n\nTalk Title: Make it Real: Generative AI for the Physical World\n\nGenerative AI systems can create high-quality digita l content\, such as\nimages\, videos\, and 3D assets. However\, most curre nt methods overlook\nan important sphere of everyday life: the physical wo rld. Generating\nblueprints for physically realizable objects—those mean t to be\nconstructed in the real world—presents novel challenges. The\nf inished object must be physically stable\, manufacturable from\nstandard c omponents\, and ideally\, buildable by automated robots. In\nthis talk\, I present BrickGPT and BrickMatic\, a step towards\nphysically realizable g eneration in the domain of LEGO bricks.\nFollowing a user prompt\, BrickGP T designs a structure made of bricks\,\nwhich is then built by a dual-arm robot system using BrickMatic\,\nresulting in a physical LEGO model. Brick GPT earned the Best Paper\nAward (Marr Prize) at ICCV 2025.\n\nPresented i n Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9122771 DTSTART;TZID=America/New_York:20260507T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260507T170000 URL:https://csd.cmu.edu/academics/doctoral-resources/dsr-schedule LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Student Review (DSR) - PL\, Security\, Systems (1-3) CLASS:PUBLIC DESCRIPTION:Talk Title: Doctoral Student Review (DSR) PL/Security/Systems\n \nProgramming Languages\, Security\, Systems - Years 1-3.\n\nPlease see em ails for additional information. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9122a2f DTSTART;TZID=America/New_York:20260507T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260507T120000 URL:https://csd.cmu.edu/academics/doctoral-resources/dsr-schedule LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Student Review - AI\, Graphics\, Theory (1-3) CLASS:PUBLIC DESCRIPTION:Talk Title: Doctoral Student Review (DSR) AI/Graphics/Theory\n\ nArtificial Intelligence\, Graphics\, Theory - Years 1-3 \n\nPlease see e mails for additional information.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9122cfa DTSTART;TZID=America/New_York:20260406T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260406T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Tyler Cloutier CLASS:PUBLIC DESCRIPTION:Speaker: TYLER CLOUTIER\, Chief Executive Officer and Co-founde r\nClockwork Labs\n\nTalk Title: Inverting the Backend: Why We Built our M MORPG Inside a\nDatabase\n\nThis talk explores what happens when you move all the infrastructure\nand logic of a fully featured\, real-time backend application into the\ndatabase. Using SpacetimeDB and BitCraft as a case s tudy\, the talk\nargues that the ideas behind stored procedures deserve a modern\nrevival as a first class database feature: bringing deployment\,\n transactions\, relational data\, and general-purpose procedural logic\nint o a single system. It covers why this architecture made sense for\nour MMO RPG\, why Postgres was insufficient\, how it simplifies the\ndesign of lar ge-scale realtime backends\, and how SpacetimeDB uses\nincremental query e valuation and high-performance OLTP execution to\nhandle everything from r eal-time chat\, to user management\, to player\nmovement\, all within the database.\n\n—\n\nTyler Cloutier is the CEO and co-founder of Clockwork Labs\, the\ndevelopers of BitCraft Online and SpacetimeDB. SpacetimeDB is a\nreal-time database system which hosts server-side programs written in\n familiar procedural languages (TypeScript\, Rust\, C#\, etc.) and serves\n as the complete backend for BitCraft Online.\n\nThis talk is part of the P ostgreSQL vs. The World Seminar Series\n\nZoom Participation.  See announ cement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91230de DTSTART;TZID=America/New_York:20260402T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260402T130000 LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI-SDM - Student Brainstorming Session CLASS:PUBLIC DESCRIPTION:Talk Title: AI-SDM Student Brainstorming Session\n\nJoin the mo nthly student-focused brainstorming sessions.\n\nLed by students of all le vels\, these sessions offer a casual forum to\ndiscuss AI-related research while exploring cross-cutting connections\nbetween various disciplines in artificial intelligence. All students\nwelcome.\n\nStudents meet regularl y to participate in informal discussion sessions\nthat delve into cutting- edge AI topics. These discussions are a\nbreeding ground for collaboration \, innovative thinking\, and\nproblem-solving from the ground up. They pro vide a stress-free forum\nfor exchanging ideas\, brainstorming new approac hes to challenges\, and\nfostering lasting connections within the AI-SDM c ommunity in an\nenvironment distinct from a traditional seminar. Lunch is provided for\nin-person attendees! Please complete the RSVP form below to help us\nwith logistics.\n\nREGISTER → please register to attend in-pe rson or on Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912343a DTSTART;TZID=America/New_York:20260408T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260408T170000 LOCATION:Gates Hillman 8102 SUMMARY:Doctoral Speaking Skills Talk - Timothy Kim CLASS:PUBLIC DESCRIPTION:Speaker: TIMOTHY KIM\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: TCO-driven Storage Provision ing for Exascale Data Centers\n\nRecent changes in data temperatures and s torage device\ncharacteristics\, both mechanical disk-drives (HDDs) and so lid-state\ndrives (SSDs)\, expand the set of deployment options for exasca le\nstorage. Until recently\, exascale storage systems followed a pattern\ nof placing most data on HDDs with smaller amounts of SSD storage used\nfo r caching and performance-critical workloads. Exascale storage\nprovisioni ng and dataset placement trade-offs have now changed.\n\nThis paper descri bes a total cost of ownership (TCO) model that\ncaptures primary aspects o f modern deployments and uses it to explore\nthe new trade-off space. Usin g capacity and performance telemetry\ninformation for 43 production datase ts+workloads at two large\nhyperscalers\, we show significant changes from prior analyses of\nworkloads and storage placement decisions across a mul titude of\nstorage device types. We also introduce a storage cluster TCO\n optimizer that identifies the lowest-TCO grouping and assignment of\ndatas ets to device types\, exposing a number of insights that can help\nguide f uture deployments and research. For example\, our analysis shows\nthat the highest-density SSDs are particularly favorable for clusters\nwith heavy AI/ML workloads but are only cost-effective at exascale\nwhen combined wit h high-density HDDs. Finally\, we use our framework to\nevaluate how stora ge provisioning and overall TCO change as a function\nof key parameters li ke device write amplification\, cluster power\nbounds\, and the maximum nu mber of device types allowed.\n\nPresented in Partial Fulfillment of the C SD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912381a DTSTART;TZID=America/New_York:20260413T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260413T163000 LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar CLASS:PUBLIC DESCRIPTION:Speaker: SAI TEDLA\, Ph.D. Student\nDept. of Electrical Enginee ring and Computer Science\nYork University\, Toronto\n\nTalk Title: Genera tive Re-Photography with Video Models\n\nI will introduce \"generative re- photography\" methods that use new\ngenerative video models to get more ou t of your photos—even the\nblurry ones. First\, I will present a method for converting\nmotion-blurred images to video. This method can even predi ct the\n\"past\" and \"future\" (right before and after the capture) of a\ nmotion-blurred image. I will then show how this method can bring\n\"histo rical scenes to life\" such as photos of soldiers landing on\nnorth side o f France during the Normandy invasion of 1944 or a boxing\nmatch between M ohammed Ali and Jurgen Blin in 1971. Then\, I will\npresent a robust post- capture refocusing method that converts a single\ndefocus-blurred image in to a focal stack spanning multiple focus\ndistances. Our work overturns th e conventional wisdom of photography\,\nsuggesting these \"corrupted image s\" can actually reveal more about the\nworld than the \"perfect\" images which have been the holy grail of\nimage processing. Additionally\, our fi ndings suggest that video models\nimplicitly understand how camera capture settings affect image\nappearance\, and I will discuss how this exciting capability could\ninspire new directions for computational photography. \ n\n—\n\nSai Tedla is a PhD student at York University\, Toronto\, superv ised by\nMichael Brown. He currently works on the intersection of computat ional\nphotography and generative models. He is a visiting student at the\ nUniversity of Toronto supervised by David Lindell and Kyros Kutulakos\,\n and will soon join the university as a Schmidt AI Postdoctoral Fellow.\nAd ditionally\, Sai is a current intern at Sony AI Japan and has\npreviously interned at Samsung AI Center Toronto and Adobe NextCam.\n\nThe VASC semin ar is generously sponsored by HeyGen\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9123c23 DTSTART;TZID=America/New_York:20260407T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260407T113000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:Doctoral Thesis Oral Defense - Catalina Vajiac CLASS:PUBLIC DESCRIPTION:Speaker: CATALINA VAJIAC\, Ph.D. Candidate\nComputer Science De partment\nCarnegie Mellon University\n\nTalk Title: Detection and Visualiz ation of Human Sex Trafficking in\nOnline Escort Advertisements\n\nHuman t rafficking (HT) for forced sexual exploitation is incredibly\npervasive\, affecting an estimated 6.3 million people at any given\ntime. The majority of victims are advertised online\, mainly through\nonline escort websites \, alongside at-will escorts. Practitioners\nwho want to help these victi ms\, including government organizations\,\ncriminologists\, social workers \, and investigators\, often manually\nscroll through these escort website s to try to find HT leads by\nlooking for known keywords\, geographic move ment\, or other known HT\nsignals indicating a person was advertised again st their will. This\nmanual process is inefficient\, as it requires lots o f time\, and\nineffective\, as traffickers change their patterns and keywo rds over\ntime to avoid detection. In addition\, since the majority of HT cases\nare part of organized crime groups\, practitioners realized a more\ nreliable HT indicator: groups of ads with nearly-identical text that\nadv ertise multiple people\, signaling larger organized activity than\nindivid ual escorts would post. These insights can be leveraged to help\nfacilitat e lead generation for practitioners\, enabling them to act\nmore quickly t o get HT victims out of exploitation.\n\nIn this thesis\, we assist practi tioners in identifying potential HT\ncases by: (1) developing scalable and explainable clustering\nalgorithms based on text for finding and summari zing organized crime\ngroups in escort ad data\, and (2) creating intuitiv e visualization\nmethods for presenting the results of these to practition ers. These\nvisualizations not only help practitioners to better understan d\npotential leads\, but they also facilitate label generation so\ndownstr eam algorithm evaluation can continue even as traffickers\nchange their pa tterns. In addition\, the methods outlined in this\nthesis have real-world impact\; they are currently being integrated by\nindustry practitioners.\ n\nThesis Committee\n\nChristos Faloutsos (Chair)\n\nRayid Ghani\n\nAdam P erer\n\nDuen-Horng Chau (Georgia Institute of Technology) \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91240c6 DTSTART;TZID=America/New_York:20260402T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260402T160000 LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Jonathan Tidor CLASS:PUBLIC DESCRIPTION:Speaker: JONATHAN TIDOR\, Assistant Professor\nDepartment of Ma thematics\nPrinceton University\n\nTalk Title: Discrete geometry\, semialg ebraic graphs\, and the\npolynomial method\n\nMany problems in discrete ge ometry can be encoded combinatorially via\na structure known as a semialge braic graph — a graph whose edge\nrelations are defined by polynomial eq uations and inequalities. These\ninclude the Erdős unit distance problem\ , incidence results of\nSzemerédi—Trotter\, and numerous other problems . In this talk\, I\nwill introduce these ideas and discuss what is known a bout the class\nof semialgebraic graphs. This includes new structural and extremal\nresults: a very strong (and quantitatively optimal) regularity l emma\nas well as improved Turán-type bounds for semialgebraic graphs. The \nproofs of these results are inspired by an idea of Bukh and Vasileuski\n and use novel polynomial method tools\, which generalize the polynomial\np artitioning machinery of Guth—Katz and of Walsh.\n\n4:00 pm → Jane St reet-sponsored tea and cookies in the Math Lounge\n(bring your mug). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9124466 DTSTART;TZID=America/New_York:20260403T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260403T143000 LOCATION:Newell-Simon 3002 SUMMARY:Doctoral Thesis Proposal - Andrew Park CLASS:PUBLIC DESCRIPTION:Speaker: ANDREW PARK\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Attacking and Designing Secu re Protocols for Database\nSearch\n\nThe increasing reliance on outsourced storage and cloud-based\ncomputing has made database accesses a fundament al problem in modern\nsecurity and cryptography. At its core\, the challen ge of accessing\ndatabases securely is to enable efficient query processin g over a\n(potentially encrypted) database without leaking private informa tion\nto untrusted parties. Large lines of work have yielded a plethora of \ndifferent solutions such as searchable encryption\, private information\ nretrieval (PIR)\, and oblivious RAM (ORAM)\, each offering various\ntrade offs among efficiency\, functionality\, and privacy. However\, these\nsolu tions have yet to find wide-scale adoption due to practical\ninefficiencie s. In addition\, their real-world implementations often\ncontain subtle le akage of the underlying private data.\n\nThis dissertation develops a unif ied perspective on secure database\nsearch by studying both attacks and de signs of new secure protocols\nfor database search. On the attack side\, w e design and implement a\nframework\, Polysys\, for modeling and analyzing leakage patterns. This\nresult demonstrates an empirical understanding of how theoretical\nprivacy guarantees behave in practice.\n\nOn the constru ction side\, we introduce new protocols for both PIR and\nORAM to motivate more real-world adoption of these protocols. For PIR\,\nwe design novel c ommunication-efficient\, fault-tolerant multi-server\nschemes that also su pport erasure-coded storage\, providing robustness\nagainst fail-stop or B yzantine servers. For ORAM\, we present two new\ndesigns for different wor kload settings. First\, we design a garbled\nRAM (GRAM) scheme which match es the interactive state of the art\nscheme and achieves practical improve ment against the prior state of\nart. Finally\, we propose to construct di stribution-aware ORAM schemes\,\nwhich provides ORAM-like security propert ies for a set distribution\nwhile providing better practical efficiency.\n \nOur results highlight the importance of accounting for leakage in\nreal- world settings and considering real-world limitations when\ndesigning secu re protocols.\n\nThesis Committee\n\nElaine Shi (Co-chair)\n\nWenting Zhen g (Co-chair)\n\nAayush Jain\n\nSeny Kamara (MongoDB)\n\nTarik Moataz (Mong oDB)\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912491d DTSTART;TZID=America/New_York:20260401T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260401T120000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Speaking Skills Talk - Catalina Vajiac CLASS:PUBLIC DESCRIPTION:Speaker: CATALINA VAJIAC\, Ph.D. Student\nComputer Science Depa rtment\nCarnegie Mellon University\n\nTalk Title: Modeling &amp\; Visualiz ation for Social Good: A Case Study\nin Human Trafficking Detection\n\nOne of machine learning's greatest promises is to help decision-makers\nmake more data-driven\, societally beneficial decisions\, as evidenced\nby coun tless social-good applications of ML used to motivate thousands\nof papers each year. We can increase the potential impact of\nsociety-facing ML res earch through participatory design\, where\nstakeholders and those impacte d by the problem provide feedback\nthroughout every step of the design\, i mplementation\, and validation of\na potential solution.\n\nThrough the le ns of one particular problem domain — flagging\npotential organized sex trafficking groups in online escort\nadvertisement data for further invest igation — this talk provides\nexamples of how continuous discussions wit h stakeholders guided our\nresearch and led to important insights.\n\nPres ented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9124c8d DTSTART;TZID=America/New_York:20260331T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260331T135000 LOCATION:Hamburg Hall 1002 and Zoom SUMMARY:Privacy Seminar - Saranya Vijayakumar CLASS:PUBLIC DESCRIPTION:Speaker: SARANYA VIJAYAKUMAR\, Ph.D. Candidate\, Computer Scien ce\nDepartment\, Carnegie Mellon University\n\nTalk Title: Sequential Patt ern Recognition Attacks Against Deployed\nTopic-Based Mechanisms\n\nPrivac y-preserving AI systems like Google’s Topics API attempt to\nprotect use r privacy through behavioral aggregation\, but fail against\nrealistic att acks. We demonstrate that a transformer-based sequential\npattern recognit ion framework achieves 33.96% re-identification\naccuracy on web browsing data and up to 95.67% on music listening\nbehavior\, compared to prior met hods achieving ≤15% accuracy. The\napproach exploits temporal consistenc y in topic assignments through\nhierarchical attention mechanisms. Standar d protections prove\ninadequate: extending observation windows from 3 to 8 epochs improves\nattack success by 90.1%\, while industry-standard 5% noi se injection\nprovides minimal protection. These results show that privacy \nmechanisms designed without considering modern machine learning\ncapabil ities systematically fail their stated objectives.\n\n—\n\nSaranya Vijay akumar is a Ph.D. candidate in Computer Science at\nCarnegie Mellon Univer sity\, advised by Christos Faloutsos and Matt\nFredrikson. Her research fo cuses on AI security\, privacy\, and agentic\nalignment. Before CMU\, she earned her A.B. from Harvard with a joint\nconcentration in Computer Scien ce and Government\, working with Cynthia\nDwork and Jim Waldo on algorithm ic fairness\, and spent three years as\na data scientist in electronic tra ding at Goldman Sachs. She has held\nresearch positions at IBM Research\, Inria\, and Fujitsu Research\, and\nis supported by the DoD NDSEG Fellowsh ip. \n\nSpring 2026 Privacy Seminars are sponsored by the Masters in Priv acy\nEngineering Program and the Carnegie Bosch Institute.\n\nIn Person an d Zoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91250bd DTSTART;TZID=America/New_York:20260429T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260429T170000 LOCATION:SCS Events - 4th Floor Commons\, Gates Hillman (and all across cam pus) SUMMARY:Meeting of the Minds CLASS:PUBLIC DESCRIPTION:Speaker: Undergraduate students present their research\n\nThe U ndergraduate Research Symposium\, or the \"Meeting of the Minds\,\"\nis a university-wide celebration of undergraduate research. All\nCarnegie Mello n undergrads engaged in research and creative projects\nare encouraged to apply / participate.\n\nThe Meeting of the Minds is a great opportunity to share your research\nwith the entire campus community. All research proje cts are welcome -\nyou don't have to be working on a SURG or SURF project to present.\n\nStudents: Register to Participate \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91253d1 DTSTART;TZID=America/New_York:20260330T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260330T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Marcel Kornacker CLASS:PUBLIC DESCRIPTION:Speaker: MARCEL KORNACKER\, Chief Technology Officer and Co-fou nder\,\nPixeltable\n\nTalk Title: Pixeltable: A DBMS for Multimodal AI App lications\n\nPixeltable is a novel OLTP DBMS specifically designed for mul timodal\nAI applications. It goes beyond the traditional roles of a dbms - data\nstorage and retrieval - by incorporating data generation into the\n table schema. Pixeltable natively supports multimodal data types\n(e.g.\, video\, images\, audio\, documents) in the context of traditional\nmulti-u ser transaction semantics. Unlike analytics-focused databases\,\nPixeltabl e also allows the user to express data generation\ndeclaratively\, via com puted columns and user-defined functions. This\nmakes database technology relevant again for the GenAI age. Pixeltable\nis fully open-source and ava ilable as a pip-installable Python\npackage\; it utilizes Postgres for loc al structured data.\n\n—\n\nMarcel  Kornacker is the CTO and co-founder of Pixeltable\, the\nfounder of Apache Impala\, and the co-creator of the Parquet file\nformat. He holds a PhD from UC Berkeley.\n\nThis talk is pa rt of the PostgreSQL vs. The World Seminar Series.\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912575f DTSTART;TZID=America/New_York:20260331T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260331T150000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Speaking Skills Talk - Edward Chen CLASS:PUBLIC DESCRIPTION:Speaker: EDWARD CHEN\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Bridging Usability and Perfo rmance: A Tensor Compiler for\nAutovectorizing Homomorphic Encryption\n\nH omomorphic encryption (HE) offers strong privacy guarantees by\nenabling c omputation over encrypted data. However\, the performance of\ntensor opera tions in HE is highly sensitive to how the plaintext data\nis packed into ciphertexts. Large tensor programs introduce numerous\npossible layout ass ignments\, making it both challenging and tedious\nfor users to manually w rite efficient HE programs.\n\nIn this work\, we present Rotom\, a compila tion framework that\nautovectorizes tensor programs into optimized HE prog rams. Rotom\nsystematically explores a wide range of layout assignments\, applies\nstate-of-the-art optimizations\, and automatically generates an\n equivalent\, efficient HE program. To unlock new avenues for\nperformance\ , Rotom introduces a lightweight conversion operator that\neasily modifies the underlying data layouts. Our evaluation\ndemonstrates Rotom scalably compiles all tensor workloads in under 5\nminutes and achieves up to 80× performance improvement over prior\nautovectorization systems.\n\nPresente d in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9125af9 DTSTART;TZID=America/New_York:20260401T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260401T143000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Speaking Skills Talk - Ruihang Lai CLASS:PUBLIC DESCRIPTION:Speaker: RUIHANG LAI\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Relax: Composable Abstractio ns for End-to-End Dynamic\nMachine Learning\n\nDynamic shape computations have become critical in modern machine\nlearning workloads\, especially in emerging large language models. The\nsuccess of these models has driven t he demand for their universal\ndeployment across a diverse set of backend environments. In this\npaper\, we present Relax\, a compiler abstraction f or optimizing\nend-to-end dynamic machine learning workloads. Relax introd uces a\ncross-level abstraction that encapsulates computational graphs\,\n loop-level tensor programs\, and external library calls in a single\nrepre sentation. Relax also introduces first-class symbolic shape\nannotations t o track dynamic shape computations globally across the\nprogram\, enabling dynamic shape-aware cross-level optimizations. We\nbuild an end-to-end co mpilation framework using the proposed approach\nto optimize dynamic shape models. Experimental results on LLMs show\nthat Relax delivers performanc e competitive with state-of-the-art\nsystems across various GPUs and enabl es deployment of emerging models\nto a broader set of emerging environment s\, including mobile phones\,\nembedded devices\, and web browsers.\n\nPre sented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n\ n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9125e93 DTSTART;TZID=America/New_York:20260413T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260413T183000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:The Alan J. Perlis (Fun) Lecture in Computer Science - David Sussil lo CLASS:PUBLIC DESCRIPTION:Speaker: DAVID SUSSILLO\, Technologist\, Neuroscientist\, and P rofessor\nMeta Reality Labs and\nAdjunct Professor\, Stanford University\n and SCS Alumnus\n\nTalk Title: Emergence: A Memoir of Boyhood\, Computatio n\, and the\nMysteries of Mind\n\nThe Alan J. Perlis (Fun) Lecture in Comp uter Science - An SCS\nDistinguished Lecture\n\nDavid Sussillo has made a career at the cutting edge of neuroscience\nand technology—yet his path there was anything but a straight line.\nBorn to drug-addicted parents in New Mexico\, he navigated a childhood\nmarked by violence and neglect. But a seed was planted at the\nunlikeliest of places—the local arcade.\n\nW hat followed is a remarkable journey of resilience and\ntransformation\, f rom the chaotic corridors of group homes to the halls\nof Columbia and Sta nford. Along the way\, Sussillo takes one on an\nilluminating tour of the century-long dance between neuroscience\,\nphysics\, and computation that has laid the groundwork for neural\nnetworks—the technology that drives modern artificial intelligence.\nAs he advances in the field\, working to demystify these networks\, he\nalso begins to pursue an answer to a more p ersonal question: why\, and\nhow\, did he succeed against all odds?\n\nDav id invites all of us on an unforgettable journey that bridges the\npersona l and the profound\, revealing how intricate complexities can\narise from simple beginnings.\n\n— \n\nDavid Sussillo spent five years in the Alb uquerque Christian\nChildren's Home and four years at the Milton Hershey S chool as a\nchild. After graduating high school\, he ultimately received a Ph.D.\nfrom Columbia University in Computational Neuroscience. Now David is\nan adjunct professor at Stanford University and has been a scientist\n at the Google Brain group (recently featured in the book Genius\nMakers) and Meta Reality Labs. In his professional pursuits\, David\nresearches br ain-machine interfaces to develop the next generation of\ncomputers. David was the recipient of a Fulbright research grant and\nis an internationall y recognized neuroscientist with over 40\npublications. He works to unders tand the ghost in the machine - how\ncells in our brain collectively give rise to the computations that\ndetermine behavior. David has been happily married these last 17 years\nto his wife\, Robin.\n\nPart of the SCS Disti nguished Lecture Series\n\n→  Copies of Dr. Sussillo's book will be ava ilable the the CMU\nBookstore prior to and after his presentation. \n\n      Copies will also be available at the auditorium (cash only\nor a C MU account number) the day of his lecture.\n\n→  He will also be avail able for booksigning in the CMU Bookstore\nfrom 11:00 am - 1:00 pm on Satu rday\, April 11. Stop by during Carnival\nto meet the author! \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91263ca DTSTART;TZID=America/New_York:20260414T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260414T133000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:AI Measurement Science and Engineering Center Speaker - Gokul Krish nan CLASS:PUBLIC DESCRIPTION:Speaker: GOKUL S KRISHNAN\, Senior Research Scientist\, Centre for\nResponsible AI\, Wadhwani School of Data Sciences Indian Institute of \nTechnology (IIT) Madras\n\nTalk Title: From Indigenous Benchmarks to Gro unded Reasoning: Towards\nEvaluating and Building Trustworthy AI Models\n\ nAs Large Language Models (LLMs) are increasingly deployed in\nhigh-stakes domains such as healthcare\, law\, and education\, the need\nfor robust\, transparent\, and fair systems is paramount. While fair AI\nsystems are d esirable to ensure people or communities  are not left\nout from AI solut ions\, explainable AI systems that provide\nunderstandable explanations ar e important to improve trustworthiness.\nWhile significant progress has be en made in assessing Responsible AI\naspects such as fairness/bias\, exist ing frameworks often rely on\nWestern-centric benchmarks and fail to captu re complex\, culturally\nspecific sociolinguistic nuances in diverse regio ns such as India.\nFurthermore\, off-the-shelf models frequently generate appealing\nnatural language explanations that sound plausible (and often\n convincing!) but lack faithfulness to the working of the models.\n\nThis t alk presents a series of approaches moving from the rigorous\nevaluation o f AI to the possibility of construction of trustworthy\nmodels. First\, I will present our work on region based fairness/bias\nevaluation techniques \, highlighting the IndiCASA framework's use of\ncontrastive embeddings to capture fine-grained demographic biases in\nthe Indian context. Second\, I will present our work on exposing and\nquantifying geographic and socioe conomic disparities in LLM-based\neducational recommenders. Next\, I will introduce LExT\, a novel metric\nthat jointly quantifies the aspects of pl ausibility and faithfulness\nof natural language explanations. Finally\, I will briefly discuss our\nongoing attempts to bridge the gap between asse ssment and generation\nof natural language explanations I also plan to giv e a glimpse of\nother ongoing technical and policy research at the Centre for\nResponsible AI (CeRAI) at IIT Madras.\n\nFaculty Host:  Ramayya Kris hnan\n\nHosted by AI Measurement Science & Engineering (AIMSEC) — CMU-N IST\nCooperative Research Center\n\nREGISTER\n\n⇒ Note:  1:1 Appointm ents with Dr. Krishnan are possible before\nand after the lecture — SI GN UP \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912683e DTSTART;TZID=America/New_York:20260330T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260330T130000 LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar - Abhi Shelat CLASS:PUBLIC DESCRIPTION:Speaker: ABHI SHELAT\, Professor\, Khoury College of Computer S ciences\,\nNorthwestern University\n\nTalk Title: Zero-knowledge proofs in your wallet\n\nI'll discuss the journey to get zero-knowledge proof proto cols used to\nconvey your private identity information when websites ask f or it. \n Specifically\, Google wallet now supports zk presentation for mobile\ndriver license IDs.  I'll discuss the problem\, how we made the\n technical choices\, and what new ideas we needed to get proofs for\nlegacy ID formats to work on the phone.\n\n—\n\nAbhi Shelat is a professor at Northeastern University and an engineer\nat Google for this project.\n\nFa culty Host:  Elaine Shi\n\nIn Person and Zoom Participation.  See announ cement.\n\n— The CyLab seminar is open only to CyLab partners and CMU fa culty\,\nstudents\, and staff \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9126baf DTSTART;TZID=America/New_York:20260401T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260401T170000 SUMMARY:SIGBOVIK: Last\, Last\, Last Final Submission Deadline CLASS:PUBLIC DESCRIPTION:Wednesday\, April 1\, 2026\, 8am – 5pm \n\nThe organizers hear your pleas\, and have benevolently decided to give\na FINAL EXTENSION for papers. Please submit ALL PAPERS by April 1. No\,\nthis is not a joke .\n\nYes\, this is the last submission extension. Any later and you will\n have to face the wrath of Harry Q. Bovik himself. Us organizers may\nlove procrastination as much as you\, but the time to prepare for the\nnight-of writing frenzy starts now.\n\nSubmit\n\nSIGBOVIK grows nearer! Remember t o clear your calendars\, as the\nconference is on April 10 at 5:00 pm in R ashid Auditorium! We simply\ncan't wait to see what everyone has discovere d.\n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9126ebd DTSTART;TZID=America/New_York:20260326T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260326T170000 LOCATION:Carnegie Museum of Art Theater\, 4400 Forbes Avenue\, 15213 SUMMARY:MFAI Workshop · Spring 2026 CLASS:PUBLIC DESCRIPTION:Speaker: Frontiers of Flows for Generative AI\n\nMathematical F oundations of AI (MFAI) Workshop · Spring 2026 -\nFrontiers of Flows for Generative AI\n\nThursday\, March 26\, 9am – Friday\, March 27\, 2026\, 5pm \n\nA growing fraction of modern generative modeling can be viewed t hrough\na common lens: learning a flow that transports a simple distributi on\ninto the data distribution. This spans deterministic dynamics\n(includ ing ODE-based flows) and stochastic dynamics (including\ndiffusion and sco re models)\, with deep connections to ideas in optimal\ntransport. This wo rkshop brings together researchers working on the\nmathematical and algori thmic foundations of these models to explore\nthe frontier of flows and ge nerative modeling\, and to identify key\nopen problems that will drive the next generation of advances.\n\nTopics include but are not limited to:\n\ nFlow-based generative models (flow matching\, diffusion/score models\,\na nd related formulations)Sampling and learning algorithmsOne-step and\nfew -step models\; distillation and consistency-style trainingOptimal\ntranspo rt perspectivesConnections with other generative\nparadigmsDiscrete flow a nd diffusion modelsApplications including\nvision\, text\, multimodal mode ling\, and scientific settings\n\nThe workshop will consist of featured ta lks\, lightning talks\, and\npanel discussions on the frontiers of flows a nd generative AI. The\ngoal is to surface emerging ideas and identify key open problems for\nthe field.\n\nSchedule\n\nThursday\, March 26:  Worksh op Day 1  — 9:00 am - 5:00 pm\n\n  9:00 am -   9:30 am — Welcome / Networking  9:30 am - 10:30\nam — Panel (kickoff): Frontiers of F lows & Generative AI10:30 am -\n11:15 am — Featured Talk 1 – Michael Albergo (Harvard)11:15 am\n- 12:00 pm — Featured Talk 2 – Mengyu Wang (Harvard)12:00 pm\n-   1:30 pm — Lunch  1:30 pm -   2:45 pm  — Featured Talk\n3 –  Andrej Risteski (CMU)  2:45 pm -   3:30\np m —  Featured Talk 4 – Aditi Raghunathan (CMU)  3:30 pm - \n 3 :45 pm — Break / Networking  3:45 pm  -  4:30\npm — Featured T alk 5 –  Danilo Comminiello (Sapienza\nUniversity of Rome)  4:30 pm -    5:00 pm — Student lightning\ntalks\n\nFriday\, March 26:  Worksh op Day 2  — 9:00 am - 5:00 pm\n\n  9:00 am - 10:00 am — Student lightning talks10:00 am - 10:30\nam — Coffee break10:30 am - 11:15 am  — Featured Talk 6 – \nMichal Balcerak (University of Zurich)11:15 am - 12:00\npm — Featured Talk 7 –  Aviral Kumar (CMU)12:00 pm -    1:30\npm — Lunch  1:30 pm -   2:45 pm — Featured Talk\n8  –  Sanjay Shakkottai (UT Austin)  2:45 pm  -  3:30\npm — Featu red Talk 9 –  Yutong (Kelly) He (CMU)  3:30pm - \n  4:00 pm — Br eak/Networking  4:00 pm -   5:00 pm — Closing\nPanel: Open Problem s + What's Next?\n\nREGISTER   |   Additional Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912735b DTSTART;TZID=America/New_York:20260326T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260326T160000 LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Zion Hefty CLASS:PUBLIC DESCRIPTION:Speaker: ZION HEFTY\, Ph.D. Student\, Department of Mathematics \,\nUniversity of Delaware\n\nTalk Title: Improving R(3\,k) in just two bi tes\n\nThe Ramsey number R(t\,k) is the smallest n such that any red-blue edge\ncoloring of the n-vertex complete graph has either a t-vertex red\nc omplete subgraph or a k-vertex blue complete subgraph. We will\ninvestigat e the history of asymptotic bounds on the extreme\noff-diagonal Ramsey num bers R(3\,k) and present a new lower bound that\nhas been conjectured to b e asymptotically tight.\n\n4:00 pm → Jane Street-sponsored tea and cooki es in the Math Lounge\n(bring your mug!). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9127681 DTSTART;TZID=America/New_York:20260323T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260323T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Filip Obradovic CLASS:PUBLIC DESCRIPTION:Speaker: FILIP OBRADOVIC\, Fouder and Chief Executive Officer\, Dataware\n\nTalk Title: TonicDB: Databases without an OS? Meet QuinineHM\ n\nWe spent years optimizing database internals\, only to have our\nperfor mance eaten up by the one thing we can’t control: the OS.\nGeneral-purpo se kernels are great for desktops\, but for\nhigh-performance data infrast ructure\, they become the bottleneck. In\nthis talk I will introduce Quini neHM\, a specialized \"Hardware\nManager\,\" as a replacement for traditio nal operating systems in\ndatabase workloads. We will begin by examining t he “Why”: how a\ngeneral-purpose OS restricts database performance. Fo llowing this\, we\nwill explore the implementation of the Hardware Manager \, detailing its\nstructure\, its distinctions from general-purpose OSes a nd unikernels\,\nand the implementation challenges faced while building su ch a system.\nFinally\, we will take a quick look at TonicDB\, our bare-me tal\nRedis-compatible store built on QuinineHM\, and see the performance\n improvements.\n\nThis talk is part of the PostgreSQL vs. The World Seminar Series.\n\n—\n\nFilip Obradovic founded and is CEO of Dataware.\n\nZoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9127a5e DTSTART;TZID=America/New_York:20260324T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260324T120000 LOCATION:Wean Hall 8220 SUMMARY:Mathematics Colloquium - Alexander Barvinok CLASS:PUBLIC DESCRIPTION:Speaker: ALEXANDER BARVINOK\, Professor of Mathematics\, Univer sity of\nMichigan\n\nTalk Title: Computing the probability of intersection \n\nGiven a finite family of events in some probability space\, we want to \ncompute (or approximate) the probability of the intersection of their\nc omplements. In the standard interpretation\, each event signifies\nsomethi ng unfortunate happening\, and we are interested in the\nprobability that none of the unfortunate events actually happen. If\nthe events are indepen dent\, the probability in question is determined\nof course by the probabi lities of the events themselves. I am planning\nto discuss what happens wh en the events are not independent\, but the\ndependencies are controlled\, for example\, by controlling the maximum\ndegree of the dependency graph of the family. It turns out then that\nthe probability in question can be approximated from the probabilities\nof intersections of subfamilies of lo garithmic size. Some quite\nnatural questions remain open\, however. \n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9127e06 DTSTART;TZID=America/New_York:20260327T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260327T140000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory x Machine Learning Seminar - Abhradeep Guha Thakurta CLASS:PUBLIC DESCRIPTION:Speaker: ABHRADEEP GUHA THAKURTA - To Be Rescheduled\, Staff Re search\nScientist\nGoogle DeepMind\n\nTalk Title: AI as a Research Partner : Advancing Theoretical Computer\nScience with AlphaEvolve\n\nWhile Large Language Models excel at competitive programming and\nmathematics\, their impact on novel research discovery remains limited.\nThis talk introduces AlphaEvolve\, an LLM-based agent that discovers\ncomplex combinatorial str uctures to advance the state-of-the-art in\nhardness of approximation and extremal combinatorics.\n\nWe present several new results achieved by Alph aEvolve\, including a\nnew metric TSP lower bound of 111/110 (improving up on 117/116) and\nstate-of-the-art lower bounds for classical Ramsey number s: R(3\,13)\,\nR(3\,18)\, R(4\,13)\, R(4\,14)\, and R(4\,15). Additionally \, AlphaEvolve\nidentified novel graph constructions that improve worst-ca se hardness\nof approximation for MAX-4-CUT\, alongside new average-case h ardness\nbounds for MAX-2-CUT and MAX-Independent Set.\n\nThe scale and in tricacy of these graphs make them highly unlikely to\nbe discovered via hu man intuition or traditional solvers alone.\nCrucially\, every discovery i ncludes a machine-verifiable certificate\nto guarantee validity. We conclu de by examining the evolving role of\nAI in mathematical research\, compar ing its discovery potential against\nboth human ingenuity and traditional computation.\n\nJoint work with Ansh Nagda (University of California\, Ber keley) and\nPrabhakar Raghavan (Google).\n\nBased on:\n\nReinforced Genera tion of Combinatorial Structures: Ramsey Numbers\nand Reinforced Generati on of Combinatorial Structures: Hardness of\nApproximation\n\n—\n\nDr. A bhradeep Guha Thakurta is a Staff Research Scientist at Google\nDeepMind. His research focuses on the intersection of Artificial\nIntelligence (AI) and Theoretical Computer Science (TCS)\, specifically\nleveraging AI to ac celerate mathematical discovery. He investigates\ncomplex problems in hard ness-of-approximation\, extremal combinatorics\,\nand cryptography that re sist traditional \"pencil-and-paper\" or\nstandard computational approache s.\n\nAlongside his work on AI and TCS\, Dr. Thakurta has contributed to\n machine learning and data privacy. His optimization algorithms\, DP-SGD\na nd DP-FTRL\, are widely used methods for differentially private model\ntra ining. He led the first deployment of differential privacy at Apple\nfor i OS 10 and the launch of Google's first differentially private ML\nmodel fo r Gboard.\n\nDr. Thakurta received his Ph.D. from Penn State University in 2013.\nHis career includes an Assistant Professorship at UC Santa Cruz\, and\nresearch positions at Apple\, Yahoo Research\, Microsoft Research\, a nd\nStanford University. He is also an inventor on over 10 US patents.\n\n Faculty Host:  Elaine Shi\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912834e DTSTART;TZID=America/New_York:20260325T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260325T130000 LOCATION:Gates Hillman 8102 SUMMARY:Doctoral Speaking Skills Talk - George Zhaoqi Li CLASS:PUBLIC DESCRIPTION:Speaker: GEORGE ZHAOQI LI\, Ph.D. Student\nComputer Science Dep artment\nCarnegie Mellon University\n\nTalk Title: Bellman-Ford in Almost- Linear Time for Dense Graphs\n\nI will present an n2 time algorithm for si ngle-source shortest paths\nwith negative real weights\, building on the b reakthrough work of\nFineman (STOC 2024). The talk will be entirely self-c ontained and only\nassume basic probability facts.\n\nBased on joint work with Jason Li\, Satish Rao\, and Junkai Zhang.\n\nPresented in Partial Ful fillment of the CSD Speaking Skills\nRequirement \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91286aa DTSTART;TZID=America/New_York:20260320T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260320T130000 LOCATION:Gates Hillman 8102 SUMMARY:Doctoral Speaking Skills Talk - Zhibo Chen CLASS:PUBLIC DESCRIPTION:Speaker: ZHIBO CHEN\, Ph.D. Student\nComputer Science Departmen t\nCarnegie Mellon University\n\nTalk Title: Coinductive Logic Programming via Compilation to the\nSemi-Axiomatic Sequent Calculus\n\nLogical framew orks based on computation-as-proof-construction\, such as\n?Prolog\, Elf\, and Celf\, have not supported infinitary objects or\nproofs. We present a logic programming language based on the\ninfinitary proof system CoLF who se semantics are defined by\ncompilation to the semi-axiomatic sequent cal culus (SAX)\, a\nproof-theoretic calculus with parallel operational semant ics. The\ncompilation proceeds through coinductive moded logic programs (C oMLP)\nto primitive rules that correspond directly to SAX processes. A key \ninsight is interpreting logic variables as communication channels\nshare d across parallel premises rather than instantiated through\nunification\, avoiding backtracking and enabling parallel execution.\n\nPresented as pa rt of the PLunch Seminar Series\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9128a22 DTSTART;TZID=America/New_York:20260323T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260323T130000 LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar - Adam J. Aviv CLASS:PUBLIC DESCRIPTION:Speaker: ADAM J. AVIV\, Associate Professor Department of Compu ter\nScience\, The George Washington University\n\nTalk Title: From Dashbo ards to Labels: Helping users manage and make\ndecisions about privacy\n\n The surveillance economy\, where tracking and collecting data are used\nfo r the purpose of advertising and other actions\, is central to many\nof th e money-making enterprises of the modern technology ecosystem.\nDue to reg ulations and other forces\, some of the largest companies\,\nsuch as Googl e and Apple\, have prioritized mechanisms for users to\nbetter manage and receive information about the kinds of data that are\nbeing collected abou t them. In this talk\, I will explore how effective\nthese mechanisms are and ask who they ultimately serve. I will present\nrecent experiments we'v e performed on Google's data dashboards and\ntheir effectiveness\, and als o present ongoing work on Apple's\napp-based privacy nutrition labels\, wh ich describe apps functionality\nwith relation to privacy. \n\n—\n\nAda m J. Aviv is an Associate Professor (with tenure) in the Department\nof Co mputer Science at the George Washington University and is the\ndirector of the GW-Usable Security (GWUSEC) Lab. Dr. Aviv has\npublished more than 90 peer-reviewed papers on computer security\,\nprivacy\, internet measureme nt\, applied cryptography\, and\, most\nnotably\, human factors in securit y and privacy. Dr. Aviv received his\nB.S.E from Columbia University and h is M.S.E. and Ph.D. from the\nUniversity of Pennsylvania. \n\nFaculty Hos t: Lujo Bauer\n\nIn Person and Zoom Participation. See announcement.\n\n → CyLab seminars are only open to CyLab Partners and current CMU\nfacult y\, staff and students. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9128e19 DTSTART;TZID=America/New_York:20260327T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260327T180000 URL:https://csd.cmu.edu/academics/doctoral/phd-open-house SUMMARY:Admitted Ph.D. Student Open House (Day 2) CLASS:PUBLIC DESCRIPTION:Talk Title: CSD Ph.D. Open House\n\nThursday\, March 26\, 9am  – Friday\, March 27\, 2026\n\nWe hope you will join us for our Computer Science Department Doctoral\nAdmitted Student Open House. There are inform ational group events and\none-on-one meetings with faculty and doctoral st udents\, along with\nother programming\, to help you evaluate and understa nd our\nactivities. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91290d4 DTSTART;TZID=America/New_York:20260410T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260410T120000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:Keynote Presentation - Raj Reddy CLASS:PUBLIC DESCRIPTION:Speaker: RAJ REDDY\, University Professor of Computer Science a nd\nRoboticsMoza Bint Nasser ChairCarnegie Mellon University\n\nTalk Title : The Future of AI and Robotics\n\nNational Robotics Engineering Week and Robotics Institute Open House\n\nThe last decade has seen extraordinary ad vances in AI. The potential\narrival of Artificial General Intelligence (A GI) has profound\nimplications for future of our society. We anticipate a world where AI\nassistants and humanoid-robots will perform most of the  tasks\nrequiring human expertise and skill at 10% of current costs. In thi s\nparadigm\, essential services—including food\, housing\, energy\,\ned ucation\, healthcare\, and transportation—will be provided via\nUniversa l Basic Services\, signaling a historic shift  from a society\nof scarcit y to one of abundance.\n\nThis transformation raises a critical concern: w idespread displacement\nof traditional labor. What is the human role when AI can do\neverything? This talk presents an  alternative scenario: a\n\" Human-in-the-Loop\" evolution. In this model\, humans  transition into\nh igh-level supervisory roles\, collaborating with AGI to train robots\nin n ovel skills and adapt them to unforeseen tasks.\n\nWe explore this as the \"Maharaja Model\" where technology serves\nhumanity so comprehensively th at work will be optional for humans.\nFinally\, we will discuss how  inst itutions like the Robotics\nInstitute must lead this transition\, developi ng the hybrid\ntechnologies and ethical frameworks necessary to bridge the gap\nbetween our current economy and a robot-assisted future.\n\n—\n\nR aj Reddy is a University Professor of Computer Science and Robotics\nand M oza Bint Nasser Chair at Carnegie Mellon University. He was an\nAssistant Professor at Stanford from 1966-69 and Faculty Member at\nCarnegie Mellon since 1969. He served as the founding Director of the\nRobotics Institute from 1979 to 1991 and the Dean of School of\nComputer Science from 1991 to 1999.\n\nHe has been active in AI research for over five decades in the a reas\nof AI\, Speech Understanding\, Image Understanding\, Robotics\,\nMul ti-sensor Fusion\, and Intelligent Agents. Dr. Reddy's current\nresearch i nterests include: Technology in Service of Society\, Voice\nComputing for the 3B semi-literate populations at the bottom of the\npyramid\, Digital D emocracy\, and Learning Science and Technologies.\n\nHe is a member of the National Academy of Engineering and the American\nAcademy of Arts and Sci ences. He served as co-chair of President\nClinton’s Information Technol ogy Advisory Committee (PITAC) from\n1999 to 2001. Dr. Reddy is the recipi ent of the Legion of Honor in\n1984\, the ACM Turing Award in 1994\, the P adma Bhushan in 2001\, the\nHonda Prize in 2005 and Vannevar Bush Award in 2006.\n\nNational Robotics Week | Robotics Institute Open House \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91295e2 DTSTART;TZID=America/New_York:20260603T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260603T163000 LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:CMU AI and Data Analytics Summit CLASS:PUBLIC DESCRIPTION:In years past\, AI Day and Data Analytics Day were separate eve nts\;\nhowever\, they've been combined them into a single\, unified experi ence\nto promote and leverage the fundamental relationship and synergy\nbe tween data analytics and AI development.\n\nThis year’s theme\, Intellig ence at Scale: Driving a Human-Centered\nFuture with AI and Data Analytics \, will guide the leadership panel\,\nkeynote speech\, and AI/DA Summit Cu p showcase\, promising an engaging\nand insightful experience for all atte ndees.\n\nSpeakers include CMU senior leadership\, faculty\, staff\, and g uests.\n\nEvent Highlights\n\nMorning Plenaries: Kick off with a leadershi p panel\, keynote address\,\nand the Summit Cup (formerly IronViz) Award Ceremony\, showcasing\ndata-driven innovation in action.Afternoon Breakout s: Choose your path\nwith technical sessions on model deployment\, hands-o n data\nstorytelling workshops\, and discussions on the impact of AI in\ne ducation.\n\nREGISTRATION\n\n→ Space is limited\, so your response is  essential\n\n→ This event is currently open to staff and faculty only. It may\nopen to students at a later date or in special circumstances. \n\ n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9129999 DTSTART;TZID=America/New_York:20260318T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260318T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Jingxun Liang CLASS:PUBLIC DESCRIPTION:Speaker: JINGXUN LIANG\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Stable Open Addressing and the Curse of Reappearance\nDependenciesWednesday\, March 18\, 2026\, 1 2 – 1pm Stable hash\ntables---hash tables that never move existing ele ments---are among the\nsimplest and most widely used hashing schemes. Two canonical examples\nare stable uniform probing and stable linear probing. Both are\ncommonly believed to support constant expected-time operations.\ nHowever\, formal proofs have long been obstructed by a subtle\ndependency issue known as reappearance dependency. While this\ndependency is widely conjectured to be essentially harmless\, it has so\nfar defeated all known analysis techniques.In this talk\, we show that\nthis belief is only part ially correct\, in a surprising way. Stable\nuniform probing—despite pro vably outperforming linear probing in the\ninsertion-only setting—can be severely affected by reappearance\ndependencies: there exists an obliviou s update sequence under which\nthe expected insertion time grows polynomia lly in the stable setting.\nIn contrast\, stable linear probing can be \"r escued\" from reappearance\ndependencies and still guarantees constant exp ected insertion time.\nSomewhat counterintuitively\, the same locality tha t makes linear\nprobing slower than uniform probing in the insertion-only setting\nturns out to be the key ingredient that allows it to overcome\nre appearance dependencies. Event Type: Seminars Room Number: In\nPerson Bu ilding: Gates Hillman 8102 Speaker's Name: JINGXUN LIANG\nSpeaker Websit e: sites.google.com… Speaker's Professional\nTitle: Ph.D. Student\, Co mputer Science Department\, Carnegie Mellon\nUniversity Talk Title: Stabl e Open Addressing and the Curse of\nReappearance Dependencies For More\nIn formation: hfleisch@andrew.cmu.edu Affiliations: Computer Science\nDepar tment (CSD) Organization(s): School of Computer Science Event\nWebsite Ti tle: Event Website Event Website URL: www.cs.cmu.edu\n[http://www.cs.cmu .edu]…\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9129e31 DTSTART;TZID=America/New_York:20260323T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260323T133000 LOCATION:Gates Hillman 7101 SUMMARY:SQUALL Seminar - Igor Kadota CLASS:PUBLIC DESCRIPTION:Speaker: IGOR KADOTA\, Assistant ProfessorDepartment of Electri cal and\nComputer EngineeringNorthwestern University\n\nTalk Title: Optimi zing Age of Information without Knowing the Age of\nInformation\n\nEnsurin g timely information delivery is essential for modern networked\nsystems r anging from smart factories to autonomous vehicles. The Age\nof Informatio n (AoI) has emerged as a key metric for quantifying\ninformation freshness \, but most AoI‑aware scheduling algorithms rely\non real‑time knowled ge of source or destination timestamps. This\nassumption breaks down in pr actical systems with unreliable links\,\ndelayed feedback\, or limited cro ss‑layer visibility.\n\nThis talk presents a framework for optimizing in formation freshness\nwithout knowing the actual AoI\, focusing on networks where sources\ngenerate packets according to general renewal processes an d where the\nbase station (BS) observes only transmission outcomes. We dev elop\nlow‑complexity MMSE estimators that infer both system time and AoI \nusing only the BS’s observable history. Leveraging these estimators\,\ nwe propose a Max‑Weight scheduling policy that operates without AoI\nkn owledge\, yet provably outperforms other well-known policies\,\nincluding the Optimal Randomized Policy\, in both theory and\nsimulations.\n\nOveral l\, the results highlight that freshness optimization is possible\neven wh en timestamps are unavailable\, offering a path toward\ndeployable AoI‑a ware scheduling in real systems.\n\n—\n\nIgor Kadota is an Assistant Pro fessor of Electrical and Computer\nEngineering at Northwestern University. Previously\, he was a\nPostdoctoral Research Scientist at Columbia Univer sity. He received\nthe Ph.D. degree from MIT LIDS and his B.Sc. degree fro m the\nAeronautics Institute of Technology (ITA) in Brazil. His research i s\non modeling\, analysis\, optimization\, and implementation of\nnext-gen eration communication networks\, with the emphasis on advanced\nwireless s ystems and time-sensitive applications.\n\nHe was a recipient of several r esearch\, teaching\, and mentoring\nawards\, including the Best Paper Awar d at IEEE INFOCOM 2018\, the Best\nPaper Award Finalist at ACM MobiHoc 201 9\, the Best Student Paper Award\nat WiOpt 2024 and WiOpt 2025\, and the M IT School of Engineering\nGraduate Student Extraordinary Teaching and Ment oring Award of 2020.\nAdditional information\n\nFaculty Host: Mor Harchol- Balter \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912a2e6 DTSTART;TZID=America/New_York:20260319T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260319T160000 LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Tracy Chin CLASS:PUBLIC DESCRIPTION:Speaker: TRACY CHIN\, Ph.D. Student\, Department of Mathematics \,\nUniversity of Washington\n\nTalk Title: Valuated Delta Matroids and Pr incipal Minors\n\nDelta matroids are a generalization of matroids that ari se naturally\nfrom combinatorial objects such as matchings\, ribbon graphs \, and\nprincipal minors of symmetric and skew symmetric matrices. In this \ntalk\, we will define valuated delta matroids and explore their\nconnect ion with principal minors of Hermitian matrices\, generalizing\nwork by Ri ncón on valuated even delta matroids and skew symmetric\nmatrices.\n\nBas ed on joint work with Nathan Cheung\, Gaku Liu\, and Cynthia Vinzant.\n\n4 :00 pm → Jane Street-sponsored tea and cookies  in the Math Lounge\n(br ing your mug). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912a665 DTSTART;TZID=America/New_York:20260326T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260326T180000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Special Guest Lecture - Kan Devnani CLASS:PUBLIC DESCRIPTION:Speaker: KAN K DEVNANI\, Chief Executive Officer\, Comtex News Network\,\nand Co-Founder and Managing Partner. Rainmaker Digital LLC\n\nT alk Title: How AI is Leveraged in the Modern Workplace - and how to\nprepa re for it\n\nKan K Devnani is a graduate of Carnegie Mellon University wit h a B.S.\nin Computer Science/Mathematics and a minor in Jazz Performance. He\nhas served as CEO of Comtex News Network since 2014 and\nCo-Founder/M anaging Partner of Rainmaker Digital LLC since 2021. His\nentrepreneurial efforts have included software companies\, a\nquantitative alternative ass et management fund and real estate\nprojects. Kan has extensive experience in technology\, running\nreal-time News engineering at FactSet Research S ystems in the early\n2000s\, building for scale\, redundancy and experienc ed at running\noperations for high performance systems. He has led develop ment of\nquantitative trading systems\, content distribution platforms\, A I\nclassification architectures with experience in back-end systems as\nwe ll as web and mobile platforms.\n\nBesides technology and content\, Kan is passionate about mentorship\,\nyouth development and music. He has served at the Fiver Children’s\nFoundation for 10 years as a Board member and is currently serving on\nthe Emeritus Board. He has served on the Board of the Downtown Soccer\nLeague for 7 years\, coaching teams and acting as a division\ncoordinator. He is an active saxophone player\, appearing regula rly at\nSaltwater Coffee and Wine Bar in NYC.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912aa55 DTSTART;TZID=America/New_York:20260316T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260316T130000 LOCATION:Group Viewing Newell-Simon 3305 and Zoom SUMMARY:Accessibility Lunch Seminar - Daniel Killough CLASS:PUBLIC DESCRIPTION:Speaker: DANIEL KILLOUGH\, Ph.D. Student\, Computer Sciences De partment\,\nSchool of Computer\, Data &amp\; Information Sciences\, Univer sity of\nWisconsin-Madison\n\nTalk Title: Toward Agent-Readable XR: From D eveloper Challenges to\nAI-Driven Access\n\nExtended reality (XR) applicat ions largely lack accessibility\nfeatures\, leaving many people with disab ilities excluded from virtual\nworlds. But what are some causes of this ga p? And what can we do about\nit in the meantime? This talk presents two co mplementary projects\naddressing XR accessibility from both sides of the p roblem: First\,\nthrough interviews with 25 XR practitioners across compan y sizes\, we\nidentify why developers struggle to implement accessibility\ , like\nfragmented guidelines\, lack of XR-specific standards\, resource\n constraints\, and toolkits that fail to meet industry adoption\nrequiremen ts (CHI 2026). Informed by this work we developed VRSight\, a\npost hoc AI system that provides real-time spatial audio descriptions\nfor blind VR u sers by combining custom object detection\, depth\nestimation\, and vision -language models without requiring any developer\nintegration (UIST 2025). \n\nTogether\, these projects reveal a deeper insight: the challenges\nfac ing blind and low vision XR users are not unique to accessibility\;\nthey expose fundamental gaps confronting any AI agent operating in\nspatial env ironments\, from understanding context-dependent affordances\nto supportin g embodied interaction. We finally premiere our ongoing\nwork extending th ese approaches to mixed reality with vision-language\nmodels and close wit h an open discussion: What might a minimal\nagent-readable XR API need to expose?\n\nIn Person Group Viewing and Zoom Participation.  See announcem ent. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912aee7 DTSTART;TZID=America/New_York:20260318T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260318T131500 LOCATION:Newell-Simon 4305 SUMMARY:SCS Fireside Chat with Ivo Stivoric CLASS:PUBLIC DESCRIPTION:Speaker: IVO STIVORIC\, Vice President at X\n\nTalk Title: Alph abet's “moonshot factory\"\n\nPlease join in as Erica Fuchs and Tom Mitc hell sit down with Ivo\nStivoric for a fireside chat to discuss his experi ences at Alphabet's\n“moonshot factory.\"   Over his decade at X\, Iv o has incubated and\nled dozens of projects and teams\, including Anori\, Bellwether\,\nTapestry\, Chorus\, and 280 Earth. Ivo's passion lies at the \nintersection of innovation and commercialization\, developing both the\n projects and teams to tackle some of the world's most pressing\nproblems a longside the novel funding mechanisms that allow these\nradical ideas to d evelop into successful companies.  We’ll hear\nabout his lessons learne d and his views on where the future is headed.\n\n—\n\nIvo Stivoric is a Vice President at X\, the moonshot factory where he\nserves on X’s Lead ership Team and runs a portfolio of projects\nfocused on finding radical s olutions to some of the world’s most\npressing problems.  For the last 10 years Ivo has incubated dozens of\nmoonshot technologies\, projects and teams\, while also forging strong\npartnerships with organizations outsid e of X to help bring these\nnascent ideas and technologies to life. Some o f the Alphabet companies\nand projects to emerge from Ivo’s portfolio in clude moonshots for\nindustrial robotics\, connectivity\, the electric gri d\, ocean health\,\nthe supply chain and waste. \n\nBefore joining X\, Iv o was Vice President of Research and Development\nat Jawbone\, where he de livered new consumer and healthcare\napplications for wearable and sensor technology. Ivo has worked in the\nwearable computing space since 1991\, i ncluding co-founding BodyMedia\nalongside X’s Captain of Moonshots\, Ast ro Teller. He has over 90\npatents to his name and was recognized as one o f 40 innovators\nbuilding the foundation of the next-gen electronics indus try by EE\nTimes. He holds an M.S. in Interaction Design and a B.F.A. in\n Industrial Design with a concentration in sculpture\, both from\nCarnegie Mellon.\n\nHost: George Darakos \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912b338 DTSTART;TZID=America/New_York:20260324T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260324T173000 LOCATION:CMU Track SUMMARY:22nd SCS Random Distance Run CLASS:PUBLIC DESCRIPTION:Speaker: The Enormous Fuzzy Dice are back!\n\nHow long will the race be? \n\n— Nobody knows! All we can say is that it will be somewh ere between\n1/2 mile and 3 miles (between 2 and 12 laps of the track). Th e exact\nnumber of laps will be specified by the sum of two ENORMOUS FUZZY DICE\nrolls - one occurring just before the start of the race\, and one j ust\nbefore the first runner finishes the number of laps specified by the\ nfirst dice roll.\n\nWe are limiting to the first 90 responses\, so regist er early to\nguarantee a spot! \n\n—The online form will still be open until 4:00 pm day-of the race to\naccept race entries\, assuming we haven ’t hit our 90-person limit. We\nwill also have race-day registration on the track from 4:00 - 4:20 pm.\n\nThe Race starts promptly at 4:30 pm \n\ n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912b6af DTSTART;TZID=America/New_York:20260316T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260316T170000 LOCATION:Steinberg Auditorium\, Baker Hall A53 SUMMARY:Statistics and Data Science Seminar - Arun Chandrasekhar CLASS:PUBLIC DESCRIPTION:Speaker: ARUN CHANDRASAEKHAR\, Professor Department of Economic s\,\nStanford University\n\nTalk Title: On tubes\; or\, Model-building fro m rich data by isolating\nconcepts\n\nArun G. Chandrasekhar\, a developmen t economist\, is a Professor of\nEconomics at Stanford University. He rece ived his B.A. from Columbia\nUniversity\, where he double majored in mathe matics and economics\, and\nhis Ph.D. in economics from MIT\, followed by a postdoctoral fellowship\nat Microsoft Research New England.\n\nHis resea rch focuses on social learning\, informal finance\,\nredistribution\, and policy design\, much of it set in rural India. To\nstudy these questions\, he combines field experiments\, economic theory\,\neconometrics\, and soc ial network analysis. A central theme of his work\nis understanding what r esearchers can know\, what we might be missing\,\nand how to learn from da ta and make decisions when models may be\nincomplete or fragile.\n\nChandr asekhar is a Research Associate at the National Bureau of\nEconomic Resear ch and an affiliated professor at J-PAL. He is the\nrecipient of a Sloan R esearch Fellowship (2018) and the Infosys Prize\nin Economics (2024).\n\n3 :30 pm - Refreshments  prior to the seminar in the 129 suite \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912ba7f DTSTART;TZID=America/New_York:20260316T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260316T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Hari Krishna Sunder CLASS:PUBLIC DESCRIPTION:Speaker: HARI KRISHNA SUNDER\, Architect\, YugabyteDB\n\nTalk T itle: YugabyteDB: Distributed PostgreSQL for Modern Apps\n\nYugabyteDB is an AI-ready\, multimodal\, distributed PostgreSQL\ndatabase. It effectivel y bridges the gap between traditional RDBMS\nreliability and NoSQL flexibi lity\, offering a cloud-native\narchitecture that is resilient by design. Developers can thus leverage\nfamiliar PostgreSQL semantics while achievin g the elastic scale and\nresilience required for mission-critical applicat ions. This talk\nexplores YugabyteDB's unique approach to fusing PostgreSQ L’s query\nengine with a distributed\, ACID-compliant storage and transa ction\nlayer.\n\n—\n\nHari Krishna Sunder is an Architect at YugabyteDB. He has spent his\ncareer working on cloud databases and was previously a founding member\nof Azure SQL Database and Azure Synapse Data Warehouse. A t Yugabyte\,\nhis work centers on advancing the core capabilities of distr ibuted\nPostgreSQL\, with a current focus on vector and hybrid indexing\,\ nmulti-tenancy\, and architecting a unified data layer for AI agents.\n\nT his talk is part of the PostgreSQL vs. The World Seminar Series\n\nZoom Pa rticipation. See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912be49 DTSTART;TZID=America/New_York:20260326T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260326T133000 LOCATION:Newell Simon 3305 and Zoom SUMMARY:AI-SDM Seminar - Melanie Mitchell CLASS:PUBLIC DESCRIPTION:Speaker: MELANIE MITCHELL\, Professor\, Santa Fe Institute\n\nT alk Title: Investigating Abstract Reasoning in Humans and\nMachinesThursda y\, March 26\, 2026\, 12 – 1:30pm\n\nState-of-the-art AI models have m atched or exceeded human performance\non reasoning benchmarks such as the Abstraction and Reasoning Corpus\,\na prominent benchmark for conceptual u nderstanding and analogy. But\ndoes high accuracy on this benchmark mean t hat these models understand\nand reason with the humanlike abstractions in tended by the task\ncreators?  In this talk I will describe an evaluation methodology\,\ninspired by experimental methods in cognitive science\, to assess and\ncompare the abstraction abilities of AI “reasoning” model s and\nhuman participants.  Our evaluations show that\, while some models \nmatch or exceed human accuracy\, their reasoning is frequently based on\ nsurface-level patterns or spurious associations\, and thus lacks\ngeneral izability.  I will speculate on what is still needed to\ncapture humanlik e abstract reasoning abilities in AI models.  \n\n—Melanie Mitchell  is a Professor at the Santa Fe Institute. Her\nresearch is at the intersec tion between artificial intelligence\,\ncognitive science\, and complex sy stems\; she has authored or edited six\nbooks and published over 100 schol arly papers in these fields. Her\n2009 book Complexity: A Guided Tour won the 2010 Phi Beta Kappa\nScience Book Award\, and her 2019 book Artificial Intelligence: A Guide\nfor Thinking Humans was named as one of the five b est books on AI by\nthe New York Times and the Wall Street Journal. Melani e’s public\noutreach on science includes a quarterly column for Science Magazine\,\na Substack newsletter on AI\, a 2024 podcast on “The Nature of\nIntelligence\,” and a free online course\, “Introduction to\nCompl exity” on the Santa Fe Institute’s Complexity Explorer\nwebsite.REGIST ER → please register to attend in-person or on Zoom\n\nEvent Type: Semi nars Room Number: In Person and Virtual - ET\nBuilding: Newell Simon 330 5 and Zoom Speaker's Name: MELANIE\nMITCHELL Speaker Website: melaniemit chell.me Speaker's Professional\nTitle: Professor\, Santa Fe Institute Ta lk Title: Investigating\nAbstract Reasoning in Humans and Machines For Mo re\nInformation: pwerns@andrew.cmu.edu Affiliations: Computer Science\nD epartment (CSD)\, Dietrich College of Humanities and Social Sciences\,\nHu man-Computer Interaction Institute (HCII)\, Language Technologies\nInstitu te (LTI)\, Machine Learning Department (MLD)\, Ray and Stephanie\nLane Com putational Biology Department (CBD) Organization(s): School\nof Computer Science Event Website Title: Event Website Event Website\nURL: events.cm u.edu…\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912c35d DTSTART;TZID=America/New_York:20260313T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260313T140000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Seminar - Rahul Ilango CLASS:PUBLIC DESCRIPTION:Speaker: RAHUL ILANGO\, Postdoctoral Researcher\, Theoretical C omputer\nScience\, School of Mathematics\, Institute of Advanced Study\n\n Talk Title: Gödel in Cryptography: Zero-Knowledge Proofs With No\nInterac tion\, No Setup\, and Perfect Soundness\n\nGödel showed that there are tr ue but unprovable statements. This\nwas bad news for Hilbert\, who hope d that every true statement was\nprovable. In this talk\, I’ll describe why Gödel’s result is\, in\nfact\, good news for cryptography. \n\nS pecifically\, Gödel’s result allows for the following strange\nscenario : a cryptographic system S is insecure\, but it is impossible\nto prove th at S is insecure. As I will explain\, in this scenario\n(defined carefully )\, S is secure for nearly all practical purposes.\n\nLeveraging this idea \, we effectively construct — under longstanding\nassumptions — a clas sically-impossible cryptographic dream object:\n\"zero-knowledge proofs fo r NP with no interaction\, no setup\, and\nperfect soundness”\" (I won ’t assume you know what this means!). As\nan application\, our result le ts one give an ordinary mathematical\nproof that a Sudoku puzzle is solvab le without revealing how to solve\nit. Previously\, it was not known how t o do this.\n\nAbout the Speaker \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912c723 DTSTART;TZID=America/New_York:20260316T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260316T120000 LOCATION:Newell-Simon 3305 and Zoom SUMMARY:PoP Seminar - Alex Lew CLASS:PUBLIC DESCRIPTION:Speaker: ALEX LEW\, Assistant Professor of Computer Science\nSc hool of Engineering &amp\; Applied Science\nYale University\n\nTalk Title: Automatic Integration and Differentiation of Probabilistic\nPrograms\n\nB y automating the error-prone math behind deep learning\, systems such\nas TensorFlow and PyTorch have supercharged machine learning research\,\nempo wering hundreds of thousands of practitioners to rapidly explore\nthe desi gn space of neural network architectures and training\nalgorithms. This ta lk will show how new programming language\ntechniques—particularly gener alizations of automatic differentiation\nand of knowledge compilation—ma ke it possible to generalize and\nextend such systems to support probabili stic models. Our tools can\nautomate the computation of expected values\, probability densities\,\nand their gradients\, as well as help users deriv e fast\, low-variance\,\nunbiased estimators of these quantities when they are too expensive to\ncompute exactly\, enabling orders-of-magnitude spee dups in downstream\noptimization and inference problems.\n\n—\n\nAlex Le w is an Assistant Professor of Computer Science at Yale. His\nresearch aim s to automate and scale up principled probabilistic\nreasoning\, drawing o n techniques from programming languages\, machine\nlearning\, Bayesian sta tistics\, and cognitive science. A key focus is\nthe theory\, design\, and implementation of probabilistic and\ndifferentiable programming languages \, which extend traditional\nprogramming languages with constructs for opt imization and inference\nover models defined as programs. Alex's work has been recognized with\nDistinguished Paper awards at POPL and LICS\, an Out standing Paper\naward at the Conference on Language Modeling (COLM)\, and a Facebook\nResearch Award.\n\nFaculty Host: Feras Saad\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912cb66 DTSTART;TZID=America/New_York:20260318T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260318T110000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Shreya Shankar CLASS:PUBLIC DESCRIPTION:Speaker: SHREYA SHANKAR\, Ph.D. Candidate\, University of Calif ornia\,\nBerkeley\n\nTalk Title: Building Effective Unstructured Data Syst ems\n\nDatabases and other data systems have successfully democratized\nda ta-oriented computation across domains\, thanks to decades of\nresearch in system internals and end-user interfaces. However\, such\nsystems center on structured (i.e.\, tabular) data\; unstructured\ndata—the vast majori ty of data—has largely been ignored. Large\nlanguage models (LLMs) now g ive us a building block for unstructured\ndata analysis\, and we face the same questions as in the early days of\ndata systems—e.g.\, how should u sers author queries? How do we\nefficiently execute queries at scale?—bu t many well-established\ntenets from traditional data systems must be revi sited. In my talk\, I\nwill present DocETL\, a system I developed for unst ructured data\nanalysis. I will discuss how we had to rethink query optimi zation\nunder these new assumptions\, optimizing user-written pipelines fo r\nboth accuracy and efficiency—as well as end-user interfaces for\nauth oring\, iterating on\, and debugging pipelines. DocETL is\nopen-source wit h 3.5k+ GitHub stars\; our hosted interface has\nsupported 4.1k+ pipelines across 30+ S&P-500 industries. Query\noptimization ideas from our work ha ve been adopted in databases such\nas Snowflake and BigQuery\, and our int erface design principles have\nbeen adopted by companies like LangChain an d OpenAI.\n\n—\n\nShreya Shankar is a fifth and final-year PhD student i n the Data\nSystems and Foundations group at UC Berkeley\, advised by Dr. Aditya\nParameswaran. She is broadly interested in data systems\, large\nl anguage models\, and human-computer interaction. Her PhD has been\nsupport ed by an NDSEG Fellowship and a Bridgewater Research\nFellowship\, and her work has been recognized with EECS Rising Stars\n(2025) and a best paper honorable mention award at UIST. Beyond her\nresearch\, Shreya authored th e curriculum and companion book for AI\nEvals for Engineers and PMs\, an i ndustry course on evaluating AI\napplications taken by 4\,000+ professiona ls from 500+ companies\,\nincluding 50+ students each from Google\, Micros oft\, OpenAI\, Meta\,\nAmazon\, Intuit\, and First American. Before her Ph D\, Shreya worked as\nthe first data/ML engineer at a startup after her un dergraduate degree\nin CS at Stanford.\n\nFaculty Host:  Andy Pavlo\n\nIn Person and Zoom Participation.  See announcement.\n\n→ Attendance at t his talk is restricted to members of the SCS\ncommunity and relevant CMU s takeholders.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912d06c DTSTART;TZID=America/New_York:20260313T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260313T190000 LOCATION:Perlis Atrium\, 3rd Floor\, Newell-Simon SUMMARY:SCS First-Year Major Declaration Day Event CLASS:PUBLIC DESCRIPTION:Speaker: Presented by the SCS Undergraduate Student Advisory Co mmittee\n\nSCS Students...\n\nCome and learn more about the concentrations \, majors and minors that\nSCS offers!\n\nSchedule\n\n5:00 pm → Visit M ajors |  SCS Passports6:00 pm →\nAnnouncements6:15 pm → Food!6:45 p m → The Time Capsule6:30 pm\n→ Raffle Winners\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912d392 DTSTART;TZID=America/New_York:20260316T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260316T110000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense CLASS:PUBLIC DESCRIPTION:Speaker: MINGKUAN XU\, Ph.D. Candidate\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: Optimization and Simulatio n of Quantum Circuits\n\nOptimizing and simulating quantum circuits at sca le are critical\nbottlenecks in quantum computing. This thesis delivers a suite of\ntools to improve them.\n\nFor quantum circuit optimization\, we first automate the discovery and\nverification of transformation rules by introducing Equivalence\nCircuit Class (ECC) sets and an efficient generat ion algorithm for\narbitrary gate sets. We then utilize the generated rule s in the\nsuperoptimizer Quartz\, which optimizes circuits using a cost-ba sed\nbacktracking search. Compared to previous rule-based methods\, Quartz \nsuccessfully escapes local minima through exhaustive search. To\nfurther improve efficiency and avoid the exponential runtime penalties\nof pure s earch\, we introduce QALM. This hybrid optimizer combines\nexhaustive sear ch with rule-based rewriting. By interleaving bounded\nsearch-based explor ation with greedy rule-based exploitation\, QALM\nescapes local minima dyn amically. It outperforms existing search-based\noptimizers in optimization quality and matches reinforcement learning\nmethods without the training overhead.\n\nWhile the prior two approaches aim for global optimization\, this\nproblem is intrinsically QMA-hard\, creating a bottleneck for large\ nprograms. To circumvent this issue and scale up\, we introduce OAC\, a\nc ut-and-meld circuit optimization algorithm. OAC cuts a circuit into\nsubci rcuits\, applies an existing oracle optimizer independently\, and\nseamles sly melds the results. This approach operates with a linear\nnumber of ora cle calls while attaining local optimality. Empirical\nevaluation shows th at OAC improves the efficiency of state-of-the-art\noptimizers by over an order of magnitude while enhancing overall\nquality.\n\nBeyond physical ex ecution\, the scalable simulation of quantum circuits\non classical hardwa re presents another major challenge. We present\nAtlas\, a distributed GPU -based simulator that hierarchically\npartitions circuits to exploit data parallelism while minimizing\ncommunication. By using integer linear progr amming to allocate\nstructurally related gates to nearby GPUs and dynamic programming for\nkernel scheduling\, Atlas runs over 2x faster than prior\ nstate-of-the-art GPU simulators.\n\nTogether\, these frameworks provide a robust toolchain\, improving both\nthe execution of quantum circuits on p hysical devices and their\nscalable classical simulation. \n\nThesis Comm ittee\n\nZhihao Jia (Co-Chair)\n\nUmut A. Acar (Co-Chair)\n\nRyan O'Donne ll\n\nYongshan Ding (Yale University)\n\nIn Person and Zoom Participation.   See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912d8a8 DTSTART;TZID=America/New_York:20260316T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260316T130000 LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar - Haojian Jin CLASS:PUBLIC DESCRIPTION:Speaker: HAOJIAN JIN\, Assistant Professor\, Halıcıoğlu Data Science\nInstitute\, University of California\, San Diego\n\nTalk Title: A World Model for Privacy: From Structured Representations\nto Personalize d Reasoning\n\nPrivacy design is often treated as inherently open-ended an d deeply\ncontextual. Small shifts in context—who requests data\, for wh at\npurpose\, under what stakes\, and with what downstream sharing—can\n flip whether a data practice feels appropriate. While this captures a\nrea l phenomenon\, it has also left us with privacy frameworks that are\ntoo a bstract and generic to operationalize\, and user preference models\nthat a re too brittle to generalize beyond the contexts in which they\nwere train ed.\n\nIn this talk\, I present an alternative view. First\, I argue that\ nprivacy design is more structured than it appears: many recurring\nprivac y-relevant decisions can be captured using a structured\,\nclosed-ended re presentation. By constructing such a representation\, we\ncan transform a significant portion of open-ended privacy design into\na closed-ended task \, significantly lowering the barrier for developers\nto create effective privacy designs. Second\, I present evidence that\nprivacy preferences can generalize across domains. In our studies\,\nindividuals’ concerns are stable over time and can be predicted from\nconsistent sensitivities to id entifiable contextual features\, enabling\nforecasts for previously unseen data practices. Building on this\, we\nare modeling the reasoning steps p eople use to reach privacy judgments\nto improve generalization\, personal ization and interpretability.\n\nTogether\, these two lines of work outlin e a research agenda for\nbuilding a “world model” of privacy: a struct ured representation\nof the privacy design decision landscape\, paired wit h models of how\nindividuals reason within it—turning privacy from ad ho c debate into\ncomputable\, inspectable design support.\n\n—\n\nHaojian Jin is an Assistant Professor at UC San Diego’s\nHalıcıoğlu Data Sci ence Institute and directs the AI Smith Lab .\nHis research sits at the i ntersection of human–computer interaction\,\nprivacy and security\, and systems\, with a focus on building usable\ntools. His work has been recogn ized with the Hellman Fellowship\, the\nUbiComp Gaetano Borriello Outstand ing Student Award\, Research\nHighlights in Communications of the ACM and GetMobile\, and best paper\nawards at CCS\, UbiComp\, and ACM Computing Re views. He received his\nPh.D. from the Human-Computer Interaction Institut e at Carnegie Mellon\nUniversity.\n\nFaculty Host: Swarun Kumar\n\nIn Pers on and Zoom Participation.  See announcement.\n\n→ CyLab seminars are o nly open to CyLab Partners and current CMU\nfaculty\, staff and students  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912dda1 DTSTART;TZID=America/New_York:20260326T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260326T173000 LOCATION:Simmons Auditorium\, Tepper Building SUMMARY:President's Lecture Series - Ted Decker CLASS:PUBLIC DESCRIPTION:Speaker: TED DECKER\, Chair\, President and Chief Executive Off icer\, The\nHome Depot\n\nTalk Title: The Home Depot's \"Fourth Growth Cha pter\": Building for the\nFuture\n\nJoin President Farnam Jahanian and Dec ker for a distinguished lecture\nand fireside chat with Ted Decker\, Chair \, President and Chief\nExecutive Officer\, The Home Depot.\n\nTed Decker became The Home Depot’s CEO in 2022 while the world was\nrecovering from a global pandemic\, yet today might be the most\nchallenging time for him at the helm of the company. While the global\npandemic ultimately provide d opportunities for The Home Depot to\nexpand its retail footprint as well as rethink the customer experience\nonline\, today he leads the company t hrough an extremely high-pressure\nenvironment in the U.S.—and the most difficult housing environment\nsince 2009. Ted will discuss how The Home D epot is continuing to grow\nits core business during a frozen housing mark et and how he is\nunlocking new ways to for the company to be innovative a nd reach new\ncustomers. This includes growing its offerings for the $1T P ro market\,\nas well as creating new opportunities through technological\n innovation\, including AI. \n\nREGISTER →Please indicate your attendan ce by Tuesday\, March 24 \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912e1a3 DTSTART;TZID=America/New_York:20260312T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260312T160000 LOCATION:Wean 6220 SUMMARY:ACO Seminar - Daniel Zhu CLASS:PUBLIC DESCRIPTION:Speaker: DANIEL ZHU\, Department of Mathematics\, Ph.D. Student \,\nPrinceton University\n\nTalk Title: Schur complements for tensors and multilinear commutative\nrank\n\nWe show that three notions of ranks for m atrices of multilinear forms\nare equivalent. This result generalizes a cl assical result of\nFlanders\, corrects a minor hole in work of Fortin and Reutenauer\,\nanswers a question of Lampert on the relation between the an alytic and\nslice ranks of trilinear forms\, and establishes a special cas e of the\nconjecture that the analytic and partition ranks of a tensor are \nequivalent.\n\nJoint work with Guy Moshkovitz\n\n4:00 pm → Jane Stree t-sponsored tea and cookies in the math lounge\n(bring your mug!). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912e50d DTSTART;TZID=America/New_York:20260311T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260311T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Zoe Xi CLASS:PUBLIC DESCRIPTION:Speaker: ZOE XI\, Ph.D. Student\, Electrical Engineering and Co mputer\nScience Department\, Massachusetts Institute of Technology\n\nTalk Title: How to Avoid Debate: Scalable AI Safety via\nDoubly-Efficient Inte ractive Proofs\n\nAs AI models continue to develop powerful capabilities\, it becomes\ncritical that we are able to verify that their output is alig ned with\nour intentions. A recent line of work focuses on verification vi a\ndebate\, a model of interactive proofs where two competing powerful\npr overs\, or AI models\, debate each other to convince a weak verifier\,\nor a human\, of the correctness of their claim. However\, debate assumes\nth at the two AI models possess equal abilities and that one of them is\ntrut hful\, which may not be realistic. \n\nIn this talk\, I will present rece nt work on single-prover interactive\nproofs for AI safety. Prior results in single-prover interactive\nproofs do not immediately carry over to the AI safety setting because\nthey do not work when the computation has acces s to an oracle\, such as\nto human judgment or an external database such a s the web. Our work\npresents doubly-efficient single-prover interactive p roofs for\noracle-aided computations (also known as relativizing proofs)\, in the\nsettings where (1) the computation is robust\, in the sense that the\noutput does not change if at most a small fraction of the answers to\ noracle queries are incorrect\, or (2) the oracle is a low-degree\npolynom ial. These results suggest that interactive verification is\npossible even without debate\, under structured or noise-tolerant\noracle access.\n\nBa sed on joint work with Liyan Chen and Yael Tauman Kalai.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912e907 DTSTART;TZID=America/New_York:20260316T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260316T110000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Audrey Cheng CLASS:PUBLIC DESCRIPTION:Speaker: AUDREY CHENG\, Ph.D. Candidate\nUniversity of Californ ia\, Berkeley\n\nTalk Title: Rethinking Transaction Scheduling for Databas e Performance\n\nThe shift to multi-tenant clouds and growing data demands exacerbate\ncontention on shared data infrastructure. In these environmen ts\,\ncontention remains a primary bottleneck to performance. While there is\nextensive research on concurrency control protocols\, these approaches \nshare a fundamental limitation: they handle conflicts only after they\nh ave materialized\, missing opportunities to improve performance by\navoidi ng conflicts altogether. My research addresses this limitation\nby revisit ing transaction scheduling. Instead of resolving conflicts\nafter they occ ur\, I focus on preventing them by intelligently\nreordering transactions before execution. I present novel schedulers\nthat leverage this insight t o improve performance and ensure fairness\non real-world workloads.\n\n— \n\nAudrey Cheng is a PhD student at UC Berkeley\, advised by Natacha\nCro oks and Ion Stoica. Her research focuses on performance\noptimization for database systems. Her work has been deployed in\nindustry databases at Met a\, PlanetScale\, and TiDB. She was named a\nRising Star in EECS and has r eceived an NSF GRFP Fellowship\, a Meta\nResearch PhD Fellowship\, a Berke ley Chancellor’s Fellowship\, and a\nVLDB Best Industry Paper Award.\n\n Faculty Host:  Dimitrios Skarlatos\n\n                          Computer Science Department\n\nIn Person and Zoom Participation.  S ee announcement.\n\n→ For SCS Community Only \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912ed3b DTSTART;TZID=America/New_York:20260311T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260311T140000 LOCATION:Newell-Simon 4201 and Zoom SUMMARY:Doctoral Thesis Proposal - Yonghao Zhuang CLASS:PUBLIC DESCRIPTION:Speaker: YONGHAO ZHUANG\, Ph.D. Student\nComputer Science Depar tment\nCarnegie Mellon University\n\nTalk Title: On Efficient Language Mod el Post Training with Attention\nDisaggregation\n\nToday's LLM training in troduces more stages to further improve the\nmodel quality\, and post-trai ning is the most important. Despite the\nlong-context workload imbalance\, post-training includes reinforcement\nlearning (RL)\, which iteratively r uns the \"rollout generation - reward\nevaluation - policy update'' pipeli ne.\n\nThis thesis proposes the concept of attention server\, where the ma in\npart of attention (core attention) is disaggregated from other\ncompon ents of the model\, and is handled by an independent cluster of\nGPUs. The first benefit of the disaggregation is independent scaling\,\nenabling a higher batch size of other components\; Besides\, the core\nattention kern el only needs a subset of the GPU resources to saturate\nits memory bandwi dth demand\, allowing the GPU to utilize the remaining\nresources for comp ute intensive tasks.\n\nThesis Committee\n\nEric Xing (Chair)\n\nTianqi Ch en\n\nZhihao Jia\n\nIon Stoica (University of California\, Berkeley)\n\nHa o Zhang (University of California\, San Diego)\n\nAdditional Information\n \nIn Person and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912f12f DTSTART;TZID=America/New_York:20260309T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260309T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Simon Hørup Eskildsen CLASS:PUBLIC DESCRIPTION:Speaker: SIMON HØRUP ESKILDSEN\, Co-founder and Chief Executiv e\nOfficer\, turbopuffer\n\nTalk Title: turbopuffer: Object Storage-native Database for Search\n\nturbopuffer is an object storage-native search eng ine. It puffs data\ninto a tiered NVMe SSD and RAM cache to achieve sub-10 ms warm query\nlatency. The storage engine minimizes roundtrips to object storage to\nlimit cold latency and maximize concurrent IOPS. It even drive s\nconsensus with object storage through conditional writes.\n\nTo build a successful\, independent database you need three\n\nconditions:\n\nA new workload that compels every company on earth to put that\nworkload in you r database\, either directly or indirectlyA new storage\narchitecture\, e nabled by recent technological advances\, that uniquely\nserves that new  workloadThe ambition to service every query plan in\nthe limit\n\nIn thi s talk\, I'll make turbopuffer's case for existence\, and argue\nthat Post gres can't save you this time.\n\n—\n\nSimon Hørup Eskildsen is the co -founder and CEO at turbopuffer.\nFormerly Principal Engineer at Shopify\, where he helped scale infra\nfrom 1K → 1M RPS.\n\nThis talk is part of the PostgreSQL vs. The World Seminar Series.\n\nIn Person and Zoom Partici pation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912f4dc DTSTART;TZID=America/New_York:20260310T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260310T135000 LOCATION:Hamburg Hall 1002 and Zoom SUMMARY:Privacy Seminar - Lorrie Faith Cranor CLASS:PUBLIC DESCRIPTION:Speaker: LORRIE FAITH CRANOR\, Director and Bosch Distinguished \nProfessor in Security and Privacy TechnologiesCyLabFORE Systems\nUnivers ity Professor\, Carnegie Mellon University\n\nTalk Title: Explaining Priva cy Concepts to Four-year-olds (and Their\nParents)\n\nIn this talk I will discuss why I wrote “Privacy\, Please!” a\npicture book about privacy\ , my sources of inspiration and information\,\nand how I went about explai ning privacy concepts to young children. I\nwill discuss how I define priv acy for this audience and how I give\nchildren the vocabulary to ask for p rivacy and examples of how they\ncan achieve it\, despite the need for adu lt supervision. I’ll talk\nabout boundaries and discuss how privacy can help children calm down\nand recharge\, be creative\, keep their bodies sa fe\, and more. Finally\,\nI’ll talk about what I’ve learned from my di scussions about\nprivacy with children and caregivers and lessons for priv acy\nengineers.\n\n→ Signed books will be available for $12 after the t alk.\n\n—\n\nLorrie Faith Cranor is Director and Bosch Distinguished Pr ofessor in\nSecurity and Privacy Technologies of CyLab and FORE Systems Un iversity\nProfessor of Computer Science and of Engineering and Public Poli cy at\nCarnegie Mellon University. She directs the CyLab Usable Privacy an d\nSecurity Laboratory (CUPS) and co-directs the Privacy Engineering\nprog ram. In 2016 she served as Chief Technologist at the US Federal\nTrade Com mission. She co-founded Wombat Security\, a security awareness\ntraining c ompany acquired by Proofpoint. She founded the Symposium On\nUsable Privac y and Security (SOUPS) and co-founded the Conference on\nPrivacy Engineeri ng Practice and Respect (PEPR). She serves on the\nCenter for Democracy an d Technology (CDT) Board of Directors\, the\nAspen Institute Cybersecurity Group\, the Future of Privacy Forum (FPF)\nadvisory board\, and the Elect ronic Privacy Information Center (EPIC)\nadvisory board.\n\nIn 2003 she wa s honored as one of the top 100 innovators 35 or younger\nby Technology Re view. She was elected to the ACM CHI Academy and named\na fellow of ACM\, IEEE\, and AAAS. She received the ACM CHI Social\nImpact Award and Lifetim e Research Award\, the International\nAssociation of Privacy Professionals Privacy Leadership Award\, and\n(with colleagues) the IEEE Cybersecurity Award for Practice. She was\npreviously a researcher at AT&T-Labs Research . She holds a doctorate\nin Engineering and Policy from Washington Univers ity in St. Louis. She\nhas authored or edited several books\, including a privacy book for\nkids . She plays soccer\, walks to work\, sews her own clothing with\npockets\, and tries not to embarrass her three young adult children.\n\nThe Spring 2026 Privacy Seminars are sponsored by the Masters in\nPrivacy Engineering Program and the Carnegie Bosch Institute.\n\nIn P erson and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912f993 DTSTART;TZID=America/New_York:20260309T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260309T163000 LOCATION:Remote Access - Zoom SUMMARY:VASC Seminar - Zhujun Shi CLASS:PUBLIC DESCRIPTION:Speaker: ZHUJUN SHI\, Assistant Professor\, Department of Physi cs and\nAstronomy\, University of Pittsburgh\n\nTalk Title: Nano-optics fo r smart sensing and display\n\nNano-optical devices provide a new way to c ontrol light at the\nsubwavelength scale\, enabling optical functionalitie s beyond\nconventional optics. By engineering the nanostructures\, we can tailor\nthe optical response as a function of space\, polarization\, wavel ength\,\nand angle of incidence -- effectively turning the optical front e nd\ninto a controllable\, programmable physical layer. This creates an\nin teresting interplay between optical design and computation: on one\nhand\, nanooptics can be incorporated and co designed within the\ncomputational pipeline\, enabling new approaches to smart sensing\,\nimaging\, and displ ay\; on the other hand\, computational methods can be\nused to discover an d optimize new classes of optical instruments that\ngo beyond intuitive\, hand designed architectures. \n\nIn this talk\, I will first introduce th e basics of nanooptics\,\nhighlighting key opportunities and current limit ations. I will then\npresent several concrete examples: nanooptics for dep th sensing\,\npolarization imaging\, and nanooptics-based new AR display\n architectures. I will conclude with a view of what it would take to\nmake these systems robust and scalable\, and where collaboration with\nthe comp uter vision community can have the most impact.\n\n—\n\nZhujun Shi is an Assistant Professor of Physics and Astronomy at the\nUniversity of Pittsb urgh. Her group explores new frontiers in light\nmanipulation using nanoph otonics. Prior to joining Pitt\, she was a\nresearch scientist at Meta Rea lity Labs. She received her B.S. in\nPhysics from Tsinghua University in 2 015 and her Ph.D. in Physics from\nHarvard University in 2020.  More...  \n\nThe VASC seminar is generously sponsored by HeyGen\n\nZoom Participati on.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef912fdef DTSTART;TZID=America/New_York:20260312T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260312T173000 LOCATION:Gates HIllman 7501 and Zoom SUMMARY:Joint Crypto Seminar / Doctoral Speaking Skills Talk - Andrew Park CLASS:PUBLIC DESCRIPTION:Speaker: ANDREW PARK\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Polysys: An Algebraic Lea kage Attack Engine\n\nIn this work\, we propose a novel framework called P olySys for modeling\nand designing leakage attacks as constraint-solving a lgorithms over\npolynomial systems. PolySys formalizes the design of attac ks using\ninvertible encodings\, structural and leakage equations\, and ef ficient\nconstraint-solving algorithms including SAT and constraint solver s. It\nis capable of modeling resolution\, known-data\, and inference atta cks\nfor common leakage patterns.\n\nTo demonstrate the practicality of ou r framework\, we implement a\nPolySys attack engine in Python and apply it to state-of-the-art query\nrecovery\, data resolution\, and query inferen ce attacks on point and\nrange multi-maps. Our results show that PolySys o utperforms all\nexisting attacks under identical assumptions\, achieving u p to 60×\nhigher recovery rates in some scenarios.\n\nWhile scalability r emains a challenge for larger datasets\, PolySys\nrepresents a promising s tep toward a general-purpose framework for\ndesigning leakage attacks. We believe future work can further enhance\nits efficiency to scale to larger and more complex workloads. \n\nThe Crypto Seminar is generously sponsor ed by LayerZero Labs.\n\nThis talk is presented in Partial Fulfillment of the CSD Speaking\nSkills Requirement.\n\nIn Person and Zoom Participation.   See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91301c7 DTSTART;TZID=America/New_York:20260309T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260309T130000 LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar - Miranda Wei CLASS:PUBLIC DESCRIPTION:Speaker: MIRANDA WEI\, Postdoctoral Research Associate\, Center for\nInformation Technology Policy (CITP)\, Princeton University\n\nTalk Title: Harm is Not an Anomaly nor a Reason to Despair\n\nSome software vul nerabilities are one-off bugs or typos in code.\nOthers\, however\, are sy mptoms of deeper design flaws embedded in a\nsystem. As we increasingly co nfront harms across digital experiences\,\nI argue that these harms are no t anomalies\, but canaries in a coal\nmine for systemic inequities. Synthe sizing feminist epistemology with\ncomputer security and privacy methodolo gies\, I apply a lens of\nsociotechnical safety to the urgent challenge of combating image-based\nsexual abuse. In this talk\, I will share empirica l insights into the\nmechanisms of digital abuse and propose interventions that bridge\ntechnical defenses and social realities. I conclude by argui ng that\nthis approach provides a grounded basis for hope: by reframing ha rm as\na measurable systemic failure of sociotechnical safety\, we move\nt owards rigorous models for intervention and repair.\n\n—\n\nMiranda Wei studies online abuse and societal factors in\nsociotechnical safety\, espe cially concerning social media\, gender\, and\ninterpersonal relationships . Their research interests lie at the\nintersections of computer security and privacy (S&P)\, human-computer\ninteraction (HCI)\, and feminist scien ce and technology studies (STS).\nThey publish in leading S&P and HCI venu es such as USENIX Security\,\nACM CHI\, IEEE S&P\, and SOUPS. Wei holds a Ph.D. from the Paul G. Allen\nSchool of Computer Science & Engineering at the University of\nWashington\, where they were a member of the Security & Privacy\nResearch Lab and Tech Policy Lab\, supported by a Google PhD\nFe llowship. They received a B.A. from the University of Chicago in\npolitica l science and computer science. In fall 2026\, they will start\nas an assi stant professor at École Polytechnique Fédérale de\nLausanne (EPFL) in Lausanne\, Switzerland.\n\nFaculty Host: Lorrie Cranor\n\nIn Person and Zo om Participation.  See announcement.\n\n→  CyLab Seminars are only ope n to CyLab Partners and current CMU\nfaculty\, staff and students.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9130657 DTSTART;TZID=America/New_York:20260325T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260325T120000 LOCATION:Remote Access SUMMARY:Re-Braiding Cybernetics and AI Symposium: Unspoken Continuities CLASS:PUBLIC DESCRIPTION:Speaker: SYMPOSIUM #1 - THE SPLIT\n\nThe Re-Braiding Cybernetic s & AI Project has the overarching goal to\nbring Cybernetics and Artifici al Intelligence back into conversation.\nThe overall intention is to explo re\, from the origins to the present\nday\, how we’ve arrived at the unp redictable and fast-moving\ntheoretical\, technical\, and business advance s that will inevitably\nimpact the future of our society. \n\nSymposium # 1 titled “The Split” will focus on the 1956 Dartmouth\nSummer Conferen ce and its immediate aftermath. Researchers had\ngathered to solidify thei r intention to move away from\nCybernetics—away from its technology and its guru\, at the\nleast—and to embrace the ascendant digital computer a nd the promise\nof its symbolic programming as a means to what they newly coined as\n“artificial intelligence.” \n\nWhy did these researchers\, who became more famous as their careers\nmatured\, want to split off from their intellectual parents? What did\nthey leave behind in Cybernetics t hat could help address the\nlimitations of AI today? And what could advan ces in AI bring to\nmodern Cybernetics?Is there a productive path to bring ing them back\ninto conversation?\n\nREGISTER → Participation is free\n \nThe Re-Braiding Cybernetics and AI Symposium Series will run\nperiodical ly until Spring 2027 and be captured in recordings and\ntranscripts and pu blished with open commentary. The Re-Braiding\nProject will culminate in a physical exhibit in April 2027 at CMU’s\nPosner Center. Books and other physical artifacts representing the\ncore concepts of each field\, throug h the personal collections of Heinz\nvon Foerster and Allen Newell\, will afford a compare-and-contrast\, as\nwell as paths toward\, the cooperative shared futures of the two\nfields.\n\nThe Re-Braiding Project is supporte d by a generous grant from Carnegie\nMellon-Architecture  Computational D esign Laboratory (CodeLab). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9130a72 DTSTART;TZID=America/New_York:20260312T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260312T153000 SUMMARY:National Robotics Engineering Center (NREC) 30-year Anniversary CLASS:PUBLIC DESCRIPTION:Speaker: SCS Faculty and StaffTalk Title: Open House with Robot ics\nSystem Demonstrations\n\nSCS Faculty and Staff are invited to celebra te NREC's 30-year\nanniversary during our open house.\n\nRSVP → request ed by March 2 (see invitation)\n\n           → Business casual at tire and closed-toe shoes.\n\nTransportation will be provided between the CMU campus and NREC. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9130d5c DTSTART;TZID=America/New_York:20260225T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260225T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Zeyu Zheng CLASS:PUBLIC DESCRIPTION:Speaker: ZEYU ZHENG\, Ph.D. Studnt\, Ph.D. Program in Algortihm s\,\nCombinatorics and Optimization\, Department of Mathematical Sciences\ ,\nCarnegie Mellon University\n\nTalk Title: The generalized trifference p roblem\n\nWe study the problem of finding the largest number T(n\,m) of te rnary\nvectors of length n such that for any three distinct vectors there are\nat least m coordinates where they pairwise differ.\n\n   For m=1\, this is the classical trifference problem which is wide\nopen.\n\n   We prove upper and lower bounds on T(n\,m) for various ranges of\nthe paramet er m and determine the\n\n   phase transition threshold on m=m(n) where T(n\,m) jumps from\nconstant to exponential in n.\n\nBy relating the linea r version of this problem to a problem on\nblocking sets in finite geometr y\, we give explicit constructions and\nprobabilistic lower bounds.\n\nJoi nt work with Anurag Bishnoi\, Bartłomiej Kielak\, Benedek Kovács\,\nZolt án Lóránt Nagy\, Gábor Somlai\, and Máté Vizer. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91310e0 DTSTART;TZID=America/New_York:20260226T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260226T173000 LOCATION:Gates Hillman 7501 and Zoom SUMMARY:Joint Crypto Seminar and Doctoral Speaking Skills Talk - Changrui M u CLASS:PUBLIC DESCRIPTION:Speaker: CHANGRUI MU\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Public-Key Encryption from t he MinRank Problem\n\nWe construct a public-key encryption scheme from the hardness of the\n(planted) MinRank problem over uniformly random instance s. This\ncorresponds to the hardness of decoding random linear rank-metric \ncodes. Existing constructions of public-key encryption from such\nproble ms require hardness for structured instances arising from the\nmasking of efficiently decodable codes. Central to our construction is\nthe developme nt of a new notion of duality for rank-metric codes.\n\nPresented in Parti al Fulfillment of the CSD Speaking Skills\nRequirement\n\nThe Crypto Semin ar is sponsored by LayerZero\n\nIn Person and Zoom Participation.  See an nouncement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9131463 DTSTART;TZID=America/New_York:20260226T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260226T160000 LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Carl Schildkraut CLASS:PUBLIC DESCRIPTION:Speaker: CARL SCHILDKRAUT\, Ph.D. Student\, Department of Mathe matics\,\nStanford University\n\nTalk Title: Ramsey Cayley graphs over non -abelian groups\n\nA conjecture of Alon states that\, for some absolute co nstant C\, every\nfinite group G possesses a Cayley graph with clique and independence\nnumber each at most C*log|G|. Recently\, Conlon\, Fox\, Pham \, and\nYepremyan have verified this conjecture for most abelian groups us ing\nmainly graph-theoretic techniques. In this talk\, I will discuss some \nrecent work of mine extending their results to many non-abelian\ngroups. In addition to combinatorial inputs from\nConlon–Fox–Pham–Yepremyan \, the techniques used are inspired by\nadditive combinatorics and expansi on in groups.\n\n4:00 pm → Jane Street-sponsored tea and cookies in the Math Lounge\n(bring your mug!) \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91317b9 DTSTART;TZID=America/New_York:20260416T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260416T150000 LOCATION:McConomy Auditorium\, Cohon University Center SUMMARY:Joint CMU-PITT Mathematical Sciences Colloquium - June Huh CLASS:PUBLIC DESCRIPTION:Speaker: JUNE E. HUH\, Professor\, Department of Mathematics\, Princeton\nUniversity and 2022 Fields Medal Laureate\n\nTalk Title: Volume Polynomials\n\nVolume polynomials constitute a distinguished class of log -concave\npolynomials with remarkable analytic and combinatorial propertie s\narising from convex bodies and projective varieties. I will introduce\n new entropy inequalities satisfied by volume polynomials\, discuss\napplic ations to the combinatorics of algebraic matroids\, introduce the\nnew cla ss of analytic matroids\, and pose several open questions.\n\nBased on var ious joint works with Lukas Grund\, Mateusz Michalek\,\nHenrik Süss\, Bot ong Wang\, and Jian Xiao. \n\nSponsored by Jane Street\n\nReception to Fo llow \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9131b03 DTSTART;TZID=America/New_York:20260223T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260223T130000 LOCATION:Newell-Simon 3305 and Zoom SUMMARY:Accessibility Lunch Seminar - Kynnedy Smith CLASS:PUBLIC DESCRIPTION:Speaker: KYNNEDY SMITH\, Ph.D. Student\nHuman-Computer Interact ion Institute\nCarnegie Mellon University\n\nTalk Title: Identifying\, Exp laining\, and Correcting Ableist Language\nwith AI\n\nAbleist language per petuates harmful stereotypes and exclusion\, yet\nits nuanced nature makes it difficult to recognize and address.\nArtificial intelligence could ser ve as a powerful ally in the fight\nagainst ableist language\, offering to ols that detect and suggest\nalternatives to biased terms. This two-part s tudy investigates the\npotential of large language models (LLMs)\, specifi cally ChatGPT\, to\nrectify ableist language and educate users about inclu sive\ncommunication. We compared GPT-4o generations with crowdsourced\nann otations from trained disability community members\, then invited\ndisable d participants to evaluate both. Participants reported equal\nagreement wi th human and AI annotations but significantly preferred\nthe AI\, citing i ts narrative consistency and accessible style. At the\nsame time\, they va lued the emotional depth and cultural grounding of\nhuman annotations. The se findings highlight the promise and limits of\nLLMs in handling cultural ly sensitive content. Our contributions\ninclude a dataset of nuanced able ism annotations and design\nconsiderations for inclusive writing tools.\n\ n—\n\nThis seminar series explores the evolution and future directions o f\naccessibility research and practice. Each week\, we discuss diverse\nfa cets of accessible and assistive technologies\, featuring speakers\nfrom a cademia\, industry\, and beyond. The seminar is organized by\nthe Accessi bility Research Group.\n\nIn Person and Zoom Participation.  See annnounc ement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9131ef7 DTSTART;TZID=America/New_York:20260224T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260224T130000 LOCATION:Gates Hillman 9115 SUMMARY:Database Seminar - Nikhil Simha CLASS:PUBLIC DESCRIPTION:Speaker: NIKHIL SIMHA\, Co-founder and Chief Technology Officer \,\nzipline.ai\n\nTalk Title: Chronon - Mixed-workload Data Processing Fra mework\n\nChronon is a data processing framework open-sourced by Airbnb. I t is\nadopted across organizations like Stripe\, Netflix\, OpenAI\, and Ub er.\nChronon was originally built for ML applications. It has since been\n adopted to power a variety of use-cases—heuristics for rule engines\,\nc ontext for LLMs\, user-facing and business-facing metrics.\n\nChronon is a dopted for its ability to generate training data at scale\nand serve featu res with very low latency with a simple\, high-level\nAPI. It abstracts aw ay the effort required to manually build batch and\nstream processing pipe lines\, indexes\, and services.\n\nThis talk will focus largely on algorit hms and optimizations inside\nChronon. We will only briefly touch upon the core concepts of the API\nand a couple of example use-cases. \n\n—\n\n Nikhil Simha Raprolu is the Co-founder & CTO at zipline.ai. Prior to\nthat he worked on the ML Infra team - where he open-sourced Chronon.\nAt Faceb ook he worked on stream processing systems\, schedulers and\ncompilers - e g.\, Stylus & Turbine. Prior to that he worked on\ndistributed data proces sing infrastructure at Amazon and Walmart\nLabs. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91322d6 DTSTART;TZID=America/New_York:20260226T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260226T153000 LOCATION:Reddy Conference Room\, Gates HIllman 4405 and Zoom SUMMARY:Doctoral Thesis Proposal - Sam Arch CLASS:PUBLIC DESCRIPTION:Speaker: SAM ARCH\, Ph.D. Student\nComputer Science Department\ nCarnegie Mellon University\n\nTalk Title: Leveraging Optimization-Enablin g Properties of\nUser-Defined Functions for Efficient Database Query Execu tion\n\nAfter decades of research\, analytical database management systems \n(DBMSs) have become remarkably effective at optimizing and executing\nSQ L queries. However\, many users write queries that are not written\nentire ly in SQL. Instead\, these queries invoke user-defined functions\n(UDFs)\, external functions written in non-SQL programming languages\nsuch as Pyth on or PL/SQL. UDFs provide software engineering benefits\nby enabling code reuse and by extending the DBMS’s capabilities to\ninclude those of the UDF language. However\, UDFs are inherently\nnon-relational\, which makes them challenging for DBMSs to reason about\nand execute efficiently. Effe ctive optimization is also challenging\nbecause UDF languages are Turing-c omplete\, allowing UDFs to be\narbitrarily complex. Although general-purpo se optimization techniques\ncan improve UDF performance (e.g.\, compilatio n and batching)\, they\ntarget arbitrary UDF code and therefore have limit ed effectiveness. We\nobserve that the most beneficial UDF optimizations ( e.g.\, memoization\nand inlining) leverage key optimization-enabling prope rties of UDFs\n(i.e.\, how users actually use them in practice).\n\nIn thi s proposal\, we present multiple techniques that leverage\noptimization-en abling properties of UDFs to improve database query\nexecution performance . First\, we observe that inlining only the\nrelevant pieces of a UDF impr oves performance\, and leverage UDF\ndecomposability to break UDFs into pi eces and hide irrelevant pieces\nthrough outlining. Next\, we observe that processing all unique UDF\ninputs simultaneously improves parallelism\, a nd leverage UDF\nredundancy to build lightweight indexes during query proc essing to\navoid repeated UDF invocations.\n\nWe propose extending our pre liminary work by observing that enabling\ninter-tuple parallelism of UDFs improves query execution performance.\nWe plan to leverage UDF pipelining\ , the observation that UDFs operate\nas a pipeline of data transformations over their inputs\, to enable\nfusion and auto-vectorization of pipeline stages. Collectively\, the\ntechniques presented in this dissertation will enable an analytical\ndatabase system to execute queries that contain UDF calls efficiently.\n\nThesis Committee\n\nAndrew Pavlo (Co-Chair)\n\nTodd C. Mowry (Co-Chair)\n\nJignesh Patel (Co-Chair)\n\nPhillip B. Gibbons\n\n Joseph M. Hellerstein (University of California\, Berkeley)\n\nAdditional Information\n\nIn Person and Zoom Participation.  See announcement. \n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91327cb DTSTART;TZID=America/New_York:20260223T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260223T170000 LOCATION:Steinberg Auditorium\, Baker Hall A53 SUMMARY:Statistics and Data Science Seminar - Sahil Loomba CLASS:PUBLIC DESCRIPTION:Speaker: SAHIL LOOMBA\, Chapman–Schmidt AI in Science Fellow\ ,\nImperial College London\n\nTalk Title: Inference and Intervention in Pa rtially Observed Social\nNetworks\n\nThe behavior of complex social system s arises jointly from their\nstructure — how individuals are connected — and their mechanism\n— how individuals influence one another. In man y settings\, these\ncomponents are only partially observed: structures are measured\nincompletely\, mechanisms may be misspecified\, and behaviors a re\nobserved only under experimentally realized treatments\, as in the\nfu ndamental problem of causal inference.\n\nIn this talk\, I will examine ho w partial observability creates\nfoundational challenges for both statisti cal inference and the design\nof interventions in social systems. From a s tatistical perspective\, I\nwill discuss how asymptotic arguments can be u sed to analytically\nestimate global connectivity properties of social net works from\nincomplete structural observations\, like egocentric samples\, enabling\ninferences about shortest paths in population-level systems.\n\ nFrom a causal and decision-theoretic perspective\, I will argue that\npar tial behavioral observability in social networks requires a\nrethinking of canonical causal estimands. I will show that there is a\nfundamental tens ion between individual-level causal contrasts and\npolicy-relevant social objectives\, and that this tension can only be\nresolved by focusing on ex pected average outcomes over a space of\ntreatment policies\, yielding a u nifying framework for causal inference\nand decision-making in networked s ocial interventions.\n\nThis perspective naturally leads to new challenges in off-policy\ncausal estimation: I will discuss how one can evaluate the effects of\nnetworked interventions different from those experimentally\n implemented\, and how these ideas connect to testing and learning\nunderly ing behavioral mechanisms. I will conclude by discussing how\nthese insigh ts inform the design of optimal experiments and effective\ninterventions i n social networks under partial observability.\n\n—\n\nSahil Loomba is a Chapman–Schmidt AI in Science Fellow at Imperial\nCollege London. He wa s previously a Schmidt Science Fellow at the\nInstitute for Data\, Systems \, and Society in the Schwarzman College of\nComputing at MIT\, and an EPS RC Doctoral Prize Fellow at Imperial\,\nwhere he also earned his PhD in Ma thematics. His doctoral thesis on\nsparse and partially observed large-sca le networks was awarded the\nYael Naim Dowker Prize. His research lies at the intersection of\nstatistics\, applied probability\, and causal inferen ce\, focusing on\nunderstanding connectivity\, behavior\, and intervention s in large-scale\nsocial systems\, with applications in public health. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9132cc6 DTSTART;TZID=America/New_York:20260225T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260225T170000 LOCATION:Steinberg Auditorium\, Baker Hall A53 SUMMARY:Statistics and Data Science Seminar - Yuhua Zhang CLASS:PUBLIC DESCRIPTION:Speaker: YUHUA ZHANG\, Postdoctoral Fellow\, Department of\nBio statistics\, Harvard University\n\nTalk Title: From Communities to Causes: Statistical Methods for\nNetwork Data\n\nNetworks are everywhere\, from p atient transfers across hospitals to\ninteractions on social platforms. Ye t such data is not fully explored.\nIf we can learn how entities connect a nd influence one another\, we can\nidentify population disparities in comp lex systems and design\ninterventions that improve social and public healt h outcomes. In this\ntalk\, I will cover two complementary themes\, modeli ng of\nnetwork-structured data and causal inference in networks\, and show how\nnetwork structure can be translated into actionable scientific and\n policy insight.\n\nIn the first part\, I will discuss community detection in networks.\nWhile many methods aim to recover community structure\, fewe r account\nfor the fact that modern networks often arise from interaction\ nprocesses\, where edges\, not nodes\, are the basic statistical units. I\ nwill present block edge exchangeable models (BEEM) for interaction\nnetwo rks with latent node-level community structure\, and show how this\nmodel enables inference for sparse network data.\n\nIn the second part\, I will discuss causal inference with unknown\ninterference networks. Interference \, the phenomenon that a unit’s\noutcome depends on other units’ treat ments\, creates major challenges\nfor identifying and estimating causal ef fects. Most existing\napproaches assume the interference network is known\ , which is often\nunrealistic in practice because such a network is typica lly latent or\nonly partially observed. To address this\, I develop a fram ework for\nidentifying and estimating heterogeneous\, group-level interfer ence\neffects without requiring a known interference network. I will\nillu strate these ideas with applications to healthcare systems\, social\nnetwo rks\, and more.\n\n—\n\nYuhua Zhang is a postdoctoral fellow working wit h Dr. Jukka-Pekka\nOnnela in the Department of Biostatistics at Harvard Un iversity. Prior\nto this\, she earned her Ph.D. from the Department of Bio statistics at\nthe University of Michigan\, supervised by Dr. Walter Demps ey and Dr.\nSebastian Zoellner. Her research interests include network ana lysis\,\ncausal inference in networks and their applications in social and \nhealth science. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9133150 DTSTART;TZID=America/New_York:20260313T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260313T124500 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:From Lab to Launch: Commercializing Your Research CLASS:PUBLIC DESCRIPTION:Speaker: CRAIG MARKOVITZ\, Director\, Swartz Center\nEntreprene ur-in-Residence and Mentor ProgramDistinguished Service\nProfessor of Entr epreneurship\, Tepper School of Businessand\, former\nPresident and CEO\, Blue Belt Technologies\n\nTalk Title: For SCS Faculty\, Post-docs\, and Ph .D. Candidates\n\nJoin the Swartz Center for Entrepreneurship and Center f or Technology\nTransfer and Enterprise Creation (CTTEC) for an  for  in formal\nlunch for School of Computer Science faculty\, Ph.D. students\, a nd\npostdoctoral fellows where we demystify the path from academic\ninno vation to commercial impact.\n\nWhy Attend?\n\nHear from Craig Markovitz \, Director of the Swartz Center\nEntrepreneur-in-Residence and Mentor Pro gram and former President and\nCEO of Blue Belt Technologies. Blue Belt wa s CMU's first spin off\ncompany to secure FDA clearance and sell a medica l robotics\ntechnology in the market.  The company's handheld\, bone cu tting\nrobotic handpiece was based on technology licensed from Robotics\nI nstitute.  Craig will provide an overview of the company's 13-year\ncomme rcialization journey and answer your questions related to deep\ntech comm ercialization.Meet and network with our faculty research\nfocused Entrepre neurs-in-Residence\, CTTEC\, and Swartz Center\nstaff.Learn about Swartz Center and CTTEC resources dedicated to\nfaculty research founders.\n\nAdd itional Information\n\nREGISTER  → by Wednesday\, March 11\n\n                   → Lunch will be provided\, please\nindicate dietar y preferences upon registration \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913351e DTSTART;TZID=America/New_York:20260415T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260415T173000 LOCATION:Cohon University Center SUMMARY:Celebration of Education CLASS:PUBLIC DESCRIPTION:Carnegie Mellon University’s Celebration of Education honors the\noutstanding educators\, advisors\, and mentors who exemplify the\nuni versity’s commitment to teaching\, innovation\, and student\nsuccess. Ea ch year\, faculty and staff are recognized through a series\nof awards tha t highlight excellence in teaching\, advising\, mentoring\,\nand education al outreach. Each spring\, the annual ceremony brings\ntogether the CMU co mmunity to celebrate these achievements and to\nrecognize the individuals whose dedication and creativity shape the\nuniversity’s culture of learn ing.\n\nThe College Teaching Awards honor exemplary teaching by faculty\nm embers within the seven colleges at Carnegie Mellon. Each college\nselects their own recipient(s) based on specific criteria and they are\nchosen th roughout the academic year\, depending on the college's award\ncycle. The most recent awardees are again recognized at\nthe Celebration of Educatio n award ceremony. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91338af DTSTART;TZID=America/New_York:20260223T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260223T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Adam Prout CLASS:PUBLIC DESCRIPTION:Speaker: ADAM PROUT\, Software Engineering\, Azure PostgreSQL t eam\,\nMicrosoft\n\nTalk Title: HorizonDB: Co-Designing PostgreSQL and Azu re for\nCloud-Native OLTPMonday\, February 23\, 2026\, 4:30 – 5:30pm A zure\nHorizonDB is a new PostgreSQL service that improves the OLTP\nperfor mance and reliability of upstream PostgreSQL through\nco‑design. By evol ving both Azure’s infrastructure and PostgreSQL\nitself\, HorizonDB enab les the two to work together more efficiently.\nThis talk introduces the a rchitecture of HorizonDB and explores\nseveral key design and implementati on decisions that enable more\nscalable\, reliable PostgreSQL in Azure.— Adam Prout is a Software\nEngineer on the Azure PostgreSQL team at Microso ft. He has spent his\ncareer designing and building SQL database systems\, from work on SQL\nServer to co‑founding MemSQL (now SingleStore)\, wher e he helped\nbuild a distributed HTAP database. Today\, he works on cloud ‑native\nPostgreSQL systems at Microsoft.This talk is part of the Postgr eSQL\nvs. The World Seminar SeriesIn Person and Zoom Participation.  See\ nannouncement. Event Type: Seminars Room Number: Virtual Presentation\n- ET Building: Remote Access - Zoom Speaker's Name: ADAM PROUT\nSpeaker W ebsite: www.linkedin.com [http://www.linkedin.com]…\nSpeaker's Professi onal Title: Software Engineering\, Azure PostgreSQL\nteam\, Microsoft Tal k Title: HorizonDB: Co-Designing PostgreSQL and\nAzure for Cloud-Native O LTP For More Information: db-www@cs.cmu.edu\nAffiliations: Computer Scie nce Department (CSD)\nOrganization(s): School of Computer Science Event W ebsite\nTitle: Event Website Event Website URL: db.cs.cmu.edu…\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9133d02 DTSTART;TZID=America/New_York:20260312T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260312T133000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:AI-SDM Seminar - Paul Lehner CLASS:PUBLIC DESCRIPTION:Speaker: PAUL LEHNER\, Senior Principal Scientist and Decision\ nScientist\nThe MITRE Corporation\n\nTalk Title: Can we accurately estimat e AI classifier accuracy without\nground truth data or trustworthy expert ratings?\n\nDr. Paul Lehner is currently a Senior Principal AI and Decisio n\nScientist in the MITRE Corporation. Dr. Lehner has directed and\npublis hed on a variety of topics including innovative research\nmethods\, reason ing under uncertainty\, AI planning and game playing\,\nhuman judgment and decision making\, predictive analysis and machine\nlearning.\n\nDr. Lehne r recently completed a four-year detail to the Intelligence\nAdvanced Rese arch Projects Activity (IARPA) where he served as the\nChief for Test and Evaluation (T&E). In previous roles in MITRE\, Dr.\nLehner was the Chief E ngineer for the Center for Enterprise\nModernization as well as for two di fferent IT research divisions.\n\nREGISTER → to attend in-person or on zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9134088 DTSTART;TZID=America/New_York:20260319T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260319T130000 LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI-SDM - Student Brainstorming Session CLASS:PUBLIC DESCRIPTION:Talk Title: AI-SDM Student Brainstorming Session\n\nJoin the mo nthly student-focused brainstorming sessions.  Led by\nstudents of all le vels\, these sessions offer a casual forum to discuss\nAI-related research while exploring cross-cutting connections between\nvarious disciplines in artificial intelligence. All students\nwelcome! \n\nStudents meet regula rly to participate in informal discussion sessions\nthat delve into cuttin g-edge AI topics. These discussions are a\nbreeding ground for collaborati on\, innovative thinking\, and\nproblem-solving from the ground up. They p rovide a stress-free forum\nfor exchanging ideas\, brainstorming new appro aches to challenges\, and\nfostering lasting connections within the AI-SDM community in an\nenvironment distinct from a traditional seminar. Lunch i s provided for\nin-person attendees! Please complete the RSVP form below t o help us\nwith logistics.\n\nREGISTER - to attend in person or on Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913440e DTSTART;TZID=America/New_York:20260223T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260223T130000 LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar - Liz Izhikevich CLASS:PUBLIC DESCRIPTION:Speaker: LIZ IZHIKEVICH\, Assistant Professor\, Department of E lectrical\nand Computer EngineeringSamueli School of Engineering\, Univers ity of\nCalifornia\, Los Angeles\n\nLow Earth Orbit (LEO) satellite networ ks are rapidly becoming a\ncritical component of the global Internet\, yet their operational\nbehavior is difficult to observe at scale. Protecting and improving\nthe performance and security of LEO networks requires visib ility into\ntheir operations and users. In this talk\, I show how data-col lecting\nsystems reveal previously unseen operational challenges in LEO\nn etworks\, including unexpected routing through third-party countries\nand the exposure of legacy systems.\n\n—\n\nLiz Izhikevich is an Assistant P rofessor in Electrical and Computer\nEngineering at UCLA. Her research foc uses on improving Internet\nperformance and security\, both on earth and i n space. Her work has\nbeen recognized nationally\, including Forbes’ 30 Under 30 in Science\n(2025) and the Internet Measurement Conference Commu nity Contribution\nAward (2022). Government agencies and industry partners rely on her\ntools to identify and mitigate online vulnerabilities\, and her\ncollaborations have improved video delivery for over a million\nsatel lite broadband users. She received her Ph.D. in Computer Science\nfrom Sta nford University in 2024 and has held research positions at\nNetflix and C ensys.\n\nFaculty Host:  Justine Sherry\n\nIn Person and Zoom Participati on.  See announcement.\n\n→ This seminar is only open to CyLab Partner s and current CMU\nfaculty\, staff and students \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91347d3 DTSTART;TZID=America/New_York:20260216T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260216T130000 LOCATION:Group Viewing Newell-Simon 3305 and Zoom SUMMARY:Accessibility Lunch Seminar - Owen Kent CLASS:PUBLIC DESCRIPTION:Speaker: OWEN KENT\, Co-founder\, ATDev\n\nTalk Title: Designin g With\, Not For: User-Centered Design as a Startup\nStrategy in Assistive Technology\n\nIn this talk\, I will share lessons from building assistive robotics\nand rehabilitation devices in a startup environment. We will ex plore\nhow participatory design influences product-market fit\, regulatory \nstrategy\, clinical validation\, and fundraising. I will argue that\ninc lusive design is not only ethically necessary\, but commercially\nadvantag eous. For entrepreneurs and researchers alike\, designing with\nusers is n ot a constraint. It is a competitive edge.\n\nAbout ATDev\n\nThe seminar i s organized by the Accessibility Research Group at\nCarnegie Mellon Univ ersity.\n\nGroup Viewing and Zoom Participation.  See announcement. \n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9134b4f DTSTART;TZID=America/New_York:20260219T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260219T160000 LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Michael Zheng CLASS:PUBLIC DESCRIPTION:Speaker: MICHAEL ZHENG\, Ph.D. Student\, Department of Mathemat ics\,\nEmory University\n\nTalk Title: A Lovász-Kneser theorem for triang ulations\n\nWe show that the Kneser graph of triangulations of a convex n- gon has\nchromatic number n - 2.\n\nJoint work with Anton Molnar\, Cosmin Pohoata\, and Daniel G. Zhu.\n\n4:00 pm → Jane Street-sponsored tea and cookies in the math lounge\n(bring your mug!) \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9134e7f DTSTART;TZID=America/New_York:20260410T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260410T200000 LOCATION:Rashid Auditorium\, Gates HIllman 4401 SUMMARY:SIGBOVIK 2026 CLASS:PUBLIC DESCRIPTION:Friday\, April 10\, 2026\, 5 – 8pm To friends and foes from all\naround\, our colleagues borne of flesh and knowing silicon\,I bid th ee\nwarning:The beast\, it stirs\, and once again\, SIGBOVIK be upon\nus.R ashid Auditorium\, The Universitie Carnegie-Mellon.April 10\, 5PM\nEastern Time\, til’ nothing’s left to bear.(and\, hopefully\, live on\nYouTub e. We’ll get back to you on that.)Our 20th colloquium of all\nthe greate st living minds\, on topics ranging from.. Algorithmic\nComplication Zero- trust Whistleblowing IA (Iowa) Quantum Philanthropy\nScience Computers Con notative Semantics of Arcane Sigils Arcane Sigils\nas a Service … and ma ny more!A deadline looms\, and close ahead\n→March the Fourth.  May it be the first of many.Submit your papers\,\nwhether you intend to give a pr esentation or not.And to those with\ngifts more visual—a presentation\, vid'eo\, a demonstration\, et\nceterawho wish to grace our stage but pen n o paper:  fear not!Your\nmoment comes. We shall unveil a path for thee an on.Assist us in a fun\nevening of Scientography\, Gistaculation\, and othe r\n%00%00%00%00%00%00%00.For those not in the know…SIGBOVIK is an\neveni ng of tongue-in-cheek academic presentations. If we have a goal\,\nit's to poke fun at our fields and provide a venue for silly ideas\nwith (often\, but not always) serious executions. SIGBOVIK has both\npublished proceedi ngs and live talks\, and everyone is welcome to\nparticipate. All subjects are welcome\, although our audience is\nprimarily made up of computer sci entists. The best way to get a feel\nfor SIGBOVIK is to look at our past p roceedings\, which you can find at\nsigbovik.org.We celebrate his name\, B OVIKLivestream for ye who can't\nmake it in person.  Event Type: Colloq uium Room Number: In Person\n(right now) Building: Rashid Auditorium\, G ates HIllman 4401\nAffiliations: Computer Science Department (CSD)\, Huma n-Computer\nInteraction Institute (HCII)\, Language Technologies Institute (LTI)\,\nMachine Learning Department (MLD)\, Ray and Stephanie Lane\nComp utational Biology Department (CBD)\, Robotics Institute (RI)\,\nSoftware a nd Societal Systems Department (S3D)\nOrganization(s): School of Computer Science\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9135310 DTSTART;TZID=America/New_York:20260218T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260218T130000 LOCATION:Gates Hillman 8102 SUMMARY:Joint Theory Lunch Seminar / Doctoral Speaking Skills Talk - Rose\n Silver CLASS:PUBLIC DESCRIPTION:Speaker: ROSE SILVER\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: History-Independent Load Bal ancing\n\nWe show that there exists a (strongly) history-independent two-c hoice\nballs-and-bins algorithm that supports both insertions and deletion s\non a set of up to m balls\, while guaranteeing a maximum load of m / n\ n+ O(1) with high probability\, and achieving an expected recourse of\n 𝑂⁡(log⁡log⁡(𝑚/𝑛)) per operation. To the best of our\nknowle dge\, this is the first history-independent solution to achieve\nnontrivia l guarantees of any sort for 𝑚/𝑛 ≥ 𝜔⁡(1)\, and is\nthe first fully dynamic solution (history independent or not) to\nachieve O(1) overl oad with o(m/n) expected recourse.\n\nJoint work with Michael A. Bender\, William Kuszmaul\, and Elaine Shi.\n\nTo appear at SODA 2026\n\nPresented as part of the Theory Lunch Seminar\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91356be DTSTART;TZID=America/New_York:20260217T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260217T135000 LOCATION:Hamburg Hall 1002 and Zoom SUMMARY:Privacy Seminar - Jacob Gursky CLASS:PUBLIC DESCRIPTION:Speaker: JACOB GURSKY\, Technical Privacy Consultant\, Privacy Company\,\nThe Hague\n\nTalk Title: RiskAI Business\n\nIn this talk\, Jaco b will discuss a recent data protection impact\nassessment (DPIA) performe d on a Dutch generative AI system developed\nfor educational settings. The service\, EduGenAI is designed to be a\nprivacy preserving alternative to United States-based companies that\ncreate a risk under the GDPR of impro per transfer of personal data to\nthird countries.\n\nIn the talk Jacob wi ll discuss their experiences navigating the\ndifferences in American and E U based legal privacy frameworks\, the\nstructure of the risk assessments produced by Privacy Company\, and how\nthese risk assessments become part of larger systemic protections for\ndata subjects through stakeholder mana gement and negotiations with\ntechnology providers.  \n\nFor anyone curi ous\, these resources will provide additional context\nfor the talk\, but the talk is not designed to require familiarity with\nthem:\n\nPrivacy Com pany DPIA on EduGenAI\,  Privacy Company DPIA on EduGenAI\nfor Surf (Eng lish language) Dutch National Model for DPIAs (Dutch\nlanguage)Privacy Co mpany DPIA on Microsoft CoPilot\n\n—\n\nJacob Gursky (they/she) is a 202 2 graduate of the CMU MSIT-Privacy\nEngineering program. Jacob specializes in privacy rights and\ncountering disinformation. They studied communicat ion and social\nnetwork systems at the University of Pennsylvania before a ttending\nCMU. During their time at the Propaganda Vertical at the Center for\nMedia Engagement at the University of Texas at Austin\, Jacob produc ed\nresearch linking the risks of digital surveillance and non-transparent \nplatforms to systemic disinformation and violent extremism.\n\nJacob now works in The Hague\, Netherlands as a technical privacy\nconsultant for Privacy Company.  At Privacy Company\, Jacob helps\ncreate data protectio n and human rights impact assessments for clients\nin the Dutch government and EU institutions writ large. These reports\nhave ranged across multipl e applications and services (video\nconferencing\, social media applicatio ns\, VR headsets\, employee\nmanagement software\, etc).\n\nSpring 2026 Pr ivacy Seminars are sponsored by the Masters in Privacy\nEngineering Progra m and the Carnegie Bosch Institute.\n\nIn Person and Zoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9135b95 DTSTART;TZID=America/New_York:20260223T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260223T133000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:The Future is Now: Black History in the Making CLASS:PUBLIC DESCRIPTION:Speaker: Panel DiscussionTalk Title: The Future is Now: Black H istory\nin the Making\n\nThe SCS Office of Community Engagement\, in colla boration with HCII\, \ninvites you to attend the School of Computer Scien ce inaugural Black\nHistory Month event.  \n\nJoin us for a panel discus sion celebrating Black excellence in\ncomputer science. We’ll spotlight pioneering Black innovators who\nshaped the field\, hear from five current Black SCS students about\ntheir experiences and perspectives\, and engage with critical questions\nabout technology’s impact on marginalized comm unities through the\nlens of scholars like Ruha Benjamin and Timnit Gebru. This event\nhonors both the contributions of Black technologists througho ut\nhistory and the ongoing work to build a more equitable tech future.\n\ nLunch will be provided.\n\nRSVP → please register if you plan to atten d. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9135f2a DTSTART;TZID=America/New_York:20260220T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260220T123000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:AI-SDM Seminar - David Alvarez-Melis CLASS:PUBLIC DESCRIPTION:Speaker: DAVID ALVAREZ-MELIS\, Assistant Professor of Computer Science\,\nJohn A. Paulson School of Engineering and Applied Sciences\, Ha rvard\nUniversity\n\nTalk Title: From Discrete Choices to Continuous Space s: Interpolating\nModels\, Data\, and Compute\n\nMany core design choices in machine learning are discrete: the number\nof layers\, parameters\, or fine-tuned models\; even datasets themselves\nare finite sets. In this tal k\, I explore what becomes possible when we\ntake continuum limits of thes e objects\, allowing us to interpolate\nbetween architectures\, model beha viors\, and data domains. Treating\ndepth as continuous enables variable t est-time compute and richer\nrepresentational power. Viewing datasets as p oints on a smooth path\nmakes it possible to generate intermediate domains and study how\nmodels generalize across them. And lifting model families into\ncontinuous parameter spaces allows controllable behavioral blending\ ,\nas well as zero-shot interpolation of model size without training or\ns toring multiple discrete variants. \n\nAcross these examples\, a common l esson emerges: turning discrete\nquantities into continuous spaces unlocks new degrees of freedom —\nfor representation\, adaptation\, efficiency\ , and ultimately for\ndesigning more flexible model and data families. The se continuous\nperspectives also offer a practical lens for data-centric d ecision\nmaking: valuing and refining data under resource constraints\, an d\nexposing redundancy\, bias\, and compressibility\, ultimately enabling\ nreliable model transfer and deployment.\n\n—\n\nDavid Alvarez-Melis obt ained a PhD in computer science from MIT\, where\nhe worked at CSAIL on va rious topics in machine learning and natural\nlanguage processing. Alvarez -Melis also holds BSc (Licenciatura) and\nMS degrees in mathematics from I TAM and Courant Institute (NYU)\,\nrespectively. During the latter\, Alvar ez-Melis worked on semidefinite\nprogramming for domain adaptation for his thesis. Between Master’s\nand PhD\, Alvarez-Melis spent a year at IBM ’s T.J. Watson Research\nCenter\, working in the Speech Recognigtion and NLP teams.\n\nREGISTER →  to attend in-person or on Zoom\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91363bd DTSTART;TZID=America/New_York:20260217T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260217T170000 LOCATION:Remote Access - Zoom SUMMARY:Center for Informed Democracy & Social - Cybersecurity Seminar -\nC hinmayi Sharma CLASS:PUBLIC DESCRIPTION:Speaker: CHINMAYI SHARMA - To Be Rescheduled\, Associate Profes sor\nFordham Law School\nFordham University\n\nTalk Title: Automating Away Bureaucratic Dissent\n\nA President who can direct the executive branch i s not necessarily a\nPresident who can make it obey. The administrative st ate has long\nresisted perfect compliance through ordinary human judgment: career\nofficials interpret statutes\, demand reasons\, surface legal ris k\, and\nslow plans that outrun law or evidence. That internal contestatio n is\nnow under coordinated pressure. Unitary-executive doctrine supplies a\nconstitutional vocabulary of supervision and removal. Political\npracti ce hollows agencies through defunding\, forced attrition\,\npoliticization \, loyalty screening\, mission destabilization\, and\nefforts to reclassif y or purge career roles that influence policy.\nAutomation supplies the fi nal step. By relocating decision making\nupstream and executing it at spee d and scale\, automation can\noperationalize a perfectly obedient bureaucr acy—implementation that\ntracks presidential preferences while bypassing the human layers that\nonce turned disagreement into friction.\n\nThis Ar ticle argues that perfect obedience is a rule-of-law problem\nbefore it is a partisan one. Rule-of-law government is not exhausted\nby legality. A l egal system that claims authority must\, at a minimum\,\nrespect individua l autonomy\, constrain arbitrary sovereign power\, and\nsustain legitimacy in the eyes of the governed. Bureaucratic dissent\nhelps do that work. It preserves autonomy within government by\nrequiring officials to reason\, deliberate\, confront competing views\,\nand offer justification rather th an simply transmit will. It preserves\nthe public’s autonomy by making a dministration legible: dissent\nsurfaces reasons\, exposes tradeoffs\, and creates the informational and\nprocedural conditions for meaningful parti cipation and critique. It\nconstrains the sovereign by forcing executive a ction to meet\nresistance inside the branch most capable of evaluating pol icy against\nlaw\, expertise\, and institutional memory—before harm hard ens into\npractice. It sustains legitimacy by keeping law responsive\, not merely\nenforceable.\n\nAutomation threatens these rule-of-law functions precisely because it\nconverts governance into specification. When decisio n paths ossify\ninto code\, procurement\, and vendor-maintained pipelines\ , officials\nbecome monitors of preauthorized “happy paths\,” not agen ts\nresponsible for reasoned application. Policy can change through a\nthr eshold tweak or model update without the public-facing signals that\nordin arily make overreach legible\; correction arrives late\, after mass\nimple mentation. The risk is compounded by automation’s unvirtuous\ncycle—ve ndor dependence\, deskilling\, normalization\, metrics\,\nlegitimacy laund ering\, and ossification—which raises switching costs\nand narrows the s pace for principled pushback. Worse\, it normalizes\nautomation\, desensit izes the public to governance without humans\, and\,\nby depriving bureauc rats and the public the opportunities for reasoned\ndeliberation\, erodes the very skills required for critical evaluation\nof sovereign power. Unde r backsliding conditions\, perfect obedience\nbecomes a governance strateg y: democratic constraint can be eroded in\npractice under the veneer of te chnical objectivity and formal\nlegality.\n\n—\n\nChinmayi Sharma is an Associate Professor at Fordham Law School. Her\nresearch and teaching focu s on open internet governance\,\ncybersecurity\, artificial intelligence\, and computer crime.\n\nShe is an advisor to the American Law Institute’ s Principles of Law\,\nCivil Liability for Artificial Intelligence and a m ember of the\nMicrosoft Responsible AI Committee. She is a Distinguished F ellow at\nthe Georgetown Center on Privacy and Technology\, as well as a\n Non-Resident Fellow at the Strauss Center\, the Center for Democracy\nand Technology\, the Atlantic Council\, and the Institute for Law & AI.\n\nShe is on the Lawfare masthead and has been quoted by the New York\nTimes\, N PR\, ProPublica\, Law360\, Bloomberg\, and Bloomberg Law. Her\nArticle cal ling for professionalization of AI engineers\, “AI’s\nHippocratic Oath \,” was featured in the New York Review and her\npapers have been recomm ended on the Legal Theory Blog. Her Article on\nopen source software secur ity\, “Tragedy of the Digital Commons\,”\nhas been included in the Hag ue's International Cyber Security\nBibliography. Before joining academia\, Chinmayi worked at Harris\,\nWiltshire & Grannis LLP\, a telecommunicatio ns law firm in Washington\,\nD.C.\, clerked for Chief Judge Michael F. Urb anski of the Western\nDistrict of Virginia\, and co-founded a software dev elopment company.\n\nREGISTER → confirmation email will be provided fo r joining the\nseminar.\n\n Zoom Participation \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9136a7e DTSTART;TZID=America/New_York:20260218T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260218T170000 LOCATION:Steinberg Auditorium\, Baker Hall A53 SUMMARY:Statistics and Data Science Seminar - Subhankar Bhadra CLASS:PUBLIC DESCRIPTION:Speaker: SUBHANKAR BHADRA\, Postdoctoral Researcher\, The Penns ylvania\nState University\n\nTalk Title: Causal learning in connected popu lations: insight\,\ninference\, and generalizability\n\nIn the interconnec ted and interdependent world of the twenty-first\ncentury\, causal inferen ce is complicated by interference\, because the\ntreatment assignments of units can affect the outcomes of other units\nvia treatment and outcome sp illover. Since outcome spillover induces\ndependence among outcomes\, clos ed-form expressions for causal effects\nand convergence rates for causal e ffect estimators are challenging and\nunavailable. \n\nIn this talk\, I w ill provide insight into causal mechanisms under\ninterference\, by presen ting closed-form expressions for causal effects\nin the presence of treatm ent and outcome spillover\, which help\ndisentangle the contributions of t reatment\, treatment spillover\, and\noutcome spillover into the causal ef fects. The main results do not\nmake assumptions about the joint probabili ty law of treatment\nassignments\, outcomes\, and connections beyond linea rity of conditional\nexpectations of outcomes and the standard assumptions of ignorability\nand positivity\, thus allowing complex dependence among outcomes and\nconnections. \n\nBuilding on the closed-form expressions\, I will present causal effect\nestimators along with rates of convergence\, obtained by controlling\ndependence and characterizing a high-probability subset of data that\naddresses collinearity issues. Finally\, I will disc uss\ngeneralizability of causal inference under interference\, which\nconc erns how causal conclusions based on dependent observations of\noutcomes c an be generalized from a sample to the population. \n\n—\n\nSubhankar B hadra is a postdoctoral researcher at The Pennsylvania\nState University. His research focuses on statistical inference in\nconnected populations\, with particular emphasis on network science and\ncausal inference under in terference. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9136eb5 DTSTART;TZID=America/New_York:20260319T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260319T183000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Distinguished Lecture: Hank Suz-Chi Wan Memorial Lecture - Blak e\nScholl CLASS:PUBLIC DESCRIPTION:Speaker: BLAKE SCHOLL\, Founder and Chief Executive Officer\, B oom\nSupersonic\n\nTalk Title: Breaking the Sound Barrier with Computer Sc ience: How a\ncomputer science alum built the first civil supersonic jet s ince\nConcorde and solved sonic boom with software\n\nBlake Scholl (SCS'01 ) always loved airplanes and hated that they\nweren't getting any faster. After a stint as a software engineer at\nearly Amazon and a few software s tartups\, Blake founded Boom in 2014\nwith the goal of bringing back super sonic passenger travel and making\nit mainstream. Last year\, Boom made hi story by flying its XB-1\naircraft faster than the speed of sound without an audible sonic\nboom—paving the way for a mainstream renaissance in supersonic\ntravel.\n\nIn this talk\, Blake will talk about his experience going from software\nto hardware—and how a software mindset unlocks a r enaissance in\nsupersonic passenger flight.\n\n—\n\nBlake Scholl founded Boom Supersonic in 2014 with the goal of making\nhigh-speed travel mainst ream and enabling a new world of human\nconnection. Blake is passionate ab out tackling big problems which the\nworld has overlooked. After seeing Co ncorde in a museum\, he became\ncaptivated by the question: why had supers onic travel\ndisappeared—and how could it return?  Driven by curiosity and\nconviction\, Blake dove into aerospace engineering from first\nprinci ples\, ultimately assembling a team that would bring the vision\nfor Boom to life. As Boom’s founder and CEO\, Blake focuses on\nuniting top talen t across disciplines to solve one of the most complex\nengineering challen ges of our time. He is passionate about tackling\noverlooked problems with global impact—and building the teams\ncapable of solving them.\n\nPrior to founding Boom\, Blake held leadership roles at Amazon and\nGroupon. He started his career at Amazon as a software engineer\, where\nhe developed software for personalized and automated merchandising\,\nand by age 24\, was responsible for initiatives that generated $300\nmillion in revenue. B lake pioneered the automation of digital ad\nbuying\, enabling advertisers to build their own “long tail” by\npromoting niche products to releva nt customers. Later\, Blake was the\nfirst employee and director of produc t development at Kleiner-backed\nmobile startup Pelago. In 2010\, he co-fo unded mobile technology\nstartup Kima Labs\, which was acquired by Groupon in 2012. Blake held l\nmultiple leadership positions under Groupon’s ow nership before\nleaving to start Boom.\n\nBlake’s love of aviation start ed early—he grew up in suburban\nCincinnati\, where his parents took him to the airport to watch Cessnas\ntake off and land. He began flying in co llege and holds a private\npilot’s license with an instrument rating. He earned a BS in\nComputer Science from Carnegie Mellon University\, lives in Denver and\ncontinues to enjoy flying in his free time with his four ch ildren.  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913738b DTSTART;TZID=America/New_York:20260216T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260216T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Marek Galovic CLASS:PUBLIC DESCRIPTION:Speaker: MAREK GALOVIC\, Chief Executive Officer and Co-founder \, TopK\n\nTalk Title: TopK: Billion-Scale Hybrid Retrieval from the Groun d Up\n\nTopK is a search engine built from the ground up for unstructured\ nretrieval. It combines dense/sparse/multi-vector search\, lexical\nsearch \, powerful filtering\, and customizable scoring capabilities in a\nsingle \, cloud-native system that scales to billions of documents with\nhigh ing est throughput and O(10ms) p99 query latencies. In this talk\,\nI'll focus on how TopK is designed on a high-level\, including our\ndisaggregated re ad-write path and distributed compaction\, and then\ndive deep into our co lumnar file format (.bob) and query engine\n(reactor)\, which we built fro m the ground up to support search at\nscale.\n\n—\n\nMarek Galovic is th e CEO and co-founder of TopK - an AI-native search\nengine. Before foundin g TopK\, Marek led data/control plane engineering\nteams at Pinecone and w orked on fraud detection and financial\nforecasting at Shopify. He holds a degree in computer science and\nartificial intelligence from CTU Prague\, where he researched game\ntheory and adversarial machine learning algorit hms applied to computer\nsecurity (published at NeurIPS).\n\nThis talk is part of the PostgreSQL vs. The World Seminar Series. \n\nZoom Participati on.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913772e DTSTART;TZID=America/New_York:20260220T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260220T143000 LOCATION:Remote Access - Zoom SUMMARY:STAMPS Seminar - Maximilian Dax CLASS:PUBLIC DESCRIPTION:Speaker: MAXIMILIAN DAX\, Principal Investigator\, Ellis Instit ute\, and\nMax Planck Institute for Intelligent Systems\, Tübingen\n\nTal k Title: Real-Time Gravitational-Wave Inference with Probabilistic\nMachin e Learning\n\nGravitational-wave (GW) astronomy promises groundbreaking di scoveries\nin the coming decades\, but its progress is bottlenecked by the \ncomputational challenges of large-scale and real-time data analysis. I\n will present DINGO\, a machine learning approach for fast and accurate\nGW inference that addresses these challenges. DINGO trains generative\nneura l networks to directly estimate probability distributions over GW\nsource parameters. I will explain the core ideas behind DINGO and\nhighlight seve ral machine learning techniques that we developed to\nadapt modern simulat ion-based inference to the challenging field of GW\ndata analysis.\n\n—\ n\nMaximilian Dax is a Principal Investigator at the ELLIS Institute\nTüb ingen and the Max Planck Institute for Intelligent Systems\, where\nhe lea ds the research group for Science and Probabilistic\nIntelligence. His res earch focuses on probabilistic machine learning\nand its application in sc ience. Before joining ELLIS\, he completed his\nPhD at the Max Planck Inst itute for Intelligent Systems and the\nUniversity of Tübingen\, and spent time at ETH Zurich and Google\nResearch.\n\nIn Person and Zoom Participat ion.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9137aea DTSTART;TZID=America/New_York:20260217T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260217T133000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral - Jiaming (Andy) Zou CLASS:PUBLIC DESCRIPTION:Speaker: JIAMING (ANDY) ZOU\, Ph.D. CandidateComputer Science\n DepartmentCarnegie Mellon University\n\nTalk Title: Improving Safety and S ecurity of Generative Models\n\nRecent advances in large language and mult imodal models have enabled\npowerful new applications\, but they also rais e critical challenges in\nsafety\, robustness\, and alignment. This thesis studies these\nchallenges through three complementary research directions . First\, we\nshow that current alignment methods remain brittle by develo ping\nadversarial attacks that reliably bypass safeguards across text\,\nm ultimodal\, and embodied systems\, demonstrating that alignment alone\ndoe s not guarantee robustness. Second\, we introduce evaluation\nframeworks a nd benchmarks that systematically measure safety failures\nin modern AI sy stems\, revealing widespread vulnerabilities in deployed\nmodels and agent s. Third\, we propose methods to improve alignment and\ncontrol\, includin g representation-level interventions\, circuit\nbreakers\, and safety pret raining\, which significantly reduce attack\nsuccess while preserving mode l capability. Together\, these\ncontributions advance our understanding of AI safety risks and provide\npractical tools for building safer and more trustworthy AI systems.\n\nThesis Committee\n\nZico Kolter (Co-chair)\n\nM att Fredrikson (Co-chair)\n\nGraham Neubig\n\nNicholas Carlini (Anthropic) \n\nIn Person and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9137eca DTSTART;TZID=America/New_York:20260324T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260324T173000 LOCATION:The Landing Hotel\, 757 Casino Drive\, Pittsburgh\, PA 15212 SUMMARY:ASPLOS 2026 CLASS:PUBLIC DESCRIPTION:Tuesday\, March 24\, 8am – Thursday\, March 26\, 2026\, 5:30 pm \n\nASPLOS\, the ACM International Conference on Architectural Support for\nProgramming Languages and Operating Systems\, is the premier academi c\nforum for multidisciplinary computer systems research spanning\nhardwar e\, software\, and their interaction. It focuses on computer\narchitecture \, programming languages\, operating systems\, and associated\nareas such as networking and storage.  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91381bb DTSTART;TZID=America/New_York:20260213T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260213T140000 LOCATION:Gates Hillman 7101 SUMMARY:Theory Seminar - Ted Pyne CLASS:PUBLIC DESCRIPTION:Speaker: TED PYNE\, Ph.D. StudentElectrical Engineering and Com puter\nScience DepartmentMassachusetts Institute of Technology\n\nTalk Tit le: When Connectivity is Hard\, Random Walks are Easy (with\nNondeterminis m)\n\nTwo well-studied problems on graphs are to 1: determine s -> t\nconn ectivity\, and 2: estimate the behavior of random walks. Currently\,\nther e is no algorithm for (1) that runs in polynomial time and\nstrongly subli near space\, and no algorithm for (2) that runs in\nnondeterministic logsp ace. We show that for every graph\, at least one\nof these problems is sol vable more efficiently than the state of the\nart. Our results build on re cent work on distinguish-to-predict\ntransformations (Li\, Pyne\, Tell) an d bootstrapping systems (Chen\,\nTell). As a consequence\, either randomiz ed linear space can be\nderandomized\, or a time- and space- efficient sim ulation of\nnondeterministic linear space holds.\n\nJoint work with Dean D oron\, Roei Tell\, and Ryan Williams (STOC\n2025). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9138503 DTSTART;TZID=America/New_York:20260216T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260216T130000 LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar - William Enck CLASS:PUBLIC DESCRIPTION:Speaker: WILLIAM ENCK\, Goodnight Distinguished Professor in Se curity\nSciences\, Co-director\, Secure Computing Institute\, Department o f\nComputer Science North Carolina State University\n\nTalk Title: Securin g the Software Supply Chain\n\nOpen source software is an economic driving force behind nearly all\nsoftware products. However\, the past half decad e has seen a surge of\nattacks targeting critical but often overlooked par ts of this software\nsupply chain. The response from industry and governme nt has been a\nfrenzy of frameworks\, regulations\, tools\, and best pract ices. But\nwhere does academia fit in?\n\nThe term \"software supply chain \" does not capture the nuance of the\nspace. For the vast majority of ope n source software\, there is no\nformal \"supplier\" and the \"chain\" is a complex interconnected graph.\nExisting tools that help developers manag e known vulnerabilities in\ntheir dependencies rely on imperfect and impre cise data. Build\nprocesses were created based on threat models that no lo nger hold.\nDevelopment environments make it too easy for developers to ma ke\nchoices that open both their projects and their own workstations to\na ttack. The problems that underlie these challenges are not foreign to\ncom puter security research\, but they require a partnership with\npractitione rs to solve. This talk will describe what we have learned\nthrough our res earch and our extensive interactions with >150\npractitioners from >50 com panies as part of the NSF-funded Secure\nSoftware Supply Chain Center (S3C 2).\n\n—\n\nWilliam Enck is the Goodnight Distinguished Professor in Sec urity\nSciences in the Department of Computer Science at the North Carolin a\nState University where he is co-director of the Secure Computing\nInsti tute (SCI) and member of the Secure Software Supply Chain Center\n(S3C2). His research spans the broad area of systems security\,\napplying a range of systems design\, program analysis\, and empirical\nstudies. He is Vice President of the USENIX Board of Directors and has\npreviously co-chaired the program committees of flagship security\nconferences including USENIX Security and IEEE S&P.\n\nFaculty Hosts: Lorrie Cranon\, Elijah Bouma-Sims \n\nIn Person and Zoom Participation.  See announcement.\n\n→ The CyLa b Seminar is open only to partners and CMU faculty\,\nstudents\, and staff . \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913899d DTSTART;TZID=America/New_York:20260211T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260211T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Louie Putterman CLASS:PUBLIC DESCRIPTION:Speaker: AARON (LOUIE) PUTTERMAN\, Ph.D. Student\, Computer Sci ence\,\nJohn A. Poulson School of Engineering and Applied Sciences\, Harva rd\nUniversity\n\nTalk Title: Breaking the \\sqrt{n} Barrier: New Parallel Algorithms for\nFinding a Matroid Basis\n\nOver 40 years ago\, Karp\, Upf al\, and Wigderson initiated the study of a\nfundamental question in paral lel computation: how many adaptive rounds\nare required to find a basis of a matroid using only polynomially many\nindependence queries (that is\, q ueries that test whether a set is\nindependent)? Their pioneering work est ablished an upper bound of\nO(√n) rounds and a lower bound of roughly n1 /3 rounds\; these bounds\nhave remained unchanged since.\n\nIn this talk\ , I will present recent progress on this question. We give\na new parallel algorithm that\, with high probability\, finds a matroid\nbasis in Õ(n3/ 7) rounds\, improving upon the classical O(√n) bound.\nFor the important special case of partition matroids\, we obtain an\noptimal Õ(n1/3) round algorithm\, essentially settling the round\ncomplexity in this setting. O ur approach introduces a new matroid\ndecomposition technique that may be of independent interest and also\nyields faster parallel algorithms for th e classic matroid intersection\nproblem.\n\nThis is joint work with Sanjee v Khanna and Junkai Song (NYU). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9138d69 DTSTART;TZID=America/New_York:20260212T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260212T160000 LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Han Huang CLASS:PUBLIC DESCRIPTION:Speaker: HAN HUANG\, Assistant Professor\, Department of Mathem atics\,\nUniversity of Missouri\n\nTalk Title: Can You Recover a Manifold from a Single Random Geometric\nGraph?\n\nConsider a manifold M that is ei ther embedded in Euclidean space or a\nRiemannian manifold. We sample poin ts X1\,…\,Xn from an unknown\nprobability measure μ on M. We observe on ly a single random graph G\non {1\,…\,n}\, where edges {i\,j} appear ind ependently with probability\np(|Xi-Xj|) for a known\, monotone decreasing connection function p.\n\nThis setting asks a basic inverse question: how much of the underlying\ngeometry and sampling measure can be recovered fro m connectivity\nalone?\n\nIn this talk I will describe the reconstruction results showing that\,\nunder natural regularity conditions\, the combinat orial structure of G\nencodes substantial geometric information.\n\nJoint work with Pakawut Jiradilok and Elchanan Mossel.\n\n4:00 pm → Jane Stre et-sponsored tea and cookies in the Math Lounge\n(bring your mug!) \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91390f3 DTSTART;TZID=America/New_York:20260209T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260209T130000 LOCATION:Group Viewing Newell-SImon 3305 and Zoom SUMMARY:Accessibility Lunch Seminar - Sohyeon Park CLASS:PUBLIC DESCRIPTION:Speaker: SOHYEON PARK\, Ph.D. Candidate\, Department of Informa tics\,\nUniversity of California\, Irvine\n\nTalk Title: Bias Beneath the Surface in Large Language Models\n\nThe rapid integration and deployment o f Large Language Models (LLMs)\nfar outpace research aimed at understandin g their impacts on human\npopulations. For autistic people specifically\, these systems are\nincreasingly used to support socially and cognitively d emanding tasks\,\nincluding drafting workplace emails\, preparing for inte rpersonal\ninteractions\, organizing daily activities\, and supporting emo tional\nwell-being. This gap highlights the need for human-centered method s\nthat examine not only how LLMs operate\, but also the underlying\npersp ectives they reflect about autistic people. In this talk\, I\npresent two studies that use distinct methodological approaches to\nexamine these unde rlying perspectives. Together\, these studies provide\npractical tools and conceptual clarity for more inclusive and\nresponsible LLM design for aut istic users\, with implications for other\npopulations as well.\n\nIn Pers on Group Viewing and Zoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913949b DTSTART;TZID=America/New_York:20260219T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260219T130000 LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI-SDM - Senior/Junior Student Panel CLASS:PUBLIC DESCRIPTION:Talk Title: What Worked\, What Didn’t\, and How to Survive Ac ademia\n\nSenior students — final year doctorate students\, postdocs — will\nanswer questions and give advice to junior students on what worked\ ,\nwhat didn’t\, and how to survive academia. \n\nOpen to all students. \n\nREGISTER \n\n→ Register if attending in person or on Zoom to assist with\nlogistics.\n\n→ Lunch is provided for in-person attendees\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91397bf DTSTART;TZID=America/New_York:20260213T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260213T170000 LOCATION:Gates Hillman 9115 and Zoom SUMMARY:Doctoral Thesis Proposal - Yash Savani CLASS:PUBLIC DESCRIPTION:Speaker: YASH SAVANI\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Controlled Generation of Fou ndation Models for Training\n\nThis thesis proposal presents controlled ge neration methods for\nfoundation models\, focusing on gradient-based steer ing to improve\ntraining and robustness. For diffusion and flow models\, I introduce\nDiffusing Differentiable Representations (NeurIPS 2024)\, whic h guides\nthe training of differentiable representations\, such as Neural\ nRadiance Fields\, by pulling back the score function through the\ndiffere ntiable render function. I also present work with Adobe\nResearch on tempo ral credit assignment for policy gradient methods\,\nenabling more effecti ve training of flow models via GRPO-style\nreinforcement learning. \n\nFo r large language models\, I present two methods for controlled\ngeneration . The first maximizes resource utilization in GRPO-style\nreinforcement le arning by selectively dropping low-variance\ntrajectories (in submission). The second\, Antidistillation Sampling\n(NeurIPS 2025)\, steers generatio n to defend against distillation\nattacks using precomputed proxy gradient s. Together\, these\ncontributions establish a unified framework for contr olled generation\nacross modalities\, with applications spanning creative content\nsynthesis\, model protection\, and efficient training.\n\nThesis Committee\n\nJ. Zico Kolter (Chair)\n\nAviral Kumar\n\nNicholas M. Boffi\n \nKrishna Kumar Singh (Adobe Research)\n\nAdditional Information\n\nIn Per son and Zoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9139bc9 DTSTART;TZID=America/New_York:20260211T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260211T163000 LOCATION:Steinberg Auditorium\, Baker Hall A53 SUMMARY:Statistics and Data Science Seminar - Josh Gilbert CLASS:PUBLIC DESCRIPTION:Speaker: JOSH GILBERT\, Ph.D. Candidate in Education Policy and Program\nEvaluation\, Graduate School of Education\, Harvard University\n \nTalk Title: Estimating heterogeneous treatment effects with item-level\n outcome data: Insights from Item Response Theory\n\nAnalyses of heterogene ous treatment effects (HTE) are common in\napplied causal inference resear ch. However\, when outcomes are latent\nvariables assessed via psychometri c instruments such as educational\ntests\, standard methods ignore the pot ential HTE that may exist among\nthe individual items of the outcome measu re. Failing to account for\n“item-level” HTE (IL-HTE) can lead to both underestimated standard\nerrors and identification challenges in the esti mation of\ntreatment-by-covariate interaction effects. We demonstrate how Item\nResponse Theory (IRT) models that estimate a treatment effect for ea ch\nassessment item can both address these challenges and provide new\nins ights into HTE generally. This study articulates the theoretical\nrational e for the IL-HTE model and demonstrates its practical value\nusing 75 data sets from 48 randomized controlled trials containing 5.8\nmillion item res ponses in economics\, education\, and health research.\nOur results show t hat the IL-HTE model reveals item-level variation\nmasked by single-number scores\, provides more meaningful standard\nerrors in many settings\, all ows for estimates of the generalizability\nof causal effects to untested i tems\, resolves identification problems\nin the estimation of interaction effects\, and provides estimates of\nstandardized treatment effect sizes c orrected for attenuation due to\nmeasurement error.\n\n—\n\nJoshua Gilbe rt is a PhD candidate in Education Policy and Program\nEvaluation at the H arvard Graduate School of Education\, where he works\nwith James Kim and L uke Miratrix. His research interests include the\nintersection of causal i nference and psychometric methods. He has over\ntwenty peer-reviewed publi cations in journals such as Developmental\nPsychology\, Journal of Educati onal Psychology\, Journal of Educational\nand Behavioral Statistics\, Beha vior Research Methods\, Psychological\nMethods\, and others. He has also c urated a collection of over 100\nitem-level datasets from randomized contr ol trials available via the\nItem Response Warehouse\, and hosted workshop s on how to conceptualize\nand analyze item-level effects. In 2025\, he wa s awarded a prestigious\n2025 Spencer / NAEd dissertation fellowship. \n\ n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913a03f DTSTART;TZID=America/New_York:20260209T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260209T163000 LOCATION:Steinberg Auditorium\, Baker Hall A53 SUMMARY:Statistics and Data Science Seminar - David Bruns-Smith CLASS:PUBLIC DESCRIPTION:Speaker: DAVID BRUNS-SMITH\, Postdoctoral Fellow\, Stanford Dat a\nScience\, Graduate School of Business\, Stanford University\n\nTalk Tit le: Two-Stage Machine Learning for Nonparametric Instrumental\nVariable Re gression\n\nThe growing access to large administrative datasets with rich\ ncovariates presents an opportunity to revisit classic two-stage least\nsq uares (2SLS) applications with machine learning (ML). We develop\nTwo-Stag e Machine Learning\, a simple and efficient estimator for\nnonparametric i nstrumental variables (NPIV) regression. Our method\nuses ML models to fle xibly estimate nonparametric treatment effects\nwhile avoiding the computa tional complexity and statistical\ninstability of existing machine learnin g NPIV approaches. Our\nprocedure has two steps: first\, we predict the ou tcomes given\ninstruments and covariates (the reduced form) and extract a basis from\nthis predictor\; second\, we predict the outcomes using the tr eatment\nand covariates\, but where the predictions are projected onto the \nlearned basis of instruments. We prove that under a testable\ncondition\ , our estimation error depends entirely on the reduced-form\nprediction ta sk\, where ML methods excel. We also develop a bias\ncorrection procedure that provides valid confidence intervals for\nscalar summaries like averag e derivatives. In an empirical application\nto California supermarket data featuring bunching at 99-ending price\npoints\, we find our machine learn ing approach is crucial for modeling\ndiscontinuities in demand at the dol lar boundary: we reduce NPIV\nestimation error nearly eight-fold compared to previous estimators and\nestimate a price elasticity that is 2.5-6 time s larger than prior\nestimates.\n\n—\n\nDavid Bruns-Smith is a Postdocto ral Fellow at Stanford Data Science\nworking with Guido Imbens. His resear ch focuses on machine learning\nmethods for causal inference with an empha sis on applications in\nmacroeconomics. David's recent methodological rese arch studies\ndebiased machine learning\, including for instrumental varia bles\nregression and reinforcement learning. He then applies these causal\ nmachine learning methods to a variety of economics questions including\nf iscal stimulus\, monetary policy\, and the evolution of income\ninequality .\n\nPreviously\, David completed his PhD in Computer Science at UC\nBerke ley\, advised by Avi Feller and Emi Nakamura. In addition to\nComputer Sci ence\, he completed core PhD courses in Economics as a\nBerkeley Opportuni ty Lab Labor Science Fellow. From 2023-2024\, he\nworked with Alex D'Amour at Google Deepmind. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913a556 DTSTART;TZID=America/New_York:20260210T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260210T130000 LOCATION:Gates Hillman 9115 SUMMARY:Doctoral Speaking Skills Talk - Hyoungjoo Kim CLASS:PUBLIC DESCRIPTION:Speaker: HYOUNGJOO KIM\, Ph.D. Student\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: No Cap\, This Memory Slaps : Breaking Through the Memory\nWall of Transactional Database Systems with Processing-in-Memory\n\nMemory channel bandwidth imposes an upper bound o n the performance of\nonline transaction processing (OLTP) on in-memory da tabase management\nsystems (DBMS). Emerging processing-in-memory (PIM) har dware has the\npotential to overcome this barrier by using small cores in DRAM chips\nthat can read and process data in situ\, thereby avoiding movi ng these\ndata across memory channels. However\, naively offloading all da tabase\ncomponents to PIM does not solve the problem due to the\ncharacter istics of software components and the limitations of PIM\nhardware.\n\nOLT Pim is the first end-to-end OLTP DBMS designed for PIM systems. We\nbuild a formalized model for the affinity of each database operation\ntowards PI M and use it to decide the partitioning of components on\ndifferent types of memory. We also design a lightweight batching\nalgorithm to overcome th e large PIM control latency while minimizing\nthe batching overhead. We im plement and evaluate OLTPim on the latest\nPIM system from UPMEM with 64 w orker threads and 2048 PIM modules. Our\nresults show that OLTPim achieves up to 1.71x throughput and up to\n6.14x less per-transaction memory chann el traffic over MosaicDB\, a\nstate-of-the-art in-memory system.\n\nPresen ted in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913a933 DTSTART;TZID=America/New_York:20260209T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260209T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Marc Brooker CLASS:PUBLIC DESCRIPTION:Speaker: MARC BROOKER\, Vice President and Distinguished Engine er\nAmazon Web Services (AWS)\n\nTalk Title: Aurora DSQL\n\nMarc Brooker i s a VP and Distinguished Engineer at AWS. During his 16\nyears at AWS\, Ma rc has worked on EC2\, EBS\, Lambda\, and most recently\nlead the team tha t launched Aurora DSQL. He is currently focused on\ninfrastructure for age ntic AI\, and the availability and security of\nour large-scale systems. B efore AWS\, Marc completed his PhD at the\nUniversity of Cape Town.\n\nThi s talk is part of the PostgreSQL vs. The World Seminar Series\n\nZoom Part icipation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913ac5f DTSTART;TZID=America/New_York:20260205T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260205T160000 LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Bernardo Subercaseaux CLASS:PUBLIC DESCRIPTION:Speaker: BERNARDO SUBERCSEAUX\, Ph.D. Student\, Computer Scienc e\nDepartment\, Carnegie Mellon University\n\nTalk Title: Breaking down gr aphs and hypergraphs into structured\npieces\, optimally and efficiently\n \nWe will consider the problem of writing an arbitrary graph as an\nedge-d isjoint union of complete bipartite graphs\, and its natural\ngeneralizati on to hypergraphs. At a very high level\, the question at\nhand is to what extent we can summarize complicated structures by\ndecomposing them into very structured pieces. I will present several\noptimal asymptotic bounds and algorithms\, with applications to graph\ncompression\, SAT solving\, c ryptographic secret sharing\, and\napproximations for the densest subgraph problem. More concretely\, our\nmain result is that every n-vertex d-unif orm hypergraph H can be\nwritten as the union of a family F of complete d- partite hypergraphs\nsuch that every vertex of H belongs to at most (n cho ose d)/(n lg n)\ngraphs in F. This improves on results of Csirmaz\, Ligeti \, and Tardos\n(2014)\, gives the best upper bound for some secret sharing questions\,\nand answers several 40-year-old questions of Chung\, Erdös\ , and\nSpencer (1983). The heart of our proof is a simple idea from word\n combinatorics\, which allows us to balance the number of d-cliques each\nv ertex belongs to.\n\nThis talk is based on joint work with Andrew Krapivin \, Benjamin\nPrzybocki\, and Nicolás Sanhueza-Matamala.\n\n4:00 pm →  Jane Street-sponsored tea and cookies  in the Math\nLounge (bring your m ug).\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913b02e DTSTART;TZID=America/New_York:20260209T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260209T130000 LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar - Jonathan Aldrich CLASS:PUBLIC DESCRIPTION:Speaker: JONATHAN ALDRICH\, Professor of Computer Science and S oftware\nEngineering\, Software and Societal Systems Department\, Carnegie Mellon\nUniversity\n\nTalk Title: (Sub)-structural Information Flow: A Si mple\, Scalable\, and\nGeneral Foundation for Security Reasoning\n\nInform ation flow is a foundational property underlying a wide variety\nof securi ty issues. Type systems are a promising approach to reasoning\nabout infor mation flow\, but the complexity of previously proposed\napproaches has li mited adoption. We propose a new foundation for\ninformation flow types: ( sub-)structural information flow. Rather than\nplacing constraints on poly morphic labels\, we build information flow\ntypes as a structural set latt ice. The resulting system is not only\nsignificantly simpler and more scal able\, it also for the first time\npositions declassification as compatibl e with noninterference\, rather\nthan in opposition to it. Our approach\, presented in recent papers at\nOOPSLA'25 and POPL'26\, can be extended to substructural labels in\norder to support capabilities\, quantitative info rmation leakage\,\nsandboxing\, authorization protocols\, and more. In add ition to security\napplications\, we are studying a novel theory of type s ystem usability\nthat emerges from our approach as well as related work in Rust\nlifetime tracking. This is joint work with Hemant Gouni and Frank\n Pfenning.\n\n—\n\nJonathan Aldrich is a Professor of Computer Science an d Software\nEngineering at Carnegie Mellon University\, where he directs t he Master\nof Software Engineering program. He is the coauthor (with Micha el\nScott) of the textbook Programming Language Pragmatics\, 5th edition.\ nHis research combines programming languages\, software engineering\, and\ nhuman-computer interaction to explore how the way we express software\naf fects our ability to engineer software at scale. A particular theme\nof mu ch of his work is improving software quality and programmer\nproductivity through better ways to express structural and behavioral\naspects of softw are design within source code. Aldrich has contributed\nto ownership\, typ estate checking\, modular and gradual verification\ntechniques\, and usabi lity in programming language and type system\ndesign. For his work specify ing and verifying architecture\, he\nreceived a 2006 NSF CAREER award\, th e 2007 Dahl-Nygaard Junior Prize\,\nand an ICSE test of time award. Outsid e the university\, he serves on\nthe ACM Publications Board and is the CTO of Noteful\, a startup\ndelivering an educational app for music theory an d note reading.\n\nIn Person and Zoom Participation.  See announcement.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913b4c8 DTSTART;TZID=America/New_York:20260202T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260202T173000 LOCATION:Remote Acccess - Zoom SUMMARY:Database Seminar - Tyler Akidau and Adam Symanski CLASS:PUBLIC DESCRIPTION:Speaker: TYLER AKIDAU and ADAM SYMANSKI\, Tyler: CTO\, Redpanda Data\nAdam: Creator of Oxla\n\nTalk Title: Redpanda Oxla or: Why Your Has hmaps are Secretly Wrecking\nYour Performance\n\nIn this talk\, we'll firs t give an overview of the Oxla analytical\ndatabase and how it fits into w hat we do at Redpanda. Then we'll dive\ninto one of the more interesting a spects of Oxla internals: combating\nmemory bandwidth performance bottlene cks in GROUP BY and JOIN via a\nspecialized\, custom hashmap implementatio n.\n\n—\n\nTyler is CTO of Redpanda Data\n\nAdam is the creator of Oxla\ n\nThis talk is part of the PostgreSQL vs. The World Seminar Series. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913b82d DTSTART;TZID=America/New_York:20260205T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260205T110000 LOCATION:Gates Hillman 8102 SUMMARY:Doctoral Speaking Skills Talk - Gabriele Oliaro CLASS:PUBLIC DESCRIPTION:Speaker: GABRIELE OLIARO\, Ph.D. Student\nComputer Science Depa rtment\nCarnegie Mellon University\n\nTalk Title: SuffixDecoding: Extreme Speculative Decoding for Emerging\nAI Applications\n\nSpeculative decoding is widely adopted to reduce latency in large\nlanguage model (LLM) infere nce by leveraging smaller draft models\ncapable of handling diverse user t asks. However\, emerging AI\napplications\, such as LLM-based agents\, pre sent unique workload\ncharacteristics: instead of diverse independent requ ests\, agentic\nframeworks typically submit repetitive inference requests\ , such as\nmulti-agent pipelines performing similar subtasks or self-refin ement\nloops iteratively enhancing outputs. These workloads result in long \nand highly predictable sequences\, which current speculative decoding\nm ethods do not effectively exploit. To address this gap\, we introduce\nSuf fixDecoding\, a novel method that utilizes efficient suffix trees to\ncach e long token sequences from prompts and previous outputs. By\nadaptively s peculating more tokens when acceptance likelihood is high\nand fewer when it is low\, SuffixDecoding effectively exploits\nopportunities for longer speculations while conserving computation\nwhen those opportunities are li mited. Evaluations on agentic\nbenchmarks\, including SWE-Bench and Text-t o-SQL\, demonstrate that\nSuffixDecoding achieves speedups of up to 5.3x\, outperforming\nstate-of-the-art methods — 2.8x faster than model-based approaches\nlike EAGLE-2/3 and 1.9x faster than model-free approaches such as\nToken Recycling.\n\nPresented in Partial Fulfillment of the CSD Speak ing Skills\nRequirement \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913bc46 DTSTART;TZID=America/New_York:20260204T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260204T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar / Doctoral Speaking Skills Talk CLASS:PUBLIC DESCRIPTION:Speaker: THOMAS DRAPER\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Efficient Online Random Sampling Via Randomness Recycling\n\nWe study the fundamental problem of using i.i.d.~coin tosses from an\nentropy source to efficiently generate r andom variables Xi ~ Pi  (i\n≥1) where (P1\, P2\,…) is a random seq uence of rational discrete\nprobability distributions subject to an arbitr ary stochastic process.\nOur method achieves an amortized expected entropy cost within ϵ ≥ 0\nbits of the information-theoretically optimal Shann on lower bound\nusing O(log(1/ϵ)) space. This result holds both pointwise in terms of\nthe Shannon information content conditioned on Xi and Pi\, and in\nexpectation to obtain a rate of 𝔼[H(P1) +… + H(Pn)]/n + ϵ bi ts\nper sample as n → ∞ (where H is the Shannon entropy). The\ncombina tion of space\, time\, and entropy properties of our method\nimproves upon the Knuth and Yao (1976) entropy-optimal algorithm and\nHan and Hoshi (19 97) interval algorithm for online sampling\, which\nrequire unbounded spac e. It also uses exponentially less space than\nthe more specialized method s of Kozen and Soloviev (2022) and Shao and\nWang (2025) that generate i.i .d.~samples from a fixed distribution.\nOur online sampling algorithm rest s on a powerful algorithmic\ntechnique called randomness recycling\, which reuses a fraction of the\nrandom information consumed by a probabilistic algorithm to reduce its\namortized entropy cost.\n\nOn the practical side\ , we develop randomness recycling techniques to\naccelerate a variety of p rominent sampling algorithms\, which include\nuniform sampling\, inverse t ransform sampling\, lookup table sampling\,\nalias sampling\, and discrete distribution generating (DDG) tree\nsampling. We show that randomness rec ycling enables state-of-the-art\nruntime performance on the Fisher-Yates s huffle when using a\ncryptographically secure pseudorandom number generato r\, and that it\nreduces the entropy cost of discrete Gaussian sampling.\n \nPresented as part of the Theory Lunch Seminar\n\nPresented in Partial Fu lfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913c0a6 DTSTART;TZID=America/New_York:20260202T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260202T130000 LOCATION:Panther Hollow Conference Room\, Mehrabian Collaborative Innovatio n\nCenter\, CIC 4105 SUMMARY:CyLab Seminar - David Kohlbrenner CLASS:PUBLIC DESCRIPTION:Speaker: DAVID KOHLBRENNER\, Assistant Professor\nand Co-Direct or\, Security and Privacy Research Lab\nPaul G. Allen School of Computer S cience and Engineering\nUniversity of Washington\n\nTalk Title: Software s ecurity to transistor physics and backMonday\,\nFebruary 2\, 2026\, 12  – 1pm Sadly\, our computers are physical\nobjects with physical limitat ions. One such limitation is that\ntransistors themselves do not have stat ic behavior and experience a\nvariety of aging behaviors.From a security p erspective\, this aging has\nsome interesting properties: aging behaviors are dependent on usage\,\nand aging behavior can cause failures in digital logic. In this talk\,\nI will discuss our ongoing security work on both d ata recovery and\nfaults on cloud systems using transistor aging.—David Kohlbrenner is\nan Assistant Professor in the Paul G. Allen School at the University\nof Washington and co-directs the Security and Privacy Research Lab.\nHis research focuses on understanding and mitigating security risks to\nsoftware that arise from hardware design. Projects with his\ncollabor ators include discovering novel side-channel mechanisms from\nCPU frequenc y (Hertzbleed) to prefetchers (Augury)\, and verifiable\nmitigations for s ide channels. He is also a lead on the RISC-V based\nKeystone TEE Framewor k project.David's work has been recognized with\nan IEEE Micro Top Pick\, two Top Picks in HES\, and Pwnie awards.\nPreviously\, he co-founded the e mbedded security company Somerset\nRecon\, and was a member of Carnegie Me llon's PPP CTF team.Faculty\nHost:  Riccardo PaccagnellaIn Person and Zoo m Participation.  See\nannouncement. Event Type: Seminars Room Number:  In Person and\nVirtual - ET Building: Panther Hollow Conference Room\, Me hrabian\nCollaborative Innovation Center\, CIC 4105 Speaker's Name: DAVID \nKOHLBRENNER Speaker Website: dkohlbre.com Speaker's Professional\nTitle : Assistant Professor\, and Co-Director\, Security and Privacy\nResearch Lab\, Paul G. Allen School of Computer Science and\nEngineering\, Universi ty of Washington Talk Title: Software security\nto transistor physics and back For More\nInformation: bethbuch@andrew.cmu.edu Affiliations: Compu ter Science\nDepartment (CSD) Organization(s): School of Computer Science \, CyLab\nEvent Website Title: Event Website Event Website\nURL: www.cyl ab.cmu.edu [http://www.cylab.cmu.edu]…\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913c5a6 DTSTART;TZID=America/New_York:20260311T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260311T170000 LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:CMU Commencement Fair - Day One CLASS:PUBLIC DESCRIPTION:Talk Title: CMU Commencement Fair\n\nAll graduating bachelor’ s\, master’s and doctoral students are\ninvited to attend Commencement F air to purchase their regalia and\ndiploma frames\, learn about alumni res ources\, enter a raffle and much\nmore. \n\nGet Exclusive In-Person Savin gs\n\nGet 10% off regalia when you purchase in-person at the Commencement\ nFairGraduation apparel and gifts are exclusive to the Commencement\nFairD iscounted Diploma Frames\n\nWhat to Bring\n\nYour CMU IDYour height (witho ut shoes)Purchasing for a friend? Please\nbring their CMU ID/Andrew ID alo ng with their height (without shoes)\n\nDiscounted Regalia PricingBachelor ’s Cap\, Gown and Tassel: $50.99\n$45.89Master’s Cap\, Gown and Tassel : $52.99 $47.69Master’s Hood:\n$71.99 $64.79Doctoral Cap\, Gown and Tass el: $65.99 $59.39 Doctoral\nHood: $97.99 $88.19 \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913c918 DTSTART;TZID=America/New_York:20260312T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260312T170000 LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:CMU Commencement Fair - Day Two CLASS:PUBLIC DESCRIPTION:Talk Title: Commencement Fair\n\nAll graduating bachelor’s\, master’s and doctoral students are\ninvited to attend Commencement Fair to purchase their regalia and\ndiploma frames\, learn about alumni resourc es\, enter a raffle and much\nmore. \n\nWhat to Bring\n\nYour CMU IDYour height (without shoes)Purchasing for a friend? Please\nbring their CMU ID/ Andrew ID along with their height (without shoes)\n\nGet Exclusive In-Pers on Savings\n\nGet 10% off regalia when you purchase in-person at the Comme ncement\nFairGraduation apparel and gifts are exclusive to the Commencemen t\nFairDiscounted Diploma Frames\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913cc41 DTSTART;TZID=America/New_York:20260212T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260212T150000 LOCATION:Remote Access - Zoom SUMMARY:AI-SDM Seminar - Iyad Rahwan CLASS:PUBLIC DESCRIPTION:Speaker: IYAD RAHWAN\, Professor and Director\, Max Planck Inst itute for\nHuman Development\, Berlin\n\nTalk Title: Science Fiction Scien ce\n\nCan we predict the social and behavioral impacts of future\ntechnolo gies\, such as Artificial Intelligence\, while they are still\nbeing devel oped in scientific labs\, or even when they are just\nimaginations in the minds of a science fiction writer? Such prediction\nwould allow us to guid e development and regulation of technologies\nbefore their impacts get ent renched. This talk describes ‘science\nfiction science’ (sci-fi-sci)\, the use of experimental methods to\nsimulate future technologies\, and co llect quantitative measures of the\nattitudes and behaviors of participant s assigned to controlled\nvariations of the future. I present various rece nt sci-fi-sci projects\naimed at anticipating the societal impacts of Arti ficial Intelligence\,\nand discuss the potential and limitations of this f orm of science.\n\n—\n\nProfessor Iyad Rahwan is director of the Max Pla nck Institute for\nHuman Development in Berlin\, where he founded and dire cts the Center\nfor Humans & Machines. He is also an honorary professor of Electrical\nEngineering and Computer Science at the Technical University of\nBerlin. Prior to moving to Berlin\, he was an Associate Professor of\n Media Arts & Sciences at the Massachusetts Institute of Technology\n(MIT). A native of Aleppo\, Syria\, Rahwan holds a PhD from the\nUniversity of M elbourne\, Australia.\n\nRahwan’s research agenda\, which he calls scien ce fiction science\,\nanticipates the impact of Artificial Intelligence on the way we think\,\nlearn\, work\, play\, cooperate and govern. His work appeared in the\nworld’s leading academic journals\, including Science a nd Nature\, and\nfeatures regularly in media outlets\, including the New Y ork Times\, The\nEconomist\, and the Wall Street Journal. His artistic and scientific\nwork was also featured in some of the world’s leading cultu ral\ninstitutions\, such as Ars Electronica\, Science Museum London and\nC ooper Hewitt Smithsonian Design Museum.\n\nREGISTER → register to atten d in-person or on Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913d076 DTSTART;TZID=America/New_York:20260128T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260128T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Jason Li CLASS:PUBLIC DESCRIPTION:Speaker: JASON LI\, Assistant ProfessorComputer Science\nDepart mentCarnegie Mellon University\n\nTalk Title: Deterministic Padded Decompo sitions and Negative-Weight\nShortest Paths\n\nA recent breakthrough of Be rnstein\, Nanongkai\, and Wulff-Nilsen\nestablished the first near-linear time algorithm for negative-weight\nsingle-source shortest paths on intege r-weighted graphs. We refine\ntheir approach and obtain the first near-lin ear time deterministic\nalgorithm for the problem. Our main ingredient is a deterministic\nconstruction of a padded decomposition on directed graphs \, which may\nbe of independent interest.\n\nThe talk will present the ent ire algorithm and proof at the level of a\ngraduate algorithms class. \n\ n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913d3af DTSTART;TZID=America/New_York:20260129T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260129T160000 LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Boris Bukh CLASS:PUBLIC DESCRIPTION:Speaker: BORIS BUKH\, ProfessorDepartment of Mathematical\nScie ncesCarnegie Mellon University\n\nTalk Title: Most frequent subsequences i n a word\n\nThere are 12 ways to erase eight letters in the word \"Mississ ippi\" to\nobtain \"sip\". More generally\, inside any long word with very few\ndistinct letters\, there are many copies of some fixed subsequence. We\nshall discuss how many is \"many\" when \"few\" is bigger than a gazil lion\nand two.\n\nJoint work with Aleksandre Saatashvili.\n\n4:00 pm →  Jane Street-sponsored tea and cookies in the Math Lounge\n(bring your mu g) \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913d6da DTSTART;TZID=America/New_York:20260210T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260210T173000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Katayanagi Distinguished Lecture - Tom Mitchell CLASS:PUBLIC DESCRIPTION:Speaker: TOM M. MITCHELL\, SCS Founders University Professor\nM achine Learning Department\, School of Computer Science\,\nand Block Cente r for Technology and Society\nCarnegie Mellon University\n\nTalk Title: Th e History of Machine Learning: How Did We Get\nHere?Tuesday\, February 10\ , 2026\, 4 – 5:30pm Machine learning is\nthe key technology underlying today's amazing artificial intelligence\nsystems.How did we get to today' s technology which now supports a\ntrillion dollar AI industry? What were the key scientific\nbreakthroughs? What were the surprises and dead-ends a long the way\, as\nseen by the researchers who created them? Who were the personalities\ninvolved\, and what were they thinking at the time? What sh ould we\nlearn from all this?This talk will explore the history of machine \nlearning based on personal experience of the speaker\, augmented by\nvid eo interviews with a dozen pioneering researchers in the\nfield.—Tom M. Mitchell is the Founders University Professor at\nCarnegie Mellon Univers ity\, where he founded the world's first Machine\nLearning Department\, an d served as Interim Dean of the School of\nComputer Science (2018-2019).  Beginning with his 1979 Ph.D. research\nhe has worked in machine learning and AI\, and he remains optimistic\nabout its future. In 2010 Mitchell wa s elected to the U.S. National\nAcademy of Engineering  \"For pioneering contributions and leadership\nin the methods and applications of machine l earning.\"  Mitchell is\nalso a member of the American Academy of Arts an d Sciences\, and a\nFellow and Past President of the Association for the A dvancement of\nArtificial Intelligence (AAAI).…About the Lecture:  The Katayanagi\nLectures recognize the best and the brightest in the field of computer\nscience and are presented by the School of Computer Science at\n Carnegie Mellon University in close cooperation with the Tokyo\nUniversity of Technology (TUT). The lectures recognize both senior and\njunior talen t.  The series were established through a gift from\nJapanese entrepreneu r and education advocate\, Mr. Koh Katayanagi\, who\nfounded TUT and other technical institutions in Japan over many\nmultiple decades.  We are del ighted to have TUT as\npartners.Rescheduled from January 27 Event Type: S CS Distinguished\nLectures Room Number: In Person Building: Rashid Audit orium\, Gates\nHillman 4401 Speaker's Name: TOM M. MITCHELL Speaker\nWebs ite: www.cs.cmu.edu [http://www.cs.cmu.edu]… Speaker's\nProfessional Ti tle: SCS Founders University Professor\, Machine\nLearning Department\, S chool of Computer Science\,\, and Block Center for\nTechnology and Society \, Carnegie Mellon University Talk Title: The\nHistory of Machine Learnin g: How Did We Get Here? For More\nInformation: scs-dls@cs.cmu.edu Affilia tions: Computer Science\nDepartment (CSD)\, Human-Computer Interaction In stitute (HCII)\,\nLanguage Technologies Institute (LTI)\, Machine Learning Department\n(MLD)\, Ray and Stephanie Lane Computational Biology Departme nt (CBD)\,\nRobotics Institute (RI)\, Software and Societal Systems Depart ment\n(S3D) Organization(s): School of Computer Science Event Website\nTi tle: Event Website Event Website URL: www.cs.cmu.edu\n[http://www.cs.cmu .edu]…\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913dc98 DTSTART;TZID=America/New_York:20260130T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260130T143000 LOCATION:Remote Access - Zoom SUMMARY:STAMPS Seminar - Viviana Acquaviva CLASS:PUBLIC DESCRIPTION:Speaker: VIVIANA ACQUAVIVA\, Professor of Physics\, Physics Dep artment\,\nCollege of Technology\, City University of New York\n\nTalk Tit le: Learning from simulations using statistics\, ML\, and AI\n\nMy researc h focuses on the process of learning from simulations using\na variety of numerical methods\, from classic statistics to machine\nlearning to genera tive AI tools. I will show a few examples from my\nAstrophysics work\, on validating cosmological simulations and\nformulating hypotheses for the ph ysical models that drive galaxy\nevolution processes. I will then move on to current research in\nclimate science\, where we are developing custom m etrics to assess\nsimilarity in climate models outputs\, and using represe ntation\nlearning to improve the reconstruction of full spatiotemporal fie lds\nfrom sparse and biased ocean data. I will conclude with some lessons\ nlearned in applying ML/AI across disciplines\, and some considerations\na nd open questions on how AI is changing the way we do science.\n\n—\n\nD r. Viviana Acquaviva is a Professor of Physics in the City University\nof New York. She received her Masters degree in Theoretical Physics\nfrom the University of Pisa and her PhD in Astrophysics from the\nInternational Sc hool for Advanced Studies in Trieste\, and held\npostdoctoral positions at Princeton University and Rutgers University\nbefore joining the faculty a t CUNY. After many years of research in\nAstrophysics with statistical too ls\, machine learning\, and AI\, she\npivoted to Climate Data Science than ks to a PIVOT fellowship\, followed\nby a PIVOT Research Award\, by the Si mons Foundation. Her current\nresearch is centered on developing new metri cs to assess the\nperformance of global climate models and on reconstructi ng full\nspatio-temporal fields\, in particular ocean carbon\, from limite d and\nbiased data. She is also working with early career scientists to\nr eflect on how generative AI tools can be responsibly incorporated in\nthe scientific workflow and on developing community tools around that\ntopic. Her textbook “Machine Learning for Physics and Astronomy”\,\npublished in 2023 by Princeton University Press\, won the 2024\nChambliss Astronomi cal Writing award from the American Astronomical\nSociety.\n\nZoom Partici pation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913e0d8 DTSTART;TZID=America/New_York:20260202T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260202T163000 LOCATION:Remote Access - Zoom SUMMARY:VASC Seminar - Chen Zhao CLASS:PUBLIC DESCRIPTION:Speaker: CHEN ZHAO\, Postdoctoral Research Fellow\, Computer Vi sion\nLabEPFL\n\nTalk Title: From Lab to Reality: Reliable 3D Vision in th e Wild\n\n\\While deep learning has revolutionized 3D computer vision\, a\ nsignificant gap remains between the performance achieved in controlled\nl aboratory settings and that in complex\, uncontrolled real-world\nenvironm ents. This talk addresses the critical challenges of\nrobustness and gener alization required to bridge this gap. In this\npresentation\, I will firs t discuss our contributions to 3D\nreconstruction\, including robust multi -view reconstruction\, physically\ngrounded 3D shape generation\, and 3D G aussian Splatting under\nsparse-view conditions. Next\, I will cover 3D in teraction with a focus\non generalizable object pose estimation. I will de monstrate how\nleveraging different types of reference information can fac ilitate\npose estimation for previously unseen objects in uncontrolled\nen vironments. Finally\, I will conclude by outlining future directions\ntowa rd multi-modal 3D understanding\, unified 3D representations\, and\nthe de velopment of 3D foundation models. \n\n—\n\nChen Zhao is a Postdoctoral Research Fellow at the Computer Vision\nLab\, EPFL\, working with Dr. Mat hieu Salzmann and Prof. Pascal Fua.\nEarlier\, he was a PhD candidate at E PFL\, supervised by Dr. Mathieu\nSalzmann and Prof. Pascal Fua. His resear ch interests lie in 3D\ncomputer vision\, with a specific focus on 3D reco nstruction\, 3D\ninteraction\, and 3D understanding.\n\nThe VASC seminar i s generously sponsored by HeyGen\n\nZoom Participation.  See announcemen t. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913e4bc DTSTART;TZID=America/New_York:20260130T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260130T140000 LOCATION:Newell-Simon 3002 SUMMARY:Doctoral Thesis Oral Defense - Caspar Oesterheld CLASS:PUBLIC DESCRIPTION:Speaker: CASPAR OESTERHELD\, Ph.D. Candidate\nComputer Science Department\nCarnegie Mellon University\n\nTalk Title: New foundational ide as in cooperative AI\n\nMy doctoral research addresses two fundamental obs tacles to beneficial\noutcomes from strategic interactions between multipl e parties:\nstrategic incentives against cooperation (as in the Prisoner's \nDilemma) and the multiplicity of strategic solutions (sometimes called\n the equilibrium selection problem). As AI systems are increasingly\ninvolv ed in consequential decision making processes on behalf of human\nprincipa ls\, understanding how to achieve desirable outcomes in\nmulti-agent AI se ttings becomes critical. My research leverages unique\nfeatures of AI syst ems -- including their transparency\,\nreproducibility\, and malleability -- to develop novel game-theoretic\napproaches that enable better\, more c ooperative outcomes.\n\n    \n\nThree primary research directions form the core of this dissertation.\nFirst\, the concept of safe (Pareto) impro vements provides a rigorous\nframework for improving outcomes without reso lving equilibrium\nselection problems. Unlike traditional solution concept s\, safe Pareto\nimprovements make qualitative assumptions about pairs of games rather\nthan individual games. This sometimes allows us to prefer pl aying one\ngame over another\, without any judgment about how each of the\ nindividual games is played. Second\, my research on so-called\nNewcomb-li ke decision problems takes inspiration from philosophical\nbranches of dec ision theory concerning\, for example\, how one should\nreason when intera cting with a copy of oneself. I investigate how\ncooperation can be achiev ed when different parties deploy similar AI\nsystems. Third\, the concept of program equilibrium explores how the\nuse of mutually transparent decis ion-making algorithms can allow for\ncooperation.\n\nThesis Committee\n\nV incent Conitzer (Chair)\n\nTuomas Sandholm\n\nFei Fang\n\nStuart Russell ( University of California\, Berkeley)\n\nBen Levinstein (University of Illi nois Urbana-Champaign / Anthropic) \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913e927 DTSTART;TZID=America/New_York:20260126T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260126T170000 LOCATION:Baker Hall A53 SUMMARY:Statistics and Data Science Seminar - Dylan Foster CLASS:PUBLIC DESCRIPTION:Speaker: DYLAN FOSTER\, Principal ResearcherMicrosoft Research\ , New\nEngland and New York City\n\nTalk Title: From Estimation to Decisio ns: Statistical Foundations for\nInteractive Learning\n\nThe prevailing re cipe of ever-larger models trained on passively\ncollected data is showing diminishing returns and faces growing\nconstraints on data. The next phas e of progress will increasingly rely\non interactive learning: systems tha t actively collect data through\nexperimentation\, from clinical trials an d A/B testing to\nrecommendation systems and scientific discovery. This ta lk presents a\nresearch program developing statistical foundations for int eractive\nlearning with modern deep models.\n\nPart I introduces the Decis ion-Estimation Coefficient\, a unifying\nframework for understanding when interactive learning is statistically\ntractable\, analogous to empirical process theory in supervised\nlearning. This theory directly yields practi cal\, industry-deployed\nalgorithms that transform any off-the-shelf estim ator into an optimal\nsequential decision-making method.\n\nPart II addres ses the modern paradigm of adapting pre-trained\nfoundation models for seq uential decision making. I introduce the\ncoverage profile—the probabili ty mass a model places on high-quality\noutputs—as a key statistical qua ntity governing post-training\nsuccess\, leading to new interventions that connect pre-training\nobjectives\, post-training signals\, and downstream performance.\n\n—\n\nDylan Foster is a principal researcher at Microsof t Research\, New\nEngland and New York City. Previously\, he was a postdoc toral fellow at\nIDSS at MIT and received his PhD from Cornell University\ , advised by\nKarthik Sridharan. His research develops statistical foundat ions for\ninteractive learning---spanning reinforcement learning\, bandits \,\nimitation learning\, and causal inference---with a focus on\nunderstan ding when complex decision-making problems can be solved\nusing standard e stimation primitives. His recent work applies this\nperspective to underst and and improve the foundation model training\npipeline\, particularly pos t-training. His work has received several\nawards\, including COLT Best Pa per (2019)\, COLT Best Student Paper\n(2018\, 2019)\, and the Cornell CS d octoral dissertation award.\n\n3:30 pm → Refreshments will be offered p rior to the seminar \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913ed9f DTSTART;TZID=America/New_York:20260126T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260126T130000 LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar - Abhi Shelat CLASS:PUBLIC DESCRIPTION:Speaker: ABHI SHELAT - To be Rescheduled\, Professor\nKhoury Co llege of Computer Sciences\nNortheastern University\,\nand Engineer\, Goog le\n\nTalk Title: Zero knowledge proofs in your wallet\n\nI'll discuss the journey to get zero-knowledge proof protocols used to\nconvey your privat e identity information when websites ask for it. \n Specifically\, Googl e wallet now supports zk presentation for mobile\ndriver license IDs.  I' ll discuss the problem\, how we made the\ntechnical choices\, and what new ideas we needed to get proofs for\nlegacy ID formats to work on the phone .\n\n—\n\nAbhi Shelat is a professor at Northeastern University and an engineer\nat Google for this project.\n\nIn Person and Zoom Participation.   See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913f0ff DTSTART;TZID=America/New_York:20260220T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260220T230000 LOCATION:Rallgos Ballroom\, Cohon University Center SUMMARY:SCS Day: A Celebration of Incredible Diversity & Talent in SCS CLASS:PUBLIC DESCRIPTION:Speaker: SCS in Wonderful\n\nFor those who don't know\, SCS Day is the pride day for the School of\nComputer Science! It is a great way t o celebrate the talents and\npassion of our fellow SCS members. The theme this year is  SCS in\nWonderland!\n\nFeaturing\n\nTalent Show and Facult y FeudArt Gallery and Photo-OpFree Food\,\nStickers\, ShirtsPaint-a-Tote-B ag\, Games by theGame Creation\nSociety\, & More!\n\nQuestions \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913f404 DTSTART;TZID=America/New_York:20260123T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260123T150000 LOCATION:Newell-Simon Hall SUMMARY:Doctoral Thesis Oral Defense - William Zhang CLASS:PUBLIC DESCRIPTION:Speaker: WILLIAM ZHANG\, Ph.D. Candidate\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: On Holistic Database Optimization via Leveraging\nSimilarity Across Actions\, Workloads\, Confi gurations\, and Scenarios\n\nModern database management systems (DBMSs) ha ve evolved to support\nincreasingly sophisticated data-intensive applicati ons\, at the cost of\nsubstantial complexity to configure them for two rea sons. First\, DBMSs\nexpose a vast configuration space with trillions of p ossibilities that\nencompass system knobs\, physical design (e.g.\, indexe s)\, and query\noptions\, amongst others. Second\, these applications are constantly\nevolving with changes in data access patterns\, query types\, load\nintensities\, hardware\, and data distributions that necessitate\nco ntinuous re-optimization.\n\nTo address these challenges\, decades of auto nomous DBMS optimization\nresearch have produced specialized tuning tools to assist human\noperators. Deploying these tools involves a complex multi -step\nworkflow where an operator (1) observes the DBMS’s behavior\, (2) \nselects tools based on the objectives and their expertise\, (3)\nconfigu res them with an isolated environment\, (4) orchestrates their\nexecution to obtain recommendations\, and (5) reviews those\nrecommendations before deployment. This cumbersome process results in\nsuboptimal configurations and slow adaptation to evolving\napplications’ workloads due to isolated specialized tools\,\ninefficient reuse of prior tuning knowledge\, and th e fallible human\nfactor.\n\nIn this dissertation\, we present techniques for addressing those\nlimitations with similarity to enable holistic datab ase optimization.\nFirst\, we present a holistic tuning tool that optimize s multiple DBMS\naspects simultaneously by using action similarity to orga nize actions\ninto neighborhoods conducive to exploration. We then present a\nframework that assists tuners in adapting to environment changes by\nl everaging workload and configuration similarity to re-mix historical\nknow ledge. Lastly\, we present a system that transforms the\nhuman-centric tun ing workflow into an agentic process by using\nscenario similarity to link the deployment context with semantic tool\ninterfaces to optimize the dep loyment.\n\nThe techniques and associated similarity definitions presented in this\ndissertation enable agentic holistic DBMS optimization over a\nd eployment’s lifetime\, improving the deployment’s performance and\nred ucing time taken to adapt to changes in upstream user applications.\n\nThe sis Committee:\n\nAndrew Pavlo (Chair)\n\nJignesh Patel\n\nVincent Conitze r\n\nImmanuel Trummer (Cornell)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913f8b3 DTSTART;TZID=America/New_York:20260123T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260123T113000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Thesis Proposal - Joshua Clune CLASS:PUBLIC DESCRIPTION:Speaker: JOSHUA CLUNE\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Leveraging Automated The orem Provers for Lean Proof\nAutomationThis proposal discusses work that b uilds towards the\ncreation of a hammer for the Lean interactive theorem p rover. Said\nwork includes the development of a proof-producing superposit ion\ntheorem prover for Lean\, a tool which interfaces with the cvc5 SMT\n solver to produce self-contained Lean proof scripts\, and a neural\npremis e selection system. The proposal culminates in the description\nof a preli minary hammer for Lean along with discussion of how to\nrefine and improve it into a more powerful and robust tool.\n\nThesis Committee:\n\nJeremy A vigad (Chair)\n\nMarijn Heule\n\nBryan Parno\n\nHaniel Barbosa (Universid ade Federal de Minas Gerais)\n\nIn-person and Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef913fc57 DTSTART;TZID=America/New_York:20260202T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260202T103000 LOCATION:Remote Access - Zoom SUMMARY:AI-SDM Seminar - Jim Jansen CLASS:PUBLIC DESCRIPTION:Speaker: JIM JANSEN\, Principal Scientist\nAI Group\nQatar Comp uting Research Institute (QCRI)\n\nTalk Title: Cipherbot: a 24/7 TA for Pr ofs\; 24/7 Tutor for Students\n\nDr. Jim Jansen has authored or co-authore d 400 or so research\npublications\, with articles appearing in a multi-di sciplinary range of\njournals and conferences. He is co-author of the book \, Understanding\nAudiences\, Customers\, and Users via Analytics – An I ntroduction to\nthe Employment of Web\, Social\, and Other Types of Digita l People Data\n(Springer Nature)\, Data-Driven Personas (Springer Nature)\ , author of\nthe book Understanding Sponsored Search: A Coverage of the Co re\nElements of Keyword Advertising (Cambridge University Press)\, author\ nof the book Understanding User - Web Interactions Via Web Analytics\,\nco -author of the book\, Web Search: Public Searching of the Web\, and\nco-ed itor of the book Handbook of Research on Weblog Analysis. Dr.\nJansen is a ACM Distinguished Speaker.\n\nDr. Jansen is a Principal Scientist in the artificial intelligence\n(AI) group of the Qatar Computing Research Instit ute (QCRI). He is a\ngraduate of West Point and has a Ph.D. in computer sc ience from Texas\nA&M University\, along with master degrees from Texas A& M (computer\nscience) and Troy State (international relations). Dr.  Jans en served\nin the U.S. Army as an Infantry enlisted soldier and communicat ion\ncommissioned officer.\n\nProfessor Jim Jansen is editor-in-chief of t he journal\, Information\nProcessing & Management (Elsevier)\, a member of the editorial boards\nof seven international journals\, interim editor-in -chief of the\njournal International Journal of Information Management\, f ormer\neditor-in-chief of the journal\, Internet Research (Emerald)\, and he\nhas served on the research committee for the Search Engine Marketing\n Professional Organization (SEMPO). He has received several awards and\nhon ors\, including an ACM Research Award\, six application development\naward s\, and a university-level teaching award\, along with other\nwriting\, pu blishing\, research\, teaching\, and leadership honors.\n\nProfessor Janse n has served as a Senior Fellow at the Pew Research\nCenter with the Pew I nternet and American Life Project and a\nuniversity expert with the Nation al Ground Intelligence Center. Jim\nhas held several jobs\, including busb oy\, dishwasher\, janitor\, cook\,\nand movie projectionist. Jim served in the US Army as an enlisted\nsoldier in the Infantry and the Signal Corps as an officer. After\nserving in the military\, he transitioned to his sec ond career in\nacademia. ​\n\nREGISTER → please register to attend in -person or on Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914011b DTSTART;TZID=America/New_York:20260122T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260122T160000 LOCATION:Wean 8220 SUMMARY:ACO Seminar - Noah Singer CLASS:PUBLIC DESCRIPTION:Speaker: NOAH SINGER\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Direct-product testers an d PCPs from coset complexes\n\n“Direct-product testers” are objects us ed in the design of (some)\nprobabilistically checkable proofs\, which\, i n turn\, play a fundamental\nrole in modern complexity theory and cryptogr aphy. We gently introduce\nthe direct-product testing problem and its rela tionship with expansion\nproperties of simplicial complexes. Then\, we dis cuss the so-called\n“Kaufman—Oppenheim coset complex” and our proof that it has\ndirect-product testing properties which were previously known only for\nless-elementary constructions. We only assume a basic backgroun d in\nfinite group theory (and no prior knowledge of theoretical computer\ nscience).\n\nBased on joint work with Ryan O’Donnell.\n\n4:00 pm → Ja ne Street-sponsored tea and cookies at in the math\nlounge (bring your mug !). \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914047d DTSTART;TZID=America/New_York:20260129T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260129T133000 LOCATION:Newell Simon 3305 and Zoom SUMMARY:AI-SDM Seminar - Chris McComb CLASS:PUBLIC DESCRIPTION:Speaker: CHRIS McCOMB\, Gerard G. Elia Associate Professor\nDep artment of Mechanical Engineering\nand Director\, Human+AI Design Initiati ve\n\nTalk Title: AI and the Battle for the Soul of Design\n\nChristopher McComb is a faculty member in Carnegie Mellon\nUniversity’s Department o f Mechanical Engineering. Previously\, he\nwas an assistant professor in t he School of Engineering Design\,\nTechnology\, and Professional Programs at Penn State. He also served as\ndirector of Penn State’s Center for Re search in Design and\nInnovation and led its Technology and Human Research in Engineering\nDesign Group.\n\nHe received dual B.S. degrees in civil a nd mechanical engineering from\nCalifornia State University-Fresno. He lat er attended Carnegie Mellon\nUniversity as a National Science Foundation G raduate Research Fellow\,\nwhere he obtained his M.S. and Ph.D. in mechani cal engineering.\n\nHis research interests include human social systems in design and\nengineering\; machine learning for engineering design\; human -AI\ncollaboration and teaming\; and STEM education\, with funding from NS F\,\nDARPA\, and private corporations.\n\nREGISTER → to attend in pers on and remotely via zoom.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9140841 DTSTART;TZID=America/New_York:20260205T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260205T130000 LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI-SDM Student Brainstorming Session CLASS:PUBLIC DESCRIPTION:Talk Title: AI-SDM Student Brainstorming Session\n\nJoin our mo nthly student-focused brainstorming sessions. Led by\nstudents of all leve ls\, these sessions offer a casual forum to discuss\nAI-related research w hile exploring cross-cutting connections between\nvarious disciplines in a rtificial intelligence.\n\nAll students welcome\n\nLunch is provided for i n-person attendees\n\nREGISTER → register to attend in-person or on Zo om\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9140b46 DTSTART;TZID=America/New_York:20260129T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260129T173000 LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:Deeper Conversations - Arthur Levine CLASS:PUBLIC DESCRIPTION:Speaker: ARTHUR LEVINE\, PresidentBrandeis University\n\n From Upheaval to Action: What Works in Changing Higher Education\n\nJoin Presi dent Farnam Jahanian as he hosts Brandeis University\nPresident Arthur Lev ine for a distinguished lecture and fireside chat.\nDr. Levine will reflec t on his decades of higher education leadership\nand discuss his upcoming book\, From Upheaval to Action: What Works in\nChanging Higher Ed.\n\nINDI CATE Your Attendance\n\n→ Please indicate your attendance by Tuesday\, J anuary 27. \n\n→ The first 100 registered attendees will receive a cop y upon\npublication in early March. \n\n► Arthur E. Levine\, a nationa lly respected higher education leader\nand researcher\, is president of Br andeis University. A 1970 Brandeis\ngraduate\, Levine previously served as president of Teachers College at\nColumbia University and the former Wood row Wilson National Fellowship\nFoundation. \n\nLevine has authored 13 bo oks\, including The Great Upheaval: Higher\nEducation’s Past\, Present a nd Uncertain Future (2021)\, and articles\nin such publications as The Wal l Street Journal\, The New York Times\nand The Washington Post. Levine’s honors include Carnegie\,\nFulbright\, Guggenheim and Rockefeller Foundat ion Fellowships. He is a\nmember of the American Academy of Arts and Scien ces.\n\n  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9140f25 DTSTART;TZID=America/New_York:20260115T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260115T133000 LOCATION:Newell Simon 3305 and Zoom SUMMARY:AI-SDM Seminar - Nuoya Xiong CLASS:PUBLIC DESCRIPTION:Speaker: NUOYA XIONG\, Ph.D. StudentMachine Learning Department Carnegie\nMellon University\n\nTalk Title: Multi-Objective and Multi-Group RLHF\n\nReinforcement Learning with Human Feedback (RLHF) is a widely use d\nfine-tuning approach that aligns machine learning models\, particularly \nLanguage Model (LM) with human preferences. There are typically\nmultipl e objectives driving the preference\, hence humans find it\neasier to expr ess per-objective comparisons rather than a global\npreference between two choices. Multi-Objective RLHF (MORLHF) aims to\nuse per-objective prefere nce feedback and achieve Pareto optimality\namong these objectives by aggr egating them into a single unified\nobjective for optimization. However\, nearly all prior works rely on\nlinear aggregation\, which rules out polic ies that favor specific\nobjectives such as the worst one. The only existi ng approach using\nnon-linear aggregation is computationally expensive due to its\nreward-based nature and the need for retraining whenever the\nagg regation parameters change. \n\nIn this work\, we address this limitation by transforming the\nnon-linear aggregation maximization problem into a s eries of\nsub-problems. Each sub-problem involves only linear aggregation\ ,\nmaking it computationally efficient to solve. We further extend our\nfr amework to handle multi-group scenarios\, where each group has\ndistinct w eights for the objectives. Our method enables achieving\nconsensus or maxi mizing the aggregated objective across all groups.\nTheoretically\, we dem onstrate that our algorithmic framework achieves\nsublinear regret. Empiri cally\, leveraging our theoretical insights\, we\npropose a nearly trainin g-free algorithm once the optimal policies for\nindividual objectives are obtained.\n\n—\n\nNuoya Xiong (熊诺亚) is a first-year PhD student at Carnegie Mellon\nUniversity in the Machine Learning Department\, advised by Prof. Aarti\nSingh. Before that\, he was an undergraduate student at II IS (Yao\nclass)\, Tsinghua University. Xiong’s recent research interest is\nlearning theory and RLHF/LLM\, especially theoretically elegant and\ni mplement-friendly algorithms for RLHF/LLM/decision-making. He is also\nint erested in reinforcement learning theory and deep learning theory.\nXiong ’s research goal is to understand the phenomenon of machine\nlearning an d provide new efficient algorithms from the theoretical\naspect\, serving as a guide for the application of machine learning.\n\nREGISTER → to a ttend in-person or on Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91413a7 DTSTART;TZID=America/New_York:20260122T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260122T170000 LOCATION:Remote Access (registration required) SUMMARY:CMU University Libraries: Fine & Rare VI CLASS:PUBLIC DESCRIPTION:Speaker: Sam LemleyTalk Title: Fine & Rare VI Rare Books and An cestral\nMachines\n\nThe University Libraries invites you to join us onlin e for another\nedition of Fine & Rare\, highlighting the inaugural exhibit ion in the\nnewly reimagined Posner Center for Special Collections and the \ncollections that inspire teaching\, research\, and public engagement\nac ross the university.  The Posner Center for Special Collections at\nCMU s erves as an interdisciplinary laboratory where historical\nmaterials and e merging research practices come together to support\ndiscovery\, creativit y\, and scholarship. \n\nIn this installment of Fine & Rare\, Curator of Special Collections Sam\nLemley introduces Rare Books & Ancestral Machines \, an exhibition\nfeaturing more than thirty rare books\, instruments\, an d early\ncomputing devices—many on public view for the first time. Organ ized\ninto five thematic sections—mechanical computing\, robotics\,\ncry ptography\, artificial intelligence\, and scholarly method—the\nexhibiti on surveys over four centuries of scientific inquiry\,\nuncovering abandon ed prototypes and long-forgotten technologies. Join\nus for an inside look at the Posner Center’s collections\, its\ninaugural exhibition\, and th e future of research and teaching in this\nreimagined space.\n\nThis is a virtual event.\n\nPlease REGISTER to receive the link to attend.\n\n► S am V. Lemley is curator of Special Collections at Carnegie\nMellon Univers ity Libraries (CMU)\, an affiliate in the Center for\nEarly Modern Print\, Networks and Performance (CMU Department of\nEnglish) and a member of the Print & Probability project. He holds a\nPh.D. in English from the Univer sity of Virginia and an MLIS with a\ncertificate of concentration in rare books and special collections\nlibrarianship from the Palmer School (NYC).  At CMU\, he administers\nacquisitions\, exhibitions\, research and instr uction in the Posner\nCenter for Special Collections of rare books\, manus cripts and early\ncalculating devices and cryptographic machines. He is a (re)founding\nmember of The Pittsburgh Bibliophiles\, a gardener\, a book collector\nand father to a four-year-old son.\n\nLemley has held research fellowships at Princeton University\nLibraries\, the Dibner Library of the History of Science and\nTechnology\, Harvard University’s Houghton Libr ary\, Rare Book School\nand Oak Spring Garden Foundation. His work has app eared in Papers of\nthe Bibliographical Society of America\, The Library\, Studies in\nBibliography\, Shakespeare Quarterly\, Proceedings of the AAA I\nConference on Artificial Intelligence\, Hyperallergic and other\njourna ls. He edited a book on the four Folios of Shakespeare’s plays\n(Penn St ate University Press). The volume accompanied a pair of\nexhibitions mount ed at CMU and the Frick Pittsburgh. \n\nLemley’s most recent release\, Rare Books & Ancestral Machines: A\nHandbook to the Posner Center for Spec ial Collections at Carnegie\nMellon University\, spotlights a treasured co llection of objects from\nthe university’s special collections to mark t he reopening of the\nPosner Center.\n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914190c DTSTART;TZID=America/New_York:20260113T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260113T135000 LOCATION:Hamburg Hall 1002 and Zoom SUMMARY:Privacy Seminar - Clement Fung CLASS:PUBLIC DESCRIPTION:Speaker: CLEMENT FUNG\, Postdoctoral Research AssociateSoftware and\nSocietal Systems DepartmentCarnegie Mellon University\n\nTalk Title: Improving the Effectiveness of ML-based Approaches for\nIndustrial Contro l Systems Security\n\nTuesday\, January 13\, 2026\, 12:30 – 1:50pm \n \nIndustrial control systems (ICS) govern critical infrastructure and\npro cesses\, such as power generation\, chemical processing\, and water\ntreat ment. To defend and protect ICS from potential harm\, researchers\ncommonl y propose techniques based on machine learning (ML) for\ndetecting anomali es in ICS process values. Despite strong results in\nresearch\, ML-based a pproaches are rarely adopted in practice for ICS\ntoday. In this talk\, I cover work that makes ML-based anomaly\ndetection more effective for secur ing ICS\, by both investigating the\nneeds of current practice and by deve loping new ML-based approaches to\nmeet these newly identified needs. Firs t\, to better understand how\nML-based anomaly detection would be used in practice for ICS\, we\ninterview practitioners that work in ICS security a nd operations to\nunderstand their needs and requirements for adopting ML into ICS\nenvironments. Second\, it is unclear if and how anomaly-detectio n\noutputs can be used to diagnose ICS anomalies\; we evaluate a variety\n of explainable.\n\n —\n\nClement Fung is a postdoctoral research associ ate at Carnegie Mellon\nUniversity and a member of the CyLab security and privacy institute.\nHis research focuses on designing and evaluating ML-ba sed approaches\nfor securing industrial control systems in practice. He re ceived his\nM.S. and B.A.Sc. from the University of British Columbia and t he\nUniversity of Waterloo respectively.\n\nIn Person and Zoom Participati on.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9141d2f DTSTART;TZID=America/New_York:20260112T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260112T110000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Thesis Proposal - Zhihao Zhang CLASS:PUBLIC DESCRIPTION:Speaker: ZHIHAO ZHANG\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: A Path Towards Efficient Large Language Model Deployment\nthrough Algorithm and System Co-Design\n \nRecent advancements in large language models have shown promising\nresul ts in diverse downstream tasks by training and test time scaling.\nHowever \, the fast-paced development of large models has posed\nsignificant chall enges to their energy cost and efficient\ndeployment. \n\nTo achieve this \, my thesis topic is centered around bridging the gap\nbetween the algori thm-system co-design space for better large model\ndeployment. More specif ically\, through:\n\nhardware-guided algorithmic explorations for efficien t large language\nmodel inference\, and LLM inference-specific system opt imizations to\nfully exploit hardware utilization. \n\nFor algorithmic im provements\, I will present two lines of research\nprojects on Speculative Decoding (SpecInfer\, RaLMSpec) and Sparse\nAttention (TidalDecode\, Less isMore).\n\nFor system optimizations\, I will present one project on LLM d eployment\nwith MegaKernel (MPK) and one ongoing project that is focusing on\ngeneralizing the megakernel runtime to support multi-LLM\ndeployment.  \n\nBenefiting from the algorithm-system co-optimizations\, the proposed \nthesis topic is expected to provide an effective solution for reducing\n the energy cost and improving the efficiency of LLM deployment in the\nrea l world.\n\nThesis Committee\n\nZhihao Jia (Chair)\n\nTianqi Chen\n\nDimit rios Skarlatos\n\nRavi Netravali (Princeton University)\n\nAdditional Info rmation\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9142164 DTSTART;TZID=America/New_York:20260112T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260112T210000 LOCATION:All Over Campus SUMMARY:Spring Term - First Day of Classes CLASS:PUBLIC DESCRIPTION:Speaker: We're Back!\n\nMonday\, January 12\, 2026\, 8am –  9pm \n\nClasses resume and we're ready to go... \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9142453 DTSTART;TZID=America/New_York:20251222T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251222T113000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Thesis Oral Defense - Nuno Sabino CLASS:PUBLIC DESCRIPTION:Speaker: NUNO SABINO\, Ph.D. Candidate\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: Improving Code-Injection V ulnerability Detection and\nConfirmation in JS Programs\n\nJavaScript appl ications face serious security risks\, including\nclient-side DOM-based Cr oss-Site Scripting (DOM-XSS) and server-side\narbitrary command injection (ACI) and arbitrary code execution (ACE).\nExploiting these vulnerabilitie s can lead to severe consequences\,\nincluding unauthorized access to sens itive data and even full server\ncompromise.\n\nDynamic taint analysis (DT A) tools have been used to identify how\nattacker-controlled input\, such as a URL\, may reach sensitive\nfunctions that lead to arbitrary code exec ution. Such propagations of\nattacker information\, termed potential flows \,can be good indicators of\nvulnerabilities. However\, existing approache s struggle to (1) generate\nconcrete inputs that exercise these flows due to limited path\nexploration\, and (2) automatically confirm vulnerabiliti es\, because\ninputs must satisfy program constraints while also triggerin g the\nintended side effects. This thesis leverages program analysis\ntech niques to address these challenges\, with tailored approaches for\nthe dis tinct requirements of server and client code.\n\nClient-side analysis is c omplicated by program behaviors dependent on\nuser interactions and URL GE T parameters. To overcome this\, we\ndeveloped a fuzzer to interact with t he target web page and we employ\ndynamic symbolic execution (DSE) to synt hesize GET parameters\nsatisfying program constraints. Relative to our rep lication of prior\nwork DOMsday\, the fuzzer alone identifies 15% more vul nerabilities in\na dataset of 44\,480 popular pages\, and the combination of fuzzing and\nDSE iden tifies 43% more vulnerabilities than DOMsday.\n\n On the server-side\, DTA-based tools miss ACI and ACE that require\ninputs with complex structure. We develop a novel type- and\nstructure-aware fuz zing technique to explore Node.js packages\, and an\nenumerator to synthes ize syntactically valid payloads for ACE\nvulnerabilities. Extending NodeM edic with these components led to\nfinding 1.7x more vulnerabilities. Fina lly\, we find that\nnon-exploitable potential flows can still indicate rea l\nvulnerabilities\, but exploitation may imply extra steps\, such as\nbyp assing sanitization or extending attacker capabilities. We\nintroduce an e xploitability metric designed to indicate proximity to\nan exploitable pat h\, and use it to guide fuzzing and confirmation\ntowards paths that are m ore likely automatically exploitable.\nIntegrating this in NodeMedic-FINE results in 1% more confirmed flows\,\nwhile saving 28% of the baseline con firmation time.\n\nThesis Committee\n\nLimin Jia (Chair)\n\nLujo Bauer\n\n Ruben Martins\n\nPedro Adão (Advisor\, Instituto Superior Técnico)\n\nRu i Maranhão (Advisor\, Faculdade de Engenharia da Universidade do\nPorto)\ n\nJosé Fragoso (Instituto Superior Técnico)\n\nCristian-Alexandru Staic u (CISPA Helmholtz Center for Information\nSecurity) \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9142968 DTSTART;TZID=America/New_York:20251215T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251215T150000 LOCATION:McWilliams Classroom\, Gates Hillman 4303 SUMMARY:5th Year Master's Thesis Presentation - Umut Olmez CLASS:PUBLIC DESCRIPTION:Speaker: UMUT OLMEZ\, Master's Student\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: Autoregressive Customizabl e Bach Chorale Generation Using\na Lightweight Hybrid Model\n\nMusic\, esp ecially Bach’s music which is a staple in the teaching of\nfundamentals of musical composition\, exhibits various patterns and\nrules throughout v arious parameters\, such as harmony\, melody and\nstructure\, that give it a defining structure when we listen to it and\nmake it different than jus t noise. Many of these parameters are\nsimplified in Bach chorales\, makin g them great for studying the\nfundamentals of tonal composition and makin g them a perfect candidate\nto be generated by a computer. Looking at both the earlier\nautoregressive methods and the recent hybrid models\, this t hesis\nproduced a lightweight and customizable hybrid model that can take in\nany initial sequence of chords and start generating a chorale on top\n of it. With an augmented symbolic dataset of 4584 chorales\, using a\nTCNN and a transformer with self attention and RoPE\, the model\ncaptures loca l voice leading patterns and more global structural cues\nto generate a ch orale autoregressively. \n\nThe final model was able to train under 1.5 h ours in most cases and is\nvery easy to customize with adjustable layer nu mbers\, model dimension\,\nkernel size and dilation\, and extra steps duri ng fusion of transformer\nand TCNN layers. The generated chorales exhibite d interesting musical\nmaterial\, some mimicking Bach’s style successful ly through 2 bar\nphrases that end sensibly\, 4 bar phrases with half cade nces and\nperfect cadences\, modulations to the correct relative key\, and \nidiomatic four-part writing patterns. But it was also observed that\nthe chorales started to degenerate especially after the first few\nphrases. T he final model can be used by students or musicians being\nintroduced to f our-part writing as an assistant and is even able to\ncreate interesting i deas for experienced composers.\n\nThesis Committee\n\nRoger Dannenberg (C hair)\n\nChris Donahue\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9142de7 DTSTART;TZID=America/New_York:20251219T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251219T150000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Speaking Skills Talk - Zhihao Zhang CLASS:PUBLIC DESCRIPTION:Speaker: ZHIHAO ZHANG\, Ph.D. Student\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: TidalDecode/LessIsMore: rob ust sparse attention for\nefficient large language model reasoning\n\nLarg e reasoning models achieve strong performance through test-time\nscaling b ut incur substantial computational overhead\, particularly\nfrom excessive token generation when processing short input prompts.\nWhile sparse atten tion mechanisms can reduce latency and memory usage\,\nexisting approaches suffer from significant accuracy degradation due\nto accumulated errors d uring long-generation reasoning. These methods\ngenerally require either h igh token retention rates or expensive\nretraining. We introduce LessIsMor e\, a training-free sparse attention\nmechanism for reasoning tasks\, whic h leverages global attention\npatterns rather than relying on traditional head-specific local\noptimizations. \n\nLessIsMore aggregates token selec tions from local attention heads with\nrecent contextual information\, ena bling unified cross-head token\nranking for future decoding layers. This u nified selection improves\ngeneralization and efficiency by avoiding the n eed to maintain\nseparate token subsets per head. Evaluation across divers e reasoning\ntasks and benchmarks shows that LessIsMore preserves—and in some\ncases improves—accuracy while achieving a 1.1× average decoding\ nspeed-up compared to full attention. Moreover\, LessIsMore attends to\n2 × fewer tokens without accuracy loss\, achieving a 1.13× end-to-end\nspe ed-up compared to existing sparse attention methods.\n\nPresented in Parti al Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914323e DTSTART;TZID=America/New_York:20260204T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260204T181500 LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:2026 Dr. Martin Luther King Jr. Keynote Lecture - Dr. Beverly Danie l\nTatum CLASS:PUBLIC DESCRIPTION:Speaker: BEVERLY DANIEL TATUM\, President Emerita\, Spelman Col lege\n\nVice Provost for Community\, Culture and Engagement Wanda Heading- Grant\ninvites you to attend the annual Dr. Martin Luther King Jr. Keynote \nLecture\, featuring\n\nDr. Beverly Daniel Tatum\, president emerita of S pelman College\, is a\nclinical psychologist widely known for both her exp ertise on race\nrelations and as a thought leader in higher education. Her thirteen\nyears as the president of Spelman College (2002-2015) were mark ed by\ninnovation and growth and her visionary leadership was recognized i n\n2013 with the Carnegie Academic Leadership Award. The author of\nsevera l books including the best-selling Why Are All the Black Kids\nSitting Tog ether in the Cafeteria? and Other Conversations About Race\n(now in a new 2017 20th anniversary edition) and Can We Talk About\nRace? and Other Conv ersations in an Era of School Resegregation\n(2007)\, Tatum is a sought-af ter speaker on the topics of racial\nidentity development\, race and educa tion\, strategies for creating\ninclusive campus environments\, and higher education leadership. In\n2005 Dr. Tatum was awarded the prestigious Broc k International Prize\nin Education for her innovative leadership in the f ield. A Fellow of\nthe American Psychological Association\, she was the 20 14 recipient of\nthe APA Award for Outstanding Lifetime Contributions to P sychology.\n\nA civic leader in the Atlanta community\, Dr. Tatum is engag ed in\neducational initiatives designed to expand educational opportunity for\nunderserved students and their families. She serves on the governing\ nboards of the Westside Future Fund\, Achieve Atlanta\, Morehouse College\ nand Smith College as well as the Georgia Power Company and the\nEducation al Testing Service.\n\nShe holds a B.A. degree in psychology from Wesleyan University\, and\nM.A. and Ph.D. in clinical psychology from the Universi ty of Michigan\nas well as an M.A. in Religious Studies from Hartford Semi nary. Over\nthe course of her career\, she has served as a faculty member at the\nUniversity of California\, Santa Barbara\, Westfield State Univers ity\,\nand Mount Holyoke College. Prior to her 2002 appointment as preside nt\nof Spelman\, she served as dean and acting president at Mount Holyoke\ nCollege. In Spring 2017 she was the Mimi and Peter E. Haas\nDistinguished Visitor at Stanford University. She is married to Dr.\nTravis Tatum\; the y are the parents of two adult sons.\n\nRSVP\n\n→ On-site registration is available as space permits. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91436d4 DTSTART;TZID=America/New_York:20260326T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260326T180000 URL:https://csd.cmu.edu/academics/doctoral/phd-open-house SUMMARY:Admitted Ph.D. Student Open House (Day 1) CLASS:PUBLIC DESCRIPTION:Talk Title: CSD Ph.D. Open House\n\nThursday\, March 26\, 9am  – Friday\, March 27\, 2026\n\nWe hope you will join us for our Computer Science Department Doctoral\nAdmitted Student Open House. There are inform ational group events and\none-on-one meetings with faculty and doctoral st udents\, along with\nother programming\, to help you evaluate and understa nd our\nactivities. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9143978 DTSTART;TZID=America/New_York:20260508T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260508T133000 URL:https://www.cs.cmu.edu/commencement SUMMARY:SCS Commencement 2026 - Master's Diploma Ceremony CLASS:PUBLIC DESCRIPTION:Talk Title: SCS Commencement 2026 - Master's Diploma Ceremony\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9143be2 DTSTART;TZID=America/New_York:20260509T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260509T140000 URL:https://www.cs.cmu.edu/commencement LOCATION:Carnegie Music Hall SUMMARY:SCS Commencement 2026 - Ph.D. Hooding and Diploma Ceremony CLASS:PUBLIC DESCRIPTION:Talk Title: SCS Commencement 2026 - Ph.D. Hooding and Diploma C eremony\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9143e98 DTSTART;TZID=America/New_York:20251212T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251212T163000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:MSCS Thesis Presentation - Shiqi Pan CLASS:PUBLIC DESCRIPTION:Speaker: SHIQI PAN\, Master's Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Architectural Disaggrega tion for LLM Serving on\nHeterogeneous Device: An Analytical Framework and Serving System\n\nModern large language models contain operations with va stly different\ncomputational characteristics: projections and MLPs are co mpute-bound\,\nwhile attention mechanisms are memory-bound. Hybrid archite ctures\ncombining sliding window attention\, linear attention\, and Mixtur e of\nExperts further complicate this operational heterogeneity. Meanwhile \,\ndatacenters deploy heterogeneous GPUs with complementary\nprofiles—H 100s excel at compute-intensive workloads while H20s\nbetter serve memory- bound operations. This creates opportunities for\noperation-level disaggre gation: matching different operations to\nspecialized hardware.\n\nHowever \, two critical gaps prevent realizing these opportunities.\nFirst\, no fr amework systematically characterizes how hybrid LLM\noperations perform on heterogeneous hardware. Second\, current serving\nsystems use rigid layer -granularity pipeline parallelism\, preventing\nspecialized placement of i ndividual operations.\n\nThis thesis addresses both gaps. We develop quant itative performance\nmodels characterizing operation-level costs\, arithme tic intensity\, and\nbottlenecks for attention variants\, MLP\, and MoE op erations\,\ndemonstrating the motivation for architectural disaggregated\n placement. Additionally\, we design and implement a flexible system\nexten ding vLLM that supports arbitrary operation-level stage\ndefinitions and n on-contiguous patterns through multi-visit execution\,\nmetadata caching\, zero-copy tensor transmission\, and tensor reordering\nfor FlashAttention compatibility.\n\nThis work provides the analytical foundation and system infrastructure\nfor operation-aware heterogeneous LLM serving\, enabling future\nresearch in automated configuration and deployment optimization.\n \nThesis Committee\n\nRashmi K. Vinayak (Chair)\n\nZhihao Jia\n\nAdditiona l Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9144333 DTSTART;TZID=America/New_York:20260205T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260205T180000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Distinguished Lecture: Alan J. Perlis Fun Lecture in Computer\n Science CLASS:PUBLIC DESCRIPTION:Speaker: STEPHEN WOLFRAM - To Be Rescheduled\, Founder and Chie f\nExecutive Officer\nWolfram Research\n\nTalk Title: Making the World Com putable: from Foundations to\nTechnology (for Humans and AIs)\n\nStephen W olfram is the creator of Mathematica\, a scientific computing\nplatform us ed by thousands of engineers and scientists worldwide. He\nis the CEO of W olfram Research\, a software and technology consulting\ncompany that serve s industry\, government\, and academia. Wolfram holds\na PhD in physics fr om the California Institute of Technology and is a\nMacArthur Fellow. He l ast visited CMU over 30 years ago. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914463a DTSTART;TZID=America/New_York:20251215T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251215T143000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Wan Shen Lim CLASS:PUBLIC DESCRIPTION:Speaker: WAN SHEN LIM\, Ph.D. Candidate\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Database Gyms: Towards Autonomous Database Tuning\n\nDatabase management systems (DBMSs) are the foundation of modern\ndata-intensive applications. But as more features a re developed to\nsupport new workloads\, they become increasingly complex and difficult\nto configure. Thus\, researchers have invested decades of e ffort into\nautonomous DBMS configuration. Recent advances in machine lear ning\n(ML) have produced tools that outperform unassisted experts in\nreal -world deployments. However\, these tools are advisory and require\nhuman expertise for deployment into database tuning pipelines.\n\nUsing these to ols involves a multi-step process where a human operator\n(1) determines a n optimization objective\, (2) selects a suitable tool\,\n(3) sets up the DBMS\, (4) runs a workload to collect telemetry\, (5)\nuses the telemetry to calibrate the tool\, and (6) operates the tool to\nobtain recommendatio ns\, which the operator must then review and apply.\nThese ad-hoc pipeline s require significant human effort to set up\,\nextend\, and deploy. Moreo ver\, interface differences make tools\ndifficult to compose and interchan ge. Thus\, despite the demonstrated\nability of database tuning tools to i mprove performance and lower\ncosts\, the expertise required to operate th em limits their adoption.\n\nThis dissertation presents the database gym\, an integrated framework\nthat systematizes and automates the DBMS configu ration pipeline.\nUnlike prior research that focused on improving tool eff ectiveness\nwith ML\, the gym targets deployment and operational challenge s by\nproviding reusable\, interoperable\, and interchangeable components that\nsimplify tool development and integration.\n\nThe gym’s design ref lects the observation that the bottleneck in\ndatabase tuning has shifted from developing better algorithms for\ntools to acquiring the training dat a needed to operate them. We\ndemonstrate how the gym's architecture accel erates and adapts\ntool-based database tuning pipelines through the system atic generation\nand utilization of training data\, enabling the augmentat ion and\norchestration of tools with end-to-end knowledge. For example\, i t\nreduces step-level overhead by skipping redundant computation during\nt elemetry generation\, thus reducing the tuning pipeline's latency. It\nals o eliminates pipeline-level repetition by reusing training data to\nadapt a tool's calibrated models across new software versions and\nhardware envi ronments. Such optimizations are enabled by the gym’s\nholistic control over the entire tuning process.\n\nThesis Committee\n\nAndrew Pavlo (Chair )\n\nJignesh Patel\n\nDavid Andersen\n\nLin Ma (University of Michigan)\n\ nIn Person and Zoom Participation. See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9144b27 DTSTART;TZID=America/New_York:20251209T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251209T133000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:5th Year Master's Thesis Presentation - Mihir Khare CLASS:PUBLIC DESCRIPTION:Speaker: MIHIR KHARE\, Master's Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Filters are not One-Si ze-Fits-All: Survey and Evaluation\nof Join Pre-Filtering in Database Mana gement Systems\n\nThe current era of data storage is defined by the widesp read adoption\nof data lakes\, and the disaggregation of storage and compu te hardware.\nModern database management systems (DBMSs) are often operati ng on\nlarge volumes of data stored in object stores (like Amazon's S3)\,\ nunmanaged file formats (like Apache's Parquet)\, or otherwise have\noutda ted or nonexistent statistics. In join-heavy analytical\nworkloads\, the t raditional approach of optimizing query plans to\nminimize the cost of joi ns breaks down if the available information to\nestimate cardinalities and costs is inaccurate. In recent years\, a\nclass of techniques known as `` join pre-filtering'' has gained focus\nas an attempt to reduce the relianc e on a good optimizer for\nminimizing join costs by reducing the inputs of joins to the minimum\nset of tuples needed to produce the output. This th esis explores the\ncurrent state-of-the-art in pre-filtering\, and conclud es that more\nwork must be done to create a strategy that is applicable ac ross a\nwide range of workloads. First\, we provide an overview of the\nfu ndamental concepts of pre-filtering\, and describe the key design\ndecisio ns of and differences between currently studied techniques. We\nthen evalu ate two pre-filtering methods\, min-max filters and Bloom\nfilters\, using a modified implementation of Dynamic Predicate Transfer\n(RPT+)\, a leadi ng contemporary technique for join pre-filtering. Our\nanalysis focuses on the performance interactions between these filter\ntypes. Finally\, we di scuss the implications of the results for modern\nDBMSs\, and future direc tions of study in this area.\n\nThesis Committee\n\nJignesh Patel (Chair)\ n\nAndy Pavlo\n\nAdditional Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9144f34 DTSTART;TZID=America/New_York:20251207T193000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251207T210000 LOCATION:WQED Studio A SUMMARY:Electronic Music - Senior Studio Concert CLASS:PUBLIC DESCRIPTION:Talk Title: Electronic Music - Senior Studio Concert\n\nElectro nic Music Studio Concert - Senior Concert\n\nProgram\n\nKatherine Wang → With You Katherine Wang → L’appel du Vide\nL’appel du Vide Alen Z hu → GestationAllen Zhu → Reflection\nFireMartin Baker → FEEL IT IN MY CORE [part one]Liam Neely → The\nReactive Room Demo \n\nNo tickets r equired. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914524a DTSTART;TZID=America/New_York:20251206T190000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251206T203000 LOCATION:WQED Studio A\, 4802 Fifth Avenue SUMMARY:Studio Concert - Electronic Music (Years 1-3) CLASS:PUBLIC DESCRIPTION:Talk Title: Studio Concert - Electronic Music\n\nElectronic Mus ic Studio Concert (Years 1-3)\n\nProgram\n\nOwen Libby → Pale Blue DotQ uin Kondis → Statue FluctuationsSean\nXue  → Song 1 - Very Cool St udio PieceStella Cedar → ill the\nSacrosanctCharles Chen → When It ’s AllEmma\nDiprima →  FacelessNoah Martin → Think/shadowJasmin e\nPapino-Wood → it hrtsCharles Chen → SegmentRowan Sullivan\n→ B laspheme\n\nNo tickets required. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9145536 DTSTART;TZID=America/New_York:20251208T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251208T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Jark Wu CLASS:PUBLIC DESCRIPTION:Speaker: JARK WU\, Original Creator\, Apache Fluss\nLead\, Flin k SQL and Fluss Teams\nAlibaba Cloud\n\nTalk Title: Apache Fluss: A Stream ing Storage for Real-Time Lakehouse\n\nMonday\, December 8\, 2025\, 4:30  – 5:30pm \n\nModern data lakehouses promise unified batch and streamin g processing\,\nyet their storage layer remains inherently batch-oriented —optimized\nfor large\, immutable files. This mismatch forces streaming workloads\nto rely on external systems (e.g.\, Kafka)\, while analytical q ueries\noperate on stale snapshots\, breaking end-to-end freshness.\n\nIn this talk\, I’ll present Apache Fluss (incubating)\, a\nlakehouse-native streaming storage system designed to bridge this gap.\nFluss rethinks str eaming storage from the ground up for analytical\nworkloads. Its core abst raction is a columnar stream built on Apache\nArrow\, enabling sub-second ingestion and high-hroughput analytical\nscans. Furthermore\, Fluss introd uces the \"Streaming Lakehouse\" concept\nthat Fluss serves as the real-ti me data layer on top of Lakehouse. It\nallows query engines to seamlessly unify both fresh streaming data in\nFluss and historical data in Lakehouse (Iceberg) to achieve truly\nreal-time data\n\nanalytics.\n\n—\n\nJark W u is the original creator of Apache Fluss and PMC member of\nApache Flink. He currently leads the Flink SQL (streaming compute) and\nFluss (streamin g storage) teams at Alibaba Cloud\, where he is\ndedicated to building a s erverless Flink cloud service. His work\nfocuses on data streaming systems for over a decade. \n\nThis talk is part of the Future Data Systems Semi nar Series.\n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9145929 DTSTART;TZID=America/New_York:20251205T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251205T113000 LOCATION:Wean Hall 5316 and Remote Call SUMMARY:Master of Science in Information Networking Thesis Presentation -\n Shyamal Vaderia CLASS:PUBLIC DESCRIPTION:Speaker: SHYMAL VADERIA\, Master's Student\, Information Networ king\nInstitute\, Carnegie Mellon University\n\nTalk Title: Cache-Efficien t Bloom Filters: Designs for 2035's Memory\nHierarchies\n\nBloom filters a re widely used in modern data systems to accelerate\nmembership tests in d atabases\, storage engines\, and networking stacks.\nOn modern CPUs\, Bloo m filter performance is typically limited by\nmemory latency rather than c omputation: each query issues K seemingly\nrandom probes over a bit array that is often much larger than the\nlast-level cache (LLC). Most prior wor k on making Bloom filters faster\nhas focused on changing the data structu re itself. We take a\ncomplementary approach and ask how far one can go by changing only the\nquery strategy.\n\nFirst\, we show that reordering the K probes for each query--either by\nfully sorting them (cache-oblivious) or by partitioning them around a\ncache-sized pivot (cache-aware)--transfo rms the aggregate access\ndistribution from uniform to strongly skewed tow ard low indices\,\ngreatly improving cache reuse without modifying the Blo om filter\nrepresentation. Second\, we refine the hash generation process to\nbehave more like a low-discrepancy sequence over the filter\, reducing \nthe fraction of queries whose probes all fall outside the LLC-sized\nreg ion.\n\nAcross a range of filter sizes for negative-query workloads\, our\ ndesigns reduce LLC-load misses by up to 8.3x and improve end-to-end\nrunt ime by up to 1.45x relative to a conventional Bloom filter\, while\nremain ing competitive with specialized cache-aware structures such as\nsplit-blo ck Bloom filters.\n\nThesis Committee\n\nDave Andersen (Chair)\n\nMichael Kaminsky\n\nIn Person and Video Call\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9145d43 DTSTART;TZID=America/New_York:20251204T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251204T173000 LOCATION:Remote Access - Zoom SUMMARY:Crypto Seminar - Yu Wei CLASS:PUBLIC DESCRIPTION:Speaker: YU WEI\, Ph.D. StudentCollege of Computing\, Georgia I nstitute\nof Technology\n\nTalk Title: Black-box Estimation of Differentia l Privacy in Machine\nLearning\n\nPrivate data publication with provable p rivacy guarantees\, such as\nDifferential Privacy (DP)\, becomes pressing as ML algorithms\nincreasingly deployed on sensitive data. Yet many widely used\nrandomized algorithms were designed before privacy became a central \nconcern\, and too complex or nonstandard for existing DP tools\, leaving \ntheir deployed privacy guarantees largely unknown. In this talk\, we\nwi ll discuss how to address this challenge by developing tools for\nanalyzin g DP and its variants in general randomized algorithms.\n\nFirst\, I will present a general black-box framework for estimating the\nDP guarantees of arbitrary randomized algorithms without requiring\naccess to their intern al code or algorithmic description. The idea is\nto recast privacy estimat ion as a binary classification problem\, which\nyields tight estimates of privacy parameters together with\nfinite-sample theoretical guarantees.\n\ nSecond\, I extend this approach to f-differential privacy\, a\nhypothesis -testing–based refinement of DP\, and add an explicit\nauditing layer on top. I will show how this framework can audit DP-SGD\nimplementations wit h comparable tightness and computational cost to\nprior methods\, while re quiring strictly less information about the\nunderlying training procedure .\n\nWe conclude with applications of our black-box estimators: recovering \nthe privacy profiles of nonstandard randomized ML methods\, such as\nran dom projection and approximate least squares\, and showing how\,\ncombined with analytical results\, they can be upgraded into DP\nmechanisms with m inimal modification.\n\n—\n\nYu Wei is a fifth year PhD student at the G eorgia Institute of\nTechnology\, advised by Professor Vassilis Zikas. He has a broad\nresearch interest in cryptography\, privacy\, and machine lea rning. His\nrecent research focuses on differential privacy and related to pics in\nsecure computation and game-theoretic cryptography. His work has\ nappeared in venues such as IEEE S&P\, USENIX Security\, and TCC.\n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91461c9 DTSTART;TZID=America/New_York:20251208T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251208T140000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Thesis Proposal - Myra Dotzel CLASS:PUBLIC DESCRIPTION:Speaker: MYRA DOTZEL\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Logical Foundations of Inter mittent Computing\n\nIntermittent computing is gaining popularity in appli cations that rely\non batteryless energy-harvesting devices which experien ce frequent and\narbitrary power failures. To enable correct program re-ex ecution\ndespite these potentially frequent and arbitrary power failures\, \nruntime support is needed to save and restore necessary state. \n\nIn t his talk\, we study the formal foundations of intermittent\ncomputing by u se of type systems to guarantee the correctness of\nprograms prior to thei r deployment\, and runtime systems to facilitate\ncorrect program executio n\, including support for sequential and\nconcurrent models of execution.  \n\nFirst\, we explore the logical underpinning of sequential\, intermit tent\ncomputing and model checkpoint\, crash\, restore\, and re-execution\ noperations as computation on crash types. We draw inspiration from\nadjoi nt logic to reason about the relationship between persistent and\ntransien t memories through (re-)execution\, checkpointing\, and\nrestoration. Usin g crash types\, we show that any correct intermittent\nexecution can be si mulated by a continuously-powered execution.\n\nSecond\, we present the fi rst provably-correct system for concurrent\,\nintermittent program executi on\, which is needed as many embedded\napplications rely on interactions w ith hardware-triggered interrupts\nand accesses to shared memory. Prior wo rk on concurrent\, intermittent\nexecution has only provided restrictive p rogramming models with no\nformal correctness guarantees. In this talk\, w e present a co-designed\nruntime system and type system that together supp ort the provably\ncorrect intermittent execution of task-based concurrent programs with\nshared memory. This system promotes a more flexible program ming model\nand supports a broader spectrum of task re-execution behaviors than is\nconsidered by prior work. We provide the first formal definition and\nproofs of correctness for concurrent\, intermittent execution.    \n\nThesis Committee\n\nLimin Jia (Co-chair)\n\nFarzaneh Derakhshan (Co-ch air\, Illinois Institute of Technology)\n\nFrank Pfenning \n\nJan Hoffman n \n\nBrigitte Pientka (McGill University) \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914667d DTSTART;TZID=America/New_York:20251205T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251205T120000 LOCATION:Gates Hillman 9115 SUMMARY:Theory Seminar - Prasanna Ramakrishnan CLASS:PUBLIC DESCRIPTION:Speaker: PRASANNA RAMAKRISHNAN\, Ph.D. Student\, Theory Group\, Computer\nScience Department\, Stanford University\n\nTalk Title: How to Appease a Voter Majority\n\nIn 1785\, Condorcet established a frustrating property of elections and\nmajority rule: it is possible that\, no matter which candidate you pick\nas the winner\, a majority of voters will prefer someone else. You\nmight have the brilliant idea of picking a small set o f winners\ninstead of just one\, but how do you avoid the nightmare scenar io where\na majority of the voters prefer some other candidate over all th e ones\nyou picked? How many candidates suffice to appease a majority of t he\nvoters? In this talk\, we will explore this question. Along the way\, we\nwill roll some dice — both because the analysis involves randomness\ nand because of a connection to the curious phenomenon of intransitive\ndi ce\, that has delighted recreational and professional mathematicians\nalik e ever since Martin Gardner popularized it in 1970.\n\nBased on joint work with Moses Charikar\, Alexandra Lassota\, Adrian\nVetta\, and Kangning W ang \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91469f3 DTSTART;TZID=America/New_York:20251215T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251215T163000 LOCATION:Newell-Simon 3305 (New Date) SUMMARY:VASC Seminar - Simon Lucey CLASS:PUBLIC DESCRIPTION:Speaker: SIMON LUCEY\, Professor\nSchool of Computer and Mathem atical Sciences\, and\nDirector\, Australian Institute for Machine Learnin g\nUniversity of Adelaide\n\nTalk Title: Should we skip attention?\n\nTran sformers are ubiquitous. They influence nearly every aspect of\nmodern AI. However\, the mechanics of their training remain poorly\nunderstood. This poses a problem for the field due to the immense\namounts of data\, compu tational power\, and energy being invested in the\ntraining of these netwo rks. I highlight a recent intriguing empirical\nresult from our group. Spe cifically\, self attention catastrophically\nfails to train unless it is p aired with a skip connection. This\ncontrasts with other components of a t ransformer that continue to\ndemonstrate good performance (albeit suboptim al) when skip connections\nare removed. In this talk\, I explore why this is the case and what\ncould be done to enhance the fundamental training ef ficiency of modern\ntransformers. We even showcase some practical cases in which removing\nself-attention completely can lead to significantly impro ved\nperformance.    \n\n—\n\nSimon Lucey\, Ph.D.\,  is the Director of the Australian Institute for\nMachine Learning (AIML) and a professor in the School of Computer and\nMathematical Sciences\, at the University o f Adelaide. He is also\nDirector of the CommBank Foundational AI Research Centre. Prior to\nthis he was an associate research professor at Carnegie Mellon\nUniversity's Robotics Institute (RI) in Pittsburgh USA\; where he spent\nover 10 years as an academic. He was also Principal Research Scient ist\nat the autonomous vehicle company Argo AI from 2017-2022. He has\nrec eived various career awards\, notably the AmCham AI Scientist of the\nyear in 2024. He is also currently a member of the Australian\nGovernment’s AI Expert Group\, and their National Robotics Strategy\ncommittee. Simon ’s research interests span AI\, machine learning\,\ncomputer vision and robotics.\n\nThe VASC seminar is generously sponsored by HeyGen\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9146e19 DTSTART;TZID=America/New_York:20251205T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251205T200000 LOCATION:Simmons Auditorium A\, Tepper Building 1101A SUMMARY:Game Creation Society: End of Semester Reslease CLASS:PUBLIC DESCRIPTION:Talk Title: Computer Game Programming - Game Studio: Interactiv ity\n\nFriday\, December 5\, 2025\, 5 – 8pm \n\nRelease showcases gam es and interactive experiences created by the\nmajority of CMU's courses a nd clubs this semester.\n\nSIGN-UP\n\n→ Food for those registered! \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91470f6 DTSTART;TZID=America/New_York:20251208T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251208T163000 LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Eliahu Horwitz CLASS:PUBLIC DESCRIPTION:Speaker: ELIAHU HORWITZ\, Google Ph.D. Fellow in Machine Learni ng and\nML Foundations\nand\, Ph.D. Candidate in Computer Science\nHebrew University of Jerusalem\n\nTalk Title: What Can We Learn from a Million Mo dels?\n\nMachine learning has transformed many fields by learning from lar ge\ncollections of data. Yet\, it is rarely applied to its own outputs: th e\nmodels themselves. Today\, with millions of publicly available models\, \na natural question arises: what can we do with so many models? In this\n talk\, I will motivate two core applications that leverage this\nuntapped potential\, demonstrating their utility in the context of\ncomputer vision : (i) identifying emerging trends in model design\, and\n(ii) reducing the need to train models from scratch through model\nrecycling. To support th ese goals\, I introduce the Model Atlas: a\nstructured graph that represen ts models\, their attributes\, and the\nweight-space transformations that interconnect them. My research into\nweight-space learning enables the con struction of this atlas by\ntreating models themselves as data and inferri ng properties such as\nfunctionality\, performance\, and lineage directly from their weights. I\nwill present key observations and methodologies tha t make weight-space\nlearning possible at scale. As a visual prelude\, you can explore the\nrepository under study.\n\n—\n\nEliahu Horwitz is a Go ogle PhD Fellow in Machine Learning and ML\nFoundations and a final-year P hD candidate in Computer Science at The\nHebrew University of Jerusalem\, advised by Prof. Yedid Hoshen. His\nresearch centers on learning represent ations of neural network weights\nand understanding model populations dire ctly in weight space. He is\nparticularly interested in how weight-space l earning can enable new\ndownstream capabilities\, such as model forensics\ , model discovery\, and\ninterpretability\, and in how treating models as data points can\nadvance broader areas of machine learning. Eliahu is also a recipient\nof the Israeli Council for Higher Education Scholarship and has\npreviously interned at Google Research.\n\n \n\nThe VASC seminar is generously sponsored by HeyGen \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914755b DTSTART;TZID=America/New_York:20251205T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251205T113000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year Master's Thesis Presentation - Samvitti Sharma CLASS:PUBLIC DESCRIPTION:Speaker: SAMVITTI SHARMA\, Master's Student\, Computer Science\ nDepartment\, Carnegie Mellon University\n\nTalk Title: SPAM Tolerance for Pauli Error Estimation\n\nThe Pauli channel is a fundamental model of noi se in quantum systems\,\nmotivating the task of Pauli error estimation. We present an algorithm\nthat builds on the reduction to Population Recovery introduced in\n[FO21]. Addressing an open question from that work\, our a lgorithm has\nthe key advantage of robustness against even severe state pr eparation\nand measurement (SPAM) errors. To tolerate SPAM\, we must analy ze\nPopulation Recovery on a combined erasure/bit-flip channel\, which\nne cessitates extending the complex analysis techniques from [PSW17\,\nDOS17] . For n-qubit channels\, our Pauli error estimation algorithm\nrequires on ly exp(n1/3) unentangled state preparations and\nmeasurements\, improving on previous SPAM-tolerant algorithms that had\n2n-dependence even for rest ricted families of Pauli channels. We also\ngive evidence that no SPAM-tol erant method can make asymptotically\nfewer than exp(n1/3) uses of the cha nnel.\n\nThesis Committee\n\nRyan O'Donnell (Chair)\n\nDavid Woodruff\n\nA dditional Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9147905 DTSTART;TZID=America/New_York:20251203T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251203T130000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Speaking Skills Talk - Nj Mukherjee CLASS:PUBLIC DESCRIPTION:Speaker: NIRJHAR 'NJ' MUKHERJEE\, Ph.D. Student\nComputer Scien ce Department\nCarnegie Mellon University\n\nTalk Title: EmuCS: An Emulato r for Computational Storage\n\nComputational Storage Devices (CSDs) are em erging as a means to reduce\ndata transfer between the host and the storag e device. Yet widespread\nadoption faces challenges due to  diverse platf orms\, diverse\nhost-side architectures and limitations of existing interf aces.\n\nIn this talk I will present a computational storage emulator\, Em uCS\,\nas both the first and second challenge can be solved by an emulator \,\nwhich is currently missing. EmuCS is easy to use and facilitates rapid \nsoftware and hardware prototyping in x86 architectures. EmuCS supports\n an interface for programming CSDs  that remove the limitations of\ncurren t interfaces. I also demonstrate how EmuCS is validated\, and how\nvarious hardware configurations and host-side support can be\nimplemented - using two use cases\, Compression and Parquet Queries.\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9147c8e DTSTART;TZID=America/New_York:20251203T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251203T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Sam Hopkins CLASS:PUBLIC DESCRIPTION:Speaker: SAM HOPKINS\, Jamieson Career Development Professor an d\nAssistant Professor\, Theory Computing GroupDepartment of Electrical\nE ngineering and Computer Science\, Massachusetts Institute of\nTechnology\n \nTalk Title: Deciding high-dimensional sub-Gaussian-ness in polynomial\nt ime\n\nGiven samples from a probability distribution\, can efficient\nalgo rithms tell whether the distribution has heavy or light tails?\nThis probl em is at the core of algorithmic statistics\, where\nalgorithms for decidi ng heavy-versus-light tailed-ness are key\nsubroutines for clustering\, le arning in the presence of adversarial\noutliers\, and more. It is easy in one dimension but challenging in\nhigh dimensions\, where a distribution c an have light tails in some\ndirections and heavy ones in others — detec ting a single direction\nwith a heavy tail hiding in an otherwise light-ta iled distribution can\nseemingly require brute-force search. In this talk\ , I describe a\nfamily of efficient algorithms for deciding whether a high -dimensional\nprobability distribution has sub-Gaussian tails\, with appli cations to\na wide range of high-dimensional learning tasks using sub-Gaus sian\ndata.\n\nBased on joint work with Ilias Diakonikolas\, Ankit Pensia\ , and Stefan\nTiegel\, appearing in STOC 2025 \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9147ff3 DTSTART;TZID=America/New_York:20251201T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251201T170000 LOCATION:Posner Hall 151 SUMMARY:Statistics and Data Science Seminar - Yandi Shen CLASS:PUBLIC DESCRIPTION:Speaker: YANDI SHEN\, Assistant Professor\, Department of Stati stics &\nData Science\, Carnegie Mellon University\n\nTalk Title: Besting Good—Turing: Optimality of Non-Parametric\nMaximum Likelihood for Distri bution Estimation\n\nWhen faced with a small sample from a large universe of possible\noutcomes\, scientists often turn to the venerable Good—Turi ng\nestimator. Despite its pedigree\, however\, this estimator comes with\ nconsiderable drawbacks\, such as the need to hand-tune smoothing\nparamet ers and the lack of a precise optimality guarantee. We\nintroduce a parame ter-free estimator that bests Good—Turing in both\ntheory and practice. Our method marries two classic ideas\, namely\nRobbins's empirical Bayes a nd Kiefer--Wolfowitz non-parametric maximum\nlikelihood estimation (NPMLE) \, to learn an implicit prior from data\nand then convert it into probabil ity estimates. We prove that the\nresulting estimator attains the optimal instance-wise risk up to\nlogarithmic factors in the competitive framework of Orlitsky and\nSuresh\, and that the Good—Turing estimator is strictl y suboptimal in\nthe same framework. Our simulations on  synthetic data a nd\nexperiments with English corpora and U.S. Census data show that our\ne stimator consistently outperforms both the Good--Turing estimator and\nexp licit Bayes procedures.\n\n—\n\nYandi Shen is an Assistant Professor in the Department of Statistics &\nData Science at Carnegie Mellon University .  In 2021\, he obtained his\nPh.D. in Statistics at the University of Wa shington\, advised by Fang\nHan and Daniela Witten. He was then a Kruskal instructor in Statistics\nat University of Chicago from 2021 to 2023\, and a postdoctoral\nresearcher hosted by Zhou Fan at the Department of Statis tics and Data\nScience at Yale University.\n\nHe is broadly interested in nonparametric and semiparametric\nstatistics\, high dimensional inference\ , and applied probability. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91483ef DTSTART;TZID=America/New_York:20251202T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251202T110000 LOCATION:Gates Hillman 8102 SUMMARY:5th Year Master's Thesis Presentation - Tianyue Ruby Sun CLASS:PUBLIC DESCRIPTION:Speaker: TIANYUE RUBY SUN\, Master's Student\nComputer Science Department\nCarnegie Mellon University\n\nTalk Title: Remote Photoplethysm ography: Spatiotemporal Architecture\n\nRemote photoplethysmography (rPPG) enables contactless measurement of\nphysiological signals such as heart r ate and respiratory rate from\nvideos\, offering a practical alternative t o traditional contact-based\nsensor measurements. Recent deep learning met hods have achieved strong\nrPPG accuracy\, but these approaches often depe nd on controlled\nsettings and struggle to generalize to real-world enviro nments with\nmotion and varying lighting. These limitations are in part du e to the\nreliance on techniques such as manual parameter tuning and the n eed\nfor large labelled datasets that are often captured under clean\ncond itions.\n\nThis research thesis presents an exploration of the end-to-end rPPG\npipeline. The primary contribution is a novel spatiotemporal\narchit ecture for rPPG that combines DINOv2\, a vision transformer\, and\nChronos \, a time series model. This represents the first multimodal\nrPPG framewo rk that leverages a combination of spatial and temporal\nrepresentations f rom foundation models for physiological measurement.\nThe two foundation m odels are kept frozen\, and a lightweight\nprediction head is trained.\n\n The proposed model achieves strong performance on the synthetic SCAMPS\nda taset for heart rate estimation\, establishing benchmarks for future\nrPPG research. On real-world datasets\, including PURE and UBFC-rPPG\,\nthe mo del demonstrates effective learning of blood volume pulse (BVP)\nwaveforms and heart rate estimation\, despite the increased errors\nreflecting the difficulty of more challenging conditions. Extensions\nof the model to res piratory rate illustrate the generalizability of\nthe architecture across different physiological measurement tasks.\nOverall\, the results show tha t foundation models can improve rPPG\nrobustness and generalization\, offe ring a promising path towards\npractical rPPG systems with applications in inpatient monitoring\,\ntelehealth\, and emergency response.\n\nIn additi on to model development\, this thesis analyzes components of\nthe full rPP G pipeline\, including signal processing and ground truth\nextraction. It is shown that common signal processing methods applied\nto the same BVP si gnal can lead to discrepancies in the estimation of\nthe scalar heart rate value. Moreover\, the method of obtaining the\nground truth from data aff ects the reported performances. These\ninsights motivate the need to furth er discuss reliable signal\nprocessing and evaluation procedures to ensure reliable comparisons\nand interpretations of rPPG methods.\n\nThesis Comm ittee\n\nArtur Dubrawski (Chair)\n\nLaszlo Jeni\n\nAdditional Information  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914893e DTSTART;TZID=America/New_York:20251210T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251210T140000 LOCATION:4300 and 4400 Corridors - 4th Floor Gates Hillman SUMMARY:SCS Fall Undergraduate Thesis & Independent Study Poster Session &\ nResearch Presentations CLASS:PUBLIC DESCRIPTION:Stop in to learn more about our students' research and on-going \nwork.  Posters and presentations will be plentiful.  Bring your\nquest ions!\n\nNote: If you are interested in getting involved in research\, one way\nto start is to see what your fellow students are working on.  The\n students can explain their work and how they got involved with the\nresear ch.  Faculty will also be there\, so there is also the ability\nto talk w ith prospective research advisors.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9148c8b DTSTART;TZID=America/New_York:20251202T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251202T113000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Thesis Proposal - Kaiyang Zhao CLASS:PUBLIC DESCRIPTION:Speaker: KAIYANG ZHAO\, Ph.D. Student\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: Architecting Memory Efficie ncy in Modern Datacenters\n\nThe proliferation of memory-intensive applica tions\, the rapid\nexpansion of memory capacity to terabyte scales\, and t he slowing of\nDRAM cost scaling have established memory as the critical b ottleneck\nin modern datacenter computing. This inefficiency manifests in two\ndimensions: the cycle efficiency lost to the virtual memory\nabstract ion and the escalating financial cost of memory.\n\nFirst\, the virtual me mory abstraction is under increasing strain. As\nmemory capacity grows whi le Translation Lookaside Buffer sizes remain\nstagnant\, address translati on overhead becomes severe\; internal\nprofiling at hyperscalers reveals t hat approximately 20% of CPU cycles\nare stalled on TLB misses. This overh ead is bound to worsen due to\ninherent TLB scaling limits\, the introduct ion of additional page table\nlevels\, vast heterogeneous memory capacity\ , and the page-level\nsecurity checks required by confidential computing. Second\, the\nfinancial cost of memory has skyrocketed. Memory now account s for\nnearly a quarter of rack power consumption and half of the Total Co st\nof Ownership of a typical datacenter server. In this proposal\, I\nadd ress these challenges through a set of operating system and\narchitectural designs.\n\nTo improve cycle efficiency\, I present two completed works.\ nContiguitas creates abundant physical memory contiguity by grouping\nunmo vable allocations in the OS and introducing hardware extensions to\nmigrat e pages previously locked for device I/O. This contiguity is\nleveraged to allocate huge pages\, reducing translation overhead and\nyielding up to 1 8% performance improvement for production workloads.\nLearned Virtual Memo ry (LVM) replaces rigid radix page tables with\nlearned indexes tailored t o the application's virtual address space.\nBy leveraging address space re gularity\, LVM reduces page walk overhead\nby an average of 44% and achiev es a 2–27% speedup in execution.\n\nTo improve cost efficiency\, I prese nt two ongoing works. Multi-Tier\ndynamically manages pages across DRAM\, CXL memory\, and SSDs to\nmaximize financial savings by utilizing cheaper tiers within a defined\nperformance loss limit. Equilibria addresses the c hallenges of\nmulti-tenant tiering\, ensuring fair memory sharing and miti gating\nnoisy-neighbor interference through flexible placement policies an d\nthrashing mitigation.\n\nTogether\, these works provide a comprehensive solution to improving\nmemory efficiency in datacenters from the perspect ives of both cycle\noverhead and financial cost.\n\nThesis Committee\n\nDi mitrios Skarlatos (Chair)\n\nPhil Gibbons\n\nTodd Mowry \n\nKim Keeton (G oogle) \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91491d0 DTSTART;TZID=America/New_York:20251202T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251202T163000 LOCATION:ASA Conference Room\, Gates HIllman 6115 SUMMARY:5th Master's Thesis Presentation - Yuchen Liang CLASS:PUBLIC DESCRIPTION:Speaker: YUCHEN LIANG\, Master's Student\nComputer Science Depa rtment\nCarnegie Mellon University\n\nTalk Title: Feedback-Driven Query Op timization: Design and\nInfrastructure\n\nQuery optimizers are critical co mponents in database management\nsystems (DBMSs) that turn a query that mi ght otherwise take hours to\nrun into one that completes in seconds. Howev er\, modern data stacks\nallow applications to generate data files (e.g.\, Parquet) outside the\nDBMS’s purview that lack statistical summaries. W ithout information\nabout data distribution\, optimizers fall back to ungr ounded guesses\nwhen computing cardinality and cost estimates needed to se lect the\nbest plan from an exponential number of candidates. \n\nThis th esis investigates the design of an adaptive query optimization\nstrategy t hat leverages query feedback to improve planning for future\nqueries. We f irst study how an optimizer should memoize query plans\nduring optimizatio n so that it can associate past traces with future\nqueries. To facilitate this investigation\, we built a Cascades-style\ntransformational search e ngine with a memo table that organizes\nequivalent sub-plans and detects d uplicates. We measured the amount of\nreuse when maintaining the optimizat ion state across queries. Our\nresults showed that future queries can reus e the cached optimization\ninformation from previous queries when they exh ibit similar subplan\npatterns. We then analyze runtime feedback collectio n methods\,\nfocusing on integration strategies and the resulting artifact s. \n\nOur study reveals a tension between the query coverage provided by \nruntime artifacts and the collection overhead they entail. Although\nmos t DBMSs can provide runtime row-count profiles at the operator\nlevel\, th e optimizer can only use this information for future queries\nthat match t he exact predicates. To address this limitation\, we\npresent a method tha t measures the true selectivity of each prefix in\nthe conjunctive predica te during execution\, which the optimizer can\nthen leverage to improve pl anning for a broader range of queries. We\nimplemented this method in Post greSQL and measured the runtime\nperformance overhead. Our experiments sho w that fine-grained\ninstrumentation incurs a constant overhead regardless of the number of\nconjuncts. We conclude the thesis by discussing the des ign choices and\ntrade-offs of integrating an extensible query optimizer s ervice with\nexisting DBMSs.\n\nThesis Committee\n\nAndy Pavlo (Chair)\n\n JIgnesh Patel\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9149671 DTSTART;TZID=America/New_York:20251201T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251201T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Prashant Singh CLASS:PUBLIC DESCRIPTION:Speaker: PRASHANT SINGH\, Senior Software Engineer\, Snowflake\ n\nTalk Title: From Storage Formats to Open Governance: The Evolution to\n Apache Polaris\n\nAs organizations build their data lakehouses on Apache I ceberg\, the\nprimary challenge shifts from managing individual files to\n orchestrating a cohesive ecosystem of tables. How can you guarantee\nconsi stency and enable complex operations when multiple data\nengines—like Sp ark\, Trino\, and Flink—need to interact with the\nsame data concurrentl y?\n\nThe answer lies in a standardized service layer\, defined by the\nIc eberg REST Catalog specification. This talk explores how this open\nAPI ac ts as the blueprint for true interoperability\, enabling smarter\nmulti-wr iter conflict resolution and unlocking powerful new\ncapabilities\, such a s enabling complex multi-statement and multi-table\ntransactions.\n\nWe th en introduce Apache Polaris\, a complete\, open-source\nimplementation tha t brings this vision of open governance to life.\nWe'll demonstrate how Po laris moves beyond basic metadata management\nto provide a robust\, centra lized control plane. Join us to learn how\nPolaris delivers essential feat ures like Role-Based Access Control\n(RBAC)\, integration with external Id entity Providers\, and intelligent\nconflict resolution\, policy managemen t establishing itself as the\ncentral nervous system for a secure\, multi- engine\, and\ntransactionally-consistent Iceberg data lakehouse.\n\n—\n\ nPrashant Singh is a Senior Software Engineer @ Snowflake\, Committer @\nA pache Iceberg™\, Apache Polaris™ (incubating) and is an active\ncontri butor to both the projects. \n\nThis talk is part of the Future Data Syst ems Seminar Series.\n\nZoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9149ab6 DTSTART;TZID=America/New_York:20251204T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251204T160000 LOCATION:CNA Room\, Wean Hall 7218 SUMMARY:ACO Seminar - Maya Sankar CLASS:PUBLIC DESCRIPTION:Speaker: MAYA SANKAR\, Institute for Advanced Study\n\nTalk Tit le: The Discrete Fundamental Group\n\nThe discrete fundamental group pi1(G ) of a graph G is an object\ninspired by the fundamental group of a topolo gical space. I will\ndefine this group and present two results that use pi 1(G) in very\ndifferent ways. First\, we show that no Cayley graph over (Z /2Z)m x\n(Z/4z)n can have chromatic number 3. Our proof is purely combinat orial\nyet generalizes the k=1 case of a pivotal result of Lovász\, which \nstates that a graph G has chromatic number at least k+3 if an\nassociate d topological space is k-connected.\n\nSecond\, I will discuss an applicat ion to the homomorphism thresholds\nof odd cycles. For r ≥ 2\, consider a family F of C2r+1-free graphs\,\neach having minimum degree linear in it s number of vertices. Such a\nfamily is known to have bounded chromatic nu mber\; equivalently\, each\ngraph in F is homomorphic to a complete graph of bounded size. We\ndisprove the analogous statement for homomorphic imag es that are\nthemselves C2r+1-free. The counterexample arises from a famil y of\ngraphs on high-dimensional spheres\, and the analysis relies on the\ ndiscrete fundamental group in a crucial way.\n\n—\n\nMaya Sankar studie s extremal graph theory\, with a particular focus on\nconnections to topol ogical combinatorics. She aims to continue\ndeveloping graph-theoretic ana logues of tools from algebraic topology\nand identifying further applicati ons for parameters that have already\nbeen introduced.\n\nThe first result is joint with Mike Krebs. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9149eff DTSTART;TZID=America/New_York:20251202T070000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251202T235500 SUMMARY:CMU Giving Day CLASS:PUBLIC DESCRIPTION:Speaker: Giving is Powerful\, especially when we do it together !\n\nBe one of those to do something extraordinary for CMU.\n\nOn December 2\, for 24 hours only\, your gift goes above and beyond. \n \n\nThe 11t h annual Giving CMU Day — where Tartans everywhere show\nsupport for Car negie Mellon’s people\, places and programs — your\ngenerosity is part of a global impact. \n\n#GivingCMUDay \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914a1f3 DTSTART;TZID=America/New_York:20251205T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251205T140000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Honghao Lin CLASS:PUBLIC DESCRIPTION:Speaker: HONGHAO LIN\, Ph.D. Candidate\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: Algorithms for Massive Dat a: Optimal Bounds\, Adversarial\nRobustness\, and Data-Driven Insights\n\n With the rapid growth of massive datasets in areas such as machine\nlearni ng and numerical linear algebra\, classical algorithms are often\nno longe r feasible. In this thesis\, we develop provably efficient\nalgorithms for various problems in these settings\, such as the\nstreaming and distribut ed model. Our contributions span three\ndirections:\n\nOptimal Bounds.  W e introduce a general technique for lifting\ndimension lower bounds for re al-valued linear sketches to polynomially\nbounded integer inputs. This le ads to the first optimal sketching\nlower bounds for discrete data streams in fundamental problems such as\nfrequency moment approximation\, operato r norm estimation\, and\ncompressed sensing. Beyond this\, we also establi sh nearly-optimal\nbounds for a variety of streaming and sketching tasks\, including\nℓ p subspace sketches for constant dimension d\, ℓ p re gression\nin the arbitrary-partition distributed model\, and graph problem s such\nas approximating the minimum cut and constructing cut sparsifiers in\nbalanced directed graphs.Adversarial Robustness.  While most\nstreami ng algorithms are studied in static worst-case models\, many\npractical sc enarios involve adaptive adversaries who generate inputs\nbased on previou s outputs. We present the first adaptive attack\nagainst linear sketches f or ℓ p-estimation over turnstile integer\nstreams. Specifically\, we sh ow that any linear streaming algorithm\nwith sketching matrix A ∈ ℤrxn can be broken using only poly(r log\nn) queries\, with high constant prob ability. This result highlights\nfundamental limits of robustness in adapt ive streaming. Furthermore\,\nwe will next introduce our recent work on an adversarially robust F2\nestimation algorithm\, based on a non-linear ske tch\, that achieves\npolylogarithmic space in turnstile streams.Learning-b ased\nAlgorithms.  Classical algorithms guarantee correctness in the wors t\ncase but often ignore structure in real-world data\, while machine\nlea rning methods leverage structure but typically lack guarantees. We\ndesign learning-based algorithms that incorporate machine learning\npredictions to adapt to input distributions\, achieving faster\nruntimes\, reduced spa ce\, or improved accuracy. Crucially\, these\nalgorithms retain rigorous w orst-case guarantees even when the\npredictions are imperfect\, bridging t he gap between theory and\ndata-driven practice.  \n\nThesis Committee\n \nDavid P. Woodruff (Chair)\n\nYang P. Liu\n\nRichard Peng\n\nJelani Nelso n (University of California\, Berkeley)\n\nIn Person and Zoom Participatio n.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914a6e6 DTSTART;TZID=America/New_York:20251205T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251205T143000 LOCATION:Remote Access - Zoom SUMMARY:STAMPS Seminar - Jonathan Lilly CLASS:PUBLIC DESCRIPTION:Speaker: JONATHAN M. LILLY\, Senior Scientist\, Planetary Scien ce\nInstitute\n\nTalk Title: Local polynomial fitting on the sphere\, a ma pping solution\nfor the Earth sciences\n\nThe problem of mapping scattered data is considered from the\nperspective of the earth sciences. A particu larly promising method is\nlocal polynomial fitting\, which involves fitti ng not only a field of\ninterest\, but also its derivatives up to some spe cified order\, in the\nvicinity of each grid point. Among other desirable properties\, this\nmethod has the virtues of simplicity and ease of applic ation. Local\npolynomial fitting is adapted for use on the sphere by recas ting it in\nterms of the coordinates of a local tangent plane. Three algor ithmic\nchoices lead to substantially improved maps. Firstly\, the use of a\nvariable bandwidth\, in which the smoothing radius is not constant but\ nvaries to incorporate a fixed number of data points\, performs well\nwith irregularly spaced data. Secondly\, first- and second-order fits\nare sho wn to offer considerably improved performance compared with a\nzeroth-orde r fit due to a property known as design adaptivity. Third\,\na generalized kernel is introduced that subsumes existing forms in the\nliterature and which allows a wider degree of possibilities.  With\nthese considerations \, the problem of mapping sea surface height from\nalongtrack satellite me asurements—an important data analysis problem\nin oceanography—is cons idered. Applying the method to a numerical\nmodel\, for which errors can b e assessed directly\, a sweep through\nparameter space is conducted to ide ntify optimal parameters. The\nresults are compared with the community sta ndard product\, and sources\nof remaining error are discussed.\n\n—\n\nJ onathan M. Lilly was born in Lansing\, Michigan\, in 1972. He received\nth e B.S. degree in atmospheric and oceanic physics from Yale\nUniversity\, N ew Haven\, Connecticut in 1994\, and the M.S. and Ph.D.\ndegrees in physic al oceanography from the University of Washington in\nSeattle\, Washington \, in 1997 and 2002\, respectively. From 2003 to 2005\nhe was a postdoctor al researcher in the Laboratoire d'Océanographie\nDynamique et de Climato logie\, Université Pierre et Marie Curie\,\nParis\, France. Since 2005 he has worked as a research scientist or\nsenior research scientist at vario us institutes\, including Earth and\nSpace Research\, NorthWest Research A ssociates\, and Theiss Research.\nSince July 2021 he has been a Senior Sci entist at the Planetary\nScience Institute in Tucson\, Arizona\, and since June 2024 he has been\na visiting scientist in the Department of Physics at the University of\nToronto in Toronto\, Canada.  Since September 2025 he is also serving\nhalf-time as Lead Oceanographer for Oceanbox AS in Tro msø\, Norway. \nHis research interests include oceanic vortex structures \, statistical\nmethods for data analysis\, Lagrangian observations\, and high-latitude\noceanography.\n\nZoom Participation.  See announcement.\n\ n  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914abdc DTSTART;TZID=America/New_York:20251210T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251210T130000 SUMMARY:CMU Faculty Dialogues: Is Mathematics Obsolete? CLASS:PUBLIC DESCRIPTION:Speaker: JEREMY AVIGAD\, Dean's Chair in Logic and Philosophy o f\nMathematics\, andProfessor\, Department of Philosophy and Department of \nMathematical SciencesDietrich College of Humanities and Social\nSciences \n\nAs AI and machine learning become more prevalent\, questions are\naris ing about the role of mathematics and traditional reasoning in\ntoday’s society. In this talk\, professor Jeremy Avigad will discuss\nthe importan ce of mathematical and symbolic reasoning in the age of AI\nand why societ y needs to be careful as it navigates the changes ahead.\n\nREGISTER → By Tuesday\, December 9\n\nThis Faculty Dialogue event is part of CMU125: The Power of\nPossibilities\, the celebration of CMU’s past\, present an d\nfuture.  Explore CMU’s legacy of achievement and innovation. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914af0a DTSTART;TZID=America/New_York:20260109T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260109T150000 LOCATION:Posner Hall SUMMARY:SCS Teaching and Education Meeting CLASS:PUBLIC DESCRIPTION:Speaker: for SCS Faculty\n\nFriday\, January 9\, 2026\, 10am  – 3pm \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914b230 DTSTART;TZID=America/New_York:20251204T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251204T133000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:AI-SDM Seminar - Ann Bostrom CLASS:PUBLIC DESCRIPTION:Speaker: ANN BOSTROM\, Weyerhaeuser Endowed Professor in Enviro nmental\nPolicyUniversity of Washington\n\nTalk Title: Converging on trust worthy AI for weather\, climate\, and\ncoastal oceanography through risk c ommunication research\n\nDespite recent uptake of AI techniques for foreca sting some weather\nphenomena\, the lack of trustworthy AI still presents a barrier to the\nadoption of AI for extreme weather phenomena. To address such complex\,\nsocietally consequential environmental science problems\, the NSF AI\nInstitute for Research on Trustworthy AI in Weather\, Climate \, and\nCoastal Oceanography (AI2ES) brings AI researchers\, meteorologist s\,\noceanographers\, and risk communication researchers together from\nac ademia\, national labs\, government agencies\, and the private sector.\nDr awing from risk communication\, judgment and decision making\, and\nmental models research\, the AI2ES risk communication team developed\nand tested a multi-method convergence research approach for the\nInstitute. We appli ed this to assess with National Weather Service\nforecasters the trustwort hiness of two prototype AI models for severe\nconvective weather predictio n. Among our findings are that AI model\nperformance\, input variables\, a nd interactivity influenced\nforecasters’ perceptions of model trustwort hiness. They also wanted\nto know under what conditions the model was like ly to fail. Centering\nforecasters and other expert users in the research process is critical\nto advancing AI and developing trustworthy AI models. Trust has long\nbeen recognized as a linchpin for effective risk communic ation\;\ntreating AI as a form of risk communication can help make AI more \neffective and ethical.\n\n—\n\nAnn Bostrom joined the Evans School fac ulty in 2007. Her research\nfocuses on risk perception\, communication\, a nd management\; and\nenvironmental policy and decision making under uncert ainty. She served\non the faculty at the Georgia Institute of Technology ( Georgia Tech)\nfrom 1992-2007\, where she served as Associate Dean for Res earch at the\nIvan Allen College of Liberal Arts and Professor in the Scho ol of\nPublic Policy. Bostrom co-directed the Decision Risk and Management \nScience Program at the National Science Foundation from 1999-2001.\nWhil e in this position she organized\, participated in\, and made\npresentatio ns at national and international meetings on research and\nscience policy\ , including but not limited to\, the Subcommittee on\nNatural Disaster Red uction and the National Earthquake Hazard\nReduction Program.\n\nShe has a uthored or contributed to numerous publications\,\nincluding Risk Communi cation: A Mental Models Approach (Cambridge\nUniversity Press\, 2002)\,  Risk Assessment\, Modeling and Decision\nSupport: Strategic Directions (B erlin: Springer\, 2008)\, and National\nResearch Council\, Institute of Me dicine\, U.S. EPA Science Advisory\nBoard\, and U.S. EPA Board of Scientif ic Counselors reports. She also\nserves on the editorial board for Risk A nalysis\, as Associate Editor\nfor the journal Journal of Risk Research\, and reviews for numerous\ntechnical journals. Bostrom has received resear ch funding from the\nNational Science Foundation\, the U.S. Environmental Protection Agency\,\nand the National Institutes of Health.\n\nBostrom is the recipient of several fellowships\, including the\nAmerican Statistical Association/National Science Foundation/Bureau of\nLabor Statistics Resea rch Associateship (1991-92)\, Fulbright Graduate\nResearch Fellowship and Lois Roth Endowment Fund grant for studies at\nthe University of Stockholm (1989-90)\, and Patricia Roberts Harris\nFellowship at Carnegie Mellon (1 988-89). She is also the recipient of\nthe 2020 Distinguished Educator awa rd from the Society for Risk\nAnalysis\, and of the 1997 Chauncey Starr aw ard for a young risk\nanalyst from the Society for Risk Analysis for her w ork on mental\nmodels of hazardous processes. Bostrom is an elected fellow of the\nAmerican Association for the Advancement of Science (AAAS)\, the\ nWashington State Academy of Sciences (WSAS)\, and the Society for Risk\nA nalysis. She is past president of the Society for Risk Analysis\, past\nCh air of the AAAS Section on Social\, Economic\, and Political Sciences\n(Se ction K)\, and is a member of the Association for Public Policy\nAnalysis and Management\, the Society for Judgment and Decision-Making\,\nand the A merican Association for Public Opinion Research.\n\nBostrom is currently s erving as an elected member on the Board of\nDirectors for the AAAS and on the Board of Directors for the\nWashington State Academy of Sciences. She is also serving on\nthe Advisory Committee for Earthquake Hazards Reduct ion\, on\nthe National Renewable Energy Laboratory Decision Science and\ nAnalysis Technical Advisory Committee\, and on the Advisory Committee\nfo r the National Center for Atmospheric Research Mesoscale &\nMicroscale Me teorology Lab.\n\nAt the University of Washington\, Bostrom serves as co-P I on\nthe Cascadia Coastal Hazards Research Coordination Network (PI Dav id\nSchmidt)\, on the steering committee for EarthLab\, as Primary Resear ch\nArea Chair for Environments and Populations in the Center for Studies \nin Demography and Ecology\, and on the Governing Board of the Program\n on Climate Change. Bostrom is also a co-PI on the NSF-funded Cascadia\nCo astlines and Peoples Hazards Research Hub (PI Peter Ruggiero)\, and\nco-l eads the risk communication research in the NSF AI Institute for\nResearc h on Trustworthy AI in Weather\, Climate\, and Coastal\nOceanography (AI2E S).\n\nBostrom holds a Ph.D. in Public Policy Analysis from Carnegie Mello n\nUniversity\, an M.B.A. from Western Washington University\, a B.A. in\n English from the University of Washington\, and completed postdoctoral\nst udies in Engineering and Public Policy at Carnegie Mellon University\nand in cognitive aspects of survey methodology at the Bureau of Labor\nStatist ics. She has also been a visiting professor at the University\nof Bergen  DICE Lab\, and is affiliated with the Center for Climate\nand Energy Deci sion Making at Carnegie Mellon University.\n\nREGISTER → register to a ttend if attending in-person or on Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914b999 DTSTART;TZID=America/New_York:20251211T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251211T130000 LOCATION:Wean Hall 1327 and Zoom - Register to Attend SUMMARY:AI-SDM - Student Brainstorming Session CLASS:PUBLIC DESCRIPTION:Speaker: AI-SDM Student Brainstorming Session\n\nStudents meet regularly to participate in informal discussion\nsessions that delve into cutting-edge AI topics. These regular\nmeetings throughout the semester p rovide a platform for students to\ndeepen understanding of specific areas and broaden knowledge by\nexploring cross-cutting connections between vari ous AI disciplines.\nDiscussions are a breeding ground for collaboration\, innovative\nthinking\, and problem-solving from the ground up. They provi de a\nstress-free forum for exchanging ideas\, brainstorming new approache s\nto challenges\, and fostering lasting connections within the AI-SDM\nco mmunity in an environment distinct from a traditional seminar.\n\nREGISTER  → register to attend in-person or on Zoom\n\n                   → lunch will be provided for in-person\nregistrants\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914bd0b DTSTART;TZID=America/New_York:20251128T070000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251128T210000 SUMMARY:Thanksgiving Break - University Closed CLASS:PUBLIC DESCRIPTION:Talk Title: THANKSGIVING BREAK - University ClosedWednesday\, N ovember\n26 – Friday\, November 28\, 2025 — no classes.\n\nThursday\, November 27 and Friday November 28\, 2025  — university\nclosed. \n\n Normal Class Schedules and Business Office Hours resume\nMonday\, Decembe r 1 \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914bfcc DTSTART;TZID=America/New_York:20251127T070000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251127T210000 SUMMARY:Thanksgiving Day - University Closed CLASS:PUBLIC DESCRIPTION:Talk Title: THANKSGIVING DAY - University ClosedWednesday\, Nov ember\n26 – Friday\, November 28\, 2025 — no classes.\n\nThursday\, N ovember 27 and Friday November 28\, 2025  — university\nclosed. \n\nNo rmal Class Schedules and Business Office Hours resume\nMonday\, December 1 \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914c25f DTSTART;TZID=America/New_York:20251201T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251201T150000 LOCATION:Gates Hillman 7101 SUMMARY:Doctoral Speaking Skills Talk - Patrick Coppock CLASS:PUBLIC DESCRIPTION:Speaker: PATRICK COPPOCK\, Ph.D. Student\nComputer Science Depa rtment\nCarnegie Mellon University\n\nTalk Title: LithOS: An Operating Sys tem for Efficient Machine Learning\non GPUs\n\nThe rapid growth of machine learning (ML) has made GPUs indispensable\nin datacenters and underscores the urgency of improving their\nefficiency. However\, balancing diverse m odel demands with high\nutilization remains a fundamental challenge. Trans parent\, fine-grained\nGPU resource management that maximizes utilization\ , energy efficiency\,\nand isolation requires an OS approach. This paper i ntroduces LithOS\, a\nfirst step towards a GPU OS.\n\nLithOS includes the following new abstractions and mechanisms for\nefficient GPU management: ( i) a novel TPC Scheduler that supports\nspatial scheduling at the granular ity of individual TPCs\, unlocking\nefficient TPC stealing between workloa ds\; (ii) a transparent kernel\natomizer to reduce head-of-line blocking a nd allow dynamic resource\nreallocation mid-execution\; (iii) a lightweigh t hardware right-sizing\nmechanism that dynamically determines the minimal TPC resources needed\nper atom\; and (iv) a transparent power management mechanism that\nreduces power consumption based upon in-flight work charac teristics.\n\nWe build LithOS in Rust and evaluate its performance across a broad\nset of deep learning environments\, comparing it to state-of-the- art\nsolutions from NVIDIA and prior research. For inference stacking\,\nL ithOS reduces tail latencies by 13× compared to MPS\; compared to the\nbe st-performing SotA\, it reduces tail latencies by 4× while improving\nagg regate goodput by 1.3×. Furthermore\, in hybrid inference-training\nstack ing\, LithOS reduces tail latencies by 4.7× compared to MPS\;\ncompared t o the best-performing SotA\, it reduces tail latencies by\n1.18× while im proving aggregate throughput by 1.35×. Finally\, for a\nmodest performanc e hit under 4%\, LithOS’s hardware right-sizing\nprovides a quarter of G PU capacity savings on average\, while for a 7%\nhit\, LithOS’s transpar ent power management delivers a quarter of GPU\ntotal energy savings on av erage. Overall\, LithOS transparently\nincreases GPU efficiency\, establis hing a foundation for future OS\nresearch on GPUs.\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914c6ec DTSTART;TZID=America/New_York:20251125T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251125T170000 LOCATION:Newell-Simon 3305 SUMMARY:Doctoral Speaking Skills Talk - Caspar Oesterheld CLASS:PUBLIC DESCRIPTION:Speaker: CASPAR OESTERHELD\, Ph.D. Student\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Safe (Pareto) impro vements — a new approach in game\ntheory\n\nOften when you make a decisi on in the real world\, the outcome (your\npayoff) is determined by some st rategic interaction between a group of\nother players\, a _game_ in game-t heoretic parlance. For instance\, the\noutcome of a government regulation in a market depends on how various\nmarket participants interact under the regulation. Unfortunately\,\nstrategic interactions are often hard to jud ge because they have\nmultiple strategic solutions (e.g.\, multiple Nash e quilibria). In the\ntalk\, I will explain this problem of strategic multip licity in more\ndetail\, as well as some existing solutions. I will then i ntroduce the\nconcept of _safe improvements_ (SIs)\, a new approach to (so metimes)\nresolving such questions. The SI approach is to analyze how pair s of\ngames relate. Sometimes this allows us to confidently reach\nconclus ions about which of two games is better (as judged by some\nspecified pref erences) without requiring predictions about either of\nthe individual gam es. For instance\, if two games are isomorphic\, it\nstands to reason that they will be played isomorphically.\n\nAfter explaining the basic idea be hind SIs\, I will focus mostly on\ncomputational questions about them. For instance\, how hard is it to\ndecide whether one given game is an SI on a nother? If we have some way\nof intervening on a game (say\, removing an a rbitrary subset of\nactions)\, how hard is it to find an intervention that induces an SI on\nthe given game?\n\nPresented as part of the Econ-CS Sem inar Series\n\nPresented in Partial Fulfillment of the CSD Speaking Skills \nRequirement \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914cb57 DTSTART;TZID=America/New_York:20251124T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251124T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Jeremy Taylor CLASS:PUBLIC DESCRIPTION:Speaker: JEREMY TAYLOR\, Head of Product\, XTDB\n\nTalk Title: Reconstructing History with XTDB\n\nXTDB is a SQL database that challenges long held assumptions about how\ndata mutates in databases. Instead of UP DATEs and DELETEs destroying\ninformation\, or forcing developers to imple ment archival strategies\,\nXTDB preserves history automatically without l eaving such decisions to\ndevelopers. Additionally XTDB implements a varia tion of the SQL:2011\nsyntax to simplify time-travel queries across two di mensions of time:\nsystem-time (what developers usually think about) and v alid-time (what\nbusiness users care about the most). XTDB is built around Apache Arrow\nand offers a FlightSQL interface in addition to Postgres wi re protocol\ncompatibility. This seminar talk will walk through the design \nconstraints and internal architecture\, which have been heavily\ninspire d by Datomic (released in 2012) and 'big data' systems.\n\n—\n\nJeremy T aylor is Head of Product for XTDB\, an open source database\ncreated to si mplify how developers manage bitemporal data and dynamic\nsystems. He deve loped XTDB in 2019 alongside a team of Lisp hackers at\nJUXT\, a company b uilt around a passion for Clojure (a functional Lisp\nfor the JVM) and a f ocus on Investment Banking. He previously worked\non eCommerce systems wit h IBM and holds a computer engineering degree.\n\nThis talk is part of the Future Data Systems Seminar Series.\n\nZoom Participation.  See announce ment. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914d085 DTSTART;TZID=America/New_York:20251202T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251202T170000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:Pittsburgh Quantum Institute Distinguished Seminar - Matthias Troye r\nand Chetan Nayak CLASS:PUBLIC DESCRIPTION:Speaker: MATTHIAS TROYER and CHETAN NAYAK\, Matthias\, Technica l Fellow\nand Corporate Vice President of Quantum\nChetan\, Technical Fell ow\nMicrosoft\n\nTalk Title: Utility-Scale Quantum Computing with Topologi cal\nQubitsTuesday\, December 2\, 2025\, 3:30 – 5pm PQI hosts Dr. Matt hias\nTroyer and Dr. Chetan Nayak from Microsoft\, who will\npresent\, Ut ility-Scale Quantum Computing with Topological\nQubits.About the Speakers ► Dr. Matthias Troyer is accountable for\narchitecting Microsoft’s qu antum computer and applications. His work\nat Microsoft is focused on acce lerating scientific discovery globally\nby bringing to bear the benefits o f a scaled\, fault tolerant quantum\nsystem to the world ins secure and re sponsible ways. He received his\nPhD from ETH Zurich in Switzerland in 199 4. Afterwards\, he spent time\nas a post-doctoral fellow at the University of Tokyo\, then returned to\nETH Zurich as a Computational Physics profes sor. He joined Microsoft\nin 2017. Dr. Troyer is also a Fellow of the Amer ican Physical Society\nand President of the Aspen Center for Physics. He i s the recipient of\nthe Hamburg Prize for Theoretical Physics and the Rahm an Prize for\nComputational Physics of the American Physical Society “ … for\npioneering numerical work in many seemingly intractable areas of\ nquantum… physics and for providing efficient sophisticated computer\nco des to the community.”► Dr. Chetan Nayak has been a researcher\nat Mi crosoft since 2005. He was born and raised in New York City\,\nwhere he gr aduated from Stuyvesant High School in 1988. He received\nhis B.A. from Ha rvard in 1992 and his Ph.D. from Princeton University\nin 1996. He was a p ost-doctoral fellow at the Institute for\nTheoretical Physics at UCSB from 1996-97. He was a Professor of\nPhysics at UCLA from 1997 through 2006 an d at UCSB from 2007 through\nthe present. He was a visiting Professor at N ihon University in Japan\nin 2002. He is a Fellow of the American Physical Society and a\nrecipient of the Outstanding Young Physicist Award from th e American\nChapter of the Indian Physics Association\, an Alfred P. Sloan \nFoundation Fellowship\, and an NSF Early Career Award. He has been the\n Principal Research Manager of Microsoft Station Q since 2014.Chetan\nhas m ade significant contributions to the theory of topological\nphases\, high- temperature superconductivity\, ‘strange metals’\, the\neffects of im purities on electronic behavior\, the quantum Hall effect\,\nand phases of periodically-driven quantum systems. In 1996\, Chetan and\nFrank Wilczek discovered the type of non-Abelian statistics associated\nwith Majorana ze ro modes\, which will be the building block of\nMicrosoft’s quantum comp uter. His subsequent work in 2005 with\nMichael Freedman and Sankar Das Sa rma sparked attempts to build a\ntopological quantum computer using the 5/ 2 fractional quantum Hall\nstate. In 2008\, he was the lead author of an i nfluential article\nsurveying the field of topological quantum computing. In 2016\, he\nrepaired and revived the concept of a “time crystal” wit h Dominic\nElse and Bela Bauer and predicted its occurrence in\nperiodical ly-driven systems\, which was experimentally verified shortly\nthereafter. The 2016 paper that he and several co-authors wrote on\nMajorana zero mod e device designs serves as a guide for Microsoft’s\nquantum effort.Under his leadership\, Microsoft’s quantum team\npublished papers demonstrati ng topological superconductivity in\nnanowires and single-shot readout of the parity state of a pair of\nMajorana zero modes. Event Type: Seminars Room Number: In Person\nBuilding: Rashid Auditorium\, Gates Hillman 4401 Speaker's\nName: MATTHIAS TROYER and CHETAN NAYAK Speaker's Professional \nTitle: Matthias\, Technical Fellow and Corporate Vice President of\nQua ntum\, Chetan\, Technical Fellow\, Microsoft Talk\nTitle: Utility-Scale Q uantum Computing with Topological Qubits Event\nPoster Title: Poster with QR Code Event Poster URL: www.cs.cmu.edu\n[http://www.cs.cmu.edu]… Aff iliations: Computer Science Department\n(CSD)\, Machine Learning Departme nt (MLD)\, Robotics Institute (RI)\nOrganization(s): School of Computer S cience\, PQI\, Department of\nPhysics Event Website Title: Event Website Event Website\nURL: calendar.pitt.edu…\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914d87f DTSTART;TZID=America/New_York:20251124T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251124T113000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year MS Thesis Presentation - Shalini Panthangi CLASS:PUBLIC DESCRIPTION:Speaker: SHALINI PANTHANGI\, Master's Student\nComputer Science Department\nCarnegie Mellon Universityt\n\nTalk Title: Consistent Formali zation of Legal Text into Logical Rules\nvia Guided Large Language Models\ n\nPrecise logical formalization of legal text helps automated compliance\ nanalysis and machine-readable legal reasoning\, which help streamline\nan d prove complex queries in industries like law and insurance.\nAchieving t his is challenging\, as legal text includes ambiguity\,\nexceptions\, and layered nuances that make it difficult to\nconsistently translate into lo gical rules. Existing large language\nmodel-based methods often generate i nconsistent predicates\, drift in\nmeaning\, and fail to capture complex l egal structures. \n\nThis thesis introduces a structured pipeline for con verting legal text\ninto Defeasible Deontic Logic and First-Order logic wi th a focus on\nkeeping key terms consistent and grounding predicates in a stable\nmanner. The approach introduces consistency through a symbol-table \nframework that constrains LLM outputs to a vocabulary of legal actors\na nd actions. Combined with clause segmentation\, multi-stage LLM\nrewriting \, and automated Z3 consistency verification\, the system\nproduces logica l rules that better maintain the intended argument\nstructure of legal sta tutes. Evaluating with multiple legal examples\nshows that this method red uces logical errors and produces\nformalizations suitable for reasoning ta sks and analysis. The results\ndemonstrate that integrating symbolic guida nce with LLM-based\nprocessing provides a path toward generating trustwort hy formal\nrepresentations of legal text.\n\nThesis Committee\n\nUmut Acar (Chair)\n\nSherry Tongshuang Wu\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914dcb3 DTSTART;TZID=America/New_York:20251118T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251118T130000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Database Seminar - Benjamin Owad CLASS:PUBLIC DESCRIPTION:Speaker: BENJAMIN OWAD\, Software Engineer\, Snowflake\n\nTalk Title: Optimizing the Table Scan Operator: I/O Minimization and\nRuntime A daptivity\n\nTable scan is a foundational operator in any analytical datab ase and\nis often the primary bottleneck for a given query. This talk prov ides\na technical deep dive into optimizations our team has developed for\ nthe table scan operator. First\, we will discuss I/O reduction\ntechnique s\, including pruning strategies to avoid reading unnecessary\ndata and st orage request coalescing to batch I/O operations\nefficiently. Next\, we w ill explore runtime adaptivity\, with features\nlike adaptive scan\, a met hodology to balance CPU and memory usage with\nI/O bandwidth\, dynamic col umn reordering\, and data encoding schemes.\n\n—\n\nBenjamin Owad is a c omputer science enthusiast based in the San\nFrancisco Bay Area working at Snowflake. He studied computer science\nat Carnegie Mellon University\, c oncentrating in systems. His interests\nspan databases\, open-source softw are\, vintage hardware\, and niche\noperating systems \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914e042 DTSTART;TZID=America/New_York:20260302T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260302T235500 LOCATION:NO CLASSES SUMMARY:Spring Break 2026! CLASS:PUBLIC DESCRIPTION:Talk Title: Spring Break - March 2-6\n\nSpring Break 2026 will run from Monday\, March 2 through Friday\, March\n6.  While the universit y business offices will remain open\, there are\nno classes for the durati on of the break.\n\nClasses resume Monday\, March 9. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914e326 DTSTART;TZID=America/New_York:20251121T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251121T153000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 (New Location) SUMMARY:5th Year MS Thesis Presentation - Sonya Simkin CLASS:PUBLIC DESCRIPTION:Speaker: SONYA SIMKIN\, Master's Student\nComputer Science Depa rtment\nCarnegie Mellon University\n\nTalk Title: Logical Relations for No n-Interference with Cryptography\n\nCryptographic primitives\, such as enc ryption\, decryption\, and key\ngeneration\, are vital to the security of many of today's applications\,\nbut are difficult to reason about in an in formation-flow setting. In\nparticular\, they conflict with the notion of non-interference\, the\nproperty that observable outputs are uninfluenced by secret inputs in\nan information-flow secure program. Not only do ciphe rtexts often rely\non secure data (and thus would be ruled out in a typica l\ninformation-flow setting)\, the use of non-determinism in their\ngenera tion may serve as a source of information-flow leak (called\nocclusion).\n \nLogical relations are a technique which can be used to prescribe\nproper ties of a program based on its computational behavior\, rather\nthan relyi ng on static well-formedness. Through this semantic\napproach\, one can no t only validate well-typed terms by showing their\ninhabitation of the log ical relation (via the \"fundamental theorem\")\,\nbut also validate ill-t yped terms which behave \"correctly\" with\nrespect to the definition of t he logical relation.\n\nIn this thesis\, we extend a version of the simply -typed lambda\ncalculus with cryptographic primitives\, and define an info rmation-flow\ncontrol type system to statically enforce safe usage of said \nprimitives. We then use the technique of logical relations to define a\n semantic verification method for non-interference. In particular\, to\ncom bat occlusion\, we define a possibilistic logical relation\, which\nconsid ers the set of all possible values an expression could evaluate\nto.\n\nTh esis Committee\n\nStephanie Balzer (Chair)\n\nRobert Harper\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914e758 DTSTART;TZID=America/New_York:20260204T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260204T170000 LOCATION:Deadline to Register: February 4\, 2026 SUMMARY:2026 3MT Competition: Registration Deadline CLASS:PUBLIC DESCRIPTION:Speaker: CMU Ph.D. Students regardless of year in programTalk T itle:\n2026 3MT Competition Registration Deadline\n\nThree Minute Thesis  (3MT) is an internationally recognized\ncompetition that challenges Ph.D. students to present a compelling\noration on their thesis and its signific ance in just 3 minutes\, in\nlanguage that anyone can understand. 3MT is n ot an exercise in\ntrivializing or ‘dumbing-down’ research but require s students to\nconsolidate their ideas\, crystallize their research discov eries and\ncapture the imagination of their audience.\n\nWho's Eligible to Compete?   All CMU Ph.D. Students regardless of\nyear in program. Wheth er it's your first year or last year\, Come one -\ncome all!\n\n \n\nWhy Compete?   3MT offers CMU doctoral candidates of any level the\nopportun ity to gain early career recognition\, connect with the campus\ncommunity\ , and win prizes of up to $3\,000. It's a great way to\npractice sharing y our research with a non specialist audience and get\nlisteners excited abo ut your area of study. Students get to hear from\nfellow Ph.D. students ac ross campus\, which can spark interdisciplinary\ncollaboration. They also get to present before staff\, faculty\, and\nalumni from a wide range of d epartments\, as well as an extensive\nalumni network. Preliminary round wi nners receive $250 cash.\n\n \n\nWhat Happens Next?   Start practicing !   All 3MT registrants will\nbe scheduled to compete in preliminary rou nds in late February and\nearly March.  Preliminary round registrations w ill open November\n2025.\n\nREGISTER → The deadline to register is Feb ruary 4\, 2026.\n\n \n\nThen What?\n\nWinners of preliminary rounds go on to compete in-person in the CMU\n3MT Championship on Wednesday\, March 25 \, 2026. Finalists will compete\nfor an additional $500-$3\,000 in cash pr izes in the Three Minute\nThesis Championship. Prizes are given for 1st ($ 3\,000)\, 2nd ($2\,000)\,\nand 3rd ($1\,000) as well as the live audience- votes for People's\nChoice ($500) and virtual-vote for the Alumni Award ($ 750).\n\nMore Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914ec27 DTSTART;TZID=America/New_York:20251117T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251117T170000 LOCATION:Posner Hall 151 SUMMARY:Statistics and Data Sciences Seminar - Ankit Pensia CLASS:PUBLIC DESCRIPTION:Speaker: ANKIT PENSIA\, Assistant Professor\, Department of Sta tistics\nand Data Science\, Carnegie Mellon University\n\nTalk Title: Cont amination Models for Linear Regression: Statistical\nand Algorithmic Limit s\n\nWe study estimation in the Gaussian linear model in the presence of\n outliers. Historically\, these outliers were modeled using the Huber\ncont amination model\, where contamination can corrupt both the\ncovariates (X) and the responses (y). In this model\, it is well known\nthat consistent estimation is not possible. This has led to interest\nin weaker contaminat ion models that still permit consistent estimation\nwhile providing some r obustness.\n\nA popular option has been to restrict the contamination to c orrupting\nonly the responses. In this talk\, I will introduce a hierarchy of\ncontamination models belonging to this family and discuss their\nresu lting statistical and algorithmic limits.\n\nBased on ongoing joint works with Ilias Diakonikolas\, Chao Gao\,\nSaminul Haque\, Daniel Kane\, Rohith Kuditipudi\,  Kabir Aladin\nVerchand\, and Dong Xie.\n\n—\n\nAnkit Pen sia is an assistant professor in the Department of Statistics\nand Data Sc ience at Carnegie Mellon University. Previously\, he was a\nResearch Fello w at the Simons Institute for the Theory of Computing\nand a Herman Goldst ine Postdoctoral Fellow at IBM Research. He\nobtained his PhD in Computer Science from the University of\nWisconsin-Madison. His current research in terests include algorithmic\nrobust statistics\, high-dimensional probabil ity\, minimax testing\, and\nalgorithmic stability. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914f01f DTSTART;TZID=America/New_York:20251120T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251120T103000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Doctoral Thesis Proposal - Cayden Codel CLASS:PUBLIC DESCRIPTION:Speaker: CAYDEN CODEL\, Ph.D. Student\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: Building a Verified SAT Too lchain in Trestle\n\nAutomated reasoning (AR) tools are versatile and prac tically efficient\npieces of software that can solve a wide variety of pro blems in\nindustry and academia. One of their strengths is their ability t o\ngenerate proofs checkable by verified software. Even if the AR tools\nt hemselves contain bugs (and they often do)\, we can still have a high\ndeg ree of confidence in the correctness of their answers.\n\nHowever\, this v erified toolchain can be extended further to include\nencodings. Usually\, AR tools solve problems that have been encoded\ninto a form that they can understand. This process of encoding can\nintroduce bugs\, meaning that t he encoded form of the problem no longer\naccurately represents the origin al. Bugs are more likely to appear\nwhen the encoding is more complicated\ , such as when auxiliary logical\nobjects are introduced in order to make the encoding smaller or easier\nfor the AR tool to manipulate. Unfortunate ly\, complicated encodings\nare becoming increasingly necessary in order t o push the boundaries of\nwhat problems AR tools can solve. As a result\, we argue that the\nstandard AR toolchain should now include verified encod ings by\ndefault.\n\nIn this thesis\, we will develop an end-to-end verifi ed toolchain for\nthe boolean satisfiability problem (SAT) in the Trestle project.\nTrestle currently has a verified SAT proof checker and good supp ort\nfor writing verified encodings\, but its encoding tools are complicat ed\nand hard to use\, and only a handful of encodings have been verified s o\nfar. As a result\, we plan to redesign how encodings are written in\nTr estle\, and to verify new encodings. We also plan to add\nsymmetry-breakin g reasoning to Trestle and add features to enable\nend-to-end theorem prov ing with the use of SAT solvers.\n\nThesis Committee\n\nMarijn Heule (Chai r)\n\nJeremy Avigad\n\nBryan Parno\n\nBenjamin Kiesl-Reiter (Amazon Web Se rvices)\n\nAdditional Information\n\nIn Person and Zoom Participation.  S ee announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914f4b7 DTSTART;TZID=America/New_York:20251117T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251117T150000 LOCATION:Remote Access - Zoom SUMMARY:Special Data Base Seminar - Anupam Datta CLASS:PUBLIC DESCRIPTION:Speaker: ANUPAM DATTA\, Principal Research ScientistSnowflake A I\nResearch LeadSnowflake\n\nTalk Title: Cortex AISQL: A Production SQL En gine for Unstructured\nData\n\nSnowflake’s Cortex AISQL is a production SQL engine that integrates\nnative semantic operations directly into SQL. This integration allows\nusers to write declarative queries that combine r elational operations\nwith semantic reasoning\, enabling them to query bot h structured and\nunstructured data effortlessly. However\, making semanti c operations\nefficient at production scale poses fundamental challenges. Semantic\noperations are more expensive than traditional SQL operations\, possess\ndistinct latency and throughput characteristics\, and their cost and\nselectivity are unknown during query compilation. Furthermore\,\nexis ting query engines are not designed to optimize semantic\noperations.\n\nT he AISQL query execution engine addresses these challenges through\nthree novel techniques informed by production deployment data from\nSnowflake cu stomers. First\, AI-aware query optimization treats AI\ninference cost as a first-class optimization objective\, reasoning\nabout large language mod el (LLM) cost directly during query planning\nto achieve 2–8A— speedup s. Second\, adaptive model cascades reduce\ninference costs by routing mos t rows through a fast proxy model while\nescalating uncertain cases to a p owerful oracle model\, achieving\n2–6A— speedups while maintaining 90 –95% of oracle model quality.\nThird\, semantic join query rewriting low ers the quadratic time\ncomplexity of join operations to linear through re formulation as\nmulti-label classification tasks\, achieving 15–70A— s peedups with\noften improved prediction quality. AISQL is deployed in prod uction at\nSnowflake\, where it powers diverse customer workloads across\n analytics\, search\, and content understanding.\n\n—\n\nAnupam Datta is a Principal Research Scientist and Snowflake AI\nResearch Lead at Snowflak e. He joined Snowflake as part of the\nacquisition of TruEra where he serv ed as Co-Founder\, President\, and\nChief Scientist from 2019-2024. Datta was on the faculty at Carnegie\nMellon University from 2007-2022\, most re cently as a tenured Professor\nof Electrical & Computer Engineering and Co mputer Science. Datta's\ncurrent research focuses on Trustworthy AI\, span ning evaluation\,\nexplainability\, fairness\, and adversarial robustness of ML models and\nGenAI applications. Specific results include early work on Shapley\nValues & gradient-based explanations\, fairness assessments\, robustness\nof classical machine learning and deep learning models for nat ural\nlanguage processing and computer vision\, and the TruLens open sourc e\nproject for evaluation and experiment tracking of GenAI apps. These\nre search results have had a significant impact on products at TruEra\nand Sn owflake. Datta has published over 100 research papers\, served as\nChair o f the National Academies Workshop on Assessing and Improving AI\nTrustwort hiness\, on the Steering Committee of of the ACM Conference on\nFairness\, Accountability\, and Transparency\, and the IEEE Computer\nSecurity Found ations Symposium\, and as an Editor-in-Chief of\nFoundations and Trends in Privacy and Security.\n\nHe received the 2018 David P. Casasent Outstandi ng Research Award from\nthe CMU College of Engineering\, a 2020 Young Alum ni Achiever Award\nfrom IIT Kharagpur\, a 2021 Google Faculty Research Awa rd\, and several\nawards for top papers at conferences. Datta obtained a B .Tech. from\nIIT Kharagpur\, and Ph.D. and M.S. degrees from Stanford Univ ersity in\nComputer Science\, where he currently teaches a course on Trust worthy\nAI.\n\nZoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914fa9e DTSTART;TZID=America/New_York:20251120T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251120T133000 LOCATION:Newell Simon 4305 and Zoom SUMMARY:AI-SDM Seminar - Tom Manzini CLASS:PUBLIC DESCRIPTION:Speaker: TOM MANZINI\, Ph.D. StudentDepartment of Computer Scie nce\n&amp\; Engineering Texas A&amp\;M University\n\nTalk Title: Bridging the Sky: Building Multi‑Source Vision Systems\nfor Real‑Time Disaster Response\n\nWhen a disaster strikes\, disaster response operations deploy multiple\nsensors to capture imagery of impacted areas so decisions can be made\nwith the best information possible. Satellites\, small uncrewed aer ial\nsystems (sUAS)\, and crewed aircraft are deployed to capture imagery of\nthe affected area. However\, once that imagery is captured\, it arrive s\nin waves\, creating data avalanches that can paralyze decision-making\n and thus the disaster response. Computer vision models are needed to\nmana ge this immense volume of data\, but in practice\, it is impossible\nto kn ow which sources will be available when decisions need to be\nmade: clouds can obscure the scene from satellites\, weather may\nprevent crewed aircr aft from flying\, and sUAS teams may be physically\nunable to access the d isaster scenes. To manage this situation\,\ncomputer vision systems that c an accurately and gracefully handle this\ndiverse real-world imagery are c ritical\, but such systems do not yet\nexist.\n\nThis talk will explore th e research and fieldwork behind the\ndevelopment and deployment of multi-s ource\, multi-scale damage\nassessment systems. It will introduce CRASAR ‑U‑DROIDs\, the largest\ndataset of its kind\, spanning 10 disasters\, 3 imagery sources\, 70.6\ngigapixels\, and 122\,502 labels\, and discuss the technical and\noperational challenges of bringing these systems to rea l‑world\ndisaster environments. The presentation will continue by coveri ng the\nfirst known operational deployment of an sUAS-based automated dama ge\nassessment system during Hurricanes Debby and Helene\, and the talk\nw ill conclude by discussing current academic efforts to enhance model\ncros s-scale capabilities to strengthen decision-making during\ndisasters.\n\n —\n\nThomas Manzini is a PhD student at Texas A&M where his work focuses on\ncomputer vision and machine learning\, specifically in increasing the \nrobustness and applicability of machine learning to support\noperational decision-making like that found in wide-area disaster\nresponse. Thomas h as a B.S. from Rensselaer Polytechnic Institute and\nan M.S. from Carnegie Mellon University. He has helped organize the AI\nfor Humanitarian Assist ance and Disaster Response workshop at NeurIPS\nand ICCV. Before returning to academia for his PhD\, Thomas worked as a\nMachine Learning Scientist at Microsoft\, where he split his time\nbetween machine learning products and collaborations with groups like\nthe CDC and the WHO in the developmen t of machine learning and data\nmanagement systems to respond to disasters like the COVID-19 pandemic.\nThomas holds Commercial Pilot Licenses for M anned (FAA Part 61) and\nUnmanned (FAA Part 107) aircraft and has more tha n a decade of\noperational experience as an Advanced Emergency Medical Tec hnician and\nFirefighter.\n\nREGISTER → to attend in-person or on Zoom\ n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef914ffcd DTSTART;TZID=America/New_York:20251117T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251117T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Benjamin Wagner CLASS:PUBLIC DESCRIPTION:Speaker: BENJAMIN WAGNER\, Vice President of Engineering\nFireb olt\n\nTalk Title: Why Powering User Facing Applications on Iceberg is Har d\n\nAbout Firebolt\n\nThis talk is part of the Future Data Systems Semina r Series.\n\nZoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91502ea DTSTART;TZID=America/New_York:20251120T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251120T130000 LOCATION:ASA Conference Room\, Gates HIllman 6115 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Joseph Reeves CLASS:PUBLIC DESCRIPTION:Speaker: JOSEPH REEVES\, Ph.D. Candidate\nComputer Science Depa rtment\nCarnegie Mellon University\n\nTalk Title: Cardinality Constraints in Boolean Satisfiability Solving\n\nAutomated reasoning is a branch of ar tificial intelligence that uses\nsearch engines called solvers to find sol utions for problems\nformulated in mathematics and logic. Automated reason ing has been\napplied across a wide range of domains from hardware and sof tware\nverification to mathematical discovery. To solve a diverse set of\n problems\, these tools make use of low-level reasoning\, namely the\nconfl ict-driven clause learning algorithm\, which is made possible by\nfirst en coding a problem into low-level logic.\n\nEncoding is pivotal to the succe ss of a solver. High-level constraints\nare transformed into low-level log ic via abstractions\, and using a\nsuboptimal set of abstractions might in crease a solver’s runtime by\nfactors of a hundred. In this thesis\, we focus on one type of\nhigh-level constraint: cardinality constraints. Card inality\nconstraints appear in any problem that requires counting\, for ex ample\,\n\"synthesize a quantum circuit with at most k swap gates\"\, or  \"each\nvalue from 1 to 9 may appear at most once in every row of a Sudok u\npuzzle\". It is standard practice to make users encode these\nconstrain ts before passing them as input to an off-the-shelf solver.\n\nWe propose an extended input format that includes cardinality\nconstraints. This allo ws us to move questions of encoding out of the\nhands of the user and into the domain of the solver. We present\nseveral techniques that leverage th e structural information provided\nby this new input format to automatical ly generate more effective\ncardinality constraint encodings.  Furthermor e\, we developed a new\nsolver that employs cardinality-specific reasoning to quickly find\nsolutions for problems from hardware synthesis and discr ete\nmathematics\n\nThesis Committee\n\nRandal Bryant (Co-chair)\n\nMarijn Heule (Co-chair)\n\nRuben Martins\n\nArmin Biere (University of Freiburg) \n\nIn Person and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91507b1 DTSTART;TZID=America/New_York:20251120T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251120T153000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:Doctoral Thesis Oral Defense - Emre Yolcu CLASS:PUBLIC DESCRIPTION:Speaker: EMRE YOLCU\, Ph.D. Candidate\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: Proof Complexity of Resolut ion-Based Systems: Lower\nBounds\, Simulations\, and Applications to SAT S olving\n\nThe satisfiability problem for propositional logic (SAT) has bee n a\ncentral topic in computer science for many decades. It is arguably th e\ncanonical NP-complete problem: reductions from many other problems in\n NP to SAT are often straightforward. As a consequence\, a reasonable\nstra tegy when trying to solve a problem in NP is to reduce it to SAT\nand to t ry to solve the resulting SAT problem instead. This strategy\nturns out to be surprisingly effective thanks to the effectiveness of\nimplementations of heuristic algorithms for SAT\, commonly known as SAT\nsolvers. Those s olvers are expected to output proofs to certify their\nanswers\, and in th is sense they are proof search algorithms. Proof\ncomplexity\, the branch of computational complexity that studies the\nlengths of proofs in proposi tional proof systems\, offers a way to\nanalyze the performance of SAT sol vers.\n\nThis thesis explores the interplay between SAT solving and proof\ ncomplexity. On the theoretical side\, we investigate the power of weak\,\ nresolution-based proof systems that incorporate restricted forms of\nthe extension rule often used in SAT solvers. We provide a complete\ncharacter ization of their relative strengths. In particular\, we\npresent a general recipe for constructing formulas that yield\nexponential separation resul ts\, showing that none of these systems\nsubsumes another in expressivity. We also establish new exponential\nlower bounds for some of those systems \, pinpointing the inherent\nlimitations of various clause addition rules. The key insights for\nthose separations come from the notion of an effect ive simulation. We\nleverage these insights to better understand the compl exity of proof\nsearch: for example\, we show that even a seemingly weaker system like\nregular resolution can effectively simulate general resoluti on\, which\nhas the corollary that significantly faster algorithms for fin ding\nregular resolution proofs would also speed up general resolution pro of\nsearch.\n\nOn the practical side\, we demonstrate how advances in SAT solving can\naid in mathematical discovery. We develop an automated approa ch to the\nCollatz conjecture\, a notorious open problem in number theory\ , by\nencoding it as a search for a termination proof in a rewriting syste m.\nWe show how the Collatz function can be expressed as a string\nrewriti ng system that terminates if and only if the conjecture holds\,\nand we us e state-of-the-art automated reasoning tools to verify\npartial results.  \n\nThesis Committee\n\nMarijn Heule (Chair)\n\nJeremy Avigad\n\nRyan O'Do nnell\n\nSam Buss (University of California San Diego) \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9150e71 DTSTART;TZID=America/New_York:20251120T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251120T173000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Doctoral Speaking Skills Talk - Quang Dao CLASS:PUBLIC DESCRIPTION:Speaker: QUANG DAO\, Ph.D. Student\nComputer Science Department \nCarnegie Mellon University\n\nTalk Title: Speeding Up Sum-Check Proving\ n\nThe sum-check protocol is a fundamental interactive proof for\nverifyin g expensive claims about multivariate polynomials over a\nfinite field. It allows an untrusted prover to demonstrate such a\nclaim to a computationa lly limited verifier\, providing high confidence\nin the validity of the c laim. Introduced over three decades ago\,\nsum-check is now at the core of the most efficient cryptographic proof\nsystems. However\, its ubiquitous usage has made sum-check a proving\nbottleneck in many applications.\n\nI will present a suite of improvements to the sum-check prover that\noutper forms classic algorithms in many settings of interest. These\nimprovements rely on two key ideas. First\, by opening the black box of\nfinite field arithmetic\, we enable fast multiplication with small\nvalues such as 64-b it integers\, which naturally arise when proving\ncorrect program executio n. Second\, we batch the prover's work across\nmany consecutive rounds int o a larger computation\, which is\nbeneficial in two distinct settings. W hen proving program execution\,\nthis approach allows us to trade expensiv e multiplication over the\nfield for cheap multiplication with small value s. In streaming\nsettings\, where the prover is limited in working memory\ , this reduces\nthe number of passes over the input\, allowing the prover to finish\nfaster with only a minimal increase in total computation.\n\nJ oint work with Zachary DeStefano\, Suyash Bagad\, Yuval Domb\, and\nJustin Thaler. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915136a DTSTART;TZID=America/New_York:20251121T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251121T130000 LOCATION:Gates Hillman 8102 SUMMARY:Doctoral Speaking Skills Talk - Harrison Grodin CLASS:PUBLIC DESCRIPTION:Speaker: HARRISON GRODIN\, Ph.D. Student\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Abstraction Functions as Types: Modular Verification of\nData Structures\n\nFriday\, November 2 1\, 2025\, 12 – 1pm \n\nSoftware development depends on the use of li braries whose public\nspecifications inform client code and impose obligat ions on the\nprivate implementation\; it follows that any approach to veri fication\nat scale must also be modular\, preserving such abstraction. Hoa re's\ninfluential methodology for such verifications uses an abstraction\n function to demonstrate the coherence between an implementation and\nits s pecification.\n\nFor all of its influence\, the Hoare methodology relies o n conventional\nseparation between implementation and specification\, prov iding no\nlinguistic support for ensuring that these conventions are obeye d. We\npresent a synthetic account of Hoare's methodology within dependent \ntype theory\, encoding the data of an abstraction function within types\ nthemselves. Accordingly\, various foundational principles are rendered\na s theorems: for example\, a noninterference theorem characterizes an\ninte rnal notion of modularity guaranteed by the theory.\n\nMoreover\, this app roach scales to permit the specification and\nverification of the cost of programs\, allowing clients to verify their\nown cost relative to a specif ication. To maintain abstraction\, the\nimplementation must be merely uppe r bounded in cost by the\nspecification\, which is achieved via a novel \" sealing\" effect. The\nresulting type theory supports modular development of programs and\nproofs in a manner that hides private details while permi tting modular\nverification of both the cost and behavior of programs.\n\n Presented as part of the PLunch Seminar\n\nPresented in Partial Fulfillmen t of the CSD Speaking Skills\nRequirement \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91518d9 DTSTART;TZID=America/New_York:20251119T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251119T130000 LOCATION:Gates Hillman 8102 SUMMARY:Joint Theory Lunch Seminar / Doctoral Speaking Skills Talk - Henry\ nFleischmann CLASS:PUBLIC DESCRIPTION:Speaker: HENRY FLEISCHMANN\, Ph.D. Student\nComputer Science De partment\nCarnegie Mellon University\n\nTalk Title: Directed Expander Deco mpositions: a Gardener's Guide\n\nExpander decompositions are well-establi shed as a central tool in\nundirected graph algorithms\, facilitating an a valanche of breakthrough\nresults over the past two decades. Recently they have also found a\nhome in directed graph algorithms\, powering recent br eakthroughs in\nmax flow and dynamic shortest path.\n\nIn this talk\, I di scuss recent work advancing the state of the art of\ndirected expander dec ompositions. I'll describe techniques from the\nundirected setting and dis cuss how to extend them to the directed\nsetting to achieve a faster direc ted algorithm. Our methods are\ngardening-inspired: turning a hedge into a neat and tidy row of bushes\nturns out to be just like finding an expande r decomposition!\n\nBased on arXiv:2507.09729. Joint work with George Li a nd Jason Li.\n\nPresented as part of the Theory Lunch Seminar\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9151ce6 DTSTART;TZID=America/New_York:20251113T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251113T153000 LOCATION:Wean Hall 7128 SUMMARY:ACO Seminar - Amanda Priestly CLASS:PUBLIC DESCRIPTION:Speaker: AMANDA PRIESTLY\, Ph.D. Student\, Theory Group\, Unive rsity of\nTexas at Austin\n\nTalk Title: Probability and Parking\n\nIn 196 6\, Konheim and Weiss introduced a deterministic model of parking\nand a r elated family of combinatorial objects called parking\nfunctions. While th ere are extensive enumerative results on parking\nfunctions\, much less is known from a probabilistic perspective. In\nthis talk\, I will discuss th ree key probabilistic frameworks that have\nbeen applied to these objects and present our contributions within\neach. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91520a1 DTSTART;TZID=America/New_York:20251113T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251113T163000 LOCATION:Wean Hall 7218 SUMMARY:ACO Seminar - John Byrne CLASS:PUBLIC DESCRIPTION:Speaker: JOHN BYRNE\, Ph.D. Candidate\, Department of Mathemati cal\nSciences\, University of Delaware\n\nTalk Title: Nonabelian Sidon set s and extremal problems on digraphs\n\nAn Sk-set is a subset of a group wh ose k-tuples have distinct\nproducts. We give explicit constructions of la rge Sk-sets in the\ngroups Sym(n) and Alt(n) and of large S2-sets in Sym(n ) x Sym(n) and\nAlt(n) x Alt(n)\, as well as some probabilistic constructi ons for\n'nice' groups. We also give various upper bounds\; in particular\ , if k\nis even and the group has a normal abelian subgroup with bounded i ndex\nthen the trivial upper bound can be improved by some constant\nmulti plicative factor. Next\, we describe some connections between\nSk-sets and extremal graph theory. We determine up to a constant\nfactor the largest possible minimum outdegree in a digraph without\ncertain directed cycles. Interestingly\, the extremal minimum outdegree\nis much larger for any one of these cycles than for the whole family.\nWe count the directed Hamilto n cycles in one of our constructions to\nimprove the upper bound for a pro blem on Hamilton paths posed by\nCohen\, Fachini\, and Körner. This talk is based on joint work with\nMichael Tait. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915247c DTSTART;TZID=America/New_York:20251118T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251118T133000 LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Benjamin Stoler CLASS:PUBLIC DESCRIPTION:Speaker: BENJAMIN STOLER\, Ph.D. Candidate\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Towards Robust Auto nomous Driving and Social Robot\nNavigation via Enhanced Data Utilization\ n\nAutonomous robots—including self-driving vehicles\, sidewalk delivery \nrobots\, and more—must navigate among humans in a safe and\nsocially-c ompliant manner. Current approaches for building and\nevaluating such auto nomous systems rely on data-driven techniques\;\nhowever\, a generalizatio n gap emerges\, as methods trained in these\ntraditional paradigms are una ble to cope with unexpected real-world\nscenarios. Therefore\, this thesis aims to develop improved\nmethodologies and evaluation settings to increa se and assess\nrobustness in autonomous navigation against these challenge s\, along\ntwo key pillars of enhanced data utilization.\n\nFirst\, we int roduce scenario characterization and repartitioning\nschemes\, for robustn ess against out-of-distribution safety-relevant\nand corner case scenarios . We create a hierarchical characterization\nmethod which leverages counte rfactual probes to find hidden\nsafety-relevant scenarios in large dataset s. We then address the\ninduced generalization gap by incorporating the ch aracterizations into\ndownstream trajectory prediction models' inductive b iases. To promote\ngreater interpretability and generalizability\, we fact orize scenarios\ninto disentangled contexts\, creating compositionally nov el test sets.\nWe then use modular architectures and auxiliary signals to implicitly\nreason over and adapt to these settings.\n\nSecond\, we design targeted scenario modification approaches\, to expose\nand address failur e cases and weaknesses of naive autonomy methods.\nFor robustness against perception errors affecting downstream motion\nprediction\, we construct a framework for converting top-down\npedestrian trajectory datasets into a more challenging first-person\nview perspective. We then develop a correct ion module to account for\nthe resulting errors\, trained end-to-end with trajectory prediction\napproaches. For robustness against adversarial\, sa fety-critical\nscenarios\, we develop a reactive\, skill-based adversary p olicy which\nleverages a learned\, multi-faceted criticality objective to perturb\nexisting scenarios. We then train ego policies in a closed-loop m anner\nagainst these generated scenarios\, demonstrating improved downstre am\nego performance. Finally\, we process and annotate unlabeled and\nunde rutilized data sources\, to learn human-like behavior from\nreal-world cra sh videos. We use these learned behavior models to\nfurther increase the r ealism of adversarially perturbed scenarios\, as\nwell as the efficacy of closed-loop ego training.\n\nThesis Committee\n\nJean Oh (Chair)\n\nSebast ian Scherer\n\nReid Simmons\n\nJonathan Francis (Bosch Center for Artifici al Intelligence)\n\nIn Person and Zoom Participation.  See announcement.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91529a4 DTSTART;TZID=America/New_York:20251117T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251117T130000 LOCATION:Panthere Hollow Room 4105\, Mehrabian Collaborative Innovation Cen ter SUMMARY:CyLab Seminar - Anupam Datta CLASS:PUBLIC DESCRIPTION:Speaker: ANUPAM DATTA\, Principal Research Scientist\, Snowflak e AI\nResearch Lead\, Snowflake\n\nTalk Title: What is your Agent's GPA? M easuring and Improving Agent\nTrustworthiness\n\nWe introduce the Agent GP A (Goal-Plan-Action) framework: an evaluation\nparadigm based on an agent ’s operational loop of setting goals\,\ndevising plans\, and executing a ctions. The framework includes five\nevaluation metrics: Goal Fulfillment\ , Logical Consistency\, Execution\nEfficiency\, Plan Quality\, and Plan Ad herence. Logical Consistency\nchecks that an agent’s actions are consist ent with its prior\nactions. Execution Efficiency checks whether the agent executes in the\nmost efficient way to achieve its goal. Plan Quality che cks whether an\nagent’s plans are aligned with its goals\; Plan Adherenc e checks if\nan agent’s actions are aligned with its plan\; and Goal Ful fillment\nchecks that agent’s final outcomes match the stated goals. Our \nexperimental results on two benchmark datasets – the public GAIA\ndata set and an internal dataset for a production-grade data agent –\nshow th at this framework (a) provides a systematic way to cover a\nbroad range of agent failures\, including all agent errors on the GAIA\nbenchmark datase t\; (b) exhibits strong agreement between human and LLM\njudges\, ranging from 80% to over 95%\; and (c) localizes errors with\n86% agreement with h uman annotations to enable targeted improvement of\nagent performance.\n\n —\n\nAnupam Datta is a Principal Research Scientist and Snowflake AI\nRe search Lead at Snowflake. He joined Snowflake as part of the\nacquisition of TruEra where he served as Co-Founder\, President\, and\nChief Scientist from 2019-2024. Datta was on the faculty at Carnegie\nMellon University f rom 2007-2022\, most recently as a tenured Professor\nof Electrical & Comp uter Engineering and Computer Science. Datta's\ncurrent research focuses o n Trustworthy AI\, spanning evaluation\,\nexplainability\, fairness\, and adversarial robustness of ML models and\nGenAI applications. Specific resu lts include early work on Shapley\nValues & gradient-based explanations\, fairness assessments\, robustness\nof classical machine learning and deep learning models for natural\nlanguage processing and computer vision\, and the TruLens open source\nproject for evaluation and experiment tracking o f GenAI apps. These\nresearch results have had a significant impact on pro ducts at TruEra\nand Snowflake. Datta has published over 100 research pape rs\, served as\nChair of the National Academies Workshop on Assessing and Improving AI\nTrustworthiness\, on the Steering Committee of of the ACM Co nference on\nFairness\, Accountability\, and Transparency\, and the IEEE C omputer\nSecurity Foundations Symposium\, and as an Editor-in-Chief of\nFo undations and Trends in Privacy and Security.\n\nHe received the 2018 Davi d P. Casasent Outstanding Research Award from\nthe CMU College of Engineer ing\, a 2020 Young Alumni Achiever Award\nfrom IIT Kharagpur\, a 2021 Goog le Faculty Research Award\, and several\nawards for top papers at conferen ces. Datta obtained a B.Tech. from\nIIT Kharagpur\, and Ph.D. and M.S. deg rees from Stanford University in\nComputer Science\, where he currently te aches a course on Trustworthy\nAI.\n\nHost: Michael Lisanti\n\nIn Person and Zoom Participation.  See announcement.\n\n→This CyLab seminar is op en only to partners and Carnegie Mellon\nUniversity faculty\, students\, a nd staff. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9152f9e DTSTART;TZID=America/New_York:20251112T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251112T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Yomou Fei CLASS:PUBLIC DESCRIPTION:Speaker: YOMOU FEI\, Ph.D. Student\, Theory of Computation Grou p\,\nComputer Science and Artificial Intelligence LaboratoryMassachusetts\ nInstitute of Technology\n\nTalk Title: Multi-Pass Streaming Lower Bounds for Approximating\nMax-Cut\n\nIn the Max-Cut problem in the streaming mode l\, an algorithm is given\nthe edges of an unknown graph G = (V\,E) in som e fixed order\, and its\ngoal is to approximate the size of the largest cu t in G. Improving\nupon an earlier result of Kapralov\, Khanna and Sudan\, it was shown by\nKapralov and Krachun that for all ε >0 \, no o(n)  mem ory streaming\nalgorithm can achieve a (1/2 + ε)-approximation for Max-Cu t. Their\nresult holds for single-pass streams\, i.e.~the setting in which the\nalgorithm only views the stream once\, and it was open whether\nmult i-pass access may help. The state-of-the-art result along these\nlines\, d ue to Assadi and N\, rules out arbitrarily good approximation\nalgorithms with constantly many passes and n1-δ space for any δ > 0\n.\n\nWe improv e upon this state-of-the-art result\, showing that any\nnon-trivial approx imation algorithm for Max-Cut requires either\npolynomially many passes or polynomially large space. More\nspecifically\, we show that for all ε > 0\, a k-pass streaming\n(1/2+ε)-approximation algorithm for Max-Cut requi res Ωε(n1/3 /k)\nspace.\n\nBased on joint work with Dor Minzer and Shuo Wang. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91533bf DTSTART;TZID=America/New_York:20251111T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251111T130000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Database Seminar - Connor McArthur CLASS:PUBLIC DESCRIPTION:Speaker: CONNOR McARTHUR\, Co-founder\, dbt Labs\n\nTalk Title: Open Data Infrastructure with Iceberg and dbt\n\nApache Iceberg is now in teroperable with most modern data platforms\nand compute systems. While Ic eberg enables powerful new capabilities\,\nreal-world adoption still prese nts challenges for many organizations.\nIn this talk\, we will unpack Iceb erg's architecture\; demonstrate a\nnovel architecture where multiple comp ute systems connect to the same\nunderlying Iceberg catalog\; and discuss the maturity and continued\ninvestment needed to ensure Iceberg's success. \n\n—\n\nConnor McArthur is the co-founder of dbt Labs. Connor has over a\ndecade of experience as an individual contributor and senior\nengineeri ng leader building enterprise-ready data transformation\ntools. Prior to f ounding dbt Labs\, he worked as a software engineer\nand engineering manag er developing product\, platform\, and\ninfrastructure at RJMetrics. He pr eviously developed the open source\nVU-LEGO real time target enabling stud ents without coding backgrounds\nto develop complex robotics systems. He h olds a bachelor's degree in\nComputer Engineering from Villanova Universit y\, \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9153760 DTSTART;TZID=America/New_York:20251110T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251110T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Cheng Chen CLASS:PUBLIC DESCRIPTION:Speaker: CHENG CHEN\, Co-founder\, Mooncake Labs\n\nTalk Title: Mooncake: Real-Time Apache Iceberg Without Compromise\n\nApache Iceberg i s great for large-scale analytics\, but it was built\nfor batch workloads. For streaming use cases\, keeping tables fresh\nmeans writing snapshots m ore often\, which creates excess small Parquet\nfiles\, bloated metadata\, and costly compaction that never ends.\nUpdates and deletes make things w orse because equality deletes push\nthe burden to query engines\, leaving readers slow and inefficient.\n\nMooncake adds a real-time layer to Iceber g. It supports streaming\nwrites and mirroring from relational databases w ith sub-second\nlatency. It also provides continuous optimization\, cachin g\, and\nindexing for fast\, user-facing analytics\, while remaining fully \ncompatible with the Iceberg spec.\n\n—\n\nCheng Chen is co-founder of Mooncake Labs (recently acquired by\nDatabricks). He now works on the Lake base team at Databricks\, helping\nintegrate Postgres with lakehouse. Prev iously\, he led the Search &\nExtension team at SingleStore and worked on various query execution\nstuffs.\n\nThis talk is part of the Future Data S ystems Seminar Series\n\nZoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9153b36 DTSTART;TZID=America/New_York:20251110T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251110T130000 LOCATION:Panther Hollow Room 4105\, Mehrabian Collaborative Innovation Cent er\nand Zoom SUMMARY:CyLab Seminar - Benedikt Bünz CLASS:PUBLIC DESCRIPTION:Speaker: BENEDIKT BÜNZ\, Assistant Professor of Computer Scien ce\,\nCourant Institute of Mathematical Sciences\, New York University\n\n Talk Title: Hash-based accumulation schemes\n\nAccumulation schemes are pr oof systems that enable combining two or\nmore proofs into one. Despite th eir simplicity\, they have broad and\npowerful applications\, including re cursive proofs\, distributed\nproving\, and streaming provers. Previously\ , accumulation schemes\nrequired homomorphism and could only be built from pre-quantum\nassumptions. I'll be discussing a recent line of work on pos t-quantum\naccumulation schemes that are secure in the random oracle model .   \n\n—\n\nBenedikt Bünz is an Assistant Professor of Computer Scie nce at NYU\nCourant. He also co-founded and is the chief scientist of Espr esso\nSystems. He researches applied cryptography\, consensus and game\nth eory\, especially as it relates to cryptocurrencies. His work focuses\non enhancing the privacy\, usability\, and security of blockchain\nprotocols. \n\nFaculty Host: Elaine Shi\n\nIn Person and Zoom Participation. See anno uncement.\n\nNote: This CyLab seminar is open only to partners and Carnegi e Mellon\nUniversity faculty\, students\, and staff. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9153f14 DTSTART;TZID=America/New_York:20251110T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251110T170000 LOCATION:Posner Hall 151 SUMMARY:Statistics and Data Science Seminar - Philippe Rigollet CLASS:PUBLIC DESCRIPTION:Speaker: PHILIPPE RIGOLLET\, Cecil and Ida Green Distinguished\ nProfessor of Mathematics Core Member\, Institute for Data\, Systems\, and \nSociety\, Massachusetts Institute of Technology\n\nTalk Title: A Mathema tical Perspective on Transformers\n\nSince their introduction in 2017\, Tr ansformers have revolutionized\nlarge language models and the broader fiel d of deep learning. Central\nto this success is the groundbreaking self-at tention mechanism. In\nthis presentation\, I’ll introduce a mathematical framework that\ncasts this mechanism as a mean-field interacting particle system\,\nrevealing a desirable long-time clustering behavior. This persp ective\nleads to a trove of fascinating questions with unexpected connecti ons\nto Kuramoto oscillators\, sphere packing\, Wasserstein gradient flows \,\nand slow dynamics.\n\n—\n\nPhilippe Rigollet is the Cecil and Ida G reen Distinguished Professor\nof Mathematics at MIT and a Core Member of t he Institute for Data\,\nSystems\, and Society. His research explores the foundations of modern\nmachine learning\, with a particular focus on optim al transport and the\nemerging theory of transformer architectures. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91542c2 DTSTART;TZID=America/New_York:20251105T173000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251105T193000 LOCATION:Gates Hillman 4102 SUMMARY:AI & Web3 Panel Discussion CLASS:PUBLIC DESCRIPTION:Speaker: Guest Presenters\n\nHow can Web3 accelerate the incuba tion\, collaboration\, and\ncommercialization of AI?\n\nAnd how AI would h elp facilitate the further use case of Web3? \n\nThis cross-disciplinary panel brings together leaders from blockchain\nfoundations\, Web3 venture acceleration\, AI research\, market mechanism\ndesign to explore how decen tralized compute\, tokenized incentives\, and\nhuman-centered design can f oster a new generation of open source AI\necosystems. The discussion will also examine how agentic AI\nresearch can be integrated into the Web3 la ndscape—enabling\nautonomous coordination\, verifiable provenance\, and collective\nintelligence across networks. \n\nNote:  Participants will r eceive priority access to GPU/LLM credits\nfrom a $1 million funding pool! \n\n► Gui Bibeau\, Head of Education\, Solana Foundation\n\n     —  Gui leads global developer education at Solana\nFoundation\, building pr ograms that empower builders to create\nhigh-performance on-chain applicat ions. His recent initiatives explore\ndecentralized computer and AI-integr ated blockchain ecosystems.\n\n► Dafu G\, General Partner\, OnePiece Lab s (Web3 Accelerator)\n\n     — Dafu is a founding partner and COO at OnePiece Labs\, an\naccelerator focused on supporting early-stage Web3 st artups. He works\nclosely with founders on go-to-market strategy\, token d esign\, and\nachieving product–market fit across blockchain infrastructu re\, AI\napplications\, and decentralized computing. Under his leadership\ ,\nOnePiece Labs has partnered on accelerator programs with leading\necosy stem players such as OKX and Solana Foundation\, and actively\nsupports hi gh-conviction builders at the intersection of AI and\ncrypto. The firm is backed by LBank’s venture arm\, which has\ndeployed over $100M into Bloc kchain × AI opportunities. Past LPs and\nnetwork supporters of OnePiece L abs include top-tier investors such as\nSequoia and Foresight Ventures.\n\ n► Tianqin Li\, Ph.D. Candidate\, Computer Science Department (CSD) \n —  AI & Web3 research \n\n     —  Tianqin researches neural-sym bolic\nAI and computational vision at CMU\, focusing on how artificial\ nsystems can learn to “see” the world through human-like perceptual\no rganization. His work also bridges AI and Web3\, exploring\ndecentralized model marketplaces\, incentive alignment\, and mechanism\ndesign for coll aborative model development and monetization.\n\n► Jiayuan Liu\,  Ph.D . Candidate\, Computer Science Department\n(CSD)  —  Agentic Market Me chanism Design\n\n     — Jiayuan’s research centers on mechanism an d market\ndesign for digital platforms\, focusing on incentive alignment\, auction\ntheory\, and decentralized economies that sustain open AI ecosys tems.\n\n► Shiyi Du\,  Ph.D. Candidate\, Computational Biology Departm ent\n(CBD)  —  AI for Science\n\n     — Shiyi Du’s research foc uses on leveraging AI agents and\ncomputer vision models to accelerate bio logical discovery. Her\nAI-for-science work forms an application layer whe re decentralized\nmechanisms could seamlessly enable collaborative\, verif iable\nscientific intelligence.\n\n► Xueying Ding\,  Ph.D. Candidate\, Machine Learning Department &\nPublic Policy  —  Transparency and Exp lainability in AI\n\n    —Xueying’s work on outlier detection and ex plainability of\nAI models directly addresses the transparency and account ability\nchallenges of modern AI systems — especially those used in\ngov ernment\, healthcare\, or finance.\n\n► Eric Lin\, HCI Design Program\, EXR Design Research Fellow\n\n     — Eric is an HCI design student and EXR Design Research\nFellow. His research focuses on human-centered de sign\, exploring how\ntechnology can better align with human behavior\, co gnition\, and\nemotion to create meaningful interactive experiences. In th e future of\nAI & Web3 space\, user centered designing will play a critica l role in\ndetermining what technology should be useful. \n\nFaculty Host :  Tai-Sing Lee \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9154916 DTSTART;TZID=America/New_York:20251111T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251111T120000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Thesis Proposal - Andy Zou CLASS:PUBLIC DESCRIPTION:Speaker: ANDY ZOU\, Ph.D. Student\, Computer Science Department \,\nCarnegie Mellon University\n\nTalk Title: Improving Security and Safet y of Generative Models\n\nGenerative models now mediate information access \, software\ndevelopment\, and mission critical workflows\, yet their secu rity and\nsafety properties lag their rapid deployment. This thesis develo ps a\ncomprehensive approach to improving the safety of aligned language a nd\nagentic systems. \n\nFirst\, we show that alignment fine tuning leave s structural\nvulnerabilities: the Greedy Coordinate Gradient attack learn s\nuniversal and transferable suffixes that trigger harmful behaviors\nacr oss open source models\, achieving high transfer rates to\nproprietary mod els and revealing shared non-robust features in model\nrepresentations. Se cond\, we advance security measurements that span\nstandardized and live e valuation. HarmBench establishes reproducible\nstatic robustness benchmark s\, while the Gray Swan Arena and the\nresulting Agent Red Teaming benchma rk capture human\, adaptive\nadversaries whose discoveries continually ref resh static tests.\nTogether they demonstrate the fragility of current age nts and provide\na continuous feed of vulnerabilities. Third\, we introduc e\nRepresentation Engineering (RepE)\, a new class of approaches that\npro be and control population-level representations encoding\nsafety-relevant concepts. \n\nWe apply RepE methods to safety-critical concepts such as h onesty and\nharmfulness. In particular\, we present Circuit Breaking\, an alignment\nalgorithm which suppresses harmful thought processes in the\nre presentation space to combat adversarial misuse. Looking forward\, we\nwil l continue scaling the capabilities of automated red teaming agents\nand d evelop environments that allow for co-evolution of attacker and\ndefense a gents. For mitigation at the model representation level\, we\nplan to exte nd RepE monitoring to contextual policy violations. We\nbelieve that treat ing safety as a property of training\, evaluation\,\nand internal computat ion yields principled mechanisms for securing\ngenerative systems.\n\nThes is Committee\n\nMatt Fredrikson (Co-Chair)\n\nZico Kolter (Co-Chair)\n\nGr aham Neubig\n\nNicholas Carlini (Anthropic)\n\nAdditional Information \n\ n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9154e0c DTSTART;TZID=America/New_York:20251104T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251104T130000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Database Seminar - Ankit Sultana CLASS:PUBLIC DESCRIPTION:Speaker: ANKIT SULTANA\, Staff Engineer\, Uber and PMC\, Apache Pinot\n\nTalk Title: Real Time Analytics Query Architecture Evolution @ U ber\n\nWe will talk about how Apache Pinot's query feature set has grown\n tremendously over the past few years and how that growth has shaped\nUber' s Real Time Analytics Query Architecture. We will dive into the\ndifferent query engines in Apache Pinot and briefly discuss our legacy\nand unique Presto over Pinot architecture.\n\n—\n\nAnkit Sultana is a Staff Eng at Uber and a PMC of Apache Pinot.\nOutside of work\, he sporadically blogs a bout systems/databases. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91551a5 DTSTART;TZID=America/New_York:20251107T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251107T143000 LOCATION:Remote Access - Zoom SUMMARY:STAMPS Research Center Seminar - Natalie Klein CLASS:PUBLIC DESCRIPTION:Speaker: NATALIE KLEIN\, AI and Advanced Predictive Modeling Te am Lead\,\nStatistics Group\, Los Alamos National Laboratory\n\nTalk Title : From Earth to Mars: Statistical Challenges in Analyzing\nRover Spectrosc opy Data\n\nNASA’s Curiosity and Perseverance rovers have collected rich \nspectroscopic data from the Martian surface using instruments such as\nC hemCam and SuperCam. These multimodal datasets (spanning LIBS\,\ninfrared\ , and Raman measurements) pose unique challenges for\ncalibration\, interp retation\, and data integration across vastly\ndifferent environments. Thi s talk will highlight statistical and\nmachine learning methods developed to meet these challenges\, including\nBayesian neural networks for uncerta inty-aware prediction\, optimal\ntransport for aligning Earth and Mars dat a\, multimodal fusion with\ninterpretability metrics\, and density-ratio w eighting for combining\nheterogeneous observations. I’ll also discuss ge nerative models for\nLIBS spectra and ongoing work using fast simulators f or model\npretraining. Together\, these advances illustrate how planetary science\ndata drive new ideas in uncertainty quantification\, domain adapt ation\,\nand the fusion of physical and statistical modeling.\n\n—\n\nDr . Natalie Klein is the AI and Advanced Predictive Modeling Team Lead\nin t he Statistics Group at Los Alamos National Laboratory\, where she\nhas bee n a staff member since 2019. Her research focuses on\nintegrating statisti cal methodology with machine learning to address\nchallenges in scientific domains such as remote sensing and planetary\nexploration. She holds a jo int Ph.D. in Statistics and Machine\nLearning from Carnegie Mellon Univers ity.\n\nZoom Participation.  See announcement. \n\nJoin the STAtistical Methods for the Physical Sciences (STAMPS)\nSeminar mailing list. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91555bc DTSTART;TZID=America/New_York:20251105T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251105T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Natalie Collina CLASS:PUBLIC DESCRIPTION:Speaker: NATALIE COLLINA\, Ph.D. Student in Computer Science\,\ nDepartment of Computer and Information Sciences\, University of\nPennsylv ania\n\nTalk Title: Learning and Incentives in Human–AI Collaboration\n\ nAs AI systems become more capable\, a central challenge is designing\nthe m to work effectively with humans. I will first consider\ncollaborative pr ediction\, motivated by a doctor consulting an AI that\nshares the goal of accurate diagnosis. Even when the doctor and AI\nhave only partial and in comparable knowledge\, repeated interaction\nenables richer forms of colla boration: we give distribution-free\nguarantees that their combined predic tions are strictly better than\neither alone\, with regret bounds against benchmarks defined on their\njoint information. I will then revisit the al ignment assumption\nitself. If an AI is developed by\, say\, a pharmaceuti cal company with\nits own incentives\, how can we encourage helpful behavi or? A natural\nscenario is that the doctor has access to multiple models\, each from a\ndifferent provider. Under a mild ‘market alignment’\nass umption—that the doctor’s utility lies in the convex hull of\nthe prov iders’ utilities—we show that in Nash equilibrium of this\ncompetition \, the doctor can achieve the same outcomes as if a\nperfectly aligned pro vider were present.\n\nBased on joint work: Tractable Agreement Protocols (STOC’25)\,\nCollaborative Prediction (SODA’26)\, and Emergent Alignme nt via\nCompetition (in submission). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9155997 DTSTART;TZID=America/New_York:20251107T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251107T110000 LOCATION:Newell-Simon 4305 SUMMARY:Special Seminar - Anurag Khandelwal CLASS:PUBLIC DESCRIPTION:Speaker: ANURAG KHANDELWAL\, Assistant Professor\, Department o f\nComputer Science\, Yale University\n\nTalk Title: Performant Disaggrega ted Shared-Memory\n\nRecent improvements in intra-rack interconnect techno logies have\ndriven advances in data center resource disaggregation\, prom ising\nbetter resource utilization\, support for hardware heterogeneity\, and\napplication resource elasticity. However\, actualizing these benefits \nwhile ensuring application performance requires support from the\noperat ing system (OS). Unfortunately\, existing approaches expose a\nhard tradeo ff between application performance\, resource elasticity\,\nand applicatio n transparency.\n\nOur approach to navigating this tradeoff space argues f or a ground-up\nredesign of the OS stack\, central to which is a performan t shared\nmemory abstraction. In this talk\, I will first talk about MIND\ , our\nmemory management subsystem for disaggregated architectures that\np laces its logic in the interconnect fabric to enable efficient\,\ndisaggre gated shared memory\, achieving performance and resource\nelasticity for r eal-world workloads without application modifications.\nI will also descri be CORD\, an extension of MIND’s in-network\napproach to enable efficien t release consistency for disaggregated\nmemory that is being incorporated into real-world hardware coherence\ninterconnects. I will then present Sp irit\, a multi-user framework for\nfair resource allocation that addresses a challenge unique to\ndisaggregated memory systems — the interdependen ce between cache and\ndisaggregated memory bandwidth resources\, where lar ger cache\nallocations can reduce an application’s need for memory bandw idth\,\nand vice versa. To this end\, Spirit employs a novel algorithm tha t\ntakes application-specific dependency between cache and network\nbandwi dth into account and ‘trades’ cache and bandwidth resources\nacross us ers at runtime to guarantee fairness.  Finally\, I will\npresent some of our ongoing work on developing other OS abstractions\nfor disaggregated ar chitectures.\n\n—\n\nAnurag Khandelwal is an Assistant Professor of Comp uter Science at\nYale University\, where his group focuses broadly on prob lems in\ncomputer systems and networks. He received his Ph.D. from UC Berk eley\,\nwhere he was advised by Ion Stoica. His work has been recognized b y an\nNSF CAREER award\, two NetApp faculty fellowships\, a Roberts Innova tion\nAward\, best paper awards at USENIX Security’20\, ISCA’23\,\nEur oSys’24\, ISCA’25\, and an IEEE Micro Top Picks selection in\n2024.\n\ nFaculty Host: Rashmi Korlakai Vinayak \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9155e18 DTSTART;TZID=America/New_York:20251103T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251103T140000 LOCATION:Newell-Simon 3001 and Zoom SUMMARY:Accessibility Lunch Seminar - Douglas Weber CLASS:PUBLIC DESCRIPTION:Speaker: DOUGLAS WEBER\, Akhtar and Bhutta Professor\, Departme nt of\nMechanical Engineering and Neuroscience Institute Carnegie Mellon\n University\n\nTalk Title: Neurotechnology for detecting and amplifying mot or\nfunction after stroke and spinal cord injury\n\nThe convergence of adv ancements in fundamental neuroscience and neural\ninterface technologies h ave enabled clinically relevant technologies\nthat measure and regulate ne ural signaling in the brain\, spinal cord\,\nand peripheral nerves. These technologies provide new capabilities for\nstudying basic mechanisms of in formation processing and control in the\nnervous system\, while also creat ing new opportunities for restoring\nfunction lost to injury or disease. N eural sensors can measure the\nactivity of motor neurons to enable direct neural control over\nprosthetic limbs and assistive technologies. Converse ly\, these neural\ninterface technologies can stimulate activity in sensor y and motor\nneurons to reanimate paralyzed muscles. Although many of thes e\napplications rely currently on devices that must be implanted into the\ nbody for precise targeting\, ultra-miniaturized devices can be injected\n through the skin or vascular system to access deep structures without\nope n surgery. This talk will focus on efforts to develop wearable and\ninject able neural interfaces for restoring or improving motor function\nin peopl e with paralysis due to stroke\, spinal cord injury\, ALS\, and\nother neu rological disorders.\n\nAdditional Info: CMU's NeuroMechatronics Labs\n\nI n Person and Zoom Participation.  See announcement.\n\nAbout the Accessi bility Research Group \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9156200 DTSTART;TZID=America/New_York:20251106T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251106T150000 LOCATION:Scaife Hall 105 SUMMARY:Course Guest Lecturer - Rolfe Schmidt CLASS:PUBLIC DESCRIPTION:Speaker: ROLFE SCHMIDT\, Research Engineer\, Signal Foundation\ n\nTalk Title: Tutorial on the Signal protocol\n\nThursday\, November 6\, 2025\, 2 – 3pm \n\nRolfe Schmidt will give a virtual guest lecture in the  networking\nclass. He will offer a tutorial on the Signal protocol and then\nprovide some thoughts on a post-quantum future. The talk should be\naccessible to a general CS audience\, so this is a great place to lear n\nabout state-of-the-art applied crypto.\n\nFaculty Host:  Justine Sherr y \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915651e DTSTART;TZID=America/New_York:20251103T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251103T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Ryan Johnson CLASS:PUBLIC DESCRIPTION:Speaker: RYAN JOHNSON\, Principal Engineer\nDatabricks\n\nTalk Title: Multi-statement Transactions in the Databricks Lakehouse\n\nRyan Jo hnson is Distinguished CMU-DB alumni and completed his\ndoctorate in 2010. \n\nAdditional Information: Delta Lake \n\nThis talk is part of the Futu re Data Systems Seminar Series.\n\nZoom Participation.  See announcement.  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9156827 DTSTART;TZID=America/New_York:20251114T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251114T150000 LOCATION:NTU attendees in Person\, other participants via Zoom SUMMARY:AI-SDM: Forum on Artificial Intelligence - AI Education and Communi ty CLASS:PUBLIC DESCRIPTION:Talk Title: Forum on Artificial Intelligence: AI Education and\ nCommunity\n\nFriday\, November 14\, 2025\, 11am – 3pm \n\nJoin in fo r a forum hosted by Navajo Technical University (NTU) in\ncollaboration wi th the NSF AI Institute for Societal Decision Making\n(AI-SDM) - a Forum on Artificial Intelligence: AI Education and\nCommunity.\n\nThis half-day event is designed for anyone interested in artificial\nintelligence\, the link between technology and communities\, and the\nfuture of responsible A I development. It will also feature keynote\nspeaker Dr. Robin Murphy\, th e founder of the field of rescue robotics\nand Professor Emeritus at Texas A&M.\n\nThis forum will focus on the responsible development and applicat ion\nof AI and feature two key pillars:\n\nHuman-Centric AI Research: Disc over how NTU faculty\, staff\, and\nstudents—supported by AI-SDM—are t ackling community challenges\nusing AI. Projects will be showcased in key areas like\ndrone technology\, language/cultural preservation\, and\nentr epreneurship. AI Education and Community: Join discussions and\nhear pane ls on how AI can serve all people. Learn about plans for\nimplementing new AI education courses and curricula aimed at building\nAI literacy. See ho w education can empower students and community\nmembers to direct these ne w technologies toward local needs and\ncommunity growth.\n\nTime Considera tons:  9:00 am -1:00 pm MST   |  11:00 am-3:00 pm\nET\n\nREGISTER → for both in person and zoom participation \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9156c17 DTSTART;TZID=America/New_York:20251121T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251121T130000 LOCATION:Wean Hall 1327 (New Date) SUMMARY:AI-SDM: Student Brainstorming Session CLASS:PUBLIC DESCRIPTION:Speaker: Brainstorming session on mental health\, AI tools\, an d the\nethics surrounding that intersection.\n\n \n\nStudents meet regula rly to participate in informal discussion\nsessions that delve into cutti ng-edge AI topics. These regular\nmeetings throughout the semester provide a platform for students to\ndeepen understanding of specific areas and br oaden knowledge by\nexploring cross-cutting connections between various AI disciplines.\nDiscussions are a breeding ground for collaboration\, innov ative\nthinking\, and problem-solving from the ground up. They provide a\n stress-free forum for exchanging ideas\, brainstorming new approaches\nto challenges\, and fostering lasting connections within the AI-SDM\ncommunit y in an environment distinct from a traditional seminar.\n\nREGISTER → register if you plan to attend in person and via zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9156f6d DTSTART;TZID=America/New_York:20251103T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251103T150000 LOCATION:Remote Access - Zoom SUMMARY:Special Talk - Marc Brooker CLASS:PUBLIC DESCRIPTION:Speaker: MARC BROOKER\, Vice President and Distinguished Engine er\nAWS\n\nTalk Title: Transactions and Coordination in Aurora DSQLMonday\ ,\nNovember 3\, 2025\, 2 – 3pm\n\nAurora DSQL is a new global\, server less\, scalable relational database\nsystem\, built at AWS. In this talk\, I’ll dive into the architecture\nof DSQL\, how it handles transactions\ , and how and why it was designed\nto minimize coordination. We’ll touch on transaction protocols\,\nisolation\, and virtualization.\n\n—Marc Br ooker is a VP and Distinguished Engineer at AWS. During his\n16 years at A WS\, Marc has worked on EC2\, EBS\, Lambda\, and most\nrecently led the te am that launched Amazon Aurora DSQL. He is\ncurrently focused on infrastru cture for agentic AI\, and the\navailability and security of our large-sca le systems. Before AWS\, Marc\ncompleted his PhD at the University of Cape Town.Faculty Host: Jignesh\nPatel\n\nZoom Participation\n\nEvent Type: T alks Room Number: Virtual Presentation - ET\nBuilding: Remote Access - Z oom Speaker's Name: MARC BROOKER Speaker\nWebsite: www.brooker.co.za [ht tp://www.brooker.co.za]… Speaker's\nProfessional Title: Vice President and Distinguished Engineer\, AWS\nTalk Title: Transactions and Coordinati on in Aurora DSQL\nAffiliations: Computer Science Department (CSD)\, Mach ine Learning\nDepartment (MLD)\, Software and Societal Systems Department (S3D)\nOrganization(s): School of Computer Science\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9157360 DTSTART;TZID=America/New_York:20251106T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251106T130000 LOCATION:Registration Required SUMMARY:AI-SDM: Practitioner Lunch and Learn CLASS:PUBLIC DESCRIPTION:Speaker: National Weather Service ExpertsTalk Title: Practition er\nLunch & Learn\n\nAre you developing AI in industries like weather\, cl imate\, or\nemergency/disaster response? Do you want your research to have a\nreal-world impact?\n\nJoint in at the Practitioner Lunch & Learn\, whe re you can connect\ndirectly with experts who could deploy that technology every day:\nmeteorologists from the National Weather Service (NWS) Foreca st\nOffice.\n\nThe panelists will discuss:\n\nReal-World Pain Points: Disc over the actual problems in the NWS where\ncurrent technology falls short. AI Opportunities: Hear firsthand where\nyour AI\, Machine Learning\, and D ata Science skills can make the\nbiggest difference in fields like disaste r forecasting and emergency\nmanagement.Career Insights: Learn what the fu ture of working in this\ncritical field looks like.\n\nBy attending\, you can ensure your AI innovations are useful and\ndeployable. This direct lin k to practitioners is invaluable for any\nstudent working in the field of AI.\n\nBonus: Free lunch will be provided for all in-person attendees!\n\n REGISTER and Submit any questions you have for the NWS panelists to\nhelp shape the conversation.\n\n→ Space is limited! \n\nDon't miss this opp ortunity to transform your theoretical knowledge\ninto practical solutions . \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915777a DTSTART;TZID=America/New_York:20251030T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251030T173000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - George Lu CLASS:PUBLIC DESCRIPTION:Speaker: GEORGE LU\, Ph.D. Student\, Department of Computer Sci ence\,\nUniversity of Texas at Austin\n\nTalk Title: Succinctness and Adap tivity in Computational Secret\nSharing\n\nSecret sharing (SS) is a founda tional cryptographic primitive with\ndiverse applications\, including secu re multiparty computation and\nconditional disclosure of secrets. While tr aditional schemes have\nprimarily emphasized information-theoretic securit y\, recent\nadvancements have increasingly leveraged computational assumpt ions to\nachieve more efficient constructions and support broader access\n policies. In this talk\, I will present two results:\n\nFirst\, I will pre sent a new construction of CSS for monotone circuits\nwhich is succinct\, where the share size depends only on the number of\nparties and is indepen dent of the circuit size. Our scheme is the\nfirst to achieve this level o f succinctness and relies on\nindistinguishability obfuscation and one-way functions.\n\nSecond\, I will describe a general transformation that enha nces any\nstatically secure CSS scheme to adaptive security. This transfor mation\nuses only one-way functions and preserves the access policies of t he\noriginal scheme\, giving the first adaptively secure CSS schemes for\n several policy classes previously not known from the same assumptions\nas the corresponding static CSS schemes.\n\n—\n\nGeorge Lu is a PhD student in cryptography at UT Austin advised by\nBrent Waters. He is broadly inte rested in theoretical crypto with a\nfocus on advanced encryption and foun dations.\n\nIn Person and Zoom Participation.  See announcement. \n\n \ n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9157bab DTSTART;TZID=America/New_York:20251105T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251105T150000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Anup Agarwal CLASS:PUBLIC DESCRIPTION:Speaker: ANUP AGARWAL\, Ph.D. Candidate\nComputer Science Depar tment\nCarnegie Mellon University\n\nTalk Title: Designing Network Control Algorithms with Performance\nGuarantees\n\nControl algorithms are ubiquit ous in networked systems—from\ncongestion control and load balancing to scheduling and caching.\nDespite their performance-critical nature\, these algorithms are\ndesigned using human intuition and heuristics\, and they frequently\nexhibit poor or unpredictable performance. This thesis envisio ns a\nmethodology for designing controllers with formally verified\nperfo rmance guarantees. We focus on congestion control algorithms\n(CCAs)—a d omain that continues to experience repeated failures\ndespite decades of r esearch.\n\nTwo main reasons make congestion control hard. First\, CCAs op erate in\ndiverse and noisy environments (e.g.\, cellular links\, policers \,\ntoken-bucket filters\, operating system jitter). Second\, they operate \nunder uncertainty—they lack direct visibility into the state of the\nn etwork or the flows they compete with. Recent work showed that we can\nmod el networks as non-deterministic\, non-stochastic automatons to\ncapture a wide range of real-world phenomena and formally verify\ncontroller perfor mance on such environments. We seek to design CCAs\nthat pass such verific ation checks. However\, this does not scale out\nof the box.\n\nWe find th at the key to making it tractable is to formally reason\nabout uncertainty in the state of the network and other flows. This\nthesis contributes two abstractions—beliefs and contracts—that\nenable such reasoning and re veal new structure in CCAs that simplifies\ntheir design and analysis. Bel iefs formalize what a CCA can infer\nabout latent network state from its o bservations. Contracts formalize\nhow flows coordinate with each other to share the network. Since flows\ncannot directly communicate\, they implici tly encode information in\nobservable congestion signals (e.g.\, delay or loss). Contracts\nformalize these communications mechanisms. Building on t hese\nabstractions\, we develop CCmatic\, a tool that automatically\nsynth esizes CCAs with verified performance guarantees. Our\nabstractions and to ols allow us to discover previously unknown\ntradeoffs\, design new CCAs t hat are on the Pareto-frontier\, and\nprovably guarantee performance even under challenging network\nconditions.\n\nThesis Committee\n\nSrinivasan S eshan (Chair)\n\nVyas Sekar\n\nJustine Sherry\n\nPhilip Brighten Godfrey ( University of Illinois Urbana-Champaign)\n\nVenkat Arun (University of Tex as at Austin)\n\nIn Person and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9158079 DTSTART;TZID=America/New_York:20251113T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251113T133000 LOCATION:Newell Simon 4305 and Zoom SUMMARY:AI-SDM Seminar - Akshay Krishnamurthy CLASS:PUBLIC DESCRIPTION:Speaker: AKSHAY KRISHNAMURTHY\, Senior Principal Research Manag er\,\nMicrosoft Research\, NYC\n\nTalk Title: The Coverage Principle in La nguage Models: From\nPre-Training to Test-Time Scaling\n\nTest-time comput e has emerged as a new axis for scaling language model\ncapabilities\, yet we lack a principled understanding of this paradigm.\nWhat are the right algorithms and trade-offs for test-time scaling?\nWhat properties of the p re-trained model enable it? And can we better\nalign pre-training recipes for test-time success?\n\nThis talk addresses these questions through a un ified lens of\ncoverage. We first show that test-time scaling strategies l ike\nbest-of-N sampling succeed if and only if the pre-trained model has\n coverage over high-quality responses. We then demonstrate that\ncoverage\, and hence best-of-N performance\, can be improved through\ndeliberate exp loration\, either purely at test time or via RL-style\npost-training. Fina lly\, we ask why pre-training via next-token\nprediction yields models wit h good coverage in the first place. We\nuncover a rich theoretical landsca pe driven by an implicit bias of the\nnext-token prediction objective\, wh ile also identifying a fundamental\nmisalignment between next-token predic tion and coverage\, raising the\npossibility of future algorithmic innovat ions.—\n\n—\n\nAkshay Krishnamurthy is a senior principal research man ager at\nMicrosoft Research\, New York City. Previously\, he spent two yea rs as\nan assistant professor in the College of Information and Computer\n Sciences at the University of Massachusetts\, Amherst and a year as a\npos tdoctoral researcher at Microsoft Research\, NYC. Before that\, he\ncomple ted his PhD in the Computer Science Department at Carnegie\nMellon Univers ity.\n\nHe is broadly interested in foundational aspects of machine learni ng\nwith a focus on interactive decision making\, reinforcement learning\, \nand\, more recently\, language modeling and generative AI. .\nPreviousl y\, he spent two years as an assistant professor in the\nCollege of Inform ation and Computer Sciences at the University of\nMassachusetts\, Amherst and a year as a postdoctoral researcher at\nMicrosoft Research\, NYC. Befo re that\, he completed my PhD in the\nComputer Science Department at Carne gie Mellon University\, advised by\nAarti Singh. I received my undergradua te degree in EECS at UC\nBerkeley.\n\nHis research interests are in machin e learning and statistics. He is\nexcited about interactive learning\, or learning settings that involve\nfeedback-driven data collection. His recen t interests revolve around\ndecision making problems with limited feedback \, including contextual\nbandits and reinforcement learning.\n\nREGISTER  → register to attend in person or on Zoom.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9158529 DTSTART;TZID=America/New_York:20251103T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251103T163000 LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Xun Huang CLASS:PUBLIC DESCRIPTION:Speaker: XUN HUANG\, Founder &amp\; CEOStealth Startup\n\nTalk Title: From Video Generation to Video World Models\n\nVideo diffusion mode ls have achieved remarkable success in content\ncreation\, yet they still fall short of simulating interactive worlds\nthat respond to users in real time. This talk examines the fundamental\nchallenges preventing these mod els from evolving into true world\nsimulators. I will present a series of works — CausVid\,\nSelf-Forcing\, MotionStream\, and State-Space World M odel — that\ncollectively mark a paradigm shift from non-causal diffusio n models to\nautoregressive–diffusion hybrids capable of streaming long- duration\nvideos with real-time interactivity. These advances move beyond\ npassive video generation toward dynamic\, immersive experiences\,\nunlock ing new possibilities across gaming\, robotics\, live video\nediting\, and augmented/virtual reality.\n\n—\n\nXun Huang was a Research Scientist a t Adobe\, NVIDIA\, as well as an\nAdjunct Professor at Carnegie Mellon Uni versity. He is currently the\nFounder and CEO of a stealth startup. He obt ained his Ph.D. from\nCornell University in 2020 under the advisement of P rofessor Serge\nBelongie. His doctoral research was recognized with the Fe llowship\nfrom NVIDIA\, Adobe\, and Snap. His research interests lie broad ly in\ndeep generative models\, with a recent focus on video world models. \n\nThe VASC seminar is generously sponsored by HeyGen\, an all-in-one\nAI -powered video generation platform that leverages advances in\ncomputer vi sion\, generative modeling\, and multimodal learning to make\nhigh-quality video creation both scalable and accessible.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915891b DTSTART;TZID=America/New_York:20251029T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251029T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Ruoxu Chen CLASS:PUBLIC DESCRIPTION:Speaker: RUOXU CHEN\, Ph.D. Student\, Department of Computer Sc ience\,\nDuke University\n\nTalk Title: Network Unreliability in Almost-Li near Time\n\nWednesday\, October 29\, 2025\, 12 – 1pm \n\nThe network unreliability problem asks for the probability that a\ngiven undirected g raph gets disconnected when every edge independently\nfails with a given p robability. Valiant (1979) showed that this\nproblem is #P-hard\; therefor e\, the best we can hope for are\napproximation algorithms. In a classic r esult\, Karger (1995) obtained\nthe first FPTAS for this problem by levera ging the fact that when a\ngraph disconnects\, it almost always does so at a near-minimum cut\, and\nthere are only a small (polynomial) number of n ear-minimum cuts. Since\nthen\, a series of results have obtained progress ively faster\nalgorithms to a quadratic-time algorithm (Karger\, 2020). In a prior\nwork (Cen\, He\, Li\, and Panigrahi\, 2024)\, we improved the ru nning time\nto m1+o(1) + Ô(n3/2). In this talk\, I will discuss our recen t result\nthat obtains an almost-linear time algorithm for the network\nun reliability problem. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9158c88 DTSTART;TZID=America/New_York:20251031T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251031T120000 LOCATION:Newell-Simon 4305 SUMMARY:Doctoral Speaking Skills Talk - Valerie Choung CLASS:PUBLIC DESCRIPTION:Speaker: VALERIE CHOUNG\, Ph.D. Student\nComputer Science Depar tment\nCarnegie Mellon University\n\nTalk Title: Casma: Addressing Memory Bottlenecks with\nCompiler-Assisted Dynamic Memory Allocation\n\nApplicati ons are increasingly bottlenecked by the memory wall. \nOptimized memory layouts have been shown to alleviate this bottleneck\,\nbut memory allocat ors typically map data objects to memory addresses\nwithout knowledge of f uture access patterns\, often resulting in\nsuboptimal layout decisions. H eterogeneous memory systems can also\nmitigate memory bottlenecks\, but th e lack of seamless and effective\nsupport from memory allocators adds comp lexity for application\nprogrammers.\n\nIn this talk\, I will present Casm a\, a general framework for automatic\ncompiler-assisted dynamic memory al location. Using Casma\, novel\ncompiler passes analyze data objects' acces s patterns\, enabling the\nmemory allocator to make smarter layout decisio ns. I will be\nhighlighting two applications of Casma\, which aim to maxim ize the\nbenefits of caching as well as heterogeneous memory.\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915900f DTSTART;TZID=America/New_York:20251024T143000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251024T153000 LOCATION:Tepper Building 1403 SUMMARY:Robotics Seminar - Nancy Pollard CLASS:PUBLIC DESCRIPTION:Speaker: NANCY POLLARD\, Professor Robotics Institute and Compu ter\nScience DepartmentCarnegie Mellon University\n\nTalk Title: Bringing Dexterity to Robot Hands in the Real World\n\nDexterous manipulation is a grand challenge of robotics\, and fine\nmanipulation skills are required f or many robotics applications that\nwe envision.   \n\nIn this overview talk\, I will discuss my view of some major factors\nthat contribute to de xterity and discuss how we can incorporate them\ninto our robots and syste ms.\n\n—\n\nNancy Pollard is a Professor in the Robotics Institute and t he\nComputer Science Department at Carnegie Mellon University. She\nreceiv ed her PhD in Electrical Engineering and Computer Science from\nthe MIT Ar tificial Intelligence Laboratory\, where she developed grasp\nand manipula tion planning algorithms for the Stanford/JPL and Utah/MIT\ndexterous hand s. Prof. Pollard spent the next few decades studying\nhuman and robot dext erity\, with emphasis on bringing human\nmanipulation strategies with perf ormance guarantees to humanoid robots\nwith dexterous hands.  She receive d the NSF CAREER award for research\non “Quantifying Humanlike Envelopin g Grasps”\,  the Okawa Research\nGrant for \"Studies of Dexterity for C omputer Graphics and Robotics\,\"\nand was a recent recipient of an NSF Co nvergence Accelerator award for\n\"Bio-Inspired Design of Robot Hands for Use-Driven Dexterity.\"   She\nhas led the development of several genera tions of dexterous soft\nrobotic hands\, is a founder of FuturHand Robotic s and leads the CMU\nFoam Hands Laboratory. \n\nSeminars are record (with speaker permission) and accessible here. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9159406 DTSTART;TZID=America/New_York:20251027T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251027T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Joyo Li CLASS:PUBLIC DESCRIPTION:Speaker: JOYO VICTOR\, Chief Architect\, SingleStore\n\nTalk Ti tle: Storage Metadata for Modern Cloud Databases\n\nIn modern database arc hitecture\, separating compute from storage\nunlocks powerful capabilities . Our tiered storage\, “bottomless”\,\nstarted by uploading files to r emote object storage. This worked well\nuntil we wanted to create database branches pointing to the same\nremote storage. One branch does not know i f it can delete a file that\nanother branch depends on.\n\nTo solve this\, we built Bottle Service\, a metadata service that tracks\nfile references across branches\, essentially a distributed\nref-counting system. But the n we realized this service could be\nleveraged for more than garbage colle ction. It became the foundation\nfor advanced features like cross-region b ranching\, disaster recovery\,\nchange data capture (CDC) out\, and Apache Iceberg egress.\n\nThis talk will walk through the problem\, the design o f Bottle Service\,\nand how it empowers SingleStore to build modern cloud- native\nfunctionality.\n\n—\n\nJoyo Li is the Chief Architect at SingleS tore\, and has been with the\ncompany for over 11 years. Joyo worked exten sively on Separation of\nStorage and Compute. \n\nThis talk is part of th e Future Data Systems Seminar Series.\n\nZoom Participation. See announcem ent.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91597e6 DTSTART;TZID=America/New_York:20251107T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251107T170000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:Center for AI-Driven Biomedical Research Seminar - Jure Leskovec CLASS:PUBLIC DESCRIPTION:Speaker: JURE LESKOVEC\, Professor\, Computer Science Departmen t\,\nStanford University\n\nTalk Title: From Relational Foundation Models to the AI Virtual Cell:\nBuilding the Next Generation of Intelligent Syste ms\n\nFoundation Models have transformed learning from unstructured data\, \nbut the world’s structured and relational data—driving decisions\nin enterprises and science—remain underutilized. I will present\nRelationa l Foundation Models (RFMs)\, a new class of pre-trained models\nthat enabl e in-context learning over relational data\, predicting\noutcomes such as engagement\, churn\, or fraud without supervision or\nfeature engineering. RFMs combine table-agnostic embeddings\,\nrelational transformers\, and S QL-like prompting to turn predictive AI\nfrom model-building into querying . Extending these ideas to biology\, I\nwill introduce the AI Virtual Cell (AIVC)—a large neural model that\nrepresents and simulates molecular an d cellular behavior\, offering a\nnew paradigm for data-driven biological discovery. Finally\, I will\ndiscuss Biomni\, a general-purpose biomedical AI agent that\nautonomously plans and executes complex scientific workflo ws across\ndiverse biomedical domains\, acting as a virtual AI biologist t hat\naccelerates research and discovery. Together\, these efforts outline a\nvision for foundation models that reason\, simulate\, and collaborate\n across structured\, relational worlds—from enterprises to living\ncells. \n\nFaculty Host:  Jian Ma \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9159bb0 DTSTART;TZID=America/New_York:20251023T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251023T160000 LOCATION:CNA Room\, Wean Hall 7218 SUMMARY:ACO Seminar - Dylan Langharst CLASS:PUBLIC DESCRIPTION:Speaker: DYLAN LANGHARST\, Postdoctoral Researcher\, Department of\nMathematical Sciences\, Carnegie Mellon University\n\nTalk Title: Gr ünbaum’s inequality for probability measures\n\nGiven a body (compact\, connected set with non-empty interior) K in\nn-dimensional  Euclidean sp ace\, a natural question is: if one\npartitions the body into two pieces a long its barycenter\, how small\ncan each piece be? By “partition along its barycenter”\, we mean \nintersecting K with a half-space whose boun dary is a hyperplane\ncontaining said barycenter. An easy observation is t hat\, if K is\nsymmetric about a point\, then each piece will have (1/2) t he total\nvolume.\n\nGrünbaum showed that\, if K is convex\, then the vol ume of each piece\nis at least (n/(n+1))n times the total volume of K. Fur thermore\, this\nconstant is sharp: there is equality if and only if K is a cone\, which\nmeans there exists a (n − 1)-dimensional convex body L a nd a vector\nb\, such that K has face L and vertex b (we say K is the conv ex hull of\nb and L). Notice the number (n/(n+1))n is greater than (1/e)\, and in\nfact approaches it as the dimension goes to infinity. That is\, t he\ngeneral situation\, using constant (1/e)\, is not much worse than the\ nsymmetric case.\n\nIn this work\, which is joint with M. Fradelizi\, J. L iu\, F. Marin Sola\,\nand S. Tang\, we are interested in generalizing Grü nbaum’s\ninequality to other measures. Our main results are a sharp ineq uality\nfor the Gaussian measure and a sharp inequality for s-concave\npro bability measures. The characterization of the equality case is of\npartic ular interest.\n\n4:00 pm→  Jane Street-sponsored tea and cookies in W ean 6220\n(bring your mug!) \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9159fae DTSTART;TZID=America/New_York:20251023T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251023T173000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Crypto Seminar - Wei-Kai Lin CLASS:PUBLIC DESCRIPTION:Speaker: WEI-KAI LIN\, Assistant Professor\, Computer Science\n Department\, University of Virginia\n\nTalk Title: MegaBlocks: Breaking th e Logarithmic I/O-Overhead Barrier\nfor Oblivious RAM\n\nHow to securely d elegate storage to an untrusted server? The access\npattern to the storage often divulges sensitive information about the\ndata\, even when the data is securely encrypted. Oblivious RAM (ORAM)\nis a compiler that generical ly transforms any access pattern to\nunintelligible but functionally equiv alent ones. When Goldreich\nintroduced ORAM in 1987\, a lower bound of Ω( log n) for ORAMs was also\nproved. In this talk\, we circumvent the logari thmic lower bound in a\nnatural asymmetric setting\, where the block size of the storage is\nlarger than the word size simulated by ORAM. We provide both a lower\nbound and an optimal ORAM scheme in the asymmetric setting. Our scheme\nis implemented and open-sourced\, and it outperforms the best -known\nORAMs in real-world settings.\n\nThis is a joint work with Gilad A sharov\, Eliran Eiluz\, and Ilan\nKomargodski.\n\nIn Person and Zoom Parti cipation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915a338 DTSTART;TZID=America/New_York:20251020T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251020T120000 LOCATION:Mehrabian Collaborative Innovation Center 2201 and Zoom SUMMARY:Crypto Seminar - Ethan Mook CLASS:PUBLIC DESCRIPTION:Speaker: ETHAN MOOK\, Ph.D. Student\, Khoury College of Compute r\nSciencesNortheastern University\n\nTalk Title: Black Box Crypto is Usel ess for Doubly Efficient PIR\n\nA (single server) private information retr ieval (PIR) allows a client\nto read data from a public database held on a remote server\, without\nrevealing to the server which locations she is r eading. In a doubly\nefficient PIR (DEPIR)\, the database is first preproc essed offline into\na data structure\, which then allows the server to ans wer any client\nquery efficiently in sub-linear  online time. Constructin g DEPIR is a\nnotoriously difficult problem\, and this difficulty even ext ends to a\nweaker notion secret-key DEPIR (SK-DEPIR)\, where the database is\npreprocessed using secret randomness and the client is given a secret\ nkey for making queries.  We currently only have constructions of\nSK-DEP IR from the Ring LWE assumption or from non-standard code-based\nassumptio ns. \n\n We show that the black-box use of essentially all generic\ncryp tographic primitives (e.g.\, key agreement\, oblivious transfer\,\nindisti nguishability obfuscation\, etc.)\, including idealized\nprimitives (e.g.\ , random oracles\, generic multilinear groups\, virtual\nblack-box obfusca tion\, etc.) is essentially useless for constructing\nSK-DEPIR. In particu lar\, in any such SK-DEPIR construction\, we can\nreplace all black-box us e of these primitives with just a black-box\nuse of one-way functions. Whi le we conjecture that SK-DEPIR cannot be\nconstructed using black-box one- way functions alone\, we are unable to\nshow this in its full generality. However\, we do show this for 2-round\nschemes with a passive server that simply outputs requested locations\nin the preprocessed data structure\, w hich is the format of all known\nschemes. Overall\, this shows that the bl ack-box use of essentially all\ncrypto primitives is insufficient for cons tructing 2-round\npassive-server SK-DEPIR\, and does not provide any benef it beyond\nblack-box one-way functions for constructing general SK-DEPIR.\ n\nIn Person and Zoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915a787 DTSTART;TZID=America/New_York:20251022T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251022T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Sabee Grewal CLASS:PUBLIC DESCRIPTION:Speaker: SABEE GREWAL\, Ph.D. Student\, Department of Computer Science\,\nUniversity of Texas at Austin\n\nTalk Title: Demonstrating an u nconditional separation between quantum\nand classical information resourc es\n\nI’ll describe a recent demonstration of unconditional quantum\nadv antage based on quantum–classical separations in one-way\ncommunication. We construct a task for which the most space-efficient\nclassical algorit hm provably requires between 62 and 382 bits of\nmemory\, yet we solve it using only 12 qubits on a trapped-ion quantum\nprocessor.\n\nWork based on : arXiv:2509.07255.\n\nJoint work with William Kretschmer\, Matthew DeCros s\, Justin A. Gerber\,\nKevin Gilmore\, Dan Gresh\, Nicholas Hunter-Jones\ , Karl Mayer\, Brian\nNeyenhuis\, David Hayes\, and Scott Aaronson. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915aac3 DTSTART;TZID=America/New_York:20251023T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251023T153000 LOCATION:Newell-Simon 4305 SUMMARY:Doctoral Thesis Oral Defense - Costin Bădescu CLASS:PUBLIC DESCRIPTION:Speaker: COSTIN BĂDESCU\, Ph.D. Candidate\nComputer Science De partment\nCarnegie Mellon University\n\nTalk Title: Improved bounds for st ate certification\, separability\ntesting\, and shadow tomography\n\nWe pr esent improved sample complexity bounds for three fundamental\nquantum inf ormation tasks: state certification\, separability testing\,\nand shadow t omography. Given measurement access to n identical copies\nof an unknown q uantum state 𝜌\, we consider:\n\ni. State Certification: The task of ve rifying 𝜌 is equal to a\nreference state sigma or at least ε-far in tr ace distance.  We\npresent a testing algorithm for state certification th at uses O(d/\nε2) copies of 𝜌.\n\nii. Separability Testing: For a bipa rtite state 𝜌 on a\nd2-dimensional system\, we prove a lower bound of  𝛺(d2/ε2) copies\nare necessary to distinguish separability from being ε-far in trace\ndistance from the set of all separable states.\n\niii. S hadow Tomography: The problem of estimating the expectation\nvalues tr( 𝜌Ai) for m observables Ai\,...\,Am  to +/-ε accuracy. We\npresent an algorithm that accomplishes this with O(log2(m)\nlog(d)/ε4) copies\, wh ich simultaneously achieves the best known\ndependence on each parameter m \, d\, and ε.\n\nThesis Committee\n\nRyan O'Donnell (Chair)\n\nAayush Jai n\n\nDavid Woodruff\n\nJohn Wright (University of California\, Berkeley)  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915aed0 DTSTART;TZID=America/New_York:20251023T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251023T135000 LOCATION:Tepper Building 4242 SUMMARY:Center for Behavioral Decision Research Seminar - Leland Bybee CLASS:PUBLIC DESCRIPTION:Speaker: LELAND BYBEE\, Assistant Professor of Finance\, Booth School of\nBusiness\, University of Chicago\n\nTalk Title: Narratives as S oftware: Training Behavioral Models with\nReasoning\n\nLeland Bybee is an assistant professor of finance at the University of\nChicago Booth School of Business. Much of his work leverages machine\nlearning and natural lang uage processing to address key questions in\neconomics and finance.  His current focus is on developing measures\nof beliefs with applications to a sset pricing and behavioral\neconomics. \n\nBefore Booth\, he received a PhD in financial economics from the Yale\nSchool of Management\, a master ’s degree in statistics from the\nUniversity of Michigan\, and completed his undergraduate degree in\neconomics at the University of Chicago.\n\nF aculty Host: John Conlon\n\nSeminars are open to all but seating is limite d.\n\nYou are welcome to bring your own lunch. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915b24e DTSTART;TZID=America/New_York:20251020T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251020T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Ian Cook CLASS:PUBLIC DESCRIPTION:Speaker: IAN COOK\, Co-founder and Chief Executive OfficerColum nar\,\nandApache Arrow PMC member\n\nTalk Title: Where We're Going\, We Do n't Need Rows: Columnar Data\nConnectivity with ADBC\n\nADBC (Arrow Databa se Connectivity) is Apache Arrow’s answer to ODBC\nand JDBC: It’s a da tabase access API and driver standard that\ndelivers data in Arrow columna r format instead of a row-oriented\nformat. ADBC is on a roll\, speeding a nd simplifying data access for\ndbt\, Databricks\, DuckDB\, Microsoft\, Sn owflake\, and more. This talk\npresents the architecture of ADBC (APIs\, d rivers\, driver managers\, and\nother tools) and discusses its role in acc elerating current and future\ndata systems.\n\n—\n\nIan Cook is an Apach e Arrow PMC member and the co-founder and CEO of\nColumnar\, a company bui lding fast universal data connectivity\ninfrastructure powered by Arrow an d ADBC. \n\nThis talk is part of the Future Data Systems Seminar Series.\ n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915b5bc DTSTART;TZID=America/New_York:20251022T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251022T133000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Thesis Proposal - Saranya Vijayakumar CLASS:PUBLIC DESCRIPTION:Speaker: SARANYA VIJAYAKUMAR\, Ph.D. Student\nComputer Science Department\nCarnegie Mellon University\n\nTalk Title: Protection Boundary Integrity: Detecting and Preventing\nSecurity Failures Across Contexts\n\n Modern computational systems deploy technical guardrails to enforce\nsecur ity\, privacy\, and safety boundaries across increasingly complex\noperati onal contexts. While effective within their design contexts\,\nthese mecha nisms exhibit systematic vulnerabilities when systems\ntransition between different operational modes: across interaction\nmodalities\, through temp oral evolution\, or when integrating neural and\nsymbolic reasoning. This dissertation investigates where\, how\, and why\nsecurity mechanisms fail at these critical transitions.\n\nFirst\, I demonstrate patterns of bounda ry failure through empirical\nanalysis across multiple domains. My cross-m odal work evaluates such\nfailures in browser-agent safety auditing (Brows erART) and\nauthenticity detection of AI-generated code (CodeFusion). Thro ugh\nBrowserART\, I show that language models refusing harmful instruction s\nin chat interfaces pursue identical harmful behaviors when deployed as\ nbrowser agents\, despite identical safety training. Through CodeFusion\,\ nI analyze visual structure and semantic content\, demonstrating that\naut henticity boundaries require reasoning across representational\nmodalities . Second\, I identify temporal vulnerabilities that emerge\nwhen security mechanisms designed for static analysis cannot adapt to\nevolving threats. I demonstrate this through MalCentroid\, tracking\nmalware family evoluti on while maintaining robustness against\nadversarial obfuscation\, and thr ough graph-based fraud detection\nsystems identifying attack patterns emer ging across temporal\ntransaction sequences. Through TRACE\, I achieve suc cessful\nre-identification against Google's Topics API by exploiting\nvuln erabilities where privacy mechanisms protecting individual\nobservations f ail when adversaries analyze sequential behavioral\npatterns.\n\nFinally\, I introduce methods to bridge neural-symbolic security\nboundaries. Throu gh SMTLayer\, I integrate satisfiability solvers\ndirectly into neural arc hitectures\, achieving substantial data\nefficiency improvements while mai ntaining formal logical guarantees.\nIn my proposed work\, I introduce ver ifiable protection mechanisms for\nlanguage models through a game-theoreti c prover-verifier framework and\ndevelop multiplicative gating architectur es enabling efficient\nlearning of complex logical structures like XOR gat es that standard\narchitectures struggle to represent. This research provi des\nfoundations for building protection mechanisms that maintain integrit y\nacross the complex operational transitions required for safe\ndeploymen t of autonomous computational systems.\n\nThesis Committee\n\nChristos Fal outsos (Co-Chair)\n\nMatt Fredrikson (Co-Chair)\n\nSarah Cen\n\nMihai Chri stodorescu (Google Research)\n\nAdditional Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915bafb DTSTART;TZID=America/New_York:20251022T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251022T113000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:5th Year MSCS Thesis Presentation - Christopher Crawford CLASS:PUBLIC DESCRIPTION:Speaker: CHRISTOPHER CRAWFORD\, Master's Student\nComputer Scie nce Department\nCarnegie Mellon University\n\nTalk Title: Morphologically- Informed Tokenizers for Languages with\nNon-Concatenative Morphology\n\nTh is paper investigates the impact of using morphologically-informed\ntokeni zers to complete the interlinear gloss annotation of an audio\ncorpus of Y olox\\'ochitl Mixtec (YM) using a combination of ASR and\ntext-based seque nce-to-sequence tools. We present two novel\ntokenization schemes that sep arate words in a nonlinear manner\,\npreserving information about tonal mo rphology as much as possible. One\nof these approaches\, a Segment and Mel ody tokenizer\, simply extracts\nthe tones without predicting segmentation . The other\, a Sequence of\nProcesses tokenizer\, predicts segmentation f or the words\, which could\nallow an end-to-end ASR system to produce segm ented and unsegmented\ntranscriptions in a single pass. We find that these novel tokenizers\nare competitive with BPE models\, and the Segment-and-M elody model\noutperforms BPE in terms of word error rate but does not reac h the\nsame character error rate. Our results suggest that nonlinear\ntoke nizers designed specifically for the non-concatenative morphology\nof a la nguage are competitive with conventional BPE models for ASR.\nFurther rese arch will be necessary to determine the applicability of\nthese tokenizers in downstream processing tasks.\n\nThesis Committee\n\nDavid Mortensen (C hair)\n\nShinji Watanabe\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915bee3 DTSTART;TZID=America/New_York:20251105T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251105T140000 LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:CMU Benefits & Fitness Fair CLASS:PUBLIC DESCRIPTION:Speaker: CMU Benefits FairTalk Title: Hosted by the Office of H uman\nResources and CMU Athletics\n\nThe Benefits & Fitness Fair is an opp ortunity to meet with our benefit\nproviders and representatives from the Office of Human Resources to\nlearn about your 2026 benefits options.\n\nA ll faculty\, staff and students are welcome to attend.\n\nHuman Resources representatives will be available to answer questions\,\nincluding one-on- one meetings.In addition\, the fair will include\nfitness information\, ra ffles and giveaways.Giant Eagle will be\nadministering flu vaccinations at no cost to faculty and staff\nmembers. To obtain your flu vaccine\, you w ill need to present your CMU\nID card. Remember to bring your Giant Eagle Advantage Card to get 10%\noff your next grocery shopping trip.\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915c23e DTSTART;TZID=America/New_York:20251015T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251015T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Tushant Mittal CLASS:PUBLIC DESCRIPTION:Speaker: TUSHANT MITTAL\, Motwani Postdoctoral Fellow\, Departm ent of\nComputer Science\, Stanford University\n\nTalk Title: A General Fr amework for Low Soundness Homomorphism Testing\n\nCan one verify a proof j ust by reading a tiny part of it? If a\nfunction is linear on most small s ubsets\, must it be close to a truly\nlinear function? Such questions recu r in theoretical computer science\,\nand the goal is to define notions of approximate structure that are\nlocally testable\, and yet\, let us deduce global structure.  \n\n  \n\nIn this talk\, I will present a general f ramework for defining\nefficiently testable notions of homomorphisms betwe en groups\, and\nprove an inverse result showing that such maps are close to genuine\nhomomorphisms.  This framework yields novel tests for a wide variety\nof groups in the low soundness (high error) regime\, where very f ew\nresults are known.\n\n  \n\nJoint work with Sourya Roy\, University of Iowa. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915c5dd DTSTART;TZID=America/New_York:20251020T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251020T130000 LOCATION:Simmons Auditorium B\, Tepper Building and Zoom SUMMARY:CyLab Seminar - Faculty Speakers CLASS:PUBLIC DESCRIPTION:Speaker: Faculty SpeakersTalk Title: Towards a Safe\, Secure\, and\nPrivacy-Preserving Robotics Ecosystem\n\nHear from leading CyLab facu lty members on the synthesis of the CyLab\nRobotics Security and Privacy W orkshop and key research directions for\nCyLab’s most ambitious research initiative to date.\n\nThe proliferation of robotic systems from controll ed environments to\nreal-world applications in homes\, factories\, and cri tical\ninfrastructure has created an urgent need for a holistic\, system-w ide\napproach to security and privacy. The CyLab Robotics Security and\nPr ivacy Workshop brought together experts to define foundational\nchallenges and outline a path forward for building a trustworthy\nrobotics ecosystem that is secure by design. The discussions\nhighlighted that securing robo tics requires moving beyond a focus on\nindividual components and addressi ng the unique complexities of this\ndomain\, including the interplay of ha rdware\, software\, AI\, and human\nfactors.\n\nSpeakers\n\nJorge Guajardo Merchan — Prinicipal Scientist and Senior Manager\nSecurity and Priva cy Research Group\, Robert Bosch LLC – Research\nand Technology CenterL imin Jia — Research Professor\,  Department\nof Electrical and Compu ter EngineeringEunsuk Kang — Associate\nProfessor\,  Software and So cietal Systems DepartmentSebastian\nScherer — Associate Research Profe ssor\, Robotics InstituteSarah\nScheffler — Assistant Professor\, Soft ware and Societal Systems\, and\nDepartment and Engineering and Public Pol icy\n\nREGISTER \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915c987 DTSTART;TZID=America/New_York:20251013T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251013T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Will Manning CLASS:PUBLIC DESCRIPTION:Speaker: WILL MANNING\, Co-founder and Chief Executive Officer\ , Spiral\n\nTalk Title: Vortex: LLVM for File Formats\n\nApache Parquet re volutionized columnar storage after its initial\nrelease in 2013\, but has largely failed to evolve since then. As a\nresult\, nearly every Tier 1 t ech company has built their own columnar\nformat to replace Parquet.\n\nEn ter Vortex\, a Linux Foundation project that currently achieves 100x\nfast er random access\, 10-20x faster scans\, and 5x higher write\nthroughput\, while maintaining roughly the same compression ratio.\nImportantly\, it ’s also designed explicitly to support decoding via\nGPU SIMT.\n\nBut Vo rtex is actually more than just a file format. Like how LLVM\nturned \"wri ting a compiler\" into \"writing a language frontend”\,\nVortex provides extensive file format infrastructure\, turning\n“writing a new file for mat” into customizing encodings and layout\nstrategies.\n\nThis talk wil l walk through how Vortex is built\, and how we moved\ndecisions from \"sp ec writer\" to \"file writer.\" We'll also cover the\ncore research founda tions (BtrBlocks\, FastLanes) behind its\nperformance\, and why designing for GPU SIMT makes CPU SIMD and random\naccess fast too.\n\n—\n\nWill Ma nning is Co-founder & CEO at Spiral\, a startup building a\nnext-generatio n multimodal warehousing system. Spiral particularly\nexcels at workloads like GPU data loading (\"making GPUs go brr\")\, in\naddition to more trad itional relational queries. Will is also the TSC\nChair of Vortex\, an Inc ubation stage project at the Linux Foundation\nthat is building an extensi ble\, state-of-the-art columnar file format.\nPrior to starting Spiral\, h e worked at Palantir for nearly 10 years.\nWhile there\, he helped create Palantir Foundry\, started Palantir's\nEuropean commercial business\, & ra n \"every engineering team that read\nor wrote bytes\". In ancient times ( ca. 2010)\, he did research work on\nreinforcement learning. \n\nThis tal k is part of the Future Data Systems Seminar Series.\n\nIn Person and Zoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915ce1f DTSTART;TZID=America/New_York:20251026T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251026T105000 LOCATION:TBA SUMMARY:Fall Tartan Preview Day: Academics & Campus Resource Fair CLASS:PUBLIC DESCRIPTION:Speaker: Undergraduate Admissions\n\nInfo Session \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915d0e5 DTSTART;TZID=America/New_York:20251026T105000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251026T120500 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Undergraduate Fall Preview Day CLASS:PUBLIC DESCRIPTION:Speaker: Part of the Fall Tartan Preview Day\n\nWe are excited you are planning a visit to the Carnegie Mellon\nUniversity campus!   Pl ease join the School of Computer for\nadditional information on our progra ms and to have your questions\nanswered.\n\nPlease fill out the event REG ISTRATION form to reserve your spot and\nselect the the sessions you would like to attend throughout your\nday.  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915d3bf DTSTART;TZID=America/New_York:20251015T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251015T130000 LOCATION:Newell-Simon 3305 and Zoom SUMMARY:NSF AI-SDM and Heinz College AI Measurement Science & Engineering\n Center Seminar - Gopal Ramchurn CLASS:PUBLIC DESCRIPTION:Speaker: GOPAL RAMCHURN\, Professor of Artificial Intelligence\ nSchool of Electronics and Computer Science\nand Fellow\, Institution of E ngineering and Technology\nUniversity of Southampton\n\nTalk Title: Emergi ng Challenges in Responsible and Trustworthy AI\n\nIn this talk I propose that AI systems need to be studied and\ndeveloped in terms of the partners hips we form with them rather than\nas systems we react to.  To ensure su ch partnerships are trustworthy\,\na multi-disciplinary approach is needed and I will provide some\nexamples of past and current work I’m undertak ing in this area\,\nfocusing on applications of multi-agent systems in soc ial care and\nedge AI systems in agriculture. I will also describe some of the\nresearch\, financing\, and policy challenges we are addressing acros s\nResponsible AI UK\, a £33m research and innovation programme that\nsta rted in 2023.  \n\n—\n\nProf. Gopal Ramchurn is a Professor of Artific ial Intelligence and\nFellow of the Institution of Engineering and Technol ogy. He is the CEO\nof Responsible AI UK\, and former Director of the UKRI Trustworthy\nAutonomous Systems hub. He is also a Co-CEO of Empati Ltd\, an AI\nstartup working on decentralised green hydrogen technologies. His\n research is about the design of Responsible Artificial Intelligence\nfor s ocio-technical applications including energy systems and disaster\nmanagem ent. His research involves applying techniques from Machine\nLearning\, HC I\, and Game Theory. He has won multiple best paper awards\nfor his resear ch and is a winner of the AXA Research Fund Award (2018)\nfor his work on Responsible Artificial Intelligence.\n\nREGISTER → For in person and r emote participation\, to assist with\nlogistics.\n\n                       Lunch will be served to those\nregistered .\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915d7da DTSTART;TZID=America/New_York:20251015T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251015T140000 LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Anat Levin CLASS:PUBLIC DESCRIPTION:Speaker: ANAT LEVIN\, Professor\nDepartment of Electrical and C omputer Engineering\nTechnion\, Israel\n\nTalk Title: Seeing Deep Inside S cattering Tissue Using Efficient\,\nNoise-Robust Wavefront Shaping\n\nScat tering limits our ability to see inside biological tissue\, as\nlight pene tration is severely distorted by tissue components with\nvarying refractiv e indices. One promising method to overcome\nscattering aberration is wave front shaping. This technique involves\nplacing a spatial light modulator (SLM) in the microscope's optical\npath to correct the wavefront emitted f rom a point deep within the\ntissue. The goal is to bring light photons fr om a single target point\nto a single sensor point\, despite tissue aberra tions. This technique\nhas the potential to revolutionize tissue imaging b y enabling high-SNR\nimaging deep within scattering biological targets. Ho wever\, estimating\nwavefront-shaping modulations in practice is challengi ng\, since the\nmodulations must be estimated in real time\, using non-inv asive\nfeedback\, and under a low photon budget.\n\nIn the first part of t his talk\, I will discuss efforts to derive\nnoise-robust score functions that can identify effective modulation\ncorrections using non-invasive fee dback. I will review previous\napproaches and introduce a new\, simple\, n oise-robust method that uses\nconfocal correction of both incoming and out going light with linear\nsingle-photon fluorescent excitation. We show tha t despite the fact\nthat we are only measuring light outside the tissue an d have no direct\nway to measure how well light has focused inside the tis sue\,\nmaximizing the single-photon confocal intensity guarantees that we\ nalso focus all light into a spot inside the tissue.\n\n \n\nGiven a scor e function\, estimating the desired modulation becomes an\noptimization pr oblem. However\, since the desired modulation depends on\nthe unknown tiss ue structure\, typical optimization strategies involve\nslow sequential sc anning\, where each modulation parameter is queried\nindependently. In the second part of this talk\, I will present a novel\napproach for rapid mod ulation optimization. This method leverages\noptical computing ideas and u ses the optical system to directly\nmeasure the gradient of the score func tion\, allowing simultaneous\nupdates of all modulation parameters from a single measurement.\n\n—\n\n \n\nAnat Levin is a Professor at the depar tment of Electrical and Computer\nEngineering\, Technion\, Israel\, doing research in the field of\ncomputational imaging. She received a Ph.D. in c omputer science from\nthe Hebrew University in 2006. During the years 2007 - 2009 she was a\npostdoc at MIT CSAIL\, and during 2009-2016 she was an A ssistant and\nAssociate Prof. at the department of Computer Science and Ap plied\nMath\, the Weizmann Inst. of Science.\n\nProf.  Levin has received numerous awards for her research\, including\nthe CVRP PAMI young researc her award in 2013\; the eurographics young\nresearcher award in 2010\; the eurographics outstanding technical\ncontributions award in 2024\; the Bla vatnik award in 2018\; and 3 ERC\ngrants.\n\nThe VASC seminar is generousl y sponsored by HeyGen\, an all-in-one\nAI-powered video generation platfor m that leverages advances in\ncomputer vision\, generative modeling\, and multimodal learning to make\nhigh-quality video creation both scalable and accessible. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915dd42 DTSTART;TZID=America/New_York:20251007T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251007T160000 LOCATION:Wean Hall 7218 SUMMARY:ACO Seminar - Alya Kuchukova CLASS:PUBLIC DESCRIPTION:Speaker: ALYA KUCHUKOVA\, Ph.D. Student\, Ph.D. Program in Algo rithms\,\nCombinatorics and Optimization\, School of Mathematics / School of\nComputer Science\, Georgia Institute of Technology\n\nTalk Title: Samp ling colorings with fixed color class sizes\n\nTuesday\, October 7\, 2025\ , 3 – 4pm \n\nClassical results by Jerrum (1995) and Salas-Sokal (199 7) show the\npossibility of efficient approximate counting and uniform sam pling of\nproper colorings with q > 2Delta colors in graphs of maximum deg ree\nDelta. For colorings with fixed color class sizes\, Kierstand and\nKo stochka (2007) reproved the existence of equitable coloring\n(colorings wh ere class sizes differ by at most 1) and provided a\npolynomial algorithm which produces such a coloring and requires only\nq > Delta colors. In thi s paper we provide efficient approximate\ncounting and uniform sampling al gorithms for colorings with fixed\ncolor class sizes that are equitable or close to equitable. The proof\nuses techniques such as zero-freeness of p artition functions\, Local\nCentral Limit Theorems\, and cluster expansion . We hope our result adds\nto the growing evidence of the possibility to e fficiently sample\nfundamental combinatorial objects\, such as colorings\, with global\nconstraints. Joint work with Will Perkins and Xavier Povill. \n\nThis talk will not assume any knowledge of statistical physics or\nsam pling!\n\n4:00 pm  → Jane Street-sponsored tea and cookies in the Math \nLounge (bring your mug!) \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915e0f9 DTSTART;TZID=America/New_York:20260311T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20260311T180000 LOCATION:Simmons Auditorium\, Tepper Building SUMMARY:Dickson Prize in Science Lecture - Aviv Regev CLASS:PUBLIC DESCRIPTION:Speaker: AVIV REGEV\, Head and Executive Vice President\, Genen tech\nResearch and Early Development\n\nTalk Title: From Cell Atlases to M edicines\, with AI\n\nSingle-cell and spatial methods have enabled the con struction of\natlases of human cells. However\, atlases cannot directly ex plain how\ncells operate or restore them to health. Mapping every genetic or\ntherapeutic possibility in the lab is impossible\, but with AI\nresear chers can generalize\, develop testable predictions\, and iterate\nbetween experiment and model. This lecture focuses on developing cell\natlases an d building a “Lab in a Loop” to deliver biological\ninsights and new p otential therapies from target discovery to clinical\ndevelopment.\n\n—\ n\nDr. Aviv Regev is the head and executive vice president of Genentech\nR esearch and Early Development (gRED)\, where she oversees all aspects\nof the company's drug discovery and early development activities. She\nalso s erved as a professor of biology at MIT and a member of the\nexecutive lead ership teams of the Broad Institute of MIT and Harvard.\n\nDr. Regev was s elected for this award because of her highly impactful\ncontributions to c omputational methodologies and resources for\nbiological discovery. Her wo rk pioneered computational methods in\nsingle cell and spatial genomics\, which are widely used in biomedical\nresearch. As a founder and co-chair o f the international Human Cell\nAtlas (HCA) project\, she is helping to bu ild a complete compendium of\nhuman cell types to advance our understandin g of disease mechanisms.\n\n—\n\nAbout the Lecture:  Awarded annually s ince 1970\, Carnegie Mellon\nUniversity's Dickson Prize in Science recogni zes substantial\nachievements or sustained progress in the fields of the n atural\nsciences\, engineering\, computer science or mathematics.\n\nREGIS TER → by Monday\, March 9 \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915e526 DTSTART;TZID=America/New_York:20251022T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251022T180000 LOCATION:McConomy Auditorium\, Cohon University Center SUMMARY:University Lecture: Carl and Amy Jones Lecture in Interdisciplinary \nScience CLASS:PUBLIC DESCRIPTION:Speaker: AMY ROBINSON STERLING\, Neuroscience Designer\, Execut ive\nDirector\, Eyewire\n\nAmy is the executive director of Eyewire\, a ga me to map the brain from\nPrinceton University. She is an international sp eaker\, science\ncommunicator and contributor for Forbes\, covering neuros cience and\ndata visualization.\n\nSterling spends most of her time crowds ourcing neuroscience\;\ncatalyzing a future where we bring the world toget her to further\nmankind's quest to better understand ourselves. Sterling i s passionate\nabout connecting people to projects they care about and has spent over\na decade orchestrating and curating crowdsourcing projects. Sh e has\nadvised the White House's Office of Science and Technology Policy i n\ninnovation methods and has facilitated numerous brain-related\ninteract ive technology experiences over the years\, ranging from VR\nshown at TED and the Tribeca Film Festival to 3D renders and\nanimations shown in major media worldwide including Times Square NYC.\nUnder her leadership\, Eyewi re has appeared in museum exhibitions in\nNew York\, San Francisco\, Istan bul\, London\, Barcelona\, Hong Kong\, and\nmore.\n\nSterling is an advoca te of side projects and strives to live a modern\nday Renaissance lifestyl e of interdisciplinary explorations and\ninteractions. She aims to surroun d herself with curious\, passionate\npeople. She founded the TEDx Music Pr oject\, a collection of the best\nlive music performances from TEDx events around the world.\n\nRSVP → by Monday\, October 20\n\nThis event is pa rt of the University Lecture Series and is free and\nopen to the public.\n \nAbout The Carl & Amy Jones Lecture in Interdisciplinary Science: \nCarl (MCS 56) and Amy Jones established the Carl & Amy Jones Endowed\nInterdis ciplinary Fund at Carnegie Mellon University to support\ninterdisciplinary study at the university. The fund supports a number\nof efforts at the un iversity including The Carl & Amy Jones Lecture in\nInterdisciplinary Scie nce.  The Jones Lecture is  presented annually\nby a researcher whose wo rk uses interdisciplinary means of inquiry to\nadvance the sciences and sc ientific knowledge. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915e980 DTSTART;TZID=America/New_York:20251008T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251008T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Kostas Stavropoulos CLASS:PUBLIC DESCRIPTION:Speaker: KOSTAS STAVROPOULOS\, Ph.D. Student Department of Comp uter\nScienceUniversity of Texas at Austin\n\nTalk Title: Efficient Learni ng Algorithms under (Heavy) Contamination\n\nIn this talk\, I will present a series of new results in supervised\nlearning from contaminated dataset s\, based on a general outlier\nremoval algorithm inspired by recent work on learning with\ndistribution shift. Specifically:\n\nWe will show that a ny function class that can be approximated by\nlow-degree polynomials with respect to a hypercontractive distribution\ncan be efficiently learned un der bounded contamination (also known as\nnasty noise). This resolves a lo ngstanding gap between the complexity\nof agnostic learning and learning w ith contamination\, even though it\nwas widely believed that low-degree ap proximators only implied\ntolerance to label noise.For any function class that admits the\n(stronger) notion of sandwiching approximators\, we obtai n near-optimal\nlearning guarantees even with respect to heavy additive co ntamination\,\nwhere far more than 1/2 of the training set may be added\na dversarially. Prior related work held only for regression and in a\nlist-d ecodable setting.\n\nThese results significantly advance our understanding of efficient\nsupervised learning under contamination\, a setting that ha s been much\nless studied than its unsupervised counterpart. As a notable\ napplication\, our framework yields the first quasipolynomial-time\nalgori thm for learning constant-depth circuits (AC⁰) under bounded\ncontaminat ion\, extending the seminal result of Linial\, Mansour\, and\nNisan on lea rning AC⁰ under adversarial label noise. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915ed7e DTSTART;TZID=America/New_York:20251120T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251120T173000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:Carnegie Mellon Graphics Colloquium - Ken Museth CLASS:PUBLIC DESCRIPTION:Speaker: KEN MUSETH\, Senior Director of High-Fidelity Physics\ nResearch\, Nvidia\; Chair\, Technical Steering Committee for OpenVDB\nund er the Academy Software Foundation\n\nTalk Title: OpenVDB\n\nAs the invent or of VDB and founder of OpenVDB\, I am excited to talk\nabout its history \, motivation\, and diverse adoption. Specifically\,\nthis lecture will co ver the underlying VDB data structure\, and its\nadoption to computer grap hics\, physics simulations and more recently\nmachine learning. Since its open-source release in 2012\, OpenVDB has\nbecome an industry standard and has been used in numerous VFX\nfranchises like \"Avatar\"\, \"Avengers\"\ , \"The Mummy\"\, \"Pirates of the\nCaribbean\"\, \"Kung Fu Panda\"\, and \"How to Train Your Dragon\". It is\nadopted by numerous commercial softwa re packages used by the\nentertainment industry\, including Houdini\, Rend erMan\, Arnold\, Blender\,\nand Unreal Engine\, just to mention a few. Ope nVDB has also found use\nin many areas outside of media and entertainment\ , including SLAM\,\nautonomous driving\, topology optimization\, semicondu ctor designs\, 3D\nprinting\, medical imaging\, rocket design\, aerial sur veillance\,\nrobotics\, and many machine learning applications. Finally\, OpenVDB was\nthe first open-source project to be adopted by the Academy So ftware\nFoundation (ASWF) and the Linux Foundation (in 2018).\n\n—\n\nKe n Museth is Sr Director of the High-Fidelity Physics Research team\nat Nvi dia and chair of the Technical Steering Committee for OpenVDB\nunder the A cademy Software Foundation. He has a PhD in quantum physics\nfrom Copenhag en University and did his postgraduate studies in\ncomputer science at Cal tech. Previously he was a Sr. Computational\nScientist at SpaceX for six y ears\, working on CFD simulations of the\nRaptor engine\, Head of Simulati on R&D at Weta FX for three years\,\nworking on James Cameron's Avatar 2\, director for FX and CFX\nsimulation teams at DreamWorks Animation for eig ht years\, Sr Software\nEngineer at Digital Domain for three years\, full tenured professor in\ncomputer graphics at Linköping University for four years\, where he\nsupervised five PhD and 15 MSc students\, and research s cientist at\nNASA's Jet Propulsion Laboratory for three years\, working on \nspace-mission design and visualization.\n\nKen invented and founded Open VDB for which he won two Academy Awards\nfrom the Academy of Motion Pictur e Arts and Sciences\; a Technical\nAchievement Award (Academy Certificate) in 2015 and a Scientific &\nEngineering Award (Academy Plaque) in 2024. A dditionally\, in 2023 he\nwas awarded the ACM SIGGRAPH Practitioner Award and was accepted into\nthe ACM SIGGRAPH Academy. Ken has served on the Tec hnical Papers\nCommittee for ACM SIGGRAPH multiple times and has 29 movie credits\,\nincluding on franchises like \"Avatar\"\, \"Avengers\"\, \"The Mummy\"\,\n\"Pirates of the Caribbean\"\, \"Kung Fu Panda\"\, and \"How to Train Your\nDragon'\".\n\nThe Carnegie Mellon Graphics Colloquium is host ed by the Carnegie\nMellon Graphics Lab and supported by Meta and Adobe.\n \nFaculty Host: Ioannis Gkioulekas \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915f290 DTSTART;TZID=America/New_York:20251009T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251009T170000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Speaking Skills Talks - Anup Agarwal CLASS:PUBLIC DESCRIPTION:Speaker: ANUP AGARWAL\, Ph.D. Student\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: Designing Congestion Contro l Algorithms with Performance\nGuarantees\n\nCongestion control algorithms (CCAs) are a key enabler (or limiter) of\nthe performance of networked ap plications on the Internet and data\ncenters. CCAs determine the rate at w hich different flows should send\ntraffic\, aiming to share network resour ces fairly and efficiently.\nDespite decades of research and their perform ance-critical nature\,\nCCAs are largely designed through heuristics and h uman intuition. As a\nresult\, our community continuously discovers new wa ys in which CCAs\nfail\, i.e.\, exhibit severe unfairness\, underutilizati on\, or\ncongestion. My work develops tools to formally reason about CCA\n performance and designs CCAs that can provably guarantee performance\neven under challenging network conditions.\n\nIn this talk\, I will present a part of my work that introduces\nabstractions to reason about CCA behavior through the mechanisms that\nCCAs use to coordinate fairness. CCAs operat e in partially observable\nenvironments: they cannot directly observe link capacities or\ncompeting flows. To share network resources fairly\, CCAs (implicitly)\ncommunicate fair shares through observable signals such as d elay or\nloss. For instance\, Reno\, a historically popular CCA\, encodes the fair\nshare as proportional to \"1/sqrt(loss rate)\". We refer to such \ncommunication mechanisms as contracts. \n\nI will show how the choice o f contracts\, one that most CCA designers\ndo not even make explicitly\, f ully determines key steady-state CCA\nperformance metrics\, exposing funda mental tradeoffs between the\nmetrics. Contracts also help understand and uncover a wide range of\nperformance failures in past CCAs. Building on th ese insights\, I will\nshare some intuition behind FRCC\, a new congestion control algorithm I\ndeveloped that provably ensures fairness on networks where existing\nCCAs starve flows.\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915f709 DTSTART;TZID=America/New_York:20251006T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251006T140000 LOCATION:Newell-Simon 3001 and Zoom SUMMARY:Accessibility Lunch Seminar - Prerna Khanna CLASS:PUBLIC DESCRIPTION:Speaker: PRERNA KHANNA\, Ph.D. CandidateDepartment of Computer\ nScienceStony Brook University\n\nTalk Title: Making Smartphones Accessibl e to Blind Users Using\nWearables\n\nBlind users interact with their smart phones predominantly using screen\nreaders such as TalkBack for Android an d VoiceOver for iOS to read out\ncontent and navigate touchscreen interfac es. However\, touchscreen\ninteractions are challenging for blind users du e to the need for\ntwo-handed interactions (conflicting with cane or guide dog use)\,\nvulnerability to shoulder-surfing attacks\, and gesture overl oading.\nThis talk explores how we can overcome these challenges by moving core\ninteractions beyond the screen and leveraging gestural input from\n commodity wearables.\n\nI will first present AccessWear\, a system-level f ramework that uses\nnovel input virtualization to seamlessly map smartwatc h gestures to\ntouch events. This approach enables blind users to control any mobile\napplication without requiring a single line of code change fro m\ndevelopers. Our central premise is that while wearable gestures enable\ ntouch-screen-free input\, existing recognition models fail for blind\nuse rs. These models are designed and trained on sighted users\, and\nfail to capture the unique motion properties of blind users. This gap\ncalls for s pecialized algorithms that reflect how blind users perform\ngestures in ev eryday settings. Next\, I will show how we characterized\nthe unique motio n properties of blind user gestures and used these\ninsights to develop a robust\, specialized recognition approach that\nworks without per-user tra ining.\n\nBuilding on this foundation\, I extend gestural input beyond sim ple\nnavigation to support complex\, fine-grained interaction tasks for\nb lind users. With GestureVoice\, I will show how gestures can be\ncombined with voice commands to support fine-grained interactions such\nas text edi ting\, which is one of the most difficult tasks for blind\nusers on mobile devices. This multimodal system reduces text editing\ntime significantly for blind users. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915fb8b DTSTART;TZID=America/New_York:20251010T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251010T100000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Thesis Proposal - Nicole Feng CLASS:PUBLIC DESCRIPTION:Speaker: NICOLE FENG\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Robust Algorithms for Windin g Numbers and Signed Distance\n\nThis thesis presents robust algorithms fo r inside-outside computation\nand curve reconstruction (via winding number s) and signed distance\ncomputation. These algorithms make geometric infer ences from imperfect\ndata\, where such imperfect data includes noisy\, in complete\, or\ninaccurate observations or representations of shapes that r esult from\neither acquisition or authoring of geometry. A theme is that\n robustness and versatility can often be achieved by processing smooth\,\ng lobally-defined functions encoding the geometry of interest\, that are\nmo re amenable to robust computation than the original\, defective curve\nor surface. For both inside-outside and signed distance computation we\ncan u nlock further control over geometry and topology by processing\nhigher-ord er derivatives of these functions. In many cases\, we can\nalso re-cast ou r algorithms\, formulated in terms of smooth functions\,\nonto different d iscretizations and geometric data structures. Another\ntheme is that robus t reconstruction and robust signed distance\ncomputation are closely relat ed problems\; towards this end\, we provide\na formalization of their rela tionship that justifies the design of our\nalgorithms.\n\nThesis Committee \n\nKeenan Crane (Chair)\n\nIoannis Gkioulekas\n\nNancy Pollard \n\nChris Wojtan (Institute of Science and Technology Austria)\n\nAdditional Inform ation \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef915ffc3 DTSTART;TZID=America/New_York:20251008T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251008T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting - October 2025 CLASS:PUBLIC DESCRIPTION:Speaker: CSD Faculty\n\nSee email announcement for details.\n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916026d DTSTART;TZID=America/New_York:20251008T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251008T180000 LOCATION:Grand Room\, Posner Hall 340 SUMMARY:Phi BETA KAPPA Society Initiation Ceremony - Fall 2025 Candidates CLASS:PUBLIC DESCRIPTION:The members of Upsilon of Pennsylvania\, the Carnegie Mellon Un iversity\nchapter of the Phi Beta Kappa Society\,\n\nhosts the initiation ceremony for seniors who have been selected for\nearly initiation into the Phi Beta Kappa Society.\n\nwith Keynote Speaker:\n\n    Connor Halloran Phillips\n\n    Assistant Teaching Professor\, Carnegie Mellon Institu te for\nStrategy & Technology\, Carnegie Mellon University\n\nReception Im mediately Following \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916057b DTSTART;TZID=America/New_York:20251006T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251006T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Jordan Tigani CLASS:PUBLIC DESCRIPTION:Speaker: JORDAN TIGANI\, Co-founder and Chief Duck-herderMother Duck\n\nTalk Title: DuckLake: Learning from Cloud Data Warehouses to Build a\nRobust \"Lakehouse\"\n\nWhen building scalable data systems\, it is ea sy to focus on the\nstorage and the compute\, but metadata a critical thir d piece that is\noften overlooked. This talk will describe how metadata st orage enables\nquery performance and helps provide transactional semantics in modern\ndata warehouses. We will then go into how the metadata story i n\npopular open data formats take us several steps backwards. We will\nthe n talk about how DuckLake makes metadata access work more closely\nto a tr aditional data warehouse\, which solves a lot of problems.\nFinally\, we'l l discuss building a SaaS service for DuckLake\, and the\ntechnical challe nges and tradeoffs involved.\n\n—\n\nJordan Togani is co-founder and chi ef duck-herder at MotherDuck\, a\nstartup providing a serverless data ware house based on the open source\nDuckDB. This is the third cloud data analy tics SaaS service he’s\nhelped create\, and hopefully this time he’s g etting it right. He\nhelped start Google BigQuery\, spent a decade working on it as\nengineer\, book author\, engineering leader\, and product leade r. Jordan\nhas also worked at SingleStore\, Microsoft Research\, the Windo ws Kernel\nteam\, and at a handful of star-crossed startups in engineering \,\nproduct\, and leadership roles.\n\nThis talk is part of the Future Dat a Systems Seminar Series.\n\nZoom Participation.  See announcement. \n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9160971 DTSTART;TZID=America/New_York:20251029T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251029T133000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:AI -SDM Seminar - Doron Cohen CLASS:PUBLIC DESCRIPTION:Speaker: DORON COHEN\, Assistant Professor\, Department of Engi neering &\nPublic Policy\, Carnegie Mellon University\n\nTalk Title: Desig ning Human-AI Decision Systems with Behavioral\nInsights: Incentives\, Tra nsparency\, and Learning\n\nBehavioral decision research shows that experi ence and learning shape\nhow people adopt and use technology and react to changes in their\nenvironment. This talk clarifies the implications of the se results for\nsafety\, compliance\, and sustained use in human-AI system s. I will\npresent experimental results on when and why humans over-rely o n\nautomation and how design choices reduce system brittleness. I\nintrodu ce the Operator’s Dilemma\, a full-information\,\nincentive-compatible c hoice task where human and automated operators\nrepeatedly choose whether to perform a costly safety check. Across\nthree studies (N = 970)\, a Seri al design that reveals the\nautomation’s choice before the human decides increased overreliance\nand accidents\, yet was preferred. In contrast\, Parallel and a\ntransparent Serial-Reveal design preserved independent che cking and\nreduced accidents. The AI system is held fixed\, so results are driven\nsolely by human incentives\, information\, and learning. A simple \nopportunity cost model explains the results and yields practical\nprinci ples for incentives\, feedback\, and transparency.\n\n—\n\nAs a behavior al scientist\, Doron Cohen focuses on the experimental and\ncomputational analysis of behavior\, behavioral contexts\, and learning.\nCohen's resear ch focuses on clarifying some of the complexities of\neveryday decision-ma king processes and how these are shaped by various\nincentive systems. By integrating computational cognitive modeling\nwith structural modeling of environments\, he aims to predict the\nlong-term impacts of policy changes and support the evaluation and\ndesign of effective interventions. His di verse body of work includes\nthe design of safety measures\, taxation syst ems\, incentivization\nprograms\, and human-computer interactions.\n\nREGI STER → to participate on Zoom or in person \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9160d94 DTSTART;TZID=America/New_York:20251012T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251012T120000 URL:https://sites.google.com/andrew.cmu.edu/acorn-2025/ LOCATION:Registration Required SUMMARY:ACO Research Network Conference (day three) CLASS:PUBLIC DESCRIPTION:Speaker: Invited SpeakersTalk Title: The Algorithms\, Combinato rics\,\nand Optimization Research Network (ACORN)\n\nThe Algorithms\, Comb inatorics\, and Optimization Research Network\n(ACORN) represents the coll aborations and connections between\nresearchers in these related fields. T his is exemplified by the joint\nPh.D> programs in ACO offered at Georgia Tech\, Carnegie Mellon\, and\nWaterloo.  Following the success of ACORN 2 023 at Georgia Tech\,\nCarnegie Mellon University will be hosting ACORN 20 25 from October 10\nto 12.\n\nREGISTER\n\nFeaturing FriezeFest 2025\, a ce lebration of Alan Frieze's achievements\non the occasion of his 80th birth day. Alan is an expert on random\ngraphs\, one of the founders of CMU's AC O program\, and has written\nhundreds of papers.  FriezeFest will be held on October 11 and 12.\n\nLearn More \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91610e2 DTSTART;TZID=America/New_York:20251011T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251011T180000 URL:https://sites.google.com/andrew.cmu.edu/acorn-2025/home?authuser=0 LOCATION:Registration Required SUMMARY:ACO Research Network Conference (day two) CLASS:PUBLIC DESCRIPTION:Speaker: Invited SpeakersTalk Title: The Algorithms\, Combinato rics\,\nand Optimization Research Network (ACORN)\n\nThe Algorithms\, Comb inatorics\, and Optimization Research Network\n(ACORN) represents the coll aborations and connections between\nresearchers in these related fields. T his is exemplified by the joint\nPh.D> programs in ACO offered at Georgia Tech\, Carnegie Mellon\, and\nWaterloo.  Following the success of ACORN 2 023 at Georgia Tech\,\nCarnegie Mellon University will be hosting ACORN 20 25 from October 10\nto 12.\n\nREGISTER\n\nFeaturing FriezeFest 2025\, a ce lebration of Alan Frieze's achievements\non the occasion of his 80th birth day. Alan is an expert on random\ngraphs\, one of the founders of CMU's AC O program\, and has written\nhundreds of papers.  FriezeFest will be held on October 11 and 12.\n\nLearn More \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9161463 DTSTART;TZID=America/New_York:20251001T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251001T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Ying Feng CLASS:PUBLIC DESCRIPTION:Speaker: YING FENG\, Ph.D. Student\, Department of Electrical\n Engineering and Computer Science\, Massachusetts Institute of\nTechnology\ n\nTalk Title: A Fast Linear Map with Uniform Gaussian-Like Averages\nMath Jax.Hub.Config({\ntex2jax: {\ninlineMath: [ ['$'\,'$']\, [\"\\\\(\"\,\"\\\ \)\"] ]\,\nprocessEscapes: true\n}\n})\;\n\nWe consider a uniform law of l arge numbers for random linear maps: Let\n$M \\in R^{m \\times n}$ be a ra ndom Gaussian matrix and let $f: R\n\\rightarrow R$ be $L$-Lipschitz. As $ m$ increases\, the average of $\\{\nf ((Mx)_i) \\}_{i \\in [m]}$ converges to its expectation uniformly over\n$x \\in R^n$. This concentration under lies several algorithmic uses of\nrandom Gaussian maps\, but their $O(mn)$ multiplication cost limits\npracticality.\n\nRecently\, Cherapanamjeri an d Nelson (STOC’22) constructed a faster\nlinear map with similar uniform concentration but asymptotically\nslower convergence rate than Gaussians. In this work\, we give a new\nlinear map that matches the convergence rat e of Gaussians while\nachieving an even faster runtime than Cherapanamjeri -Nelson. As\napplications\, we obtain improved algorithms for $\\ell_2 \\r ightarrow\n\\ell_1$ embeddings\, kernel approximation\, adaptive distance\ nestimation\, etc.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9161840 DTSTART;TZID=America/New_York:20251003T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251003T173000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:Doctoral Thesis Proposal - Honghao Lin CLASS:PUBLIC DESCRIPTION:Speaker: HONGHAO LIN\, Ph.D. Student\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Advances in Algorithms for M assive Data: Optimal Bounds\,\nAdversarial Robustness\, and Data-Driven In sights\n\nWith the rapid growth of massive datasets in areas such as machi ne\nlearning and numerical linear algebra\, classical algorithms are often \nno longer feasible. In this thesis proposal\, we develop provably\neffic ient algorithms for various problems in these settings\, such as\nstreamin g and distributed model. Our contributions span three\ndirections:\n\nOpti mal Bounds. We introduce a general technique for lifting dimension\nlower bounds for real-valued linear sketches to polynomially bounded\ninteger in puts. This leads to the first optimal sketching lower bounds\nfor discrete data streams in fundamental problems such as frequency\nmoment approximat ion\, operator norm estimation\, and compressed\nsensing. Beyond this\, we also establish nearly-optimal bounds for a\nvariety of streaming and sket ching tasks\, including ℓ p subspace\nsketches for constant dimension d \, ℓ p regression in the\narbitrary-partition distributed model\, and graph problems such as\napproximating the minimum cut and constructing cut sparsifiers in\nbalanced directed graphs.Adversarial Robustness. While mo st streaming\nalgorithms are studied in static worst-case models\, many pr actical\nscenarios involve adaptive adversaries who generate inputs based on\nprevious outputs. We present the first adaptive attack against linear\ nsketches for ℓ p-estimation over turnstile integer streams.\nSpecifica lly\, we show that any linear streaming algorithm with\nsketching matrix A ∈ ℤrxn can be broken using only poly(r log n)\nqueries\, with high co nstant probability. This result highlights\nfundamental limits of robustne ss in adaptive streaming. Learning-based\nAlgorithms. Classical algorithms guarantee correctness in the worst\ncase but often ignore structure in re al-world data\, while machine\nlearning methods leverage structure but typ ically lack guarantees. We\ndesign learning-based algorithms that incorpor ate machine learning\npredictions to adapt to input distributions\, achiev ing faster\nruntimes\, reduced space\, or improved accuracy. Crucially\, t hese\nalgorithms retain rigorous worst-case guarantees even when the\npred ictions are imperfect\, bridging the gap between theory and\ndata-driven p ractice.  \n\nThesis Committee\n\nDavid P. Woodruff (Chair)\n\nYang P. L iu \n\nRichard Peng\n\nJelani Nelson (University of California\, Berkeley )\n\nIn Person and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9161d4a DTSTART;TZID=America/New_York:20251005T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251005T170000 LOCATION:Registration Required SUMMARY:Workshop on Neural Simulation-Based Inference (Day Two) CLASS:PUBLIC DESCRIPTION:Speaker: Presented by: CMU Statistical Methods for the Physical \nSciences Research Center\n\nThe  STAtistical Methods for the Physical S ciences (STAMPS@CMU)\nresearch center is organizing a weekend workshop on neural\nsimulation-based inference on 4-5 October 2025. The workshop will take\nplace in person at the CMU campus in Pittsburgh.  A webcast for\nre mote participants will also be available.\n\nNeural simulation-based infer ence (neural SBI) has led to a recent\nparadigm shift in statistical infer ence across fields ranging from\nastronomy and particle physics to climate and environmental science.\nThe goal of this workshop is to bring togethe r researchers in neural\nSBI who have so far been largely siloed within di sjoint communities.\nWe expect that this will lead to cross-pollination of ideas across\nthese communities to facilitate the next advances in neural SBI and\nits applications.\n\nThe workshop will feature talks by leading researchers in neural SBI\,\nposter contributions by junior researchers\, and plenty of\nopportunities for interaction between the participants.\n\n Confirmed speakers include\n\nKyle Cranmer (University of Wisconsin-Madiso n)Gaia Grosso\n(Massachusetts Institute of Technology)Patrick Heimbach (Un iversity of\nTexas at Austin)Lukas Heinrich (Technical University of Munic h)Brian\nNord (Fermilab)Laurence Perreault-Levasseur (University of\nMontr eal)Barnabas Poczos (Carnegie Mellon University)Brian Reich\n(North Caroli na State University)Bingjie Wang (Pennsylvania State\nUniversity)Larry Was serman (Carnegie Mellon University)Minge Xie\n(Rutgers University)Andrew Z ammit-Mangion (University of Wollongong)\n\nREGISTER\n\n→ Seats are lim ited so please register as soon as possible.\n\nPoster Submissions deadlin e: September 26In-Person Registration\ncutoff:  September 26Remote Regist ration cutoff:  October 1\n\n—\n\nSTAMPS@CMU develops foundational stat istical methodology that\naddresses emerging open problems in fundamental physics\, environmental\nand climate sciences. Our goal is to promote trus tworthy scientific\ndiscovery that advances science and informs policy dec isions. We\nachieve this by fostering mutually beneficial and sustained\nc ollaborations between data scientists and physical scientists to\nleverage scientific expertise\, build trust in interpretable methods\,\nand transf er knowledge across the mathematical and physical\nsciences. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9162213 DTSTART;TZID=America/New_York:20251010T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251010T130000 SUMMARY:AI-SDM Student Brainstorming Session - with Guest Speaker - H. Chad \nLane CLASS:PUBLIC DESCRIPTION:Speaker: H. CHAD LANE\, Professor of Educational Psychology and \nComputer Science\, Director\, NSF-IES INVITE AI Institute\, University o f\nIllinois\, Urbana-Champaign\n\nTalk Title: Practical Uses of GenAI for Academics\n\nDr. Chad Lane's research focuses on developing educational\nt echnologies that are as engaging as popular social media\, video\ngames\, and movies. He aims to create intelligent technologies for\nlearning and b ehavior change by blending techniques from the\nentertainment industry wit h those from artificial intelligence and\nintelligent tutoring systems. As a Professor of Educational\nPsychology\, Curriculum and Instruction\, and in the Siebel School of\nComputing and Data Science at the University of Illinois\,\nUrbana-Champaign\, his work explores the application of AI and \nentertainment technologies to improve learning experiences\,\nparticular ly in informal settings. His teaching involves graduate and\nupper-level u ndergraduate courses on topics such as educational\ntechnologies\, AI\, in formal learning\, and design\, where he encourages\nstudents to adopt an i nterdisciplinary perspective by blending\npsychological and educational th eories with emerging technological\ndevelopments.\n\nREGISTER → to atte nd in-person or on Zoom\n\nBrainstorming Session: Students meet regularly to participate in\ninformal discussion sessions that delve into cutting-e dge AI topics.\nThese regular meetings throughout the semester provide a p latform for\nstudents to deepen understanding of specific areas and broade n\nknowledge by exploring cross-cutting connections between various AI\ndi sciplines. Discussions are a breeding ground for collaboration\,\ninnovati ve thinking\, and problem-solving from the ground up. They\nprovide a stre ss-free forum for exchanging ideas\, brainstorming new\napproaches to chal lenges\, and fostering lasting connections within the\nAI-SDM community in an environment distinct from a traditional\nseminar. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91625e2 DTSTART;TZID=America/New_York:20251023T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251023T130000 LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI-SDM - Student Brainstorming Session CLASS:PUBLIC DESCRIPTION:Talk Title: Register to Attend\n\nStudents meet regularly to pa rticipate in informal discussion\nsessions that delve into cutting-edge A I topics. These regular\nmeetings throughout the semester provide a platfo rm for students to\ndeepen understanding of specific areas and broaden kno wledge by\nexploring cross-cutting connections between various AI discipli nes.\nDiscussions are a breeding ground for collaboration\, innovative\nth inking\, and problem-solving from the ground up. They provide a\nstress-fr ee forum for exchanging ideas\, brainstorming new approaches\nto challenge s\, and fostering lasting connections within the AI-SDM\ncommunity in an e nvironment distinct from a traditional seminar.\n\nREGISTER → to attend (in-person and virtual options)\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916290d DTSTART;TZID=America/New_York:20250929T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250929T130000 LOCATION:Panther Hollow Room 4105\, Mehrabian Collaborative Innovation Cent er SUMMARY:CyLab Seminar - Paul Pearce CLASS:PUBLIC DESCRIPTION:Speaker: PAUL PEARCE\, Associate Professor\nSchool of Cybersecu rity and Privacy\nGeorgia Tech University\n\nTalk Title: Expectation vs Re ality: How Network Abstractions Impact\nInternet Security\n\nInternet scan ning is a critical tool for security research\; it has\ndriven development of new security protocols\, found and tracked\nvulnerabilities\, and disc overed censorship globally. Such scanning is\nbuilt on top of numerous net working abstractions that obscure network\npaths\, infrastructure\, and pr otocols. Unfortunately\, these\nabstractions mask underlying behaviors tha t can significantly impact\nInternet security. In this talk\, we will expl ore the interplay between\nnetwork infrastructure abstractions and two pro blem spaces: Internet\ncensorship and the IPv6 Internet. We begin by explo ring the role\nrouting and infrastructure have on our understanding of Int ernet\ncensorship\, showing that our mental model for censorship is out of \nsync with the reality of networks. This disconnect both leads to\nincorr ect understanding of censorship\, as well as providing\nopportunities to d evelop evasion technology. Next\, we will explore the\nsecurity challenges introduced by the deployment of IPv6\, with\nabstractions and mental mode ls rooted in IPv4. We begin by discussing\nthe development of our IPv6 Int ernet scanning tool\, 6Sense\, showing it\nis effective at scanning the IP v6 Internet. Leveraging 6sense we then\nexplore the security challenges fa ced by IPv6 networks stemming from\nglobal addressing and accessibility\, which in turn lead to network\nvulnerabilities.\n\n—\n\nPaul Pearce is c urrently an Associate Professor at Georgia Tech and is\na founding member of the School of Cybersecurity and Privacy. His\nresearch explores attacks \, vulnerabilities\, and mitigations in online\nservices and systems and a cross the Internet. He builds Internet-scale\nmeasurement systems and desi gns new empirical methods aimed at\ndiscovering real-world complex and uns een adversarial behaviors and\nvulnerabilities\, frequently exploring netw ork infrastructure and the\ninterplay between abstractions and security go als. He has received an\nNSF CAREER award\, an IMC Community Contribution award\, distinguished\nrecognition at the IEEE Symposium on Security and P rivacy\, and an ACM\ndoctoral dissertation runner-up award. Prior to joini ng Georgia Tech\,\nPaul was a Visiting Researcher with Facebook's Site Int egrity group\,\nand he completed his Ph.D. at UC Berkeley advised by Vern Paxson. His\nresearch has been supported by the National Science Foundatio n\, the\nOffice of Naval Research\, Cisco\, and DARPA.\n\nFaculty Host:  Justine Sherry\n\nIn Person and Zoom Participation. See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9162e12 DTSTART;TZID=America/New_York:20250929T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250929T173000 LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Vinoth Chandar CLASS:PUBLIC DESCRIPTION:Speaker: VINOTH CHANDAR\, Founder and Chief Executive Officer\n Onehouse\n\nTalk Title: Apache Hudi: A Database Layer over Cloud Storage f or Fast\nMutations and Efficient Queries\n\nData lakes emerged as a way to store vast amounts of data as files and\nobjects on infinitely scalable c loud storage\, with processing done on\nscalable distributed compute engin es. However\, this architecture lacks\nmany of the capabilities of traditi onal databases\, such as efficient\nmutations\, indexing\, and transaction management. Apache Hudi was\ncreated as the first \"lakehouse\" project\, to bridge this gap by\nintroducing a database-like abstraction on top of file-based data\nlakes.\n\nThis talk will explore Hudi’s design choices and tradeoffs across\nmetadata management\, indexing\, storage layout\, an d concurrency\ncontrol—decisions that enable fast incremental reads and writes\nwhile significantly reducing processing costs and query latency. W e\nwill also share practical guidance for using Hudi effectively in\nmoder n data platforms and highlight open challenges the community is\nactively tackling\, from scaling metadata to supporting emerging AI and\nunstructur ed data workloads.\n\n—\n\nApache  is the original creator of Apache H udi\, a system that brings\ndatabase-like primitives on top of data lakes. He is the founder and\nCEO of Onehouse\, where he focuses on making lakeh ouse infrastructure\nopen and cost-effective. Previously\, he was principa l engineer at\nConfluent working on Kafka/ksqlDB. He led the data architec ture during\ngrowth years at Uber and also lead engineer on the Voldemort key-value\nstore at LinkedIn. His work spans distributed storage\, stream\ nprocessing\, and real-time data infrastructure. \n\nThis talk is part of the Future Data Systems Seminar Series.\n\nZoom Participation.  See anno uncement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9163258 DTSTART;TZID=America/New_York:20250924T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250924T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Ashwin Padaki CLASS:PUBLIC DESCRIPTION:Speaker: ASHWIN PADAKI\, Ph.D. Student\, Department of Computer and\nInformation Science\, School of Engineering and Applied Science\,\nU niversity of Pennsylvania\n\nTalk Title: Sparse Navigable Graphs for Neare st Neighbor Search\n\nNavigability captures the ability of a complex netwo rk to support\nefficient\, decentralized search. The concept has a rich hi story\, from\nMilgram’s \"six degrees of separation\" to Kleinberg’s c omputational\nmodel of small-world phenomena. Over the past decade\, navig able graphs\nhave also emerged as an important principle behind state-of-t he-art\nheuristics for nearest neighbor search. In this talk\, I will pres ent a\nperspective on this development through the lens of approximation\n algorithms: given a dataset and a distance function\, how efficiently\ncan we construct the (approximately) sparsest navigable graph?\n\nBased on jo int work with Sanjeev Khanna and Erik Waingarten. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91635da DTSTART;TZID=America/New_York:20250925T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250925T160000 LOCATION:CNA Room\, Wean Hall 7218 SUMMARY:ACO Seminar - József Balogh CLASS:PUBLIC DESCRIPTION:Speaker: JÓZSEF BALOGH\, University of Illinois Urbana-Champai gn\n\nTalk Title: Maximal independent sets in the middle two layers of the \nBoolean lattice\n\nLet B(2d-1\, d) be the subgraph of the hypercube 𝒬 2d-1 induced by\nits two largest layers. Duffus\, Frankl and Rödl propose d the problem\nof finding the asymptotics for the logarithm of the number of maximal\nindependent sets in B(2d-1\, d). Ilinca and Kahn determined th e\nlogarithmic asymptotics and reiterated the question of what their\norde r of magnitude is. We determine the number of maximal independent\nsets in B(2d-1\,d) and describe their typical structure. The proof uses\na new va riation of Sapozhenko's Graph Container Lemma\, a new\nisoperimetric lemma \, a theorem of Hujter and Tuza on the number of\nmaximal independent sets in triangle-free graphs and a stability\nversion of their result by Kahn and Park\, among other tools.\n\nJoint work with Ce Chen and Ramon Garcia. \n\n4:00 pm → Jane Street-sponsored tea and cookies in Math Lounge\n(bri ng your mug) \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9163946 DTSTART;TZID=America/New_York:20250926T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250926T170000 URL:https://www.cmu.edu/ai-sdm/research/human-ai-workshop/index.html LOCATION:Cohon University Center SUMMARY:Human-AI Complementarity for Decision Making Workshop - Day 2 CLASS:PUBLIC DESCRIPTION:Speaker: Human-AI Complementarity for Decision MakingTalk Title :\nSponsored by: NSF AI Institute for Societal Decision Making (NSF\nAI-SD M)\n\nWorkshop Motivation and Goals\n\nThis annual workshop explores the c oncept of Human-AI\nComplementarity—a condition where humans and AI syst ems working\ntogether outperform either working alone. Our 2025 theme focu ses on\nflexible Human-AI teams: systems that align with human values\,\nw ithstand unexpected behaviors\, and remain robust even under failure.\n\nK ey goals of the workshop include:\n\nDelivering cutting-edge instruction o n achieving Human-AI\ncomplementarityCreating common knowledge around emer ging research\nchallengesGenerating new ideas and concrete proposals for f uture\nresearch\n\nWho Should Participate?\n\nWe welcome contributions fro m multiple disciplines—decision science\,\ncognitive science\, computer science\, machine learning\, and beyond.\nParticipants may be:\n\nTutorial Instructors: Delivering state-of-the-art educational\nsessionsStudents: P resenting interactive posters and engaging in\ntutorialsPresenters: Provid ing brief\, targeted insights on key\nresearch topics\n\nTopics of Interes t\n\nSessions will focus on the flexibility and dynamics of Human-AI\ninte ractions for decision making\, including but not limited to:\n\nThe role o f AI agents in shaping human decision confidence and\ncalibrationAI's infl uence on trust\, coordination\, and\ncollaborationAddressing undesirable o r failure-prone AI behaviorsAnd\nothers.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9163d7a DTSTART;TZID=America/New_York:20250922T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250922T173000 SUMMARY:Database Seminar - Russell Spitzer CLASS:PUBLIC DESCRIPTION:Speaker: RUSSELL SPITZER\, Iceberg\n\nTalk Title: An Extremely Technical Overview of how the Apache\nIceberg(TM) Planning Implementation Actually Works\n\nWhat are you trying to tell me? That I can read data fas t? No\, User.\nI’m trying to tell you that when you are ready\, you won ’t have to.\n\nEveryone’s heard about how fast Apache Iceberg and mayb e you’ve\neven heard a few notes about “predicate pushdown” and “f ile\nmetrics” but you’ve been left wanting more. You want to know the\ nnitty gritty of how a predicate from a query engine is actually\ntransfor med and applied to Iceberg metadata. In this talk we will show\nyou just t hat\, we’ll work through the actual code of the Iceberg\nproject showing how the metadata is read\, how predicates are\ntransformed\, and finally how file tasks are actually broken up and\nsent to execution engines. We ’ll talk in detail about all of the\nproperties which control Iceberg pl anning and see the classes in which\nthose arameters actually take effect. This is definitely more detail\nthan any user of an Apache Iceberg table would actually need to know\,\nbut don’t you want to join the small grou p of developers who know\nhow it actually works?\n\n—\n\nRussell Spitzer received his Ph.D from UCSF in after performing a lot\nof comparisons of protein binding sites. Following that\, he became\ndeeply invested in dist ributed computing and joined Datastax\, a\ncompany using Apache Cassandra. While working at Datastax he was a key\ncontributor to the DataStax Spark -Cassandra Connector and also worked\non many other Apache projects. After leaving Datastax he worked at\nApple growing the then nascent Apache Iceb erg project where he worked\non data file management and advancing the tab le format. Currently\nRussell is working on OSS software at Snowflake and is a PMC member of\nthe Apache Iceberg project\, and a PPMC member of the Apache Polaris\n(Incubating) project. \n\nThis talk is part of the Future Data Systems Seminar Series. \n\nZoom Participation.  See announcement.  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91641b6 DTSTART;TZID=America/New_York:20251023T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251023T180000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Distinguished Industry Lecture - Mark Russinovich CLASS:PUBLIC DESCRIPTION:Speaker: MARK RUSSINOVICH\, Chief Technology Officer\, Microsof t Azure\n\nTalk Title: From Vibes to Production: The Art\, Discipline\, an d\nPitfalls of Vibe Coding\n\nIn this presentation\, Mark Russinovich\, CT O of Microsoft Azure\, will\nexplore how AI copilots and “vibe coding” are reshaping software\ndevelopment. He’ll show how natural language c oding can accelerate\ncreativity and lower barriers\, while also confronti ng the hard truths:\nAI-generated code isn’t automatically production- ready. This talk\nexamines how to harness vibe coding’s speed and power while still\napplying the architectural discipline\, testing\, and securit y practices\nneeded to ship software that is robust\, reliable\, and ready for the\nreal world. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91644f0 DTSTART;TZID=America/New_York:20251106T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251108T203000 SUMMARY:Carnegie Mellon Homecoming Weekend CLASS:PUBLIC DESCRIPTION:This year’s celebration will mark CMU’s 125th anniversary a nd\ncelebrate the university’s legacy of past\, present and future\nimpa ct.  Plus you’ll be able to enjoy all your favorite Homecoming\ntraditi ons — like the football game versus Franklin & Marshall and\nthe 75th An nual Alumni Awards — all with a CMU125 twist!\n\nREGISTRATION   |   Schedule of Events \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916478f DTSTART;TZID=America/New_York:20250923T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250923T133000 LOCATION:Gates Hillman 8115 SUMMARY:Doctoral Thesis Proposal - William Zhang CLASS:PUBLIC DESCRIPTION:Speaker: WILLIAM ZHANG\, Ph.D. Student\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: On Holistic Database Optim ization via Leveraging\nSimilarity Across Actions\, Workloads\, Configurat ions\, and Scenarios\n\nModern database management systems (DBMSs) have ev olved to support\nincreasingly sophisticated data-intensive applications\, at the cost of\nsubstantial complexity to configure them for two reasons. First\, DBMSs\nexpose a vast configuration space with trillions of possib ilities that\nencompass system knobs\, physical design (e.g.\, indexes)\, and query\noptions\, amongst others. Second\, these applications are const antly\nevolving with changes in data access patterns\, query types\, load\ nintensities\, hardware\, and data distributions that necessitate\ncontinu ous re-optimization.\n\nTo address these challenges\, decades of autonomou s DBMS optimization\nresearch have produced specialized tuning tools to as sist human\noperators. Deploying these tools involves a complex multi-step \nworkflow where an operator (1) observes the DBMS’s behavior\, (2)\nsel ects tools based on the objectives and their expertise\, (3)\nconfigures t hem with an isolated environment\, (4) orchestrates their\nexecution to ob tain recommendations\, and (5) reviews those\nrecommendations before deplo yment. This cumbersome process results in\nsuboptimal configurations and s low adaptation to evolving\napplications’ workloads due to isolated spec ialized tools\,\ninefficient reuse of prior tuning knowledge\, and the fal lible human\nfactor.\n\nThis proposal presents techniques for addressing t hose limitations\nwith similarity to enable holistic database optimization . First\, we\npresent a holistic tuning tool that optimizes multiple DBMS aspects\nsimultaneously by using action similarity to organize actions int o\nneighborhoods conducive to exploration. We then present a framework\nth at assists tuners in adapting to environment changes by leveraging\nworklo ad and configuration similarity to re-mix historical knowledge.\n\nWe prop ose to extend our preliminary work by transforming the\nhuman-centric tuni ng workflow into an agentic process through scenario\nsimilarity. We will first investigate contextualizing deployments and\ncreating semantic tool interfaces. We will then design an orchestrator\nthat learns to select and deploy relevant tuning tools to obtain\nvalidated recommendations. With t hese efforts\, the agentic process\nwill enable holistic DBMS optimization throughout its lifetime. \n\nThesis Committee:\n\nAndrew Pavlo (Chair)\n \nJignesh Patel \n\nVincent Conitzer \n\nImmanuel Trummer (Cornell Unive rsity)\n\nAdditional Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9164c76 DTSTART;TZID=America/New_York:20250918T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250918T160000 LOCATION:CNA Room\, Wean Hall 7218 SUMMARY:ACO Seminar - Olha Silina CLASS:PUBLIC DESCRIPTION:Speaker: OLHA SILINA\, Ph.D. Student\, Department of Mathematic al\nSciences\, Carnegie Mellon University\n\nTalk Title: Perfect matchings \, strongly connected orientations\, and\nlattice theory\n\nConsider a fam ily of subsets F = {Fi ⊆ T } over a finite ground set\nT. A (unweighted ) packing problem is asking what is the largest number\nof pairwise disjoi nt sets in F. In cases where F has a nice enough\npolyhedral description\, one may use tools such as linear programming\nduality and the theory of b locking polyhedra to derive min-max type\nrelations. Some examples of this include Menger’s theorem\, max-flow\nmin-cut\, etc. By relaxing integra lity conditions on the coefficients\nof the packing\, one can look for fra ctional packings or packings where\nwe allow subtraction. Both directions lead to lattice theory.\n\nIn this talk\, we consider two packing conjectu res and their lattice\ntheoretical relaxations: one in the setting of perf ect matchings in a\ngraph and another in the setting of strongly connected orientations in\na digraph. While the two theories share many similaritie s in terms of\npolyhedral descriptions and structural properties\, there a re still\ncertain distinctions leading to different lattice theoretical re sults.\n\nBased on joint works with Ahmad Abdi\, Gerard Cornuejols\, and S iyue\nLiu. \n\n4:00 pm → Jane Street-sponsored tea and cookies\, Math L ounge Wean\n6220 (bring your mug). \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916505a DTSTART;TZID=America/New_York:20250919T114500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250919T235500 LOCATION:5th Floor Pausch Bridge Entry\, Gates Hillman SUMMARY:HAPPY BIRTHDAY SMILEY! CLASS:PUBLIC DESCRIPTION:Speaker: Happy 43rd Birthday! :-)\n\nStop by at 11:47 am as we celebrate :-)\n\nMeet Scott Fahlman (Le Pére Smiley) who gave the world t he Smiley :-)\non September 19\, 1982.   \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9165342 DTSTART;TZID=America/New_York:20250924T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250924T130000 LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI-SDM: Student Brainstorming Session CLASS:PUBLIC DESCRIPTION:Talk Title: AI Institute for Societal Decision Making: Student\ nBrainstorming Session\n\nStudents meet regularly to participate in inform al discussion\nsessions that delve into cutting-edge AI topics. These reg ular\nmeetings throughout the semester provide a platform for students to\ ndeepen understanding of specific areas and broaden knowledge by\nexplorin g cross-cutting connections between various AI disciplines.\nDiscussions a re a breeding ground for collaboration\, innovative\nthinking\, and proble m-solving from the ground up. They provide a\nstress-free forum for exchan ging ideas\, brainstorming new approaches\nto challenges\, and fostering l asting connections within the AI-SDM\ncommunity in an environment distinct from a traditional seminar.\n\nREGISTER  → to attend in person or rem otely\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9165689 DTSTART;TZID=America/New_York:20251002T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251002T133000 LOCATION:Remote Access - Zoom SUMMARY:AI-SDM Seminar - Deb Roy CLASS:PUBLIC DESCRIPTION:Speaker: DEB ROY\, Professor of Media Arts and Sciences\nDirect or\, Center for Constructive Communication\nMedia Lab\nMassachusetts Insti tute of Technology\n\nTalk Title: A Better Way to Listen at Scale\n\nSocia l media platforms reward conflict and outrage\, drowning out\ngenuine conv ersation with noise\, bots\, and misinformation. Our public\ndiscourse is increasingly filled with performative rather than\nauthentic speech\, maki ng it harder to understand what people actually\nthink and feel – and ha rder still to build the trust that healthy\ncommunities and democracies re quire.\n\nIn 2017\, after years of studying social media\, we asked: How c an\ntechnology help people listen to one another\, understand across\ndivi des\, and solve problems together? At the MIT Center for\nConstructive Com munication\, working with the nonprofit Cortico\, we\nhave developed a sys tem of AI-powered tools and human-led practices\nfor systematic listening that combines practices from facilitated\ndialogue\, qualitative research\ , and community organizing. In our\napproach\, AI scaffolds human voice an d agency rather than replacing\nthem.\n\nThe approach has been deployed by Cortico in partnership with over 275\norganizations across the United Sta tes—including schools\, cities\,\nand local community groups. Communitie s use the approach to surface\nunderheard perspectives\, identify shared t hemes\, and generate media\noutputs that are used to build public understa nding\, shape policies\,\nand make decisions.\n\nThis talk distills what w e have learned\, demonstrates the approach in\npractice\, and outlines act ive research in AI-assisted sensemaking and\ndeliberation. The work addres ses three linked challenges posed by\nrapidly advancing AI: scaling conver sation without eroding trust\;\nharnessing AI’s power while keeping cont rol in the hands of\ncommunities\; and designing tools that strengthen\, r ather than\ndiminish\, human agency.\n\n—\n\nDeb Roy is professor of Med ia Arts and Sciences at MIT where he\ndirects the MIT Center for Constru ctive Communication (CCC). He\nleads research in designing human-AI syste ms that foster dialogue\,\nlistening\, and deliberation in ways that build civic muscle. Roy is\nalso co-founder and unpaid CEO of Cortico\, a clos ely affiliated\nnonprofit collaborator of CCC that develops\, operates and supports a\nconversation platform designed to surface underheard voices a nd\nperspectives and create scalable dialogue networks.\n\nPreviously\, Ro y was a visiting professor at Harvard Law\nSchool (2021-22)\, and served as executive director of the MIT Media\nLab (2019-2021)\, where CCC is based. Earlier\, while on leave from\nMIT\, Roy co-founded and was CEO of  Bluefin Labs\, a media analytics\ncompany that analyzed the interactions between television and social\nmedia at scale. Bluefin was acquired by T witter in 2013\, Twitter’s\nlargest acquisition to date. From 2013-2017 Roy served as Twitter’s\nchief media scientist.\n\nCurrently Roy serves on the board of the Knight First Amendment\nInstitute and the FRONTLIN E advisory council. He previously served\non the Knight Commission on Tr ust\, Media\, and Democracy and\nthe Aspen Institute’s Commission on Information Disorder.\n\nRoy is the author of over 185 academic papers i ncluding a study of\nthe spread of false news that was the cover story\n of Science magazine in 2018 and cited as one of the most\ninfluential  academic publications of the year. His\n2023 essay in The Atlantic des cribes his journey from studying\nsocial media to creating dialogue networ ks\, and his\n2024 Atlantic essay explores ways to tackle truth decay. Roy’s\nwidely viewed TED talk Birth of a Word presents his pioneering\ nresearch on his son’s language development that led to new ideas in\nme dia analytics.\n\nA native of Canada\, Deb was born and raised in Winnipeg and spent\nlarge parts of his childhood in Calcutta. He received his Bach elor of\nApplied Science from the University of Waterloo and PhD in Media Arts\nand Sciences from MIT.\n\nREGISTER → to participate in person or r emotely\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9165c6c DTSTART;TZID=America/New_York:20251212T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251212T170000 URL:https://csd.cmu.edu/academics/doctoral-resources/dsr-schedule LOCATION:Gates and Hillman Centers SUMMARY:DSR - General Meeting F25 CLASS:PUBLIC DESCRIPTION:Talk Title: Doctoral Student Review General Meeting\n\nFall 202 5 Doctoral Student Review.\n\nAll areas - students years 1-3.\n\nPlease re fer to emails sent to faculty and students for details.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9165f5d DTSTART;TZID=America/New_York:20251211T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251211T170000 URL:https://csd.cmu.edu/academics/doctoral-resources/dsr-schedule LOCATION:Gates and Hillman Centers SUMMARY:DSR - AI/Graphics/Theory F25 CLASS:PUBLIC DESCRIPTION:Talk Title: Doctoral Student Review (DSR) AI/Graphics/Theory\n\ nFall 2025 Doctoral Student Review.\n\nArtificial Intelligence\, Graphics\ , & Theory students years 4-n\n\nPlease refer to emails sent to faculty an d students for details.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916623e DTSTART;TZID=America/New_York:20251211T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251211T120000 URL:https://csd.cmu.edu/academics/doctoral-resources/dsr-schedule LOCATION:Gates and Hillman Centers SUMMARY:DSR - PL/Security/Systems F25 CLASS:PUBLIC DESCRIPTION:Talk Title: Doctoral Student Review (DSR) PL/Security/Systems\n \nFall 2025 Doctoral Student Review.\n\nProgramming Languages\, Security\, & Systems students years 4-n\n\nPlease refer to emails sent to faculty an d students for details.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916651f DTSTART;TZID=America/New_York:20250915T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250915T163000 LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Yutong Bai CLASS:PUBLIC DESCRIPTION:Speaker: YUTONG BAI\, Postdoctoral Researcher\nBerkeley AI Rese arch Lab\nUniversity of California\, Berkeley\n\nTalk Title: Whole-Body Co nditioned Egocentric Video Prediction\n\nWe train models to Predict Ego-ce ntric Video from human Actions\n(PEVA)\, given the past video and an actio n represented by the relative\n3D body pose. By conditioning on kinematic pose trajectories\,\nstructured by the joint hierarchy of the body\, our m odel learns to\nsimulate how physical human actions shape the environment from a\nfirst-person point of view. We train an auto-regressive conditiona l\ndiffusion transformer on Nymeria\, a large-scale dataset of real-world\ negocentric video and body pose capture. We further design a\nhierarchical evaluation protocol with increasingly challenging tasks\,\nenabling a com prehensive analysis of the model's embodied prediction\nand control abilit ies. Our work represents an initial attempt to\ntackle the challenges of m odeling complex real-world environments and\nembodied agent behaviors with video prediction from the perspective of\na human. \n\n—\n\nYutong Bai is currently a Postdoc Researcher at UC Berkeley (Berkeley\nAI Research)\ , advised by Prof. Alexei (Alyosha) Efros\, Prof. Jitendra\nMalik\, and Pr of. Trevor Darrell. Prior to that\, she obtained her PhD\nin Computer Scie nce at Johns Hopkins University advised by Prof. Alan\nYuille. She has int erned at Meta AI (FAIR Labs) and Google Brain\, and\nwas selected as a 202 3 Apple Scholar and an MIT EECS Rising Star. Her\nwork was nominated for t he CVPR 2022 Best Paper Award.  \n\nHer research aims to build intellige nt systems from first\nprinciples—systems that do not merely fit pattern s or follow\ninstructions\, but that gradually develop structure\, abstrac tion\, and\nbehavior through learning itself. She is interested in how\nin telligence emerges not from handcrafted pipelines or task-specific\nheuris tics\, but from exposure to behaviorally rich\, understructured\nenvironme nts where models must learn what to attend to\, how to reason\,\nand how t o improve. This involves designing learning systems that are\nnot narrowly optimized for a single goal\, but that can self-organize\nand grow increa singly competent through interaction\, experience\, and\ncomputation. Whil e she sees scale as a powerful tool\, she does not\nview it as the whole s olution: larger models open up capacity\, but\nwhat fills that capacity— and how it forms—is just as important.\nHer research explores how to use scale to amplify the right\nsignals—not just data quantity\, but the st ructural richness of\nbehavior and the dynamics of learning itself. \n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9166a03 DTSTART;TZID=America/New_York:20251010T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251012T150000 LOCATION:Simmons Auditorium\, Tepper Building SUMMARY:ACO Research Network Conference CLASS:PUBLIC DESCRIPTION:The Algorithms\, Combinatorics\, and Optimization Research Netw ork\n(ACORN) represents the collaborations and connections between\nresear chers in these related fields. This is exemplified by the joint\nPh.D> pro grams in ACO offered at Georgia Tech\, Carnegie Mellon\, and\nWaterloo.  Following the success of ACORN 2023 at Georgia Tech\,\nCarnegie Mellon Uni versity will be hosting ACORN 2025 from October 10\nto 12.\n\nREGISTER\n\n Featuring FriezeFest 2025\, a celebration of Alan Frieze's achievements\no n the occasion of his 80th birthday. Alan is an expert on random\ngraphs\, one of the founders of CMU's ACO program\, and has written\nhundreds of p apers.  FriezeFest will be held on October 11 and 12.\n\nLearn More \n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9166d49 DTSTART;TZID=America/New_York:20250911T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250911T173000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Crypto Seminar - Nikhil Vanjani CLASS:PUBLIC DESCRIPTION:Speaker: NIKHIL VANJANI\, Ph.D. StudentDepartment of Electrical and\nComputer EngineeringCarnegie Mellon University\n\nTalk Title: Fully Adaptive Decentralized MA-ABE: Simplified\,\nOptimized\, ASP Supported\n\n We revisit decentralized multi-authority attribute-based encryption\n(MA-A BE) through the lens of fully adaptive security – the most\nrealistic se tting in which an adversary can decide on-the-fly which\nusers and which a ttribute authorities to corrupt. Previous\nconstructions either tolerated only static authority corruption or\nrelied on highly complex “dual syst em with dual-subsystems” proof\ntechnique that inflated ciphertexts and keys.\n\nOur first contribution is a streamlined security analysis showing –\nperhaps surprisingly – that the classic Lewko–Waters MA-ABE sche me\n[EUROCRYPT 2011] already achieves full adaptive security\, provided it s\ndesign is carefully reinterpreted and\, more crucially\, its security\n proof is re-orchestrated to conclude with an information-theoretic\nhybrid in place of the original target-group-based computational step.\nBy dispe nsing with dual subsystems and target-group-based assumptions\,\nwe achiev e a significantly simpler and tighter security proof along\nwith a more li ghtweight implementation. Our construction reduces\nciphertext size by 33 percent\, shrinks user secret keys by 66 percent\,\nand requires 50 percen t fewer pairing operations during decryption –\nall while continuing to support arbitrary collusions of users and\nauthorities. These improvements mark a notable advance over the\nstate-of-the-art fully adaptive decentra lized MA-ABE scheme of Datta\net al. [EUROCRYPT 2023]. We instantiate the scheme in both\ncomposite-order bilinear groups under standard subgroup-de cision\nassumptions and in asymmetric prime-order bilinear groups under th e\nMatrix-Diffie–Hellman assumption. We further show how the\nKowalczyk –Wee attribute-reuse technique [EUROCRYPT 2019] seamlessly\nlifts our co nstruction from “one-use” boolean span programs (BSP)\nto “multi-use ” policies computable in NC1\, resulting in a\nsimilarly optimized const ruction over the state-of-the-art by Chen et\nal. [ASIACRYPT 2023].\n\nGoi ng beyond the Boolean world\, we present the first MA-ABE\nconstruction fo r arithmetic span program (ASP) access policies\,\ncapturing a richer clas s of Boolean\, arithmetic\, and combinatorial\ncomputations. This advancem ent also enables improved concrete\nefficiency by allowing attributes to b e handled directly as field\nelements\, thereby eliminating the overhead o f converting arithmetic\ncomputations into Boolean representations. The co nstruction – again\npresented in composite and prime orders – retains decentralization\nand adaptive user-key security\, and highlights inherent barriers to\nhandling corrupted authorities in the arithmetic setting.\n\ nThis is joint work with Pratish Datta (NTT Research) and Junichi\nTomida (NTT Research).\n\nIn Person and Zoom Participation.  See announcement.  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9167241 DTSTART;TZID=America/New_York:20250912T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250912T163000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:Doctoral Thesis Proposal - Bailey Miller CLASS:PUBLIC DESCRIPTION:Speaker: BAILEY MILLER\, Ph.D. Student\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: Stochastic Geometry Primit ives\n\nNumerical computing on complex geometry faces two core challenges: \nrepresenting geometry and performing computation on it.\nDiscretization —voxels\, meshes\, global solves—works until geometry\nis too detailed or uncertain to resolve. To overcome these\nlimitations\, we propose a co mplementary paradigm—stochastic graphics\nprimitives (SGPs)—that use r andomness to avoid discretization in\nboth representation and computation. \n\nFirst\, we’ll survey SGPs in graphics today: Monte Carlo rendering a s\nan algorithmic primitive that interacts with geometry via local\nquerie s\, and participating-media models as distributional\nrepresentations that replace explicit particle interactions with\nfree-flight sampling. Buildi ng on these ideas\, we’ll show how the\nsame principles extend beyond li ght transport to Monte Carlo PDE\nsolvers that handle a range of boundary conditions using only local\ngeometric queries\, and stochastic solid repr esentations that move past\nmicroparticle media to prior-free\, uncertaint y-aware geometry. These\nprimitives are modular and differentiable\, enabl ing inverse\nreconstruction and optimization-driven shape design without r emeshing.\n\nCrucially\, we’ll position these methods as general-purpose \nprimitives: black-box operators for physics simulation (elliptic and\ntr ansport PDEs)\, geometric computation (harmonic coordinates\,\ndistance-dr iven queries\, shape optimization)\, and machine learning\n(differentiable PDE layers or neural PDE surrogates supervised by\nstochastic operators). In this view\, SGPs provide a common API in\nplace of meshes and global s olves\, so the same primitives serve\nsimulation\, geometry processing\, a nd learning.\n\nFinally\, we’ll outline current limits—hyperbolic and nonlinear\nPDEs—and a path forward via hybrid neural–Monte Carlo metho ds that\niteratively refine a neural surrogate under supervision while\npr eserving the geometric scalability and robustness of Monte Carlo PDE\nsolv ers. I’ll close with practical\, more expressive stochastic\nsurface mod els and a roadmap toward more general-purpose SGPs.\n\nThesis Committee\n\ nIoannis Gkioulekas (Chair)\n\nKeenan Crane\n\nNicholas Boffi\n\nRavi Rama moorthi (University of California San Diego)\n\nMathieu Desbrun (École Po lytechnique / Inria Paris-Saclay)\n\nAdditional Information\n\nIn Person a nd Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916771c DTSTART;TZID=America/New_York:20250908T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250908T130000 SUMMARY:Graphics Seminar - Gilbert Bernstein CLASS:PUBLIC DESCRIPTION:Speaker: GILBERT BERNSTEIN\, Assistant ProfessorPaul G. Allen S chool of\nComputer Science and EngineeringUniversity of Washington\n\nTalk Title: Treating 3d Design and Fabrication as Programming Language\nand So ftware Engineering Problems\n\nOften\, when I tell people “I work on Pro gramming Languages and\nComputer Graphics” I get a kind of befuddled lo ok.  If the other\nperson is feeling polite\, maybe I get a friendly ques tion\, “What\ndoes that mean?”  I’m glad you asked\, dear reader an d listener. \nIt could mean designing Domain-Specific Languages (DSLs) fo r\ncomputations requiring high-performance in\nComputer Graphics—langua ges that automatically retarget programs to\ndifferent parallel platforms.   Or it could mean designing languages\nwith automatic differentiation fo r applications in inverse rendering\,\nin optimization\, or in simulation. \n\nIn this talk\, I’d like to tell you about a less obvious\ninterpreta tion of “Programming Languages and\nComputer Graphics”—3d design to ols and fabrication pipelines. \nIn particular\, I will discuss work on C omputer-Aided Design and\nKnitting Machines.\n\nAs many of you know from w ork by Jenny Lin and Jim McCann\, we can view\nknitting machine programs s imultaneously as a plan of execution (an\noperational view) and as a speci fication of the resulting knitted\nobject.  I will discuss our most recen t collaborative work\, led by\nNat Hurtig that provides a decision procedu re for equivalence of\nknitting machine programs.  This will take us on a delightful detour\nthrough the version of category theory that is more ab out drawing cool\npictures of knots\, rather than tying your brain in knot s.\n\nThen\, (or maybe first? TBD) I will discuss work led by Felix Hahnle in\non helping 3D CAD users “debug” their “programs.”  It may\nsu rprise you to learn\, but at its heart\, parametric CAD is about\n“reify ing” the 3d modeling process into a “program” or\n“history” that then becomes the true object of the designer’s\nactivity.  As a conseq uence\, CAD programs create the opportunity for\nnovel kinds of bugs.  I will describe how we can design\nuseful debuggers for CAD programs by ta king advantage of this\nperspective.\n\n— \n\nGilbert Bernstein is an A ssistant Professor at the University of\nWashington. He has published broa dly at SIGGRAPH\, POPL\, PLDI\, ASPLOS\,\nOOPSLA\, ICFP\, and UIST. Howeve r\, his website is very out of date\, as\nits SSL certificate\, for which he apologizes. He enjoys knitting and\ncooking in his personal time.\n\nFa culty Host:  Jim McCann\n\nIn Person and Zoom Participation.  See announ cement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9167b88 DTSTART;TZID=America/New_York:20250919T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250919T150000 LOCATION:Remote Access - Zoom SUMMARY:SCS Town Hall for Faculty and Staff CLASS:PUBLIC DESCRIPTION:Talk Title: Faculty & Staff Town Hall\n\nFriday\, September 19\ , 2025\, 2 – 3pm \n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9167e74 DTSTART;TZID=America/New_York:20250910T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250910T130000 LOCATION:Gates Hillman 8102 SUMMARY:Joint Theory Lunch Seminar / Speaking Skills Talk - Noah Singer CLASS:PUBLIC DESCRIPTION:Speaker: NOAH SINGER\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: A quadratic classical spe edup for planted kXOR\n\nA recent work of Schmidhuber et al. (QIP\, SODA\, & Phys. Rev. X 2025)\nexhibited a quantum algorithm for the noisy planted kXOR problem\nrunning quartically faster than all known classical algorit hms. \n\nIn this work\, we design a new classical algorithm that is\nquad ratically faster than the best previous one\, in the case of large\nconsta nt k. Thus for such k\, the quantum advantage of Schmidhuber et\nal. becom es only quadratic.\n\nOur algorithm\, which also works in the semirandom c ase\, combines tools\nfrom sublinear-time algorithms (essentially\, the bi rthday paradox) and\npolynomial anticoncentration.\n\nPresented as part of the Theory Lunch Seminar\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91681dc DTSTART;TZID=America/New_York:20251009T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251009T180000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:Raj Reddy Artificial Intelligence Lecture - Torsten Hoefler CLASS:PUBLIC DESCRIPTION:Speaker: TORSTEN HOEFLER\, Professor of Computer Science\, Dire ctor\,\nScalable Parallel Computing Lab\, Computer Science Department\, ET H\nZürich\n\nTalk Title: From Large Language Models to Reasoning Language Models -\nThree Eras in The Age of Computation.\n\nWe will explore the fa scinating evolution of Large Language Models\n(LLMs) and their transformat ive journey through the lenses of\ncomputation and optimization. We begin by tracing the origins of LLMs\,\nhighlighting how advances in computation and optimization were pivotal\nin their development. We then delve into t he key optimizations that\nhave achieved a staggering 1\,000x cost reducti on\, making LLMs widely\naccessible even on portable devices. Moving forwa rd\, we address the\nlimitations of human-generated data and introduce the concept of\nconstructive hallucination in LLMs. This technique allows for the\ngeneration of new hypotheses and their validation through reasoning\ nchains\, pushing the boundaries of knowledge creation. Next\, we provide\ nan overview of the technology fundamentals and early successes of\nreason ing models\, such as OpenAI's o1 and o3 preview. These models\,\nwhile sig nificantly enhancing computational capabilities\, also\nexponentially incr ease computational demands. Finally\, we conclude by\npresenting our ambit ious Ultra Ethernet effort\, which aims to\nestablish the interconnect sta ndard for future AI workloads. This\ninitiative is crucial in meeting the growing demands at the system\nlevel\, ensuring seamless and efficient ope ration in the age of\nreasoning models.\n\n—\n\nTorsten Hoefler is a Pro fessor of Computer Science at ETH Zurich\, a\nmember of Academia Europaea\ , and a Fellow of the ACM\, IEEE\, and ELLIS.\nHe received the 2024 ACM Pr ize in Computing\, one of the highest honors\nin the field. His research i nterests revolve around the central topic\nof \"Performance-centric System Design\" and include scalable networks\,\nparallel programming techniques \, and performance modeling. Torsten won\nbest paper awards at the ACM/IEE E Supercomputing Conference SC10\,\nSC13\, SC14\, SC19\, SC22\, SC23\, SC2 4\, HPDC'15\, HPDC'16\, IPDPS'15\, and\nother conferences. He published hu ndreds of peer-reviewed scientific\nconference and journal articles and au thored chapters of the MPI-2.2\nand MPI-3.0 standards. He received the IEE E CS Sidney Fernbach Award\,\nthe ACM Gordon Bell Prize\, the ISC Jack Don garra award\, the Latsis\nprize of ETH Zurich\, as well as the German Max Planck-Humboldt Medal.\nAdditional information\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9168660 DTSTART;TZID=America/New_York:20250904T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250904T173000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - Victor Kemmoe CLASS:PUBLIC DESCRIPTION:Speaker: VICTOR YOUDOM KEMMOE\, Ph.D. StudentBrown University\n \nTalk Title: Lattice-Based Accumulator and Application to Anonymous\nCred ential Revocation\n\nAn accumulator is a cryptographic system for compactl y representing a\nset of elements such that every element in the set has a short\nmembership witness.  A dynamic accumulator\, furthermore\, allows \nelements to be added to and deleted from the accumulator.  Camenisch\na nd Lysyanskaya (CRYPTO’02) constructed the first dynamic\naccumulator un der the strong-RSA assumption and showed how it can be\nused to enable rev ocation of anonymous credentials.  In this talk\, I\nwill present a commu nication-efficient cryptographic accumulator based\non the Module-SIS assu mption\, which is an accumulator that allows\nadding elements from a set w ithout the need to update membership\nwitnesses\, and show how it can be u sed in the context of anonymous\ncredential revocation.\n\nJoint work with Anna Lysyanskaya and Ngoc Khanh Nguyen.\n\nRelated Reading\n\n—\n\nVict or Youdom Kemmoe is a rising fifth-year Ph.D. student in computer\nscience at Brown University working under the supervision of Anna\nLysyanskaya. B efore joining Brown\, he obtained both a Master's and a\nBachelor's in com puter science from Kennesaw State University\,\nGeorgia.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9168a36 DTSTART;TZID=America/New_York:20250908T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250908T130000 LOCATION:Panther Hollow Room 4105\, Mehrabian Collaborative Innovation Cent er SUMMARY:CyLab Seminar - Dahlia Malkhi CLASS:PUBLIC DESCRIPTION:Speaker: DAHLIA MALKHI\, Professorand Head of Foundations of Fi nancial\nTechnology Research LabDepartment of Computer ScienceUniversity o f\nCalifornia Santa Barbara\n\nTalk Title: From Libra to Space: Research A dvances in Scaling\nByzantine Consensus\n\nThe Facebook Libra (Diem) proje ct aimed to create a global payment\nsystem using a consortium-operated bl ockchain. Libra\, with Professor\nMalkhi as CTO\, sought to be a regulator y-friendly financial system.\nAlthough Diem never launched\, its technolog y had a big impact on the\nindustry.\n\nLibra used a Byzantine Consensus m ethod called HotStuff. HotStuff's\nkey breakthrough\, linearity\, tackled Byzantine Agreement issues in\nmany settings. It unlocked tight upper boun ds for partially\nsynchronous\, asynchronous\, synchronous\, and flexible models. Moreover\,\nDiemBFT\, the first large-scale HotStuff implementatio n for Libra/Diem\,\nnow powers the Aptos blockchain. Several platforms\, l ike Cypherium\,\nFlow\, Celo\, Espresso Systems\, and Pocket Network\, hav e also adopted\nHotStuff variants as their core consensus engines.\n\nToda y\, SpaceComputer.IO employs HotShot\, a production-grade HotStuff\nvarian t\, to create a network of tamper-proof satellites. This talk\nwill end wi th the technical challenges of achieving decentralization\nwith security t hat can be verified from space.\n\n—\n\nDahlia Malkhi is a Professor in the Department of Computer Science of\nUCSB since 2024. She heads the Foun dations of Financial Technology\nlab. Her research over two decades spans broad aspects of reliability\nand security of distributed systems\, recent ly with focus on\nblockchains and advances in financial technology. Her wo rk resulted in\nover 200 publications as well as a strong impact on comput ing\ntechnology.\n\nMalkhi is sought-after industrial consultant\, current ly advising\nvarious projects\, including Space Computer\, Lyquor Labs\, N ubit|Bitcoin\nThunderbolt\, Espresso Systems\, and Chainlink Labs. Formerl y\, she\nserved as Distinguished Scientist of Chainlink Labs (2022-2025). From\n2019 to 2022\, she was the CTO at the Diem Association\, and Lead\nR esearcher at Novi Financial. In 2014\, she co-founded VMware Research\nand became a Principal Researcher at VMware until 2019. Prior to that\,\nMalk hi was a partner principal researcher at Microsoft Research\,\n2004-2014\; an Associate Professor of the Hebrew University of\nJerusalem\, and a sen ior researcher at AT&T Labs.  \n\nFaculty Host:  Elaine Shi\n\nIn Perso n and Zoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9168ee7 DTSTART;TZID=America/New_York:20250903T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250903T130000 SUMMARY:Theory Lunch Seminar -Yonggang Jiang CLASS:PUBLIC DESCRIPTION:Speaker: YONGGANG JIANG\, Max Planck Institute\n\nTalk Title: P arallel (1+ε)-Approximate Mincost Flow in Almost Optimal\nDepth and Work\ n\n We present a parallel algorithm for computing (1+ ε) approximate\nm incost flow on an undirected graph with m edges\, where capacities\nand c osts are assigned to both edges and vertices. Our algorithm\nachieves m1+ o(1) work and no(1) depth when ε > 1/polylog(m) making\nboth the work an d depth almost optimal\, up to a subpolynomial factor.\n\nPrevious algorit hms with m1+o(1) work required Ω(m) depth\, even for\nspecial cases of m incost flow with only edge capacities or max flow\nwith vertex capacities. \n\nOur key technical contribution is the first construction of\nlength-co nstrained flow shortcuts with (1+ ε) length slack\, no(1)\ncongestion sla ck\, and no(1) step bound. This provides a strict\ngeneralization of the i nfluential concept of (n{o(1)\,ε)-hopsets\n[Cohen94]\, allowing for addit ional control over congestion. Previous\nlength-constrained flow shortcuts [HHLRS24] incur a large constant in\nthe length slack\, which would lead to a large approximation factor. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916921d DTSTART;TZID=America/New_York:20250904T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250904T150000 LOCATION:Newell-Simon 3305 SUMMARY:5th Year Thesis Presentation - Frank Chen CLASS:PUBLIC DESCRIPTION:Speaker: FRANK CHEN\, Master's Student\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: Towards Utilizing Cached C ontext for Faster and Smarter\nCode Agents\n\nRecent advances in Large Lan guage Models (LLMs) have enabled the\ndevelopment of coding agents that ca n autonomously perform tasks such\nas code generation\, debugging\, patch validation\, etc.\nIndustry-proprietary systems (e.g.\, Colab AI\, Claude Code\, Cursor\nAgent) and open-source frameworks (e.g.\, Gemini Cli\, SWE- Agent\,\nAutoCodeRover\, Agentless) have demonstrated both practicality an d\npopularity in real-world software engineering workflows. Despite these\ nsuccesses\, existing agentic frameworks face two common challenges:\nprov iding accurate and complete context information and ensuring\nlow-latency patch generation under heavy workloads. Although prior\nwork has proposed partial solutions addressing either context\nretrieval or latency\, little attention has been paid to the joint\noptimization of both aspects. Ideal ly\, joint optimization should\nenhance both performance and speed without incurring additional cost\,\nwhich is particularly critical in modern fas t-paced\, iterative\nsoftware engineering environments. \n\nThis thesis i nvestigates integrating existing state-of-the-art\napproaches to these cha llenges\, specifically RepoGraph for smart\ncontext retrieval on the front end and CacheBlend for fast inference on\nthe backend. To verify feasibili ty of an integration\, we then evaluate\nperformance of the frontend acros s multiple LLMs under realistic\ncode-agent scenarios\, and measure the la tency improvement of the\nbackend against systems such as vLLM and SGLang on trace generated by\nfrontend under the same benchmark. The results high light the\ntrade-offs between cost\, efficiency\, and performance and argu e for the\nnecessity of integrated solutions that achieve balance between the\nfactors mentioned. Finally\, we suggest a preliminary design for an\n end-to-end system that combines the benefits of RepoGraph and\nCacheBlend via a simple adapter module\, along with optimizations in\nthe original al gorithms. Overall\, the findings suggest promising\ndirections toward buil ding a robust and production-ready coding agent\nthat is both fast and hig h-performing.\n\nThesis Committee\n\nRashmi K. Vinayak (Chair)\n\nZhihao J ia\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91696ac DTSTART;TZID=America/New_York:20250908T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250908T180000 LOCATION:Newell-Simon 3305 SUMMARY:Computer Science Master's Program Information Session CLASS:PUBLIC DESCRIPTION:Are you interested in continuing your education and obtaining a \nComputer Science Master's Degree?  \n\nAttend this information session !  \n\nProfessor Peter Steenkiste will discuss The 5th-Year Master of Sc ience\n(Research) Program\n\nThe Fifth Year MS - Research Program is a 12- month research-focused\nprogram for graduates of the SCS bachelor’s prog rams.  Students\ncomplete four Ph.D.-level classes while working on thesi s research\nwith a research advisor in SCS.  The program has application\ ndeadlines in October and January.\n\nProfessor Dave Eckhardt  will discu ss The Master of Science in\nComputer Science (MSCS) Program\n\nThe Master of Science in Computer Science (MSCS) Program offers\nstudents with a Bac helor’s degree the opportunity to improve their\ntraining with advanced study in Computer Science (research is\noptional). We cater to students wi th basic analytic skills and a\nstrong aptitude for mathematics\, programm ing and logical reasoning. \nStudents with SCS undergraduate minors are w elcome to apply. \nApplication deadlines are December for students planni ng to start in\nAugust\, and July for students planning to start in Januar y.\n\nAdditional Information\n\n5th-year Program HandbookMSCS Program Hand book\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9169a3e DTSTART;TZID=America/New_York:20250918T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250918T133000 LOCATION:Newell Simon 4305 and Zoom SUMMARY:AI-SDM Seminar - Suresh Venkatasubramanian CLASS:PUBLIC DESCRIPTION:Speaker: SURESH VENKATASUBRAMANIAN\, Professor of Data Science and\nComputer Science and Professor of Humanities\nCo-Director\, AI Resear ch Institute on Interaction for AI Assistants\n(ARIA)\nBrown University\n\ nTalk Title: Are We Winning Yet? Frames\, Measurements\, and Tools for AI\ nGovernance\n\nSuresh Venkatasubramanian directs the Center for Technologi cal\nResponsibility\, Reimagination\, and Redesign (CNTR) with the Data\nS cience Institute at Brown University\, and is a Professor of Computer\nSci ence and Data Science. Suresh's background is as a computer\nscientist and his current research interests lie in algorithmic\nfairness\, and more ge nerally the impact of automated decision-making\nsystems in society. He is the Co-Director of the newly established AI\nResearch Institute on Intera ction for AI Assistants (ARIA).\n\nSuresh recently finished a stint in the Biden-Harris administration\,\nwhere he served as Assistant Director for Science and Justice in the\nWhite House Office of Science and Technology P olicy. In that capacity\,\nhe helped co-author the Blueprint for an AI Bi ll of Rights.\n\nPrior to Brown University\, Suresh was at the University of Utah\, where\nas an assistant professor he was the John and Marva Warno ck Assistant\nProfessor. He has received a CAREER award from the NSF for h is work in\nthe geometry of probability\, a test-of-time award at ICDE 201 7 for his\nwork in privacy\, and a KAIS Journal award for his work on audi ting\nblack-box models. His research on algorithmic fairness has received\ npress coverage across the globe\, including NPR’s Science Friday\,\nNBC \, and CNN\, as well as in other media outlets. He is a past member\nof th e Computing Community Consortium Council of the CRA\, spent 4\nyears (2017 -2021) as a member of the board of the ACLU in Utah\, and is\na past membe r of New York City’s Failure to Appear Tool (FTA)\nResearch Advisory Cou ncil\, the Research Advisory Council for the First\nJudicial District of P ennsylvania and the Utah State Auditor's\nCommission on protecting privacy and preventing discrimination. He was\nrecently named by Fast Company to their AI20 list of thinkers shaping\nthe world of generative AI.\n\nREGIST ER → to attend in-person or on Zoom\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9169e6b DTSTART;TZID=America/New_York:20250904T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250904T160000 LOCATION:CNA Room\, Wean Hall 7281 SUMMARY:ACO Seminar - Dylan Altschuler CLASS:PUBLIC DESCRIPTION:Speaker: DYLAN ALTSCHULER\, Postdoctoral Researcher\, Departmen t of\nMathematics Sciences\, Carnegie Mellon University\n\nTalk Title: Dim ension reduction\, universal approximation\, and\nnonlinear spectral gaps\ n\nWhen can a low–regularity object\, such as a metric space or a graph\ ,\nbe embedded into a high regularity object\, such as a normed space?\nTh is is a fundamental question in metric geometry\, specifically the\nRibe p rogram\, with wide-ranging applications in algorithm design\,\ngeometric g roup theory\, and functional analysis. Traditional\napproaches to such pro blems rely on heavy machinery from analysis and\ngeometry. We will introdu ce a recent program—joint with P. Dodos\, K.\nTikhomirov\, and K. Tyros —towards developing direct combinatorial and\nprobabilistic methods for studying (random) graph embeddings. Some\nhighlight results include the re solution of a long–standing question\non the asymptotics of Bourgain’s metric dimension reduction modulus\,\nas well as a solution to an outstan ding problem of Jon Kleinberg.\n\nNo prior knowledge of metric geometry wi ll be assumed\; the first\nportion of the talk will aim to give a high-lev el overview of some of\nthe key definitions\, techniques\, and questions i n the field.\n\n4:00 pm → Jane Street-sponsored tea and cookies in Wea n 6220. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916a1e8 DTSTART;TZID=America/New_York:20250916T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250916T140000 LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:Study Abroad Fair CLASS:PUBLIC DESCRIPTION:Speaker: Learn more about study abroad opportunities at CMU\n\n All students are welcome to attend the Study Abroad fair to learn more\nab out study abroad opportunities at CMU\, talk with program providers\nabout Study Abroad Programs and speak with current exchange students\nabout our University Exchange Partners. The fair focuses on\nopportunities for unde rgraduate students.\n\nRepresentatives from the Office of International Ed ucation Study\nAbroad Office\, CMU colleges and departments will also be o n hand to\ntalk about study abroad processes\, timelines\, and deadlines. More than\n30 study abroad university exchange partners and program provid ers\nwill also be in attendance to answer your questions.\n\nLearn More an d Register  |  More about Study Abroad \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916a512 DTSTART;TZID=America/New_York:20251003T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251003T180000 LOCATION:The Schenley Park Trails under the Panther Hollow Bridge SUMMARY:CS 45th Annual Pretty Good Race CLASS:PUBLIC DESCRIPTION:Speaker: Break out Your Running and Walking Shoes!\n\nThe 44th annual running of the Pretty Good Race and Coach Tomayko 5K\nWalk will tak e place Friday\, 30 August 2024!   Join in at 4:30 p.m.\nunder the Panth er Hollow Bridge in Schenley Park. \n\nAll members of the SCS Community ( and SEI!) are welcome to\nparticipate. \n\nRegistration is required.  Wa tch for details.\n\nHeld every September\, this is a great opportunity to meet your\nfriends\, clock your racing scores\, run/walk the course or sim ply cheer\non and support your favorite \"athletes\"!  The course is idea l for\npros and novices. Bets between students and their advisors are\ntra ditional and honored.  The race began 45 years ago as part of the\nComput er Science Department IC. \n\nIt's a fun\, refreshing and challenging eve nt - join in as your skill\nlevels permit! \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916a876 DTSTART;TZID=America/New_York:20250827T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250827T130000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - William He CLASS:PUBLIC DESCRIPTION:Speaker: WILLIAM HE\, Ph.D. StudentComputer Science DepartmentC arnegie\nMellon University\n\nTalk Title: How to Count Using Taylor Series \n\nI will tell you about a cool approach to approximate counting using\na pproximations of Taylor series. Based on works by Shearer\, Dobrushin\,\nS cott-Sokal\, Barvinok\, Patel-Regts\, Bencs-Regts. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916ab69 DTSTART;TZID=America/New_York:20250905T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250905T130000 LOCATION:Panther Hollow 4105\, Mehrabian Collaborative Innovation Center SUMMARY:ECE Graduate Seminar - Alex Aiken CLASS:PUBLIC DESCRIPTION:Speaker: ALEX AIKEN\, lcatel-Lucent Professor of Computer\nScie nceStanford University\n\nTalk Title: Efficiency in Parallel Runtime Syste ms\n\nWhat makes a parallel runtime system \"efficient\"?  We present a\n metric\, minimum effective task granularity (METG)\, that differs in\nimpo rtant ways from previous approaches to measuring the performance\nof runti me systems.  We show that METG is generally applicable and\ncan be used t o meaningfully compare a wide variety of parallel runtime\nsystems from sc ientific computing and data analytics.  We also\npresent recent results i n improving the METG of a class of task-based\nruntime systems\, and show that we come close to reaching the minimum\npossible METG.\n\n—\n\nAlex Aiken is the Alcatel-Lucent Professor of Computer Science at\nStanford. Al ex received his Bachelor's degree in Computer Science and\nMusic from Bowl ing Green State University in 1983 and his Ph.D. from\nCornell University in 1988. Alex was a Research Staff Member at the\nIBM Almaden Research Cen ter (1988-1993) and a Professor in the EECS\ndepartment at UC Berkeley (19 93-2003) before joining the Stanford\nfaculty in 2003. His research intere sts are in areas related to\nprogramming languages. He is an ACM Fellow\, a recipient of ACM\nSIGPLAN's Programming Languages Achievement Award and Phi Beta Kappa's\nTeaching Award\, and a former chair of the Stanford Comp uter Science\nDepartment. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916af3f DTSTART;TZID=America/New_York:20250828T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250828T173000 LOCATION:Gates Hillman 8115 and Zoom (New Room) SUMMARY:Crypto Seminar - Ziyi Guan CLASS:PUBLIC DESCRIPTION:Speaker: ZIYI GUAN\, Ph.D. Student\nTheory Group\nSchool of Com puter and Communication Sciences\nÉcole polytechnique fédérale de Lausa nne (EPFL)\n\nTalk Title: On the Security of Succinct Arguments from Proba bilistic\nProofs\n\nSuccinct arguments are fundamental cryptographic primi tives with\nwide-ranging applications. A common approach to build succinct \narguments is from probabilistic proofs\, dating back to Kilian’s\nprot ocol that combines a PCP and a Merkle tree.\n\nIn this talk\, I will prese nt the tightest bound on the regular\nsecurity of Kilian’s protocol and show how to obtain similar bounds\nfor more general argument systems\, suc h as those based on polynomial\ncommitment schemes. I’ll conclude with r esults that achieve\npost-quantum security and Fiat-Shamir security for ge neral classes of\narguments.\n\n—\n\nZiyi Guan is a fourth-year Ph.D. st udent at the EPFL theory group\,\nadvised by Alessandro Chiesa and Mika G öös. She received a B.S. in\nComputer Science and a B.S. in Mathematical Sciences from Carnegie\nMellon University in 2020. She works on theoretic al computer science\,\nparticularly complexity theory\, probabilistic proo f systems\,\ncryptography\, and quantum computing.\n\nIn Person and Zoom P articipation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916b31d DTSTART;TZID=America/New_York:20250829T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250829T113000 LOCATION:Traffic21 Classroom and Zoom SUMMARY:Doctoral Thesis Oral Defense - Nirav Atre CLASS:PUBLIC DESCRIPTION:Speaker: NIRAV ATRE\, Ph.D. Candidate\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: Refining Classical Abstract ions of Network Subsystems\n\nWe reason about computer systems via models of their behavior —\nwhether implicit mental models\, or explicit mathem atical models. These\nmodels are the linchpins of our decision-making abil ity\, e.g.\, in\nformulating service-level agreements (SLAs)\, or tenderin g performance\nclaims. Unfortunately\, a growing disconnect between how sy stems are\nmodeled and how they are actually deployed has engendered a cla ss of\nproblems I call model incongruity: circumstances where a model's\np rediction deviates significantly from real-world behavior. Model\nincongru ities are highly pervasive in modern systems\, resulting in\nexpensive per formance anomalies\, scalability bottlenecks\, and security\nvulnerabiliti es.\n\nIn this thesis\, we argue that many incongruities observed in pract ice\ntoday are not a fundamental limitation of our modeling capabilities\, \nbut rather artifacts of using the wrong models. We show that: (a)\nassum ptions centrally underpinning contemporary models of network\nsubsystems h ave drifted far from deployment realities\; (b) these\nassumptions are fre quently violated in the field\, subverting the\noperator's expectations ab out key metrics in highly unexpected ways\;\nand\, (c) making modest model refinements not only yields designs with\nstate-of-the-art performance\, attack resilience\, and scalability\, but\nalso enables us to make rigorou s mathematical guarantees about the\nresulting system's behavior.\n\nWe ex emplify this point using case studies of three ubiquitous network\nsubsyst ems. First\, I will describe \"delayed hits\"\, an incongruity\narising in high-performance caching systems which breaks the textbook\ncaching princ iple that maximizing cache hit-rate also minimizes\nlatency\, and causes e very existing caching algorithm to make\nlatency-suboptimal decisions\; in this context\, I will introduce\nMinimum-AggregateDelay (MAD)\, a turnkey augmentation to existing\nalgorithms that makes them aware of delayed hit s\, yielding 5-35% lower\nrequest latencies. Second\, I will describe \"al gorithmic complexity\nattacks\" (ACAs)\, a highly potent class of Denial-o f-Service attacks\narising from transient workload incongruity\; in this c ontext\, I will\nintroduce SurgeProtector\, an adversarial scheduling fram ework that\nprovably protects network dataplanes against ACAs\, resulting in 90-99%\nreduction in harm for the same volume of attack traffic. Finall y\, I\nwill describe BBQ\, a system borne out of addressing design incongr uity\nin hardware packet schedulers which\, for the first time\, makes it\ nfeasible to deploy packet scheduling at line-rate on modern switches\nand SmartNICs.  \n\nThesis Committee\n\nJustine Sherry (Chair)\n\nVyas Seka r \n\nWeina Wang \n\nBrighten Godfrey (University of Illinois Urbana-Cha mpaign)\n\nIn Person and Zoom Participation. See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916b865 DTSTART;TZID=America/New_York:20250829T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250829T170000 LOCATION:5th Floor Wean Hall Patio SUMMARY:CMU Math Club Kickoff and Sculpture Build CLASS:PUBLIC DESCRIPTION:Speaker: “Knight’s Move”: A Mathematical Art Installation \n\n“Knight’s Move”: A mathematical art installation brought to you\ nby the CMU Department of Mathematical Sciences and Studio Infinity. \n\n This sculpture will be constructed in its full 8-foot tall glory on\nthe a fternoon of Friday\, August 29\, on the Wean Hall patio (near the\nfront o f La Prima).  All interested CMU community members are invited\nto help t hroughout the day. The more people that come\, the\nmerrier.  \n\nIf you ’d like to participate\, feel free to fill out the Interest\nForm\, or j ust show up on the day of: we’ll need all the hands we can\nget.  There will be free food probably.\n\nLEARN MORE \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916bbb7 DTSTART;TZID=America/New_York:20251001T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251001T130000 LOCATION:McConomy Auditorium\, Cohon University Center SUMMARY:CMU Andy Awards Ceremony & Reception CLASS:PUBLIC DESCRIPTION:Staff and faculty are invited to join in at the Andy Awards\, t he\nannual celebration recognizing staff members and teams for their\nmean ingful contributions to Carnegie Mellon University. \n\nAndrew Carnegie ’s 1900 declaration \"My heart is in the work” has\nbecome a motto for the Carnegie Mellon University community. Much like\nthe Carnegie steel t hat built America\, CMU staff are the building\nblocks upon which this uni versity thrives. The Andy Awards\, named for\nAndrew Carnegie and Andrew Mellon\, aims to recognize the legacy that\nstaff members forge through th eir hard work.\n\nIndividual staff members and teams of colleagues whose w ork has had a\nsignificant impact on the university are recognized in six categories:\nCommitment to Excellence\,  Commitment to Students\, Innovat ive and\nCreative Contributions\, Spirit\, Teamwork and Collaboration\, Co mmitment\nto a Vibrant Community and Commitment to Community Engagement.  Learn\nmore\n\nNo Registration Required \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916bf25 DTSTART;TZID=America/New_York:20251007T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251007T193000 LOCATION:By Invitation Only SUMMARY:CyLab Security and Privacy Institute: 3rd Annual Secure Blockchain\ nSummit CLASS:PUBLIC DESCRIPTION:TDespite advances\, many research questions still need to be an swered\nto ensure blockchain protocols and applications are ready for\nwid espread use. As the industry continues to see more consumer\napplications\ , such as decentralized finance (DeFi)\, securing the\nfundamental researc h elements of the technology\, especially from a\nsecurity\, privacy\, eth ics\, and societal impact perspective\, is of\nutmost importance.\n\nThis event will bring together experts from academia and industry to\ndiscuss t he future of blockchain research\, technology\, and\napplications\, focusi ng on a variety of topics\, including\ncrypto-economics\, applied cryptogr aphy\, programming languages\,\nsecurity and privacy\, policy and usabilit y\, ethics and equity\, and\nmore.\n\nThe Third Annual Secure Blockchain S ummit is co-located with the\nSeventh International Conference on Advances in Financial Technologies\n(AFT’25)\, a premier venue for presenting th e latest developments in\ntechnologies related to novel financial infrastr ucture such as\ncryptocurrencies and their applications\, blockchains\, an d exchanges.\n\nBy Invitation Only \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916c2bd DTSTART;TZID=America/New_York:20250828T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250828T120000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Joshua Williams CLASS:PUBLIC DESCRIPTION:Speaker: JOSHUA NATHANIEL WILLIAMS\, Ph.D. Candidate\nComputer Science Department\nCarnegie Mellon University\n\nTalk Title: Understandin g Representations of Humans in Generative\nImage Modeling Through Discrete Counterfactual Prompt Optimization\n\nText-to-image (T2I) models are wide ly used generative systems\, making\nit essential to understand how they r epresent human subjects.\nComparing generated images across carefully desi gned prompts can\nreveal representational patterns\, some of which reflect harmful biases\nrequiring intervention. Existing approaches often rely on fixed prompt\ntemplates or identity categories\, which are useful for ben chmarking\nbut risk blind spots shaped by researchers’ assumptions.\n\nT his thesis introduces methods grounded in counterfactual and\ncontrastive analysis to uncover representational asymmetries and harms\nbeyond predefi ned categories. We show that effective explanations for\nclassifiers must account for the underlying data distribution\; without\nthis\, analyses ri sk spurious conclusions. To address this\, we adapt\nthe graphical model u nderlying counterfactual explainability and\npropose a new distribution-aw are metric.\n\nBuilding on these insights\, we further develop distributio nally\ninformed approaches to prompt optimization in T2I settings. Our\nfr amework incorporates multiobjective optimization across language\nmodels w ith distinct tokenizers and embeddings\, enabling richer\nexploration of r epresentational behaviors. Finally\, we present an\nunsupervised strategy for surfacing candidate prompts that reveal\npreviously undocumented asymm etries. By linking the linguistic\npatterns of generative models to their visual outputs\, we advance\nmethods for diagnosing biases and targeting s pecific representational\nbehaviors in training and evaluation.  \n\nThe sis Committee\n\nZico Kolter (Chair)\n\nHoda Heidari\n\nAditi Raghunathan\ n\nSarah Laszlo (Visa)\n\nIn Person and Zoom Participation.  See announce ment.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916c719 DTSTART;TZID=America/New_York:20250826T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250826T143000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Thesis Oral Defense - Hojin Park CLASS:PUBLIC DESCRIPTION:Speaker: HOJIN PARK\, Ph.D. Candidate\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: Cost-Efficient Storage and Caching in Public Clouds\n\nAs modern data-intensive workloads increasingl y migrate to the public\ncloud\, managing the resulting costs has emerged as a pressing\nchallenge despite the operational simplicity and elasticity that cloud\nenvironments offer. Although many efforts in cost optimizatio n have\nfocused on computation\, storage-related costs have received\ncomp aratively less attention despite being a significant portion of\ntotal clo ud spending. In particular\, two categories dominate\nstorage-related cost s in public cloud: the cost of deploying and\noperating storage clusters i n the cloud\, and the cost of accessing\ndata across geographically distri buted regions or clouds. These\nchallenges cannot be effectively addressed by existing optimization\ntechniques developed for on-premise environment s\, since they often\noverlook the unique characteristics of public clouds \, including\nelastic resource provisioning\, diverse cost-performance tra de-offs\,\nand dynamic and unique access patterns found in cloud object st orage\nworkloads.\n\nThis dissertation addresses these challenges by propo sing a\ncost-efficient approach to designing storage and caching systems t hat\nare cloud-aware\, elastic\, and adaptive to workload behavior. It\nin troduces three systems that target key aspects of cloud storage cost\nopti mization. First\, Mimir reduces the cost of the deployment of\nstorage clu sters by automatically selecting cost-effective\ncombinations of virtual m achines and block storage types\, based on\nprofiling workload characteris tics and benchmarking available resource\noptions. Second\, Macaron reduce s cross-region and cross-cloud data\naccess costs by auto-configuring a ca che with a tiered storage\narchitecture that leverages low-cost object sto rage and dynamically\nresizes the cache based on workload changes. Third\, Macaron+ builds on\nMacaron by introducing a cost-aware prefetching techn ique that\nanalyzes object-level access patterns to reduce latency in work loads\nwith high cold miss ratios\, while preserving cost-efficiency.\nTog ether\, these systems demonstrate that by tailoring automated\nresource se lection\, adaptive configuration\, and predictive techniques\nto the chara cteristics of the public cloud\, it is possible to\nsignificantly reduce t he cost of storing and accessing data.\n\nThesis Committee\n\nGeorge Amvro siadis (Co-chair)\n\nGregory R. Ganger (Co-chair)\n\nJignesh M. Patel\n\nC arlo Curino (Microsoft Research) \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916cbf4 DTSTART;TZID=America/New_York:20250827T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250827T130000 LOCATION:Wean Hall 4110 and Zoom SUMMARY:Engineering and Public Policy Talk - Anurag K. Srivastava CLASS:PUBLIC DESCRIPTION:Speaker: ANURAG SRIVASTAVA\, Raymond J. Lane Professorship and\ nChairComputer Science and Electrical Engineering DepartmentWest\nVirginia University\n\nTalk Title: Augmenting Human Operators’ Decision-Making f or\nCyber-Resilient Energy Systems\n\nEnsuring the resilience of the cyber -physical-human electric grid is\nessential for maintaining power to criti cal loads such as hospitals\,\nairports\, and emergency services\, especia lly during extreme weather\nevents and cyber disruptions. Human decision-m aking plays a central\nrole in this resilience\, guiding real-time actions and adaptive\nstrategies when the grid is under stress. The increasing in tegration\nof Distributed Energy Resources (DERs)\, and Grid-Enhancing\nTe chnologies (GETs) has improved grid flexibility and automation but\nhas al so introduced new cyber and operational challenges. Extreme\nweather and c yber events\, can strain grid infrastructure\, and requires\ncoordinated s trategies. The resilience of electric grids depends on\nmultiple factors: (a) identifying weather and cyber threats and\nplanning for resilience\, ( b) utilizing advanced tools for real-time\nanalysis and automated control\ , (c) applying metrics to guide\ninvestment and operational decisions\, (d ) enhancing operator training\nand cognitive flexibility\, and (e) employi ng a testbed to validate\nresilience tools.\n\nThis talk will present resi lience objectives\, resilience metrics for\ngrid operations\, physics-awar e machine learning techniques for anomaly\nand event detection\, cognitive flexibility assessment\, and a testbed\nfor validating cyber-physical-hum an resilience strategies\, providing\ninsights that can inform both techno logy development and energy policy\nto ensure a secure\, reliable\, and re silient electric grid.\n\n—\n\nAnurag K. Srivastava holds the Raymond J. Lane Professorship and\nserves as Chairperson of the Computer Science and Electrical\nEngineering Department at West Virginia University. Additiona lly\, he\nis an adjunct professor at Washington State University and a sen ior\nscientist at the Pacific Northwest National Lab. He earned his Ph.D.\ nin electrical engineering from the Illinois Institute of Technology in\n2 005. Dr. Srivastava's research focuses on data-driven algorithms and\ntool s for cyber-resilient electric energy systems. His impactful\nresearch pro jects have resulted in the implementation of tools at\nutility control cen ters\, supported by over $66M in funding from\nentities such as the US Dep artment of Energy\, National Science\nFoundation\, Siemens Corporate Resea rch\, Electric Power Research\nInstitute\, Schweitzer Engineering Lab\, Po wer System Engineering\nResearch Center\, Office of Naval Research\, and v arious National Labs.\nOver the years\, he has held visiting positions at organizations\nincluding Réseau de transport d´électricité in France\, RWTH Aachen\nUniversity in Germany\, PEAK Reliability Coordinator\, Idaho National\nLaboratory\, PJM Interconnection\, Schweitzer Engineering Lab ( SEL)\, GE\nGrid Solutions\, Massachusetts Institute of Technology\, and Mi ssissippi\nState University. He is an IEEE Fellow\, recipient of IEEE PES Pete\nSauer Educator Award\, numbers of best papers award\, leading multip le\nIEEE technical subcommittee/ WGs (Power System Operation\, Resiliency\ ,\nMicrogrid\, voltage stability\, distributed optimization)\, and the\nau thor of over 400 technical publications\, 3 books\, and 3 patents.\n\nREGI STER\n\nThis seminar is presented jointly by the Center for Climate and En ergy\nDecision Making and the Carnegie Mellon Electricity Industry Center.  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916d197 DTSTART;TZID=America/New_York:20250814T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250814T153000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year Master's Thesis Presentation - Jinqi Chen CLASS:PUBLIC DESCRIPTION:Speaker: JINQI CHEN\, Master's StudentComputer Science\nDepartm entCarnegie Mellon University\n\nTalk Title: Towards Effortless High-Perfo rmance Kernel Development for\nLLM Workloads\n\nRecent advances in large l anguage models (LLMs) have pushed GPU\nhardware to its limits\, requiring highly optimized kernels for\ncompute- and bandwidth-intensive operations such as normalization\,\nmatrix multiplication\, attention\, and inter-GPU communication.\nHowever\, achieving state-of-the-art efficiency often dem ands deep\nlow-level expertise\, slowing development and limiting accessib ility.\n\nThis thesis presents TIR+\, a multi-level compiler framework tha t\nunifies high-level productivity and low-level optimization within a\nsi ngle compilation and runtime infrastructure. TIR+ spans from a\nPython-bas ed tiling DSL\, enabling rapid kernel prototyping\, to a\nhardware-centric intermediate representation (IR)\, offering\nfine-grained control over me mory\, parallelism\, and specialized\ninstructions. Between these extremes \, it provides optimized tensor\nlibraries and reusable primitives inspire d by CUTLASS and CuTe.\nCrucially\, TIR+ is distributed-aware\, supporting multi-GPU execution\nwith built-in communication management and compute –communication\noverlap. We demonstrate the capability of TIR+ through k ey LLM\nkernels\, including LayerNorm/RMSNorm\, GEMM\, FlashAttention-styl e\nattention\, and combined compute–communication kernels. Among these\n cases\, TIR+ delivers near–state-of-the-art throughput with\nsignificant ly less development effort than hand-tuned CUDA\,\ndemonstrating a unified and scalable path toward hardware-aware kernel\noptimization for current and future AI workloads.\n\nThesis Committee\n\nTianqi Chen (Chair)\n\nZhi hao Jia\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916d5a5 DTSTART;TZID=America/New_York:20250819T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250819T120000 LOCATION:Gates Hillman 7101 SUMMARY:5th Year Master's Thesis Presentation - Lauren Sands CLASS:PUBLIC DESCRIPTION:Speaker: LAUREN SANDS\, Master's Student\nComputer Science Depa rtment\nCarnegie Mellon University\n\nTalk Title: Typed Closure Conversion with Sum Types for Analyzing\nHigher Order Functions in Resource Aware ML \n\nResource Aware ML (RaML) is a tool that statically infers resource\nbo unds for OCaml programs. However\, RaML often cannot analyze\npartially ap -plied higher-order functions which are common in OCaml\ncode. In this the sis\, we present a closure conversion transformation\nthat rewrites high-e r-order programs into first-order ones analyzable\nby RaML. Unlike tradi-t ional closure conversion which wraps functions\nwith environments repre-se nted using existential types\, we use sum\ntypes\, which allows RaML to ac cess internal types and infer bounds. We\nshow that our transformation is well-typed and preserves semantics and\ncost\, ensuring that analysis of t he transformed program remains sound.\n\nThesis Committee\n\nJan Hoffman ( Chair)\n\nStephanie Balzer\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916d943 DTSTART;TZID=America/New_York:20250811T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250813T173000 LOCATION:Locations Across Campus SUMMARY:University-Wide Graduate Student Orientation CLASS:PUBLIC DESCRIPTION:Speaker: Welcome Graduate Students!Talk Title: Presented by the Office\nof Graduate and Postdoctoral Affairs\n\nMonday\, August 11\, 9am  – Wednesday\, August 13\, 2025\, 5:30pm \n\nTo welcome new graduate s tudents to the Carnegie Mellon community\nand help support the transition to graduate study\, the Office\nof Graduate and Postdoctoral Affairs\,  in partnership with many\ncolleagues across campus\, offers University-Wid e New Graduate\nStudent Orientation.\n\nAttendance at the University-Wid e New Graduate Student Orientation is\nhighly encouraged but is not mandat ory. For those unable to attend in\nperson\, key events on Monday\, August 11 will be available via\nlivestream.\n\nFull Schedule:  Master's and Ph .D. Students \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916dc85 DTSTART;TZID=America/New_York:20250807T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250807T130000 LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talk -Ezra Hoch CLASS:PUBLIC DESCRIPTION:Speaker: EZRA HOCH\, Software EngineerJane Street\n\nTalk Title : Depot: Multi-DC storage for AI/ML workloads\n\nAs the importance of AI/M L workloads grows\, so does the number of\nGPUs. Both power and availabili ty constraints led Jane Street to run\nGPU workloads in multiple data-cent ers.Managing datasets across the\nestate becomes a challenge\, especially in Jane Street's agile and\ndynamic environment.\n\nI'll talk about Depot\ , a storage metadata layer that we're building to\naddress Jane Street's u se cases\; what issues we've seen with using\nNFS's directory-structure as a metadata layer\, the API trade-offs\nwe've considered\, what API we lan ded on (a middle-ground between S3\nand a filesystem) and Depot's architec ture.\n\n—\n\nEzra Hoch is a software engineer at Jane Street\, working on\ndistributed storage systems. His previous roles at Google include\nTL- ing GCP’s file solutions\, developing GCP’s networking stack and\nlead ing Effingo\, Google’s global replication system. Prior to\nGoogle\, he was chief architect at Elastifile\, a startup developing a\nscale-out SSD- optimized filesystem (acquired by Google). His focus is\non large-scale di stributed infrastructure systems. He got his BSc in\nComputer Science\, MS c and PhD in distributed algorithms\, from the\nHebrew university of Jerus alem. \n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916e03f DTSTART;TZID=America/New_York:20250814T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250814T130000 LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talk - Vikramraj Sitpal & Ranjit\nN oronha CLASS:PUBLIC DESCRIPTION:Speaker: VIKRAMRAJ SITPAL and RANJIT NORONHA\, Oracle\n\nTalk T itle: Securing Modern Databases with Hardware-Assisted Memory\nIsolation: Leveraging MPK (x86) and MTE (ARM) in Oracle DB\n\nModern applications lik e Oracle Database are built atop a vast\,\never-evolving ecosystem of clos ed and open-source components\, and\nother complex software interactions. While this extensibility drives\ninnovation\, it also exposes critical ris ks: malicious or buggy\ncode—whether from third-party libraries or arbit rary user code\nexecution (think UDFs) – can attempt to steal or corrupt sensitive\ndata\, compromise shared memory\, or tamper with fundamental s tructures.\nTraditional defenses such as virtualization\, containers\, and \nsoftware-based memory protections (e.g.\, mprotect()) offer only\npartia l solutions\, often adding performance overhead and lacking the\ngranulari ty needed to protect against in-process threats.\n\nTo address these chall enges\, bleeding-edge hardware features are\nredefining the landscape of a pplication memory safety. On x86\nplatforms\, Memory Protection Keys (MPK) provide mprotect-like\nfunctionality directly in userspace\, allowing app lications to assign\nprotection \"keys\" to memory pages and efficiently u pdate permissions\nwith minimal overhead. Oracle leverages MPK\, for examp le\, in its\nMultilingual Engine (MLE). In AI vector search infrastructure \, it\nenables the secure execution of custom distance function calculatio ns\nwritten in JavaScript. Similarly\, ARM’s Memory Tagging Extension\n( MTE) enables fine-grained memory “coloring\,” tagging memory\nallocati ons and associating them with pointers to detect and prevent\nunsafe acces ses.\n\nIn this talk\, we present how Oracle Database leverages both MPK o n x86\nand MTE on ARM to enforce fine-grained\, dynamic memory isolation. We\ndiscuss the integration of these hardware technologies to protect\ncri tical resources such as shared memory\, effectively reducing the\nattack s urface posed by untrusted or compromised code. We’ll\nhighlight performa nce impacts\, security advantages\, and the\ntransformative role these fea tures play in fortifying Oracle Database\ndeployments everywhere. \n\n— \n\n► Vikramraj Sitpal is a Principal Member of Technical Staff (PMTS)\ nat Oracle\, focusing on database kernel development within the Database\n organization. He has had the opportunity to work on various modules\,\ninc luding memory\, synchronization\, resource management\, and I/O. With\nexp erience across operating systems\, storage systems\, networks\,\ndistribut ed systems\, and database technologies\, he enjoys\ncollaborating with oth ers to solve challenging technical problems and\ncontribute to Oracle’s ongoing innovation. He holds a Master’s\ndegree in Computer Systems from Carnegie Mellon University (MSIN\n’21).\n\n► Ranjit Noronha is a Sen ior Manager at Oracle. He manages ports of\nthe database to different plat forms such as AIX\, Linux on Z-series\,\nSolaris\, HP/UX and Linux on ARM. He holds a PhD in Computer Science\nand Engineering from The Ohio State U niversity.\n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916e536 DTSTART;TZID=America/New_York:20250821T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250821T130000 LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talk - Chris Carlon CLASS:PUBLIC DESCRIPTION:Speaker: CHRIS CARLON\, Software Engineer\nGoogle\n\nTalk Title : Rapid Storage: Low Latency Object Storage on Google Cloud\n\nObject stor age provides immense scalability\, but high latency can make\nit hard to u se for performance-sensitive workloads. This talk\nintroduces Rapid Storag e\, which brings filesystem-like performance to\nobject workloads on Googl e Cloud. We'll describe how Rapid Storage\nmoves from traditional stateles s requests to a stateful\, session-based\nprotocol.\n\nRapid Storage's pro tocol amortizes the cost of object storage\nmetadata\, enabling low latenc y data operations. We'll discuss benefits\nto workloads with dependent rea ds\, like selecting from columnar\nformats\, and how the protocol recovers quickly from transient issues\nwithout sacrificing latency.\n\nRapid Stor age also builds on Google systems like Spanner and Colossus\nto bring appe ndable objects to Google Cloud. We'll discuss how Rapid\nStorage provides consistency across multiple write sessions\, resulting\nin a reliable buil ding block for distributed systems.\n\n—\n\nChris Carlon is a software e ngineer at Google\, working on distributed\nstorage systems. Before Rapid Storage\, Chris was a software engineer\nat Jane Street\, and before that he helped design and launch Google's\nAlloyDB. Chris graduated from Duke U niversity with a BS in Computer\nScience/BA in Mathematics. \n\nZoom Part icipation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916e935 DTSTART;TZID=America/New_York:20250807T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250807T123000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Madhusudan Reddy Pittu CLASS:PUBLIC DESCRIPTION:Speaker: MADHUSUDHAN REDDY PITTU\, Ph.D. Candidate\, Ph.D. Prog ram in\nAlgorithms\, Combinatorics and Optimization\, Computer Science\nDe partment\, Carnegie Mellon University\n\nTalk Title: Fairness\, Diversity\ , Explainability\, and Robustness for\nAlgorithmic Decision-Making\n\nThis thesis investigates foundational algorithmic challenges that\narise when embedding fairness\, diversity\, explainability\, and\nrobustness into com putational decision-making. As machine learning\nsystems\, resource alloca tion mechanisms\, and data analysis pipelines\nincreasingly influence crit ical decisions\, it is essential that these\nsystems uphold not only effic iency and accuracy but also ethical and\nstructural guarantees. However\, enforcing these principles introduces\ncomplex trade-offs and computationa l difficulties.\n\nWe address five core problems that capture different fa cets of\nalgorithmic decision-making under structural and informational\nc onstraints: (1) determinant maximization under matroid constraints\,\nmode ling the selection of diverse and representative subsets\; (2)\napproximat ion of the weighted Nash Social Welfare objective\, a\nfairness-centric fo rmulation in indivisible resource allocation\; (3)\nconstrained subspace a pproximation\, which enforces group-level\nrepresentation in data summariz ation\; (4) explainable clustering\,\nwhich trades off interpretability an d clustering quality using\ndecision trees with axis-aligned threshold cut s\; and (5) combinatorial\noptimization using comparison oracles\, which e nables robust\ndecision-making in uncertain or preference-driven environme nts.\n\nEach of these problems introduces structural constraints that\ncha llenge conventional algorithmic techniques. We develop new\nframeworks tha t combine combinatorial methods\, convex and non-convex\nrelaxations\, con vex geometry\, and probabilistic methods. The resulting\nalgorithms offer improved approximation guarantees\, shed light on key\ntrade-offs between fairness\, interpretability\, and performance\, and\nsupport the developme nt of more equitable\, interpretable\, diverse\, and\nreliable algorithmic systems. These contributions have broad\nimplications in machine learning \, economics\, data summarization\, and\nhuman-in-the-loop decision-making .\n\nThesis Committee\n\nDavid P. Woodruff (Co-Chair)\n\nAnupam Gupta (Co- chair)\n\nPrasad Tetali\n\nMohit Singh (Georgia Institute of Technology)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916ee0f DTSTART;TZID=America/New_York:20250730T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250730T152000 URL:https://www.cmu.edu/data/community/data-day/ LOCATION:Rangos Auditorium\, Cohon University Center SUMMARY:Carnegie Mellon University’s 4th Annual Data Analytics Day CLASS:PUBLIC DESCRIPTION:Talk Title: 4th Annual Data Analytics Day\n\nSponsored by the O ffice of the Vice Provost\, the Office of Finance\,\nand the Office of the CIO\, at this year's Data Analytics Day we will\nexplore Advancing Data A nalytics through Community Engagement and\nEmpowerment. This year’s core theme is directly connected to key\ninitiatives such as data literacy\, d ata governance\, data franchise\ndevelopment\, and the innovative use of d ata analytics in university\nstrategic and operational planning. \n\nSpea kers include CMU senior leadership\, faculty\, staff\, and guests. \n\nA detailed schedule will be available soon. Y\n\nou will also have a chance to explore all the innovative\nvisualizations submitted by our IronViz Cha llenge finalists. This\nannual data visualization contest will be held on Tuesday\, July 29.\n\nEvent website and registration information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916f19c DTSTART;TZID=America/New_York:20250911T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250911T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting - September 2025 CLASS:PUBLIC DESCRIPTION:Speaker: CSD FacultyTalk Title: CSD Faculty Meeting\n\nSee emai l announcement for details.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916f437 DTSTART;TZID=America/New_York:20250801T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250801T163000 LOCATION:Gates Hillman 7101 and Zoom SUMMARY:5th Year Master's Thesis Presentation - Russell Emerine CLASS:PUBLIC DESCRIPTION:Speaker: RUSSELL EMERINE\, Master's Student\nComputer Science D epartment\nCarnegie Mellon University\n\nTalk Title: Simulating Voting Sys tems\n\nMuch of the past work on voting systems focuses on ranked voting\n systems\, which have a number of limitations such as Arrow's Theorem.\nIn this paper we consider ranked voting systems as well as the less\ncommonly used class of rated voting systems. The systems differ in\nthat ranked vo ting systems only allow the voter to order the\ncandidates\, while in rate d voting systems the voter can score each\ncandidate independently. In 200 0\, Warren Smith evaluated ranked and\nrated voting systems under a Monte Carlo simulation model of voter\nutilities and behaviors. We replicate Smi th's results with a wider\nselection of voting systems and voter utility d istributions and\nconclude that range voting\, a rated system where each c andidate can be\nindependently scored in the range [0\, 1]\, has the best performance.\n\nThesis Committee\n\nDaniel Sleator (Chair)\n\nFei Fang\n\n Additional Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916f7e5 DTSTART;TZID=America/New_York:20250801T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250801T143000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year Master's Thesis Presentation - Aksara Bayyapu CLASS:PUBLIC DESCRIPTION:Speaker: AKSARA BAYYAPU\, Master's Student\nComputer Science De partment\nCarnegie Mellon University\n\nTalk Title: Towards Agentic LLMs f or Hardware-Aware Kernel Generation\n\nRecent advances in large language m odels have furthered the\ndevelopment of agentic systems: pipelines that i nterleave planning\,\nexecution\, and iterative refinement. The following thesis demonstrates\nthe design\, implementation\, and evaluation of such agentic systems so\nthey can be leveraged in underutilized ways\, from pro ductivity tasks\nto refining kernel code. First\, we propose MLC Assistant \, which\nshowcases a general-purpose Chrome extension agent that integrat es\nLLMs with diverse web APIs. Beyond constructing a web agent\, we aimed \nto extend agentic systems to the more specialized and impactful\ndomains \, such as creating kernels through an LLM Kernel Agent. This\nagent embod ies an iterative decision-making agent that generates\,\nbenchmarks\, and refines kernels\, illustrating the core principles of\nagent architecture\ , feedback loops\, and sequential optimization. \n\nFindings will show th at there is a lack of thorough benchmarking and\nevaluation metrics for ke rnels\, making iterative\, generative\nimprovement difficult. The third co mponent of the following paper\nfocuses on creating a lightweight\, extens ible benchmarking suite\,\nFlashInfer Bench\, which resolves this issue. T he suite captures\nexecution traces to systematically evaluate and compare low-level\nkernel implementations for performance and correctness in mode l\ninference workloads. Thus\, FlashInfer Bench proposes a self-improving\ ,\ncommunity-driven platform for developing and deploying hardware-aware\, \nhigh-performance kernels. Together\, these contributions emphasize the\n central importance of agent design\, tool orchestration\, and integrated\n feedback loops in unlocking autonomous\, high-performance agents.\n\nThesi s Committee\n\nTianqi Chen (Chair) \n\nZhihao Jia\n\nAdditional Informati on \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916fc25 DTSTART;TZID=America/New_York:20250728T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250729T163000 LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:CyLab Robotics Security and Privacy Workshop - Registration Require d CLASS:PUBLIC DESCRIPTION:Talk Title: CyLab Robotics Security and Privacy Workshop\n\nAs robotics and autonomous systems technology becomes increasingly\nintegrate d into critical infrastructure sectors such as emergency\nservices\, defen se\, health and social care\, and manufacturing\, the need\nfor a secure a nd private ecosystem is more urgent than ever. However\,\nexisting approac hes to building robotic systems treat safety\,\nsecurity\, and privacy as an afterthought. For example\, the current\nlandscape of robotics middlewa re\, such as ROS\, has limitations in\nreal-time readiness\, usability\, a nd widespread implementation\, leading\nto potential security and privacy vulnerabilities. AI and ML are\nfoundational to the next generation of adv anced autonomous robots\, but\npose novel difficulties in assuring their s afety and robustness in\nphysical environments. \n\nThe CyLab Robotics Se curity and Privacy Workshop will convene leading\nresearchers and experts from across academia and industry to\ndiscuss strategic approaches to bui lding trusted middleware and\ntoolchains to foster a secure\, privacy-pres erving robotics ecosystem\nthat is safe and trustworthy by design. We wil l bridge this gap by\nuniting experts in robotics\, AI\, cybersecurity\, a nd privacy\,\ncollaborating with partners to meet the demands of industry\ , academia\,\nand government. \n\nSCHEDULE  |  ADDITIONAL INFORMATION\n \nNOTE →This is an invitation-only event.  Please fill our this FORM\ni f you are interested in attending.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef916ffe9 DTSTART;TZID=America/New_York:20250728T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250728T133000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year Master's Thesis Presentation - Nouha Tiyal CLASS:PUBLIC DESCRIPTION:Speaker: NOUHA TIYAL\, Master's StudentComputer Science\nDepart mentCarnegie Mellon University\n\nTalk Title: Multiplexed Expansion Micros copy for Drug Response\nPrediction in MIBC\n\nExpansion microscopy (ExPath ) enables nanoscale resolution of tissue\narchitecture using conventional microscopes\, offering a powerful\nalternative to traditional histopatholo gy. In this thesis\, we present\na deep learning pipeline that leverages E xPath imaging combined with a\nbiologically informed\, four-channel multip lexed staining panel: DAPI\,\nTelC\, CENPB\, and WGA to classify tissue ty pes and predict chemotherapy\nresponse in muscle-invasive bladder cancer ( MIBC). We propose that\nnuclear morphology\, when captured at high resolut ion and enriched by\nchromatin and membrane-specific markers\, contains su fficient\ninformation to compete with H&E and generalize across diagnostic and\nprognostic tasks. To test this hypothesis\, we construct a\npreproce ssing pipeline that transforms 16-bit 4-channel TIFF WSIs into\nnormalized \, pseudo-RGB 1024×1024 patches compatible with\nImageNet-pretrained mode ls. We evaluate multiple architectures\n(ResNet34\, ResNet50\, ViT-tiny\, EfficientNet) and demonstrate that\nResNet-based models trained on ExPath outperform simulated non-ExPath\nbaselines and DAPI-only variants by a sig nificant margin. Through\ncontrolled ablation experiments\, we quantify th e contribution of each\nchannel and find that multiplexing substantially b oosts classification\naccuracy. Our models achieve 89.52% tissue classific ation accuracy and\n0.9 ROC-AUC for drug response prediction. Furthermore\ , we observe\ncross-cancer generalizability when applying MIBC-trained mod els to\nlung carcinoma ExPath images. This work establishes the feasibilit y of\ncompact\, multiplexed\, ExPath-driven classification pipelines as a\ nviable alternative to costly multi-modal diagnostics. It offers an\nearly step toward a DAPI-first foundation model for computational\npathology\, with potential to scale across cancer types and tissue\nconditions using minimal staining and high-content imaging.\n\nThesis Committee\n\nRussell S. Schwartz (Chair)\n\nMin Xu\n\nAdditional Information \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9170464 DTSTART;TZID=America/New_York:20250730T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250730T153000 LOCATION:Gates Hillman 9115 SUMMARY:5th Year Master's Thesis Presentation - Zhijie Xu CLASS:PUBLIC DESCRIPTION:Speaker: ZHIJIE XU\, Master's Student\nComputer Science Departm ent\nCarnegie Mellon University\n\nTalk Title: Decomposing Complexity: An LLM-Based Approach to\nAutomating Software Engineering Tasks\n\nTask decom position in software engineering enables the division of\ncomplex engineer ing tasks into manageable components\, facilitating\nmodularization and co llaborative development. This thesis investigates\nthe practical applicati on of large language model (LLM) task\ndecomposition in software engineeri ng\, focusing on how complex GitHub\nissues can be systematically broken d own into smaller\, more solvable\nsubtasks.\n\nTo understand existing deco mposition practices\, this work creates and\nanalyzes a dataset focused on task decomposition for ten Apache\nprojects. The dataset analysis reveals operational patterns of task\ndecomposition in real-world open-source sof tware (OSS) projects\,\nproviding insights into how experienced developers naturally decompose\ncomplex tasks and identifying key characteristics of effective\ndecomposition strategies.\n\nBuilding on these insights\, a de composition component is integrated\ninto SWE-agent\, enabling structured transformation of GitHub issues\ninto subtasks. This component helps to ac hieve a 24% performance\nimprovement over the non-decomposed baseline on t he SWE-bench verified\ndataset. The results demonstrate the effectiveness of LLM task\ndecomposition in tackling software engineering problems\, pot entially\nimproving both automatic system performance and human developer\ nproductivity.\n\nThesis Committee\n\nCarolyn Rosé (Chair)\n\nMichael Hil ton\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9170858 DTSTART;TZID=America/New_York:20250729T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250729T143000 LOCATION:McWilliam Classroom\, Gates Hillman 4304 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Jun-Ting Hsieh CLASS:PUBLIC DESCRIPTION:Speaker: JUN-TING HSIEH\, Ph.D. Candidate\nComputer Science Dep artment\nCarnegie Mellon University\n\nTalk Title: Algorithms and Explicit Constructions via Spectral\nTechniques\n\nSpectral methods have become ub iquitous in computer science. By\nanalyzing the eigenvalues and eigenvecto rs of matrices naturally\nassociated with a graph\, such as its adjacency matrix\, one can extract\nuseful information about the graph's structure. Such methods have\nyielded the best-known results for a wide range of foun dational\nproblems.\n\nIn this talk\, we apply this \"spectral lens\" to p rove new results in\ngraph theory\, design algorithms\, and construct expl icit vertex\nexpanders.\n\nIn the first part of this talk\, we present alg orithms for both\nrefuting semi random constraint satisfaction problems an d recovering\nsolutions in planted ones\, both utilizing spectral informat ion of the\nunderlying hypergraph. Moreover\, we give algorithms to find l arge\nindependent sets in spectral expanders.\n\nIn the second part of thi s talk\, we introduce the tripartite line\nproduct to construct constant-d egree vertex expanders. First\, we\nobtain explicit unique-neighbor expand ers by instantiating the product\nusing Ramanujan graphs – the optimal s pectral expanders. Then\, by\nreplacing Ramanujan graphs with the incidenc e graphs of Ramanujan\ncubical complexes\, we obtain the first explicit lo ssless vertex\nexpanders. \n\nThesis Committee\n\nPravesh K. Kothari (Cha ir\, Carnegie Mellon University/Princeton\nUniversity) \n\nRyan O'Donnell \n\nJason Li\n\nVenkatesan Guruswami (University of California\, Berkeley) \n\nDavid Steurer (ETH Zürich)\n\nIn Person and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9170c87 DTSTART;TZID=America/New_York:20250728T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250728T143000 LOCATION:Gates Hillman 7501 SUMMARY:5th Year Master's Thesis Presentation - Archan Das CLASS:PUBLIC DESCRIPTION:Speaker: ARCHAN DAS\, Master's Student\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: Analyzing Student Debuggin g Using Programming Process Data\n\nBACKGROUND: Debugging is an important part of the software development\nworkflow. Understanding the cognitive pr ocess through which\nprogrammers debug is important for improving the tech niques and\ninstruction of debugging. Previous research has used a variety of\nmethodologies for studying the debugging process\, including concurre nt\nverbal protocols\, quantitative analyses\, and neural imaging. One\nfr ontier in this research is the use of process data to study\ndebugging. Th is process data consists of logs collected from\nintegrated development en vironments (IDEs) that describes the process\nby which programmers work on code.\n\nAIM: We aim to: a) create a framework for analyzing process data \ncaptured from an IDE\, b) analyze the process data to observe patterns\n across student debugging behavior\, and c) use the collected data to\niden tify efficient and inefficient habits exhibited by students while\ndebuggi ng.\n\nDATA: We collected process data across three exercises from 315\nst udents in an introductory programming class. This data consists of\nan eve nt log of every keystroke\, code execution\, and submission\nattempt stude nts made while working on their exercises.\n\nMETHODS: We extracted a time line of cognitive phases from the process\ndata for each student and valid ated our model with a panel of experts.\nWe tested the effect of two behav ioral features (use of print\nstatements and time in locate-error phase) a s a novel measure of\nstudent struggle in debugging (count of run-program events). We also\nobserved patterns across the subject population of our e xtracted\ncognitive phases\n\nRESULTS: We found that the frequency of prin t statements had a\npositive correlation with debugging struggle across al l exercises.\nIncreased time spent in locate-error phase had a statistical ly\nsignificant impact on student debugging struggle in some exercises\,\n but not others. Subjects tended to perform faster and more focused\nrepair s to their code later in debugging episodes. Finally\, debugging\nstruggle had a weak negative correlation with average exam scores in\nthe course.\ n\nCONCLUSION: Results suggest that students should be encouraged to\nspen d more time reasoning about their code while debugging. Process\ndata also shows promise as a tool for evaluating and giving feedback\non the studen t debugging process. In addition\, our framework can be\nwidely useful for experiments on student debugging behavior.\n\nThesis Committee\n\nDavid K osbie (Chair)\n\nMark Stehlik\n\nRoy Maxion\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9171157 DTSTART;TZID=America/New_York:20250729T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250729T123000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Arjun Teh CLASS:PUBLIC DESCRIPTION:Speaker: ARJUN TEH\, Ph.D. Candidate\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Computational Lens Design\n\ nContemporary lens design\, therefore\, presents a multifaceted\noptimizat ion challenge. A typical compound lens system is\ncharacterized by both co ntinuous parameters\, such as surface shape and\nthicknesses\, and discret e choices\, such as the number of elements and\ntypes of material. This mi xed discrete-continuous parameter space\ncreates a complex design landscap e where the performance of the lens\nis highly sensitive to all of the cho ices of parameters. This design\nspace has been historically hard for desi gners to navigate\,\nnecessitating assistance from theoretical and computa tional tools that\nhelp guide the search for performant designs. Yet\, eve n with these\ntools\, design remains time consuming and requires a great d eal of\ndesigner input. In parallel\, the graphics community has developed \nmethods for differentiable rendering\, which enable gradient-based\nopti mization of image based tasks. These methods have been\nsuccessfully appli ed to a variety of problems and are a great\ncandidate for application to lens design. However\, there are key\ndifferences in lens design from gene ral rendering that make directly\napplying these methods to lens design ch allenging.\n\nThe purpose of this thesis is to develop methods that levera ge ideas\nand techniques from differentiable rendering to address the chal lenges\nof lens design\, by developing a set of theoretical and computatio nal\ntools tailored to the unique requirements of lens design. Firstly\, w e\nbuild a method for calculating the unbiased gradient of light\nthroughp ut with respect to lens parameters\, enabling the optimization\nof lens sp eed. Secondly\, we devise Markov chain Monte Carlo (MCMC)\nmethod that com bines gradient-based optimization of continuous\nparameters with discrete mutations that change the number of elements\nin a lens system. Lastly\, w e derive a constant memory method for\ncalculating gradients of ray paths through gradient-index (GRIN)\nmaterials allowing for optimization of GRIN lenses\n\nThesis Committee\n\nIoannis Gkioulekas (Co-chair)\n\nMatthew O' Toole (Co-chair)\n\nJames McCann\n\nBernd Bickel (ETH Zürich)\n\nIn Perso n and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91715bd DTSTART;TZID=America/New_York:20250722T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250722T130000 LOCATION:Panther Hollow\, Room 4105\, Mehrabian Collaborative Innovation Ce nter\nand Zoom SUMMARY:CyLab Seminar - Christof Paar CLASS:PUBLIC DESCRIPTION:Speaker: CHRISTOF PAAR\, Co-founding Director\nMax Planck Insti tute for Security and Privacy\, Bochum\, Germany\nand\, Affiliated Profess or\, Electrical and Computer Engineering\nUniversity of Massachusetts Amhe rst\n\nTalk Title: Cybersecurity Beyond Software and Hardware\n\nClassical physical-layer security (PLS) has been studied for nearly\nhalf a century . In the past\, the main objective had been to realize\ninformation-theore tically secure communication. Yet\, those schemes\nhaven’t gained much p ractical relevance. We argue that it is more\npromising to apply PLS to pr oblems for which there is no solution\nusing digital-only schemes. In this talk we will look at a pretty wild\napplication for such a PLS solution.\ n\nA vexing problem in nuclear disarmament agreements is the secure\nremot e monitoring of nuclear warheads in storage. In collaboration\nwith Prince ton's School of Public and International Affairs\, we have\ndeveloped an e ntirely new approach. Our solution is a special instance\nof a virtual-pro of-of-reality scheme\, based on radio-frequency\nsignals. The basic idea i s to build a challenge-and-response protocol\n— which are widely used in cryptography\, e.g.\, for authentication in\nmobile communication — bas ed on wireless technology. \n\n—\n\nChristof Paar is co-founding direct or of the Max Planck Institute for\nSecurity and Privacy in Bochum\, Germa ny\, and is research professor at\nUMass Amherst. His research interests i nclude hardware security\,\napplied cryptography and physical-layer securi ty. At WPI\, he\nco-founded CHES\, the Conference on Cryptographic Hardwar e and Embedded\nSystems\, and is co-spokesperson of the Cluster of Excelle nce “CASA\n– Cyber Security in the Age of Large-Scale Adversaries” ( financed\nby DFG\, the “German NSF”). He is Fellow of the IEEE and the IACR\nand has given invited talks at Cambridge\, Harvard\, MIT\, Oxford\, \nStanford and Yale.\n\nFaculty Host: Swarun Kumar\n\nIn Person and Zoom P articipation.  See announcement.\n\n→The CyLab Seminar is open only to partners and Carnegie Mellon\nUniversity faculty\, students\, and staff.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9171a02 DTSTART;TZID=America/New_York:20250723T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250723T153000 LOCATION:Gates Hillman 9115 SUMMARY:5th Year Master's Thesis Presentation - Yueqi Song CLASS:PUBLIC DESCRIPTION:Speaker: YUEQI SONG\, Master's Student\nComputer Science Depart ment\nCarnegie Mellon University\n\nTalk Title: Towards Unified Interfaces for Generalist Agent In Diverse\nEnvironments\n\nRecently\, large languag e models (LLMs) have enabled agents that can\nperceive\, reason\, and act in increasingly complex environments. Yet\ntoday's agents remain constrain ed by the interfaces they rely on\,\nhampering generalization. This master thesis advances the goal of a\nunified agent framework.\n\nExamining web agents\, we found that web browsing agents\, though\nintuitive to humans a s they simulate human behaviours by browsing the\nweb\, are less effective and efficient. Thus\, we proposed an API-based\nweb agent that calls APIs through code generation\, and demonstrated\nsuperior performance compared to browsing agents. Building on this\, we\nfurther proposed a hybrid web agent that could interleave API calling\nand web browsing\, broadening the agent's interface and allowing it to\noperate more effectively and effici ently in diverse environments.\n\nBeyond web agents\, we aim to extend the unified interfaces to\ngeneralist agents across diverse environments. To this end\, we curated\na large-scale unified training dataset that spans c oding\, web tasks\,\nand general agentic tasks. The agent trained on this dataset achieved\nstate-of-the-art (SOTA) performance on benchmarks testin g a variety of\ntasks\, marking a step towards unified interface for gener alist agents.\n\nAlongside a unified framework\, strong reasoning abilitie s are crucial\nfor agents to make correct decisions\, plan\, and execute t asks based on\nusers' goals. We thus introduced VisualPuzzles\, a benchmar k that could\nevaluate models' multimodal reasoning abilities in a knowled ge-light\nenvironment\, which could provide guidance on the future develop ment of\nmodels with strong multimodal reasoning capabilities.\n\nLast but not the least\, to serve people around the world\, agents need\nto unders tand and generate multilingual content. Thus\, we proposed and\ntrained Pa ngea\, a multilingual model that achieved SOTA results on\nmultilingual be nchmarks. Together\, these contributions pave a path\ntowards unified inte rfaces for generalist agents in diverse\nenvironments\, providing the conc eptual\, empirical\, and engineering\nfoundations for the next generation of generalist AI agents.\n\nThesis Committee\n\nGraham Neubig (Chair)\n\nD aniel Fried\n\nAdditional Information \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9171e92 DTSTART;TZID=America/New_York:20250724T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250724T130000 LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talk Series - Daniel Berger CLASS:PUBLIC DESCRIPTION:Speaker: DANIEL S. BERGER\, Principal ResearcherMicrosoft Azure \n\nTalk Title: Carbon Negative Datacenters Require A Marathon\, Not A\nSp rint\n\nLarge cloud providers like Google and Microsoft promise significan t\ncarbon emission reductions over the next five years. Drawing on my\nexp erience prototyping and deploying sustainable cloud building\nblocks\, thi s talk will offer a practitioner's view on our progress and\nthe challenge s ahead. While we have key wins and learnings\, achieving\nsustainable clo ud computing requires a holistic strategy since no\nsingle aspect dominate s a cloud’s carbon emissions. We must tackle\noperational and embodied c arbon across AI\, general-purpose compute\,\nstorage\, and networking\, sp anning software and hardware stacks. This\ntalk identifies key open strate gic decisions and invites the HotCarbon\ncommunity to help us find the ans wers.\n\n—\n\nDaniel Berger is a Principal Researcher at Microsoft Azure . His work\nfocuses on improving the efficiency\, sustainability\, and rel iability\nof cloud platforms by applying measurements\, analytical models\ , and\nhardware-software prototyping to cloud server design. Daniel's rece nt\nwork has received an IEEE Top Picks in 2025 and an honorable mention\n in 2024\, an ACM ASPLOS 2023 distinguished paper award\, USENIX OSDI\n2023 best paper award\, 2021 ACM SOSP Best Paper Award.\n\nZoom Participation.   See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9172261 DTSTART;TZID=America/New_York:20250717T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250717T130000 LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talk Series - Ioan Stefanovici CLASS:PUBLIC DESCRIPTION:Speaker: IOAN STEFANOVICI\, Principal ResearcherFuture AI\nInfr astructure TeamMicrosoft Research Cambridge\n\nTalk Title: A Tale of Two E xtremes: Storing Humanity's Knowledge in\nGlass\, and AI's Data in New Mem ory\n\nThis talk will address two important and timely problems at opposit e\nends of the data temperature spectrum in today's hyperscaler\nenvironme nts: archiving humanity's knowledge for eternity in glass and\nstoring AI' s data in efficient memory.\n\nThe first part will focus on Project Silica \, which is the first cloud\nstorage system for archival data underpinned by quartz glass\, an\nextremely resilient media that allows data to be lef t in place\nindefinitely\, thereby eliminating the inefficient\, wasteful\ , and\ncostly migrations of data required with today's magnetic storage\nt echnologies. The hardware and software of Silica have been\nco-designed an d co-optimized from the media up to the service level\nwith sustainability as a primary objective. This design follows a\ncloud-first\, data-driven methodology underpinned by principles derived\nfrom analysing the archival workload of a large public cloud service.\nSilica can support a wide rang e of archival storage workloads and\nushers in a new era of sustainable\, cost-effective storage.\n\nThe second part will introduce Managed-Retentio n Memory: a new class\nof memory for the AI era. AI clusters today are the largest uses of\nHigh Bandwidth Memory (HBM). However\, HBM is suboptimal for AI\nworkloads for several reasons\; it is overprovisioned on write\np erformance\, but under provisioned on density and read bandwidth\, and\nal so has significant energy per bit overheads. It is also expensive\,\nwith lower yield than DRAM due to manufacturing complexity. We propose\na new c lass of memory: Managed-Retention Memory (MRM)\, which is\noptimized to st ore key data structures for AI inference workloads. We\nbelieve that MRM m ay finally provide a path to viability for\ntechnologies that were origina lly proposed and optimised to support\nStorage Class Memory (SCM). These t echnologies traditionally offered\nlong-term persistence (10+ years) but p rovided poor IO performance\nand/or endurance. MRM makes different trade-o ffs\, and by understanding\nthe workload IO patterns\, MRM foregoes long-t erm data retention and\nwrite performance for better potential performance on the metrics that\nare important for these workloads.\n\n—\n\nIoan St efanovici is a Principal Researcher in the Future AI\nInfrastructure team at Microsoft Research Cambridge. His research\nfocuses on inventing the in frastructure for AI systems and he is\nparticularly interested in novel\, disruptive memory and storage\ntechnologies. Previously\, he spent 8+ year s working on Project Silica\,\nwhich developed the first-ever archival sto rage technology designed\nand built from the ground up for the cloud\, by using femtosecond\nlasers to store data in glass\, and polarization micros copy + ML to\nread it back. He received his PhD in Computer Science from t he\nUniversity of Toronto in 2016\, and was also a Research Fellow at\nCor pus Christi College\, at the University of Cambridge (2016-2018).\nIoan li ves in London (UK) with his wife\, their grumpy 13-year old cat\,\nand 2 e ver-curious and playful Maine Coons.\n\nZoom Participation.  See announce ment. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91727a4 DTSTART;TZID=America/New_York:20250717T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250717T110000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year Master of Science in Computer Science Thesis Presentation CLASS:PUBLIC DESCRIPTION:Speaker: FAISAL ABDELMONEM\, Master's Student\nComputer Science Department\nCarnegie Mellon University\n\nTalk Title: Safe Kernel Extensi bility and Instrumentation With\nWebAssemblyThursday\, July 17\, 2025\, 10  – 11am Extending kernel\nfunctionality dynamically is essential for m odern workloads in\nobservability\, profiling\, and security\, and is beco ming increasingly\npopular for implementing low-latency\, kernel-bypass lo gic in\nhigh-performant systems. However\, existing mechanisms like kernel \nmodules or eBPF come with steep learning curves\, limited\nexpressivenes s\, or tightly constrained environments. WebAssembly\n(Wasm)\, with its st rong isolation guarantees\, portable semantics\,\nformally defined specifi cation with machine-checked proofs\, and low\nmemory footprint\, presents a compelling alternative for safe\,\nruntime-extensible logic inside the k ernel.This work explores Wasm as\na foundation for safe and flexible kerne l extensibility. We present an\nearly prototype that allows users to load and unload Wasm binaries\ninto the kernel and hook them into system calls for interception and\ninstrumentation. This prototype serves as an initial step toward\nrethinking kernel extensibility using Wasm as a secure and\n language-agnostic execution layer\, enabling safer and more accessible\nin -kernel customization.Thesis CommitteeAnthony Rowe (Chair)Benjamin\nTitzer Event Type: Master's Thesis Presentation Room Number: In\nPerson Buildi ng: Traffic21 Classroom\, Gates Hillman 6501 Speaker's\nName: FAISAL ABD ELMONEM Speaker's Professional Title: Master's\nStudent\, Computer Scienc e Department\, Carnegie Mellon University Talk\nTitle: Safe Kernel Extens ibility and Instrumentation With WebAssembly\nEvent Poster Title: Poster Event Poster URL: www.cs.cmu.edu\n[http://www.cs.cmu.edu]… For More Inf ormation: tracyf@cs.cmu.edu\nAffiliations: Computer Science Department ( CSD)\nOrganization(s): School of Computer Science\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9172c5c DTSTART;TZID=America/New_York:20250718T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250718T153000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Jeff Xu CLASS:PUBLIC DESCRIPTION:Speaker: JEFF XU\, Ph.D. Candidate\nComputer Science Department \nCarnegie Mellon University\n\nTalk Title: Spectral Techniques for Averag e-Case Complexity and Beyond\n\nIn recent years\, algorithmic challenges a cross diverse areas including\nstatistical physics\, machine learning and cryptography have centered\naround statistical inference problems\, i.e.\, computational problems\nwith average-case inputs. For many of these probl ems\, the best-known\nefficient algorithms are often suboptimal\, giving r ise to information\nvs. computation gaps\, discrepancies between what is t heoretically\npossible given the amount of information and what can be att ained via\nefficient algorithms. One fundamental question is how we can pr ovide\nrigorous evidence of hardness to show that such gaps are\ninsurmoun table for efficient computation. \n\nIn this talk\, I will demonstrate th at many of these questions boil\ndown to the study of random matrices that have entries being\npolynomials of the underlying input. More concretely\ , I will highlight\nthe recent advances in the past few years that lead to a significantly\nmore refined understanding of these ostensibly complicat ed matrices\,\nand some intriguing questions around them that still remain after\nyears of attacks.  The sharper understanding of these matrices\nu ltimately allows us to provide rigorous evidence via the lens of the\nSum- of-Squares (SoS) algorithms\, a hierarchy of semidefinite\nprogrammings. U nlike several other popular models in the average-case\nsetting (eg. low-d egree polynomials/statistical-query/ overlap-gap).\nSum-of-Squares algorit hms are known to capture various optimal\nalgorithms in both the average a nd worst-case setting\, and therefore\nprovide one of the strongest form o f hardness in average-case\ncomplexity.\n\nThesis Committee\n\nPravesh K. Kothari (Chair )\n\nAayush Jain\n\nRyan O’Donnell\n\nMadhur Tulsiani (To yota Technical Institute at Chicago /  University\nof Chicago)\n\nIn Pers on and Zoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91730b3 DTSTART;TZID=America/New_York:20250718T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250718T113000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Long Pham CLASS:PUBLIC DESCRIPTION:Speaker: LONG PHAM\, Ph.D. Candidate\nComputer Science Departme nt\nCarnegie Mellon University\n\nTalk Title: Hybrid Resource-Bound Analys es of Programs\n\nResource-bound analysis aims to infer symbolic bounds of worst-case\nresource usage (e.g.\, running time and memory) of programs.\ nApplications of resource analysis include job scheduling and\nprevention of side-channel attacks. Different resource-analysis\ntechniques have comp lementary strengths and weaknesses. (Automatic)\nstatic resource analysis\ , which analyzes the source code of programs\,\nis sound: if it successful ly infers a cost bound\, it is guaranteed to\nbe a valid bound. However\, due to the undecidability of resource\nanalysis in general\, every static analysis technique is incomplete:\nthere exists a program that the analysi s technique cannot handle.\nMeanwhile\, data-driven analysis\, which stati stically analyzes cost\nmeasurements obtained by running programs on many inputs\, can infer a\ncandidate cost bound for any program. However\, it d oes not guarantee\nsoundness of inference results.\n\nTo overcome limitati ons of individual analysis techniques\, this thesis\ndevelops hybrid resou rce analysis\, which integrates two complementary\nanalysis techniques via a user-adjustable interface. The user first\nspecifies which analysis tec hniques should analyze which code\nfragments and quantities. Hybrid analys is then performs its\nconstituent analysis techniques on their respective code fragments and\nquantities. Finally\, their inference results are comb ined into an\noverall cost bound. Hybrid resource analysis retains the str engths of\nconstituent analyses while mitigating their respective weakness es.\n\nThe thesis introduces two hybrid-resource-analysis techniques: Hybr id\nAARA and resource decomposition. They adopt distinct designs of an\nin terface between constituent analyses\, posing a trade-off in the\nflexibil ity of hybrid analysis. Hybrid AARA integrates static resource\nanalysis-- -Automatic Amortized Resource Analysis (AARA)---with\ndata-driven resource analysis via a type-based interface. On the other\nhand\, resource decomp osition integrates different pairs of static\,\ndata-driven\, and interact ive resource analyses via a\nnumeric-variable-based interface.\n\nIn addit ion to hybrid resource analysis\, I discuss theoretical results\nof resour ce analysis: (i) the undecidability of resource analysis\; and\n(ii) the p olynomial-time completeness of Conventional AARA. I also\ndescribe newly d eveloped Bayesian data-driven resource analysis\, which\nstatistically inf ers cost bounds by Bayesian inference. Finally\, I\npresent the optimizati on of probabilistic program-input generators by\na genetic algorithm\, sho wing that its output generator is more\neffective in triggering high compu tational cost than randomly\ngenerated inputs.\n\nThesis Committee\n\nJan Hoffmann (Chair)\n\nFeras Saad\n\nMatt Fredrikson\n\nFrancois Pottier (Inr ia\, Paris)\n\nIn Person and Zoom Participation.  See announcement. \n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9173609 DTSTART;TZID=America/New_York:20250710T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250710T130000 LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Talk - Suhas Jayaram Subramanya CLASS:PUBLIC DESCRIPTION:Speaker: SUHAS JAYARAM SUBRAMANYA\, Senior ResearcherCosmosDB\n GroupMicrosoft\n\nTalk Title: Efficient and Responsive Job-Resource Co-ada ptivity for\nDeep Learning Workloads in Large Heterogeneous GPU Clusters\n \nThursday\, July 10\, 2025\, 12 – 1pm \n\nExisting cluster scheduler s face many limitations in scheduling\nadaptive deep learning training job s on large heterogeneous GPU\nclusters – many are not heterogeneity-awar e\, few are\nadaptivity-aware\, and none scale to large clusters without s acrificing\nallocation fidelity or cluster efficiency. Emerging clusters f urther\ncomplicate this problem with larger\, more heterogeneous resources \nrunning more increasingly diverse jobs with more dimensions of\nadaptivi ty.  \n\nThis thesis develops new scheduling approaches and algorithms t hat can\n(1) scale to emerging clusters with hundreds of thousands of GPUs and\nmany GPU types\, (2) quickly optimize high-fidelity allocations for\ nadaptive DL training jobs with low scheduler overhead\, and (3)\nefficien tly adapt to changing cluster conditions to improve goodput on\nthe limite d GPU resources.\n\nWe first introduce Sia — a round-based scheduler tha t efficiently\noptimizes adaptive jobs in a heterogeneous cluster with man y GPU\ntypes. Sia uses GPU resources judiciously to gather information on\ njob-GPU fit-levels using a mix of online and offline profiling\, and\ncon tinuously co-optimizes the GPU resources allocated to jobs and\ntheir exec ution parameters at runtime to maximize cluster-wide\ntraining progress. U sing job traces derived from real-world data\ncenters\, we find that Sia ’s allocations are fair and efficient\, and\nare quickly computed using an efficient formulation\, even for 1000-GPU\nclusters.\n\nSecond\, we int roduce continual optimization — a new paradigm that\nexplicitly models t he slow evolution of resource-allocation problems\nat scale to reduce solv er runtime for quick responses to changes in\njobs or resources. We then i ntroduce COpter\, our approach to continual\noptimization that (a) efficie ntly updates the optimization problems\nfor job and resource changes using a differential interface\, (b)\nimplements a factorization-free warm-star ted LP solver to benefit from\nslowly-evolving nature of the allocations\, and (c) implements\nlightweight heuristics to recover feasible integral s olutions with\nnegligible quality loss. In our evaluations\, COpter speeds up Sia\nscheduler policy by a few orders of magnitude on clusters with te ns of\nthousands of GPUs without sacrificing job completion times and\nmak espan.\n\nThird\, COpter is easily applied to resource-allocation problems in\nother domains (e.g. shard load-balancing\, WAN traffic engineering) a nd\nwe see 57 − 83 × reductions in solver runtimes.\n\n—\n\nSuhas Jay aram Subramanya is a Senior Researcher in the CosmosDB group\nat Microsoft where he works on infrastructure for semantic retrieval.\nAt CMU\, he was advised by Greg Ganger and worked on schedulers for\ndeep learning cluste rs. \n\nZoom Participation.  See announcement. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9173b26 DTSTART;TZID=America/New_York:20250704T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250704T230000 SUMMARY:Independence Day - July 4th Observation 2025 CLASS:PUBLIC DESCRIPTION:Speaker: The University will be ClosedNormal Class Schedules an d\nBusiness Office hours will resume Monday\, July 7.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9173d7e DTSTART;TZID=America/New_York:20250703T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250703T160000 LOCATION:Reddy Conference Room\, Gates HIllman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Brian Hu Zhang CLASS:PUBLIC DESCRIPTION:Speaker: BRIAN HU ZHANG\, Ph.D. Candidate\, Computer Science De partment\,\nCarnegie Mellon University\n\nTalk Title: New Solution Concept s and Algorithms for Equilibrium\nComputation and Learning in Extensive-Fo rm Games and Beyond\n\nComputational game theory has led to significant br eakthroughs in AI\ndating back to the start of AI as a discipline. For exa mple\, it has\nbeen instrumental in enabling superhuman AI from recreation al games\nsuch as two-player zero-sum games chess\, go\, and heads-up poke r to\nmultiplayer games such as six-player poker and Hanabi\, and even in\ ngames involving human language such as Diplomacy. It has also\nempowered a growing range of non-recreational applications\, such as\ntrading\, mach ine learning robustness and safety\, negotiation\, conflict\nresolution\, mechanism (e.g.\, auction) design\, information design\,\nsecurity\, polit ical campaigning\, and self-driving cars. \n\nThis thesis pushes the boun dary on computational game theory\,\nespecially in imperfect-information s equential (extensive-form) games\,\nwhich are most prevalent in practical applications both in zero-sum\ngames and beyond. We will present new theor etical concepts and\nframeworks\, state-of-the-art and often provably opti mal algorithms for\ncomputing and learning equilibria\, and new ways to ap ply such\nalgorithms to real-world problems\, including problems in econom ics\nsuch as mechanism and information design. We will also draw\nconnecti ons to the broader literature on optimization\, yielding new\nand more eff icient algorithms for solving variational inequalities.\n\nThesis Committe e\n\nTuomas Sandholm (Chair)\n\nVincent Conitzer\n\nJ. Zico Kolter\n\nKevi n Leyton-Brown (University of British Columbia)\n\nIn Person and Zoom Part icipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9174174 DTSTART;TZID=America/New_York:20250703T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250703T230000 SUMMARY:Community Appreciation Day - Summer 2025 CLASS:PUBLIC DESCRIPTION:Speaker: University will be ClosedNormal Class Schedules and Bu siness\nOffice Hours will resume Monday\, July 7.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91743e3 DTSTART;TZID=America/New_York:20250702T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250702T170000 LOCATION:Reddy Conference Room\, Gates HIllman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Sara McAllister CLASS:PUBLIC DESCRIPTION:Speaker: SARA McALLISTER\, Ph.D. Candidate\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Toward Sustainable Datacenters through Efficient Data\nRetrieval\n\nDatacenters are projected to account for 33% of the global carbon\nemissions by 2050. As datacenter s increasingly rely on renewable\nenergy for power\, the majority of datac enter emissions will be\nembodied — emissions from lifecycle stages incl uding acquiring raw\nmaterials\, manufacturing\, transportation\, and disp osal. To reach the\nambitious emission reduction goals set by both compani es and\ngovernments\, datacenters need to reduce emissions throughout thei r\noperations\, including (and particularly relevant for this thesis) the\ nstorage system. Unfortunately\, while data storage and retrieval\nsystems are large contributors to embodied emissions\, reducing their\nembodied e missions have largely been overlooked.\n\nThis dissertation addresses how to reduce emissions in data retrieval\nfor large-scale storage systems. Th ese storage systems can reduce\ntheir carbon footprint by enabling storage devices to have longer\nlifetimes and use denser media. However\, storage hardware's IO limits\ncombined with software's unnecessary additional IO often severely\nrestrict emission reductions\, or at worse cause increased emissions.\nThus\, this thesis focuses on reducing IO in several parts of the\nstorage stack to enable efficient and sustainable data retrieval.\n\ nFirst\, this dissertation addresses the sustainability of flash\ncaching\ , a critical layer in datacenter storage systems that is\nlimited by flash write endurance. This improvement results from two\ncaching systems: Kang aroo and FairyWREN. Together\, these caches\ndramatically reduce writes by over 28x\, allowing flash devices to use\ndenser flash for longer lifetim es\, ultimately reducing emissions.\nThen\, this thesis discusses enable m ore sustainable bulk storage\,\nwhere bandwidth limitations prevent deploy ment of denser HDDs.\nDeclarative IO\, a new interface for distributed sto rage\, empowers the\nstorage system to eliminate duplicate IO accesses in maintenance tasks\nthrough exposing the time- and order-flexibility in mai ntenance tasks.\nThis work enables deployment of larger HDDs\, further red ucing\nemissions from storage systems.\n\nThesis Committee\n\nGregory R. G anger (Co-Chair)\n\nNathan Beckmann (Co-Chair)\n\nGeorge Amvrosiadis\n\nDa niel Berger (Microsoft Azure/University of Washington)\n\nMargo Seltzer (U niversity of British Columbia)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917488a DTSTART;TZID=America/New_York:20251104T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251104T170000 SUMMARY:Democracy Day CLASS:PUBLIC DESCRIPTION:Speaker: No Classes before 5:00 pm\n\nNo Classes are schedule d uring the day.\n\nEvening classes after 5:00 pm will continue to meet. \n \n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9174aff DTSTART;TZID=America/New_York:20250702T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250702T160000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Siddharth Prasad CLASS:PUBLIC DESCRIPTION:Speaker: SIDDHARTH PRASAD\, Ph.D. Candidate\, Computer Science\ nDepartment\, Carnegie Mellon University\n\nTalk Title: Mechanism Design a nd Integer Programming in the Data Age\n\nThis thesis focuses on improving computational and economic aspects of\nmechanism design\, and on improvin g critical components of integer\nprogramming algorithms. Various marketpl aces in the world today\, from\nspectrum allocation to strategic sourcing to display advertisements to\nfinancial exchanges and more\, benefit from carefully engineered rules\nto govern the efficient exchange of items. Mec hanism design offers a\nprincipled way to design the rules to such market- based systems in\norder to implement desired market outcomes subject to st rategic\nself-interested participants. It is the prominent approach to man y\nmarket design problems and has been deployed in the real world with\nhi gh impact. On the computational front\, integer programming is the\ngo-to method for solving discrete optimization problems that arise in\nmarket de sign applications and beyond.\n\nWithin mechanism design\, our focus is on the design of better\nmechanisms that take advantage of any and all infor mation available to\nthe mechanism designer. Our new mechanisms provably g eneralize and\nimprove the state of the art\, and significantly expand the scope of\nwhat forms of information can be expressed and used to boost\np erformance. We apply our advances in mechanism design to\ncombinatorial ma rkets where bidders have complex\, combinatorial\npreferences over a rich space of outcomes. Here\, our new combinatorial\nauctions directly improve over existing designs that have been used to\nconduct high-stakes auction s around the world.\n\nWithin integer programming\, our focus is on the th eory and practice of\ncutting planes\, which are one of the most critical components of\ninteger programming solvers. We invent new cutting planes t hat deliver\nstrong theoretical and practical performance\, and develop a\ ncomprehensive generalization theory for data-driven parameter\nconfigurat ion within the branch-and-cut algorithm. \n\nIn both areas\, we fundament ally advance the classical state of\nknowledge and introduce new data-driv en perspectives\, all in support\nof the thesis that high performance—e. g.\, revenue\, social welfare\,\nrun-time\, memory\, etc.—in marketplace s can only be fully realized by\na synergy of approaches in mechanism desi gn\, integer programming\, and\nmachine learning.\n\nThesis Committee\n\nM aria-Florina Balcan (Co-Chair)\n\nTuomas Sandholm (Co-Chair)\n\nGérard Co rnuéjols\n\nCraig Boutilier (Google)\n\nPeter Cramton (University of Mary land)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9174ff1 DTSTART;TZID=America/New_York:20250724T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250724T160000 LOCATION:Rangos\, Cohon University Center SUMMARY:3rd Annual AI Day 2025: AI With Purpose CLASS:PUBLIC DESCRIPTION:Explore AI with a Purpose.\n\nAre you curious about how AI can transform your everyday tasks? Join\nus as we highlight the latest innovat ions in generative AI and\nshowcase practical strategies to enhance produc tivity across various\nfields. Let's delve into AI with a purpose at this year's AI Day!\n\nThis event offers tailored sessions for all skill levels .\n\nKeynote Speaker\n\nElizabeth Reilley — Executive Director\, AI Acce leration\, with\nEnterprise Technology at Arizona State University\n\nDr. Elizabeth Reilley is Executive Director of AI Acceleration within\nEnterpr ise Technology at Arizona State University. In her previous\nrole at ASU\, she focused on Data and Analysis within Enterprise\nTechnology. With over 15 years of experience in higher education\, Dr.\nReilley has held roles across Academic Affairs\, Information\nTechnology\, and Strategy and Polic y. She holds a PhD in Educational\nResearch\, Measurement\, and Evaluation \, as well as an MBA from the\nUniversity of North Carolina at Greensboro. She earned her BA in\nPhilosophy and Community and Justice Studies from G uilford College.\n\nREGISTER\n\nSchedule at a Glance\n\n8:30 am   — Ch eck-in and Breakfast9:00 am   — Welcoming\nRemarks9:15 am   — AI L andscape at CMU11:00 am —\nKeynote12:00 pm — Lunch1:00 pm   — W orkshops4:00 pm   —\n Conclusion\n\nHosted by the Carnegie Mellon Uni versity Office of the CIO \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91753f7 DTSTART;TZID=America/New_York:20250805T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250805T120000 LOCATION:Newell-Simon Hall SUMMARY:SCS Workshop Series on GenAI in Computing Education CLASS:PUBLIC DESCRIPTION:Speaker: SCS FacultyTalk Title: GenAI in Computing Education -\ nWorkshop Three\n\nTuesday\, August 5\, 2025\, 10am – 12pm \n\nSCS is holding a four week workshop series for SCS Faculty on using\nGenAI to co mplement and improve computing education. Faculty are\nstrongly encouraged to attend to learn about the current state of the\nart in GenAI and to ex amine how some SCS faculty are already using\nGenAI in their courses (for their own teaching and research\, to adjust\nhow students do work in the c ourse\, to update curricula for the\ncourse\, etc.). The general theme for each session include:\n\nWorkshop 1: Tuesday\, July 22  (10:00 AM-12:00 Noon)  — What is\nthe current state of the art of the most-used GenAI tools? What can\nand can't they do?Workshop 2: Tuesday\, July 29  (10:0 0 AM-12:00\nNoon)  —  How faculty can use GenAI in our own jobs (incl uding\nfor education and research efforts).► Workshop 3: Tuesday\, Augu st\n5  (10:00 AM-12:00 Noon)  — How to teach students how to use\nGe nAI appropriately including concerns for bias\, misinformation\,\netc.Work shop 4: Tuesday\, August 12  (10:00 AM-12:00 Noon)  — How\nto evalua te/assess students in the age of GenAI\, in instances where\nthey should a nd should not use these tools.\n\nIn-person attendance is recommended sinc e there will be activities in\nthe workshops to apply what you are learnin g to your upcoming\ncourse(s).\n\nPlease RSVP by Friday\, July 11\n\n→ RSVP and indicate your interest for each workshop\, as well as to\nvolunte er to talk or present an example from your work or class that\nuses GenAI in one of the workshop sessions. \n\n→ The workshops will be recorded\, but in person attendance is\nstrongly recommended.\n\nFaculty Hosts:  Zi co Kolter\, Tom Cortina \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917584e DTSTART;TZID=America/New_York:20250812T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250812T120000 LOCATION:Newell-Simon Hall SUMMARY:SCS Workshop Series on GenAI in Computing Education CLASS:PUBLIC DESCRIPTION:Speaker: SCS FacultyTalk Title: GenAI in Computing Education —\nWorkshop Four\n\nTuesday\, August 12\, 2025\, 10am – 12pm \n\nSC S is holding a four week workshop series for SCS Faculty on using\nGenAI t o complement and improve computing education. Faculty are\nstrongly encour aged to attend to learn about the current state of the\nart in GenAI and t o examine how some SCS faculty are already using\nGenAI in their courses ( for their own teaching and research\, to adjust\nhow students do work in t he course\, to update curricula for the\ncourse\, etc.). The general theme for each session include:\n\nWorkshop 1: Tuesday\, July 22  (10:00 AM-12 :00 Noon)  — What is\nthe current state of the art of the most-used G enAI tools? What can\nand can't they do?Workshop 2: Tuesday\, July 29  (1 0:00 AM-12:00\nNoon)  —  How faculty can use GenAI in our own jobs (i ncluding\nfor education and research efforts).Workshop 3: Tuesday\, August 5 \n(10:00 AM-12:00 Noon)  — How to teach students how to use GenAI \nappropriately including concerns for bias\, misinformation\,\netc.►Wor kshop 4: Tuesday\, August 12  (10:00 AM-12:00\nNoon)  — How to evalu ate/assess students in the age of GenAI\, in\ninstances where they should and should not use these tools.\n\nIn-person attendance is recommended sin ce there will be activities in\nthe workshops to apply what you are learni ng to your upcoming\ncourse(s).\n\nPlease RSVP by Friday\, July 11\n\n→ RSVP and indicate your interest for each workshop\, as well as to\nvolunt eer to talk or present an example from your work or class that\nuses GenAI in one of the workshop sessions. \n\n→ The workshops will be recorded\ , but in person attendance is\nstrongly recommended.\n\nFaculty Hosts:  Z ico Kolter\, Tom Cortina \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9175c82 DTSTART;TZID=America/New_York:20250722T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250722T120000 LOCATION:Newell-Simon Hall 3305 and Zoom SUMMARY:SCS Workshop Series on GenAI in Computing Education CLASS:PUBLIC DESCRIPTION:Speaker: Workshop One\n\nTuesday\, July 22\, 2025\, 10am –  12pm \n\nSCS will hold a four week workshop series for SCS Faculty on usi ng\nGenAI to complement and improve computing education. Faculty are\nstro ngly encouraged to attend to learn about the current state of the\nart in GenAI and to examine how some SCS faculty are already using\nGenAI in thei r courses (for their own teaching and research\, to adjust\nhow students d o work in the course\, to update curricula for the\ncourse\, etc.). The ge neral theme for each session include:\n\n► Workshop 1: Tuesday\, July 2 2 (10:00 AM-12:00 Noon)  — What\nis the current state of the art of t he most-used GenAI tools? What can\nand can't they do?Workshop 2: Tuesday\ , July 29 (10:00 AM-12:00\nNoon)  —  How faculty can use GenAI in our own jobs (including\nfor education and research efforts).Workshop 3: Tues day\, August 5\n(10:00 AM-12:00 Noon)  — How to teach students how to use GenAI\nappropriately including concerns for bias\, misinformation\,\n etc.Workshop 4: Tuesday\, August 12 (10:00 AM-12:00 Noon)  — How\nto evaluate/assess students in the age of GenAI\, in instances where\nthey sh ould and should not use these tools.\n\nIn-person attendance is recommende d since there will be activities in\nthe workshops to apply what you are l earning to your upcoming\ncourse(s).\n\nPlease RSVP by Friday\, July 11\n \n→ RSVP and indicate your interest for each workshop\, as well as to\nv olunteer to talk or present an example from your work or class that\nuses GenAI in one of the workshop sessions. \n\n→ The workshops will be reco rded\, but in person attendance is\nstrongly recommended.\n\nFaculty Hosts :  Zico Kolter\, Tom Cortina \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91760bf DTSTART;TZID=America/New_York:20250729T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250729T120000 LOCATION:Newell-Simon Hall SUMMARY:SCS Workshop Series on GenAI in Computing Education CLASS:PUBLIC DESCRIPTION:Speaker: SCS FacultyTalk Title: GenAI in Computing Education\n — Workshop Two\n\nTuesday\, July 29\, 2025\, 10am – 12pm \n\nSCS i s holding a four week workshop series for SCS Faculty on using\nGenAI to c omplement and improve computing education. Faculty are\nstrongly encourage d to attend to learn about the current state of the\nart in GenAI and to e xamine how some SCS faculty are already using\nGenAI in their courses (for their own teaching and research\, to adjust\nhow students do work in the course\, to update curricula for the\ncourse\, etc.). The general theme fo r each session include:\n\nWorkshop 1: Tuesday\, July 22  (10:00 AM-12:00 Noon)  — What is\nthe current state of the art of the most-used GenA I tools? What can\nand can't they do?► Workshop 2: Tuesday\, July 29  (10:00 AM-12:00\nNoon)  —  How faculty can use GenAI in our own jobs (including\nfor education and research efforts).Workshop 3: Tuesday\, Augu st 5 \n(10:00 AM-12:00 Noon)  — How to teach students how to use Gen AI\nappropriately including concerns for bias\, misinformation\,\netc.Work shop 4: Tuesday\, August 12  (10:00 AM-12:00 Noon)  — How\nto evalua te/assess students in the age of GenAI\, in instances where\nthey should a nd should not use these tools.\n\nIn-person attendance is recommended sinc e there will be activities in\nthe workshops to apply what you are learnin g to your upcoming\ncourse(s).\n\nPlease RSVP by Friday\, July 11\n\n→ RSVP and indicate your interest for each workshop\, as well as to\nvolunte er to talk or present an example from your work or class that\nuses GenAI in one of the workshop sessions. \n\n→ The workshops will be recorded\, but in person attendance is\nstrongly recommended.\n\nFaculty Hosts:  Zi co Kolter\, Tom Cortina \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9176563 DTSTART;TZID=America/New_York:20250716T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250716T133000 LOCATION:Newell-Simon 4305 SUMMARY:SCS Staff Lunch & Learn CLASS:PUBLIC DESCRIPTION:Talk Title: Unlocking the Power of Procurement\n\nThe SCS Staff Dean's Advisory Committee invites you to our second SCS\nStaff Lunch & Le arn. The series seeks to promote a sense of community\nby gathering to lea rn about rotating topics.  Hosted by the Staff\nRetention and Engagement Subcommittee of SDAC\, the sessions enable\nstaff to learn from each other and hopefully provide opportunities to\nget to know your co-workers a lit tle better in a friendly and\nsupportive learning environment.\n\nBen Swau ger\, Business Operations Assistant for the SCS Dean's office\nwill share his knowledge and awareness regarding a variety of the\nbusiness related processes throughout all of SCS.\n\nJoin us for an engaging session that d emystifies procurement. We\nwill uncover the mysteries of purchasing and show you how smart\nprocurement makes everyone's job easier. From getting the best value\nto staying on the right side of policy\, I’ll walk you through the\ndo’s\, don’ts\, and “wait\, we can do that?” Whether you're new\nto the process or want a refresher\, this session will give yo u\npractical insights to help you work more efficiently and confidently.\n \nAll SCS staff are welcome to attend.\n\n→ Bring your lunch and join us . Coffee and tea will be provided.\n\nRSVP \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917693c DTSTART;TZID=America/New_York:20251126T070000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251126T210000 SUMMARY:Thanksgiving Break - No Classes CLASS:PUBLIC DESCRIPTION:Talk Title: No Classes - Thanksgiving Break\n\nWednesday\, Nove mber 26 – Friday\, November 28\, 2025 — no classes.\n\nThursday\, Nov ember 27 and Friday November 28\, 2025  — university\nclosed. \n\nNorm al Class Schedules and Business Office Hours resume\nMonday\, December 1  \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9176bf3 DTSTART;TZID=America/New_York:20251024T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251024T160000 LOCATION:Tepper Building\, Carnegie Mellon University SUMMARY:Pittsburgh TechFest 2025 CLASS:PUBLIC DESCRIPTION:Speaker: Presented by the Pittsburgh Technology Council\n\nPitt sburgh TechFest is a one-day event designed for technologists\,\ndeveloper s\, and engineers from Southwestern Pennsylvania to convene\,\nexplore and elevate. Hosted by the Pittsburgh Technology Council\, the\nevent feature s keynotes\, technical talks\, and panel discussions\ncovering a broad ran ge of topics in AI\, software development\,\nsecurity\, and career growth. The conference aims to provide a platform\nfor learning\, networking\, an d collaboration among professionals.\n\nWhat you can expect in programming :\n\nPresent Problem – Solution – ImpactStorytelling with DataLive Dem o\nand DiscussionEngage with Big IdeasInteractive Demo\, Challenge\, or\nT ech WorkshopUnconference – take away the structure and engage in an\ninf ormal exchange of information and ideas between technologists\n\nTopics ar e listed below and all development stacks are welcome.\n\nUX as an Agile R oleUsing AI tools in Business and DevelopmentKeeping\nit Human (AI Related Ethics)Career Progression and Managing UpHome\nLabs and DIY TechData Scie nceMapping and RoboticsDeveloper Tools for\nImmediate UseFront End Develop ment (HTMX\, Angular\, React\, Svelte\, Vue\,\netc.)Diversity\, Equity\, a nd Inclusion in EngineeringPersonal Privacy\nand SecuritySocial Web (Fediv erse)\n\nREGISTER\n\nKeynote Speaker (and complete schedule of events)\n\n Mike Rainey — Mike is a research scientist in the CMU Computer\nScience Department\, School of Computer Science. Before CMU\, he spent\ntime at In diana University\, INRIA-Paris’s DeepSea/Gallium groups\,\nand the Max P lanck Institute for Software Systems\, after finishing his\nPh.D. at the U niversity of Chicago. Mike’s research sits at the\nboundary of programmi ng-language design and high-performance\ncomputing. He co-created Heartbea t Scheduling and Oracle-Guided\nScheduling\, ways to keep parallel tasks b alanced without wasting CPU\ncycles\, and the Task Parallel Assembly Langu age\, which carries those\nideas down to the OS/assembly-language level. L ately he’s been\ncollaborating on LoCal/Gibbon\, a type-safe framework f or working\ndirectly with serialized data\, while also publishing on effic ient\ngraph algorithms\, data structures\, and language-support for multic ore\nparallelism. His work shows up in venues like PLDI\, POPL\, PPoPP\,\n ASPLOS\, ICFP\, and SC\, and his open-source prototypes power experiments\ nwith tomorrow’s multicore software.\n\nJoin in!  TechFest Student\, F aculty & Professional Attendees from\nall industries can engage in:\n\nIns piring keynote sessions.A rich lineup of technical talks\, panel\ndiscussi ons\, and interactive demos covering topics like AI\, software\ndevelopmen t\, security\, UX\, ethics\, DIY tech\, data storytelling\, and\nmore. Eng aging unconference-style sessions—an open format designed\nfor informal tech conversations and peer learning. All attendees will\nbe encouraged to submit their ideas throughout the day for this\nexciting & immersive port ion of the event.\n\nCall for Proposal is Open\n\n→ The submission deadl ine is 10 July 2025\n\n→  All speakers will be notified of their accept ance on 17 July\n2025\n\n → Submit Your Proposal HERE.\n\nQUESTIONS?\n\ nDetails:\n\nThis is an all day conference\, enabling attendees to partici pate all\nday or to join in a particular conference sessions throughout th e day\,\nas schedules permit.CMU Faculty Registration fee: $30.00 under th e PTC\nMember price (includes continental breakfast\, lunch buffet\, confe rence\nrefreshments and end of TechFest happy hour networking event)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917718f DTSTART;TZID=America/New_York:20251028T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251028T170000 LOCATION:By Invitation Only SUMMARY:2025 CyLab Partners Conference - Day One CLASS:PUBLIC DESCRIPTION:Speaker: Faculty\, Student & Industry Speakers\n\nTuesday\, Oct ober 28\, 8am – Wednesday\, October 29\, 2025\, 4:05pm \n\nThe CyLab A nnual Partners Conference highlights the latest research in\nsecurity and privacy with an interactive forum between faculty\,\nstudents\, and indust ry.  We are excited to welcome our guests back to\ncampus for this year ’s two-day event\, which will include more than\n30 faculty and student presentations. All sessions will consist of\nbrief talks\, followed by act ive dialogue with our attendees.\n\nBy invitation.\n\n→Attendance is lim ited to invited guests\, representatives of CyLab's\npartners and CMU CyLa b faculty\, staff\, and students.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91774d6 DTSTART;TZID=America/New_York:20251030T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251030T180000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Distinguished Lecture: Bruce Nelson Memorial Lecture - Lalitesh \nKatragadda CLASS:PUBLIC DESCRIPTION:Speaker: LALITESH KATRAGADDA\, Founder and Chief Executive Offi cer\,\nIndihood Creator\, Google Map Maker Co-founder\, Google India\n\nTa lk Title: The Last Assignment\n\nWhen i graduated from CMU\, Raj asked us to go forth and build\ntechnology for the poor\, illiterate\, oppressed an d differently abled.\nRed meanwhile had drilled into me “If you haven't done everything\,\nyou ain't done anything”.\n\nIt took me a while to ge t serious about these. Along the way\, I learnt\na couple of things that s eem self-evident\, but rarely acknowledged.\nHave we made any product that works for all 8 Billion of us? Maybe\ntechnology matters\, far far more t han even we technologists\nacknowledge. And all people are capable. Having no proof\, nor theory\,\nwill share experiences.\n\nIn an inquiry led by building stuff\, peppered predominantly with\nfailed attempts\, i stumbled on a few green shoots\, only one became\nuseful for more than a billion\, another became vital for a couple of\nhundred million. The last few years has been an attempt to build\ntechnology that makes such platforms substa ntially easier for people\nto create.\n\nThis is likely to be my only lect ure at CMU. It will be filled with\ntechnical snippets. It will be a simpl e message to the students who\nare of my kids' age. It will be a report to my gurus and teachers:\nThat i am still at it. Though not successful. Not yet…\n\n—\n\nLalitesh Katragadda is a pioneer in population scale pla tforms. His\ncreation\, Google Map Maker\, tripled the world's digital map s\, mapping\n4 Billion people in 187 countries. The UN and aid agencies us ed these\nuser generated maps in hundreds of disasters to assist\, rescue\ nmillions.\n\nLalitesh co-founded Google's first international engineering center\,\nretired as Google’s Head of India products having invented\, led\nproducts including Crowdsourced YouTube Subtitles\, Crowdsourced\ntra nslation corpus\, Indic handwriting & Google Transliteration.  He\nis no w focused on creating technology that works for all 8 Billion of\nus\, as founder Indihood\, Swaja Labs\, co-founder Avanti Finance. As\nadvisor MEI TY\, DOT\, NHA\, DST\, NPCI\, NHAI\, NSE SCOT\, MoF\, IIT-Bombay\;\nLalite sh has architected India’s Open API\, AP Fiber Grid\, Aarogya\nSetu\, In dia’s public digital platform for Covid\, used by 200\nMillion\; co-desi gned eSign\, GST\, DigiLocker\, UPI 3.0\, GNSS Tolling.\n\nLalitesh has sp oken at TED\, the Brookings institute and the Whitehouse\non how the inter net will revolutionize the lives of the next 5\nBillion.  Prior to Google \, Lalitesh briefly served on the faculty of\nCarnegie Mellon where he fou nded the Lunar Rover Initiative and later\nfounded Sphereo\, a robotics st artup which became Google’s first\nacquisition in 2002. Lalitesh receive d his PhD and MS from CMU in\nRobotics and Computer Science\; MS from Desi gn Division\, Stanford\; MS\nin Aerospace\, Iowa State and his B-Tech from IIT-Bombay. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9177a22 DTSTART;TZID=America/New_York:20250730T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250730T152000 LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:Carnegie Mellon University’s Data Analytics Day CLASS:PUBLIC DESCRIPTION:Speaker: Sponsored by the Office of the Vice Provost\, the Offi ce of\nFinance\, and the Office of the CIOTalk Title: Advancing Data Analy tics\nthrough Community Engagement and Empowerment\n\nSponsored by the Off ice of the Vice Provost\, the Office of Finance\,\nand the Office of the C IO\, at this year's Data Analytics Day we will\nexplore Advancing Data Ana lytics through Community Engagement and\nEmpowerment. This year’s core t heme is directly connected to key\ninitiatives such as data literacy\, imp roving data governance\, building\na strong data franchise\, and the innov ative use of data analytics in\nuniversity strategic and operational plann ing.\n\nOur distinguished keynote speaker\, Steve Turk\, Chief Data and\nA nalytics Officer for JP Morgan Global Banking\,  will share insights\non \"Advancing Data Analytics through Community Engagement and\nEmpowerment.\ "\n\nSpeakers from CMU senior leadership\, faculty\, staff\, and guests\,\ nincluding\n\nStan Wadell — Chief Information Officer (Event Host)Ange la\nBlanton — Chief Financial OfficerMartial Hebert — Dean\, Schoo l\nof Computer Science Martial HebertTheresa Mayer — Vice President\nf or ResearchKelli Shuman — Associate Vice President and Chief\nHuman Re sources Officer\n\nReady to advance your data analytics knowledge?\n\nVisi t Data Analytics Day  for a detailed schedule.\n\nREGISTER\n\nYou will a lso have a chance to explore all the innovative\nvisualizations submitted by our IronViz Challenge finalists. This\nannual data visualization contes t will be held on Tuesday\, July 29. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9177e43 DTSTART;TZID=America/New_York:20250804T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250808T170000 LOCATION:Gates Hllman Center SUMMARY:Crash Course in AI for Teachers CLASS:PUBLIC DESCRIPTION:Speaker: Crash Course in Artificial Intelligence &amp\; Learnin g\nMaterial Design SessionMonday\, August 4\, 8:30am – Friday\, August 8\,\n2025\, 5pm Carnegie Mellon University School of Computer Science is \noffering a week-long workshop\, Crash Course in Artificial\nIntelligenc e along with a Learning Material Design Session to\nsupport high school educators looking to gain familiarity with AI and\nto offer AI-related ed ucational activities to their students.Our\nweek-long workshop will cover an introduction to a broad range of\nartificial intelligence topics includ ing: Machine Learning Deep\nLearning Search Algorithms Information Retriev al Recommender\nSystemsEach topic will include a hands-on activity to help you connect\nthe topic to a real-world application. On the last two days\ , we will\nform teams of CMU faculty and teachers to work together to crea te\nAI-related activities for your classrooms. Teachers will gain:\nhands -on skills to implement and apply AI algorithms\, access to all of\nthe wr itten materials and activities presented in the course\,\nadditional activ ities and content that they helped design to be\ntailored to their classro oms and their students’ interests\, a\nprofessional network of Pittsburg h area STEM teachers and Carnegie\nMellon faculty who can act as resources and collaborators for future\nprojects.Flyer with Additional Information Event\nType: Conference/Workshop Room Number: In Person Building: Gates \nHllman Center Speaker's Name: Crash Course in Artificial Intelligence\n & Learning Material Design Session For More\nInformation: mhyde2@andrew.c mu.edu Affiliations: Computer Science\nDepartment (CSD) Organization(s):  School of Computer Science Event\nWebsite Title: Event Website and Regi stration Event Website\nURL: www.cs.cmu.edu [http://www.cs.cmu.edu]…\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91782eb DTSTART;TZID=America/New_York:20250818T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250818T170000 LOCATION:McWilliams Classroom\, Gates Hillman 4303 SUMMARY:Incoming Doctoral Student Introductory Course CLASS:PUBLIC DESCRIPTION:Speaker: CSD Introductory Course (IC)Talk Title: FIRST DAY\n\nM onday\, August 18 through  August 31\, 2025\n\nEvent Website: https://www .cs.cmu.edu/~csd-ic/ (authentication\nrequired)\n\nContacts:   Matthew S tewart | Charlotte Yano \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9178649 DTSTART;TZID=America/New_York:20250901T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250901T230000 SUMMARY:Labor Day Observance CLASS:PUBLIC DESCRIPTION:Speaker: The University will be ClosedMonday\, September 1\, 20 25\,\n8am – 11pm Normal Business Hours and Class times resume Tuesday\ ,\nSeptember 1. Event Type: Fun Speaker's Name: The University will be\n Closed Affiliations: Carnegie Mellon University\, Computer Science\nDepar tment (CSD)\, Human-Computer Interaction Institute (HCII)\,\nLanguage Tech nologies Institute (LTI)\, Machine Learning Department\n(MLD)\, Partnershi ps\, Ray and Stephanie Lane Computational Biology\nDepartment (CBD)\, Robo tics Institute (RI)\, Software and Societal\nSystems Department (S3D) Orga nization(s): School of Computer Science\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91788fd DTSTART;TZID=America/New_York:20250904T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250904T180000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:Hans J. Berliner Lecture in Artificial Intelligence - Yejin Choi CLASS:PUBLIC DESCRIPTION:Speaker: YEJIN CHOI\, Dieter Schwarz Foundation Professor and S enior\nFellow\nDepartment of Computer Science\, and the Institute for Huma n-Centered\nArtificial Intelligence\nStanford University\nand Distinguishe d Research Scientist\, NVIDIA\n\nTalk Title: Bending Scaling Laws with Bri ghter Algorithms\n\nScaling laws tell us that \"more is more\" — brute-f orce scaling of\ndata and compute leads to smooth improvements in AI capab ilities.\nHowever\, this approach is becoming increasingly unsustainable i n\npractice\, creating a need for algorithmic innovations that can bend\nt he scaling laws and achieve more compute-efficient progress. In this\ntalk \, I will discuss our recent work in this direction\, including\ngradient- based methods for synthetic data generation\, prolonged\nreinforcement lea rning that can unlock stronger reasoning capabilities\nfrom smaller models \, symbolic search algorithms for test-time\nreasoning\, test-time trainin g that extracts additional learning even\nduring inference\, and a new tok enization algorithm that enables better\nand faster inference. \n\n—\n\ nYejin Choi is the Dieter Schwarz Foundation Professor and Senior\nFellow at the Department of Computer Science at Stanford University\nand the Stan ford Institute for Human-Centered Artificial Intelligence\n(HAI). She is a lso a Distinguished Research Scientist at NVIDIA\, and\nwas previously Pro fessor at the University of Washington and Senior\nDirector at the Allen I nstitute for AI (Ai2). Choi is a MacArthur\nFellow (class of 2022)\, AI205 0 Senior Fellow (class of 2024)\, and was\nnamed among Time100 Most Influe ntial People in AI in 2023. She is a\nco-recipient of 2 Test-of-Time award s and 9 Best and Outstanding Paper\nAwards at top AI conferences including ACL\, ICML\, NeurIPS\, ICCV\, CVPR\,\nand AAAI. She received the Borg Ear ly Career Award (BECA) in 2018\, won\nthe inaugural Alexa Prize Challenge in 2017\, and was named one of IEEE\nAI's 10 to Watch in 2016. Choi was a main stage speaker at TED 2023\nand has delivered keynote talks at confere nces across AI disciplines\nincluding ACL\, CVPR\, ICLR\, MLSys\, VLDB\, W ebConf\, and AAAI. Her current\nresearch interests focus on inference-time scaling for neural language\nmodels\, large and small reasoning models\, symbolic methods for neural\nlanguage models\, alternative training recipe s for language models\,\nsynthetic data generation for generative AI\, and pluralistic\nalignment. \n\nFaculty Hosts:  Vincent Conitzer\, Aditi Ra ghunathan\n\nAbout the Lecture:  The Hans J. Berliner Lecture in Artifici al\nIntelligence has been established in tribute to Hans J. Berliner\,\n(C S'74)  in recognition of the significant and critical\naccomplishments as faculty\, researcher\, advisor\, and exemplary\ncolleague and friend to m any.  This endowed lecture is presented by\nthe Computer Science Departme nt\, in conjunction with the SCS\nDistinguished Lecture Series\, and will let us reflect on Hans'\ncontributions and all they enabled. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9178e6a DTSTART;TZID=America/New_York:20250909T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250909T190000 LOCATION:Cohon University Center SUMMARY:STEM Career Fair - Day One CLASS:PUBLIC DESCRIPTION:Tuesday\, September 9\, 2025\, 12 – 7pm \n\nFor all studen ts across campus pursuing technical studies and roles.\nStudents from the College of Engineering\, School of Computer Science\,\nMellon College of S cience\, Dietrich College of Humanities and Social\nScience\, Tepper Schoo l of Business\, Heinz College of Information\nSystems and Public Policy\, and College of Fine Arts are welcomed.\n\nDaily Schedule:\n\n12:00 - 2:00 pm  -  All Students  2:00 - 3:00 pm  -  Employer\nBreak  3:00 - 5:00 pm  -  Undergraduate Students Only  5:00 - 7:00\npm  -  All Students \n\nDay One — Tuesday\, September 9 \n\n→  STEM Plus | Broadening the STEM Talent Pipeline (w/ special\nfocus on Applied Engineering)\n\nDa y Two — Wednesday\, September 10\n\nDay Three — Thursday\, Septemb er 11 - STEM Software & Tech\n\n→  Days 2 and 3  are dedicated to emp loyers hiring for computer\nscience\, software engineering\, data science\ , cybersecurity\, IT\, AI/ML\,\nand other technology-driven roles.  They are designed to connect\ncompanies with highly skilled students in compute r science\, data\, and\nemerging technologies.\n\nEmployer Registration \ n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917921e DTSTART;TZID=America/New_York:20250910T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250910T190000 LOCATION:Cohon University Center SUMMARY:STEM Career Fair - Day Two CLASS:PUBLIC DESCRIPTION:For all students across campus pursuing technical studies and r oles.\nStudents from the College of Engineering\, School of Computer Scien ce\,\nMellon College of Science\, Dietrich College of Humanities and Socia l\nScience\, Tepper School of Business\, Heinz College of Information\nSys tems and Public Policy\, and College of Fine Arts are welcomed.\n\nDaily S chedule\n\n12:00 - 2:00 pm  -  All Students\n\n  2:00 - 3:00 pm  -  E mployer Break\n\n  3:00 - 5:00 pm  -  Undergraduate Students Only\n\n  5:00 - 7:00 pm  -  All Students\n\nDay One — Tuesday\, September 9\ n\n→  STEM Plus | Broadening the STEM Talent Pipeline (w/ special\nfoc us on Applied Engineering)\n\nDay Two — Wednesday\, September 10\n\nDa y Three — Thursday\, September 11 - STEM Software & Tech\n\n→  Day s 2 and 3  are dedicated to employers hiring for computer\nscience\, soft ware engineering\, data science\, cybersecurity\, IT\, AI/ML\,\nand other technology-driven roles.  They are designed to connect\ncompanies with hi ghly skilled students in computer science\, data\, and\nemerging technolog ies.\n\nEmployer Registration \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91795e9 DTSTART;TZID=America/New_York:20250911T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250911T190000 LOCATION:Cohon University Center SUMMARY:STEM Career Fair - Day Three CLASS:PUBLIC DESCRIPTION:For all students across campus pursuing technical studies and r oles.\nStudents from the College of Engineering\, School of Computer Scien ce\,\nMellon College of Science\, Dietrich College of Humanities and Socia l\nScience\, Tepper School of Business\, Heinz College of Information\nSys tems and Public Policy\, and College of Fine Arts are welcomed.\n\nDaily S chedule\n\n12:00 - 2:00 pm  -  All Students\n\n  2:00 - 3:00 pm  -  E mployer Break\n\n  3:00 - 5:00 pm  -  Undergraduate Students Only\n\n  5:00 - 7:00 pm  -  All Students\n\nDay One — Tuesday\, September 9\ n\n→  STEM Plus | Broadening the STEM Talent Pipeline (w/ special\nfoc us on Applied Engineering)\n\nDay Two — Wednesday\, September 10\n\nDa y Three — Thursday\, September 11 - STEM Software & Tech\n\n→  Day s 2 and 3  are dedicated to employers hiring for computer\nscience\, soft ware engineering\, data science\, cybersecurity\, IT\, AI/ML\,\nand other technology-driven roles.  They are designed to connect\ncompanies with hi ghly skilled students in computer science\, data\, and\nemerging technolog ies.\n\nEmployer Registration \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91799bb DTSTART;TZID=America/New_York:20250916T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250916T123000 URL:https://db.cs.cmu.edu/events/industry-affiliates-program-visit-2025-day -2/ LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Database Group Industry Affiliate's Day (day 2) CLASS:PUBLIC DESCRIPTION:Talk Title: Database Group — Industry Affiliate's Day\n\nCarn egie Mellon University Database Group’s Industry Affiliates\nProgram (I AP) seeks to foster a dynamic partnership between academia\nand the databa se practitioners at early-stage companies. As a hub for\ngroundbreaking ad vancements in database systems\, CMU-DB IAP has\ncultivated a unique ecosy stem that brings together leading\nresearchers\, forward-thinking students \, and industry pioneers. This\nprogram serves as a bridge between theoret ical exploration and\nreal-world application\, propelling the field of dat abases into new\nfrontiers while providing invaluable insights and solutio ns to our\nindustry affiliates.\n\nThe CMU-DB Group hosts an annual on-cam pus Visit Day for our sponsors\neach fall semester. The goal is to bring t ogether researchers\,\nstudents\, and industry partners to discuss ongoing and future\nprojects\, as well as CMU-DB’s educational and outreach eff orts.\n\nThe second day is held in the Gates-Hillman Center\, shifts focus to\nthe industry side\, featuring a series of informative sessions\nprese nted by member companies. These sessions offer companies the\nopportunity to showcase their latest innovations\, products\, and\nchallenges in the d atabase space\, while also highlighting potential\ncareer opportunities fo r students. Attendees\, including faculty\,\nstudents\, and other particip ants\, can engage directly with company\nrepresentatives to learn about re al-world applications of database\ntechnologies\, industry trends\, and th e skills sought after in the\nfield. This day serves as a valuable network ing platform\, fostering\nstronger relationships between academia and indu stry.\n\nAll database students\, enthusiasts\, and connoisseurs at CMU are \nwelcome to join.\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9179dd7 DTSTART;TZID=America/New_York:20250925T081500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250925T200000 URL:https://www.cmu.edu/ai-sdm/research/human-ai-workshop/index.html LOCATION:Cohon University Center SUMMARY:2025 Workshop on Human-AI Complementarity for Decision Making CLASS:PUBLIC DESCRIPTION:Speaker: Human-AI Complementarity for Decision MakingTalk Title :\nSponsored by: NSF AI Institute for Societal Decision Making (NSF\nAI-SD M)\n\nWorkshop Motivation and Goals\n\nThis annual workshop explores the c oncept of Human-AI\nComplementarity—a condition where humans and AI syst ems working\ntogether outperform either working alone. Our 2025 theme focu ses on\nflexible Human-AI teams: systems that align with human values\,\nw ithstand unexpected behaviors\, and remain robust even under failure.\n\nK ey goals of the workshop include:\n\nDelivering cutting-edge instruction o n achieving Human-AI\ncomplementarityCreating common knowledge around emer ging research\nchallengesGenerating new ideas and concrete proposals for f uture\nresearch\n\nWho Should Participate?\n\nWe welcome contributions fro m multiple disciplines—decision science\,\ncognitive science\, computer science\, machine learning\, and beyond.\nParticipants may be:\n\nTutorial Instructors: Delivering state-of-the-art educational\nsessionsStudents: P resenting interactive posters and engaging in\ntutorialsPresenters: Provid ing brief\, targeted insights on key\nresearch topics\n\nTopics of Interes t\n\nSessions will focus on the flexibility and dynamics of Human-AI\ninte ractions for decision making\, including but not limited to:\n\nThe role o f AI agents in shaping human decision confidence and\ncalibrationAI's infl uence on trust\, coordination\, and\ncollaborationAddressing undesirable o r failure-prone AI behaviorsAnd\nothers.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917a1e1 DTSTART;TZID=America/New_York:20251004T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20251004T170000 LOCATION:Registration Required SUMMARY:Workshop on Neural Simulation-Based Inference (Day One) CLASS:PUBLIC DESCRIPTION:Speaker: Presented by: CMU Statistical Methods for the Physical \nSciences Research Center\n\nThe  STAtistical Methods for the Physical S ciences (STAMPS@CMU)\nresearch center is organizing a weekend workshop on neural\nsimulation-based inference on 4-5 October 2025. The workshop will take\nplace in person at the CMU campus in Pittsburgh.  A webcast for\nre mote participants will also be available.\n\nNeural simulation-based infer ence (neural SBI) has led to a recent\nparadigm shift in statistical infer ence across fields ranging from\nastronomy and particle physics to climate and environmental science.\nThe goal of this workshop is to bring togethe r researchers in neural\nSBI who have so far been largely siloed within di sjoint communities.\nWe expect that this will lead to cross-pollination of ideas across\nthese communities to facilitate the next advances in neural SBI and\nits applications.\n\nThe workshop will feature talks by leading researchers in neural SBI\,\nposter contributions by junior researchers\, and plenty of\nopportunities for interaction between the participants.\n\n Confirmed speakers include\n\nKyle Cranmer (University of Wisconsin-Madiso n)Gaia Grosso\n(Massachusetts Institute of Technology)Patrick Heimbach (Un iversity of\nTexas at Austin)Lukas Heinrich (Technical University of Munic h)Brian\nNord (Fermilab)Laurence Perreault-Levasseur (University of\nMontr eal)Barnabas Poczos (Carnegie Mellon University)Brian Reich\n(North Caroli na State University)Bingjie Wang (Pennsylvania State\nUniversity)Larry Was serman (Carnegie Mellon University)Minge Xie\n(Rutgers University)Andrew Z ammit-Mangion (University of Wollongong)\n\nREGISTER\n\n→ Seats are lim ited so please register as soon as possible.\n\nPoster Submissions deadlin e: September 26In-Person Registration\ncutoff:  September 26Remote Regist ration cutoff:  October 1\n\n—\n\nSTAMPS@CMU develops foundational stat istical methodology that\naddresses emerging open problems in fundamental physics\, environmental\nand climate sciences. Our goal is to promote trus tworthy scientific\ndiscovery that advances science and informs policy dec isions. We\nachieve this by fostering mutually beneficial and sustained\nc ollaborations between data scientists and physical scientists to\nleverage scientific expertise\, build trust in interpretable methods\,\nand transf er knowledge across the mathematical and physical\nsciences. \n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917a66f DTSTART;TZID=America/New_York:20250626T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250626T130000 URL:https://pdl.cmu.edu/talk-series/index.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Talk - Yaron Minsky CLASS:PUBLIC DESCRIPTION:Speaker: YARON MINSKY\, Head of Technology\, Jane Street\n\nTal k Title: The Saga of Multicore OCaml\n\nIn December 2022\, after nearly a decade of development\, OCaml 5 was\nreleased with a multi-core capable ga rbage collector. This was an\nexciting milestone\, finally making it possi ble to write shared-memory\nparallel programs in OCaml. The new runtime wa s designed to be easy to\nadopt: it didn’t disturb OCaml’s FFI\, and p erformance was meant to\nbe only a couple of percentage points slower in s ingle-core mode.\nDespite those promising beginnings\, moving to the new r untime was\nharder than we expected. Indeed\, We’ve only managed to swit ch to it\nthis year\, after 18 months of research and engineering effort t o\novercome a significant number of serious performance problems we\nuncov ered. This talk is essentially about technology transfer: about\nwhat it t akes to move an academic result from theory into practice.\nWe’ll talk a bout some of what we learned\, both about GC design\, and\nabout how to an alyze and fix subtle performance regressions in a\ncomplex ecosystem.\n\n —\n\nYaron Minsky got his BA in Mathematics from Princeton and his PhD i n\nComputer Science from Cornell focusing on distributed systems. He\njoin ed Jane Street in 2003\, where he founded the firm's quantitative\nresearc h group. He introduced OCaml\, a statically typed functional\nprogramming language\, to the company and managed the transition to\nusing OCaml for a ll of its core infrastructure\, turning Jane Street\ninto the world's larg est industrial user of the language.  He's been\ninvolved in many differe nt aspects of Jane Street's technology stack\,\nincluding machine learning infrastructure\, distributed systems design\,\nincremental programming sy stems\, hardware synthesis\,  trading and\nrisk systems\, developer tools \, and user-interface toolkits. \n\nZoom Participation.  See announcemen t.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917aad3 DTSTART;TZID=America/New_York:20250625T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250625T160000 URL:https://www.cylab.cmu.edu/events/2025/06/25-seminar-ohm.html LOCATION:Mehrabian Collaborative Innovation Center 2101 and Zoom SUMMARY:CyLab Seminar - Paul Ohm CLASS:PUBLIC DESCRIPTION:Speaker: PAUL OHM\, Professor of Law and Chief Data Officer\, G eorgetown\nUniversity Law Center\n\nTalk Title: Gradient Dissent: A Lawyer Reads the ML Literature\n\nLife is too short for most lawyers and legal s cholars to dig through\narXiv\, searching through an unruly haystack of ma chine learning papers\nfor the few needles of most relevance to legal prac tice and theory.\nProfessor Paul Ohm\, a legal scholar and erstwhile compu ter scientist\,\nhas been doing this work so other lawyers don't have to. In this talk\,\nhe offers a guided tour of recent landmark papers in machi ne learning\,\nconnecting them to important\, interesting\, and emerging q uestions in\nthe law. He will also preview two of his current projects: On e\nexplores the legal implications of government surveillance of\nconversa tions with chatbots. The other argues that generative AI may\ndrive down t he cost of regulatory compliance\, paving the way for new\nand aggressive forms of regulation. \n\n— \n\nPaul Ohm is a Professor of Law at the G eorgetown University Law Center\nin Washington\, D.C. In his research\, se rvice\, and teaching\, Professor\nOhm builds bridges between computer scie nce and law\, utilizing his\ntraining and experience as a lawyer\, policym aker\, computer programmer\,\nand network systems administrator. His resea rch focuses on information\nprivacy\, computer crime law\, surveillance\, technology and the law\, and\nartificial intelligence and the law. Profess or Ohm has published\nlandmark articles about the failure of anonymization \, the Fourth\nAmendment and new technology\, and broadband privacy. His w ork has\ndefined fields of scholarly inquiry and influenced policymakers a round\nthe world. He has testified before committees of both houses of\nCo ngress and advised numerous government agencies including the\nFederal Tra de Commission\, Federal Communications Commission\, and\nseveral state Att orneys General. A strong believer in the importance\nof public service\, P rofessor Ohm has served in numerous roles in\nfederal government. He start ed his legal career as a federal\nprosecutor in the U.S. Department of Jus tice’s legendary Computer\nCrime and Intellectual Property Section. More recently\, he served as a\nSenior Policy Advisor for Privacy to the Feder al Trade Commission and\nto the U.S. Commission on Evidence-Based Policyma king as an Obama\nWhite House appointee. Professor Ohm received a law degr ee from UCLA\nand degrees in Computer Science and Electrical Engineering f rom Yale\nUniversity. He writes thousands of lines of Python code each yea r and\nlives his life in Emacs and org-mode\, although he holds no ill-wil l\ntowards those who prefer vi. \n\nFaculty Host:  Lorrie Cranor \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917affb DTSTART;TZID=America/New_York:20250623T101500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250623T121500 LOCATION:Frick Fine Arts Building Auditorium\, 650 Schenley Drive\, Pittsbu rgh PA SUMMARY:Vera C. Rubin Observatory First Look Watch Party CLASS:PUBLIC DESCRIPTION:The Vera C. Rubin Observatory will conduct the 10-year Legacy S urvey\nof Space and Time (LSST) to map the changing Universe.  From findi ng\nnearby asteroids to discovering exploding stars in distant galaxies\na nd massive black holes billions of light years away\, Rubin\nObservatory's LSST will provide an unprecedentedly rich movie of the\nsky and enable a new era of astronomical discovery.   \n\nThe first images from Rubin Obs ervatory will be released on 23 June \n2025.  Join in to learn about why scientists in Pittsburgh are\nexcited about LSST’s panoramic view of th e changing sky and to take\npart in the unveiling of the first public imag es from this\nground-breaking observatory\, which has a mission to open th e skies for\neveryone.   \n\nThe event is also open to everyone\, but pl ease  — REGISTER  —if\nyou plan on attending.  \n\n— \n\nNSF –DOE Vera C. Rubin Observatory\, funded by the U.S. National\nScience Fo undation and the U.S. Department of Energy’s Office of\nScience\, is a g roundbreaking new astronomy and astrophysics\nobservatory under constructi on on Cerro Pachón in Chile\, with first\nlight expected in 2025. It is n amed after astronomer Vera Rubin\, who\nprovided the first convincing evid ence for the existence of dark\nmatter. Using the largest camera ever buil t\, Rubin will repeatedly\nscan the sky for 10 years and create an ultra-w ide\,\nultra-high-definition\, time-lapse record of our Universe.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917b3ec DTSTART;TZID=America/New_York:20250619T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250619T110000 LOCATION:Gates Hillman 8102 SUMMARY:Guest Talk - Abhishek Shetty CLASS:PUBLIC DESCRIPTION:Speaker: ABHISHEK SHETTY\, FODSI Postdoctoral Fellow\, Massachu ssets\nInstitute of Technology \, Catherine M. and James E. Allchin\nEarly -Career Assistant Professor in Computer Science\, Georgia\nInstitute of Te chnology (incoming)\n\nTalk Title: When does the past predict the future: Recent perspectives\nfrom Smoothed Online Learning\, Abstention and Optima l PAC Learning\n\nHow one makes use of data that has been collected in the past to\ninform us about decisions in the future is perhaps one of the mo st\nfundamental questions in computer science. Given its fundamental\nnatu re\, several frameworks have been devised to understand this\nquestion\, b ut a key assumption lies at the heart of these is\nindependence. Through t he years several frameworks have been proposed\nto circumvent this require ment\, many of these suffer from\nintractability either statistically or c omputationally. \n\nIn this talk\, we will survey a few recent techniques that have been\nproposed to handle these issues. Most of the talk will fo cus on\nsmoothed online learning which provides tools to design statistica lly\nand computationally efficient algorithms even when the data\ndistribu tion changes over time. Further\, we will see how this ties\ntogether with the notion of abstention in online learning\, which too\nleads to surpris ing statistically efficient algorithms. Time\npermitting\, we will briefly touch upon how analyzing generalization\nunder dependent data sheds light on optimal algorithms even when the\ndata is independent.   \n\n— \ n\nAbhishek Shetty is an incoming Catherine M. and James E. Allchin\nEarly -Career Assistant Professor in the School of Computer Science at\nGeorgia Tech and is currently FODSI Postdoctoral Fellow at MIT\, hosted\nby Sasha Rakhlin\, Ankur Moitra and Costis Daskalakis. He graduated\nfrom the depar tment of EECS at UC Berkeley advised by Nika Haghtalab.\nHis interests lie at the intersection of machine learning\, theoretical\ncomputer science a nd statistics\, especially aimed at understanding how\nfundamental algorit hmic techniques can contribute to modern machine\nlearning. His research h as been awarded with the Apple AI/ML\nfellowship and the American Statisti cal association SCGS best student\npaper. Faculty Host:  Andrej Risteski\ n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917b892 DTSTART;TZID=America/New_York:20250619T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250619T203000 SUMMARY:Juneteenth: Freedom Day Observation CLASS:PUBLIC DESCRIPTION:Speaker: The University is ClosedNormal Class Schedules and Off ice\nHours resume Friday\, June 20.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917bb02 DTSTART;TZID=America/New_York:20250617T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250617T183000 URL:https://www.cmu.edu/cce-office/events/2025/juneteenth.html?utm_campaign =2025-05-22+Piper&utm_source=announce&utm_medium=all&utm_content=campus_0_ 0&utm_term=0_0&utm_id=save+the+date+for+juneteenth+events LOCATION:Rangos Ballroom 1/2\, Cohon University Center SUMMARY:2025 Juneteenth Keynote Lecture CLASS:PUBLIC DESCRIPTION:Speaker: Dr. CLARENCE LANG\, Susan Welch Dean\, College of th e Liberal\nArts\, and \, Professor of African American Studies\, Penn  State\nUniversity\n\nTalk Title: Echoes of Freedom & JubilationClarence La ng is the Susan\nWelch Dean of the College of the Liberal Arts and prof essor of\nAfrican American studies at Penn State. A specialist in\nAfr ican American urban history and social movements\, he has\nspearheaded i nitiatives that expand student access and career\nreadiness while advancin g the recruitment\, retention\, and professional\ndevelopment of diverse f aculty and staff. Before joining Penn State\nin 2019\, Lang held posts a t the University of Kansas\, including\nDean’s Professor\, chair of Afri can and African American Studies\, and\ninterim director of the Hall Cente r for the Humanities. An\nOrganization of American Historians Distinguishe d Lecturer\, he is the\nauthor of “Grassroots at the Gateway: Class Po litics and Black\nFreedom Struggle in St. Louis\, 1936–75” and  “Black America\nin the Shadow of the Sixties\,” and co‑editor of thr ee volumes on\nanticommunism\, A. Philip Randolph\, and Black urban hi story.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917beab DTSTART;TZID=America/New_York:20250617T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250617T153000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:Doctoral Thesis Proposal - Daiyaan Arfeen CLASS:PUBLIC DESCRIPTION:Speaker: DAIYAAN ARFEEN\, Ph.D. Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Designing Scalable DNN Training Systems to Overcome\nAlgorithmic Constraints\n\nLLM training req uires massive amounts of compute due to large model\nand dataset sizes\, s o it is not unusual to train LLMs on tens or\nhundreds of thousands of GPU s to complete training in a reasonable\namount of time (days or weeks). Ho wever\, GPU failures (which are\ncommon at these scales) and data-dependen cies (introduced by the\ntraining algorithms) can lead to severe GPU under utilization.  \n\nIn this talk\, we present distributed LLM training sys tems which are\nefficient and fault-tolerant at these scales. We first pre sent\nNonuniform-tensor-parallelism (NTP)\, a technique which increases th e\nfault-tolerance of tensor-parallel training\, thereby reducing the\nbla st-radius of GPU failures. NTP enables scale-up training with\nlittle-to-n o loss in training efficiency from realistic rates of GPU\nfailures. Next we present PipeFill\, a system for recovering GPU\nutilization (lost due t o scale-out training) by filling pipeline\nbubbles with third-party latenc y-insensitive jobs. We will discuss how\nPipeFill could be extended to sup port filling pipeline bubbles with\nonline inference jobs\, which are late ncy-sensitive.\n\nThesis Committee\n\nGreg Ganger (Chair)\n\nZhihao Jia\n\ nPhillip B. Gibbons\n\nDheevatsa Mudigere (NVIDIA)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917c2a1 DTSTART;TZID=America/New_York:20250613T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250613T153000 LOCATION:Remote Access - Zoom SUMMARY:CRA Research Policy Virtual Town Hall - Faculty - registration\nreq uired CLASS:PUBLIC DESCRIPTION:The Trump Administration's recently unveiled Fiscal Year 2026 b udget\nrequest poses a significant threat if passed into law\, especially for\nthose who rely on National Science Foundation funding. This\ndevelopm ent\, coupled with prior administration actions such as funding\nfreezes\, award cancellations\, proposed cuts in facilities and\nadministrative cos ts\, and federal research agency layoffs\, creates a\ncritical situation f or the computing research community.\nFaculty are invited to join in for t his 30 minute Zoom webinar\, to we\nhear from members of CRA's Office of G overnment Affairs about the\nPresident FY26 budget request\, how CRA is re sponding\, and answer your\nquestions to better understand what's going on .\nREGISTER\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917c61c DTSTART;TZID=America/New_York:20250612T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250612T130000 URL:https://pdl.cmu.edu/talk-series/2025/061225.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Talk - Pat Helland & Daniel May CLASS:PUBLIC DESCRIPTION:Speaker: PAT HELLAND and DANIEL MAY\, Salesforce\n\nTalk Title: Yours\, Mine\, and Ours: Efficient Set Reconciliation in O(n\nlog n) of t he SET DIFFERENCE\n\nWe will discuss and explain a new algorithm that empo wers efficient\nreconciliation of sets. Extremely large sets can be reconc iled in O(n\nlog n) of the SET DIFFERENCE\, not the underlying size of the sets. The\nalgorithm is a variant of erasure codes (familiar to the datab ase\ncommunity) and fountain codes (familiar to the data communications\nc ommunity). This opens new avenues for solutions based on repairing\nsets t hat do not even yet exist! When distributed systems agree in\nadvance what items belong in a set\, different participants can add\nitems to the set effectively performing replica repair over future\ncontent. \n\nWe will e xplain the set reconciliation algorithm presented at SIGCOMM\n2024 in Augu st within a paper titled \"Practical Rateless Set\nReconciliation\" by Yan g et al and how it can accomplish such\nefficiency. Many disparate researc h opportunities are opened by this\nalgorithm including replica repair (fa ster than Merkle Trees)\,\nimproved gossip protocols\, scientific computat ions including detecting\nsmall differences in large genomes\, management of cloud based control\nplanes\, and possibly even improvements to multi-p hase protocols used\nfor distributed systems. We hope to conclude with the audience\nbrainstorming about even more possible applications. \n\n—  \n\nPat Helland has been building distributed systems and databases since\ n1978 at companies including Tandem\, Microsoft\, and Amazon. He is\nconst antly curious about emerging trends in technology and their\nimplications on systems. Pat has been working on database technology\nat Salesforce sin ce 2012. \n\nDaniel May works at Salesforce and has been analyzing and im proving\nthe performance of complex systems for more than 7 years. He spen ds\nmuch of his spare time voraciously consuming technical papers\, largel y\nabout distributed systems. \n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917ca81 DTSTART;TZID=America/New_York:20250605T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250605T130000 URL:https://pdl.cmu.edu/talk-series/index.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory (PDL) Talk - Sara McAllister CLASS:PUBLIC DESCRIPTION:Speaker: SARA McALLISTER\, Ph.D. Candidate\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Taking off the comp ute-colored glasses: Storage is vital\nto datacenter sustainability\n\nBy 2050\, datacenters are expected to account for over 20% of global\ncarbon emissions. Most of the emissions will be embodied (from\nmanufacturing\, t ransporting\, and disposing of datacenter hardware).\nUnfortunately\, the vast majority of research on reducing datacenter\nembodied emissions focus es on compute\, even though the majority come\nfrom storage. My research s tarts to remedy this gap through rethinking\nstorage system design to grea tly reduce embodied emissions. In this\ntalk\, I will first introduce how IO bottlenecks limit storage's\nsustainability. I will then present how my projects\, Kangaroo and\nFairyWREN\, show that overcoming flash's write l imitations enables\nnear-optimal emissions for flash caching. Finally\, I will discuss how\nto curb hard disk drive's IO bottlenecks with Declarativ e IO to enable\ndesner\, lower emissions drives in bulk storage. \n\n—  \n\nSara McAllister is a final-year PhD student at CMU\, working with\nN athan Beckmann and Greg Ganger. She researches how to create more\nsustain able datacenters\, particularly for caching and storage systems.\nHer work includes a focus on improving efficiency and sustainability\nthrough hard ware‑software co‑design and grounding design choices\nin mathematical modeling. Her work has appeared at OSDI\, SOSP\, and\nICML\, including rec eiving a Best Paper Award at SOSP 2021. She is a\n2021 NDSEG fellow\, a 20 23 EECS Rising Star\, and a 2025 Siebel Scholar.\n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917ce96 DTSTART;TZID=America/New_York:20250605T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250605T120000 LOCATION:Gates Hillman 7101 SUMMARY:Doctoral Speaking Skills Talk - Jeff Xu CLASS:PUBLIC DESCRIPTION:Speaker: JEFF XU\, Ph.D. Student\, Computer Science Department\ , Carnegie\nMellon University\n\nTalk Title: Smooth Tradoff for Tensor PCA in Polynomial Time\n\nTensor Principal Component Analysis (PCA) is a cano nical problem in\nhigh-dimensional statistical inferences. Prior works hav e shown the\nsuccess of various spectral algorithms essentially matching t he\nconjectured computational thresholds for this problem. \n\nIt is also known that (at least) in the sub-exponential time regime\,\nsimilar to ra ndom CSP refutation\, Tensor PCA exhibits a smooth\ntradeoff in signal-str ength and runtime: increasing run-time allows\none to detect a weaker sign al. However\, it is not clear whether such\ntradeoff appears in the polyno mial-time regime. Recently\, the work of\nBandeira et al. makes partial pr ogress by establishing a smooth\ntrade-off in a “limited” polytime reg ime via techniques from free\nprobability. \n\nIn this talk\, we give a s elf-contained combinatorial argument to\nestablish a smooth tradeoff in th e full runtime. \n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917d218 DTSTART;TZID=America/New_York:20250602T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250602T113000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Doctoral Thesis Proposal - Margarida Ferreira CLASS:PUBLIC DESCRIPTION:Speaker: MARGARIDA FERREIRA\, Ph.D. Student\, Computer Science\ nDepartment\, Carnegie Mellon University\n\nTalk Title: Synthesis of State ful Programs from Execution Traces\n\nExecution traces are a valuable sour ce of information in modern\ncomputing systems. They are continuously coll ected and used for system\ndebugging\, monitoring\, and optimization. They capture behavior across\ndiverse scenarios\, from routine operations to e dge cases. This thesis\ninvestigates how execution traces can serve as spe cifications for\nprogram synthesis\, enabling reverse engineering and anal ysis of\ncomplex systems and automation of traditionally manual tasks with out\nexplicit user input.\n\nThis proposal presents three synthesis framew orks\, Abagnale\, Syren\,\nand HyGLAD\, that illustrate the challenges of this problem on multiple\napplications and how we overcome them. Abagnale reverse-engineers the\nbehavior of congestion control algorithms (CCAs) fr om network traces.\nNetwork traces contain no information about the implem entation of the\nCCA\, displaying only the effects of their executions in the network.\nThus\, Abagnale must simulate each candidate solution in the same\nnetwork conditions to assess if they exhibit the same behavior. To\ ncapture all different behaviors\, we work with traces showing hundreds\no f executions\, making trace filtering and parallelization paramount to\nAb agnale's viability. Syren allows users to generate arbitrary\nprograms fro m partial traces that contain some of the function calls\nmade by the prog ram. Syren uses optimizing rewrites to introduce\ncontrol flow in the prog ram. These optimizing rewrites track the data\nused in the functions visib le in the trace\, which is then used to\ngenerate function calls not visib le in the trace using an\nexample-based syntax-guided synthesizer. HyGLAD synthesizes\nregex-based anomaly filters that flag deviations from a syste m's\nexpected behavior from execution logs. In this case\, our goal is not \nto reverse-engineer the system itself but to synthesize a model of its\n execution.\n\nAs future work\, we propose to develop a fourth synthesis ap proach to\nautomate data-aware business processes. We will use logs collec ted\nfrom human-executed processes as traces and synthesize implementation s\nthat model the task logic\, filtering out inconsistencies and errors\nu navoidable in human-generated logs.\n\nThesis Committee\n\nRuben Martins ( Co-chair)\n\nInês Lynce (Co-Chair\, Instituto Superior Técnico)\n\nJusti ne Sherry\n\nFraser Brown\n\nJoão F. Ferreira (Instituto Superior Técnic o)\n\nNate Foster (Cornell University)\n\nAdditional Information\n\nIn Per son and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917d737 DTSTART;TZID=America/New_York:20250530T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250530T130000 LOCATION:Newell-Simon 4305 SUMMARY:Doctoral Speaking Skills Talk - Sanjith Athlur CLASS:PUBLIC DESCRIPTION:Speaker: SANJITH ATHLUR\, Ph.D. Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Okapi: Decoupling Data Striping and Redundancy Grouping in\nCluster File Systems\n\nThe Okapi cl uster file system decouples how data is spread across\ndisks (data stripin g) for IO efficiency from how data is erasure coded\ntogether (redundancy grouping) for durability. Existing systems couple\nthese two mechanisms’ configurations\, inducing significant\ninefficiencies. Decoupling allows grouping to be configured based on\nreliability and space efficiency goals \, while simultaneously allowing\nstriping to be configured based on perfo rmance goals. Decoupling also\nallows redundancy scheme changes from one E C scheme to another (e.g.\,\nto react to data temperature or disk failure rate changes) to occur\nwithout having to re-write data.  \n\nEvaluation of an Okapi prototype shows that decoupling can be\naccomplished with\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917dab2 DTSTART;TZID=America/New_York:20250527T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250527T140000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Suhas Jayaram Subramanya CLASS:PUBLIC DESCRIPTION:Speaker: SUHAS JAYARAM SUBRAMANYA\, Ph.D. Candidate\, Computer Science\nDepartment\, Carnegie Mellon University\n\nTalk Title: Efficient and Responsive Job-Resource Co-adaptivity for\nDeep Learning Workloads in Large Heterogeneous GPU Clusters\n\nExisting cluster schedulers face many limitations in scheduling\nadaptive deep learning training jobs on large h eterogeneous GPU\nclusters – many are not heterogeneity-aware\, few are\ nadaptivity-aware\, and none scale to large clusters without sacrificing\n allocation fidelity or cluster efficiency. Emerging clusters further\ncomp licate this problem with larger\, more heterogeneous resources\nrunning m ore increasingly diverse jobs with more dimensions of\nadaptivity.\n\nThis thesis develops new scheduling approaches and algorithms that can\n(1) sc ale to emerging clusters with hundreds of thousands of GPUs and\nmany GPU types\, (2) quickly optimize high-fidelity allocations for\nadaptive DL tr aining jobs with low scheduler overhead\, and (3)\nefficiently adapt to ch anging cluster conditions to improve goodput on\nthe limited GPU resources .\n\nWe first introduce Sia — a round-based scheduler that efficiently\n optimizes adaptive jobs in a heterogeneous cluster with many GPU\ntypes. S ia uses GPU resources judiciously to gather information on\njob-GPU fit-le vels using a mix of online and offline profiling\, and\ncontinuously co-op timizes the GPU resources allocated to jobs and\ntheir execution parameter s at runtime to maximize cluster-wide\ntraining progress. Using job traces derived from real-world data\ncenters\, we find that Sia ’s allocations are fair and efficient\, and\nare quickly computed using an efficient for mulation\, even for 1000-GPU\nclusters.\n\nSecond\, we introduce continual optimization — a new paradigm that\nexplicitly models the slow evolutio n of resource-allocation problems\nat scale to reduce solver runtime for q uick responses to changes in\njobs or resources. We then introduce COpter\ , our approach to continual\noptimization that (a) efficiently updates the optimization problems\nfor job and resource changes using a differential interface\, (b)\nimplements a factorization-free warm-started LP solver to benefit from\nslowly-evolving nature of the allocations\, and (c) impleme nts\nlightweight heuristics to recover feasible integral solutions with\nn egligible quality loss. In our evaluations\, COpter speeds up Sia\nschedu ler policy by a few orders of magnitude on clusters with tens of\nthousand s of GPUs without sacrificing job completion times and\nmakespan.\n\nThird \, COpter is easily applied to resource-allocation problems in\nother doma ins (e.g. shard load-balancing\, WAN traffic engineering) and\nwe see 57 − 83 × reductions in solver runtimes.\n\nThesis Committee\n\nGregory Ga nger (Chair) Zhihao Jia Virginia Smith Amar Phanishayee\n(Meta Platforms I nc.)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917dfbf DTSTART;TZID=America/New_York:20250526T000000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250526T235900 SUMMARY:Memorial Day (2025) - University Closed CLASS:PUBLIC DESCRIPTION:The University is closed and there are no classes.\nRegular off ice hours and schedules resume\, Tuesday\, May 27.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917e227 DTSTART;TZID=America/New_York:20250523T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250523T170000 SUMMARY:CMU Community Appreciation Day (May 2025) CLASS:PUBLIC DESCRIPTION:A day off for faculty and staff. \nLimited campus operations w ill be open\, such as Dining Services\,\nUniversity Libraries and the Coho n Center —the hours of operation\nwill be noted on relevant university w ebsites. \n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917e832 DTSTART;TZID=America/New_York:20250520T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250520T110000 LOCATION:Reddy Conference Room\, Gates HIllman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Mingjie Sun CLASS:PUBLIC DESCRIPTION:Speaker: MINGJIE SUN\, Ph.D. Candidate\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Hidden Properties of La rge Language Models\n\nLarge Language Models (LLMs) are deep learning mode ls trained to\nunderstand and generate natural language. Over the course o f my PhD\,\nthese models have profoundly transformed the field of machine\ nlearning. Despite their remarkable success\, most of our interactions\nwi th LLMs remain largely black-box\, leaving key questions about their\ninte rnal mechanisms and behaviors under-explored.\n\nThis thesis investigates previously overlooked hidden properties of\nLLMs across three dimensions: internal weight structure\, activation\npatterns\, and output behaviors. F irst\, we demonstrate that the weight\nspace of LLMs is intrinsically spar se and present a principled pruning\napproach capable of extracting effici ent sparse subnetworks directly\nfrom pre-trained models. Next\, we reveal the existence of structured\nactivation outliers in LLMs\, which we call \"massive activations\".\nThese activations\, despite their rarity\, are e xceptionally high in\ntheir magnitudes. We establish their strong connecti on to the\nself-attention mechanism and propose a novel attention formulat ion\nthat mitigates these extreme outliers. Finally\, we characterize the\ nidiosyncrasies of LLM outputs\, showing that generations from different\n models can be distinguished with remarkably high accuracies. We\nfurther i dentify the specific signatures that underlie these\ndifferences. Collecti vely\, these findings provide an alternative\nperspective on modern founda tion models. \n\nThesis Committee\n\nJ. Zico Kolter (Chair)\n\nGraham Neu big\n\nAditi Raghunathan\n\nKaiming He (Massachusetts Institute of Technol ogy)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917ed09 DTSTART;TZID=America/New_York:20250519T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250519T153000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Doctoral Thesis Proposal - Alexander Koujianos Goldberg CLASS:PUBLIC DESCRIPTION:Speaker: ALEXANDER KOUJIANOS GOLDBERG\, Ph.D. Student\, Compute r Science\nDepartment\, Carnegie Mellon University\n\nTalk Title: Improvin g Decision-Making from Distributed Human\nEvaluations\n\nSocially importan t decisions—from scientific funding\, to college\nadmissions and job hir ing—rely on ratings or rankings supplied by\nmultiple human evaluators. These judgments are prone to noise\, bias\,\nand strategic manipulation\, and there is seldom an objective ground\ntruth against which to determine their quality. The goal of this\nthesis is to understand and mitigate such errors in distributed human\nevaluation in order to make better decisions . Towards this end\, we\nboth conduct controlled experiments in review pro cesses and\ndevelop principled algorithms with provable guarantees. \n\ nIn particular\, we conduct large-scale experiments at peer\nreview confe rences to expose sources of error in evaluation and\nidentify opportunitie s for improvement. Then\, we develop a method for\nselecting top candidat es on the basis of uncertain evaluations\,\nproviding a principled instant iation of a \"peer review lottery.\"\nFinally\, we design privacy-preserv ing algorithms for releasing\nanonymized time-series and graph data\, whic h can enable more\ntransparency into review processes while preserving par ticipant\nanonymity.\n\nThesis Committee\n\nGiulia Fanti (Co-chair)\n\nNih ar B. Shah (Co-chair)\n\nTom Mitchell\n\nJohn Ioannidis (Stanford Universi ty)\n\nAdditional Information\n\nIn Person and Zoom Participation.  See a nnouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917f171 DTSTART;TZID=America/New_York:20250515T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250515T160000 LOCATION:Gates Hillman 6501 SUMMARY:Doctoral Thesis Oral Defense - Meng-Chieh (Jeremy) Lee CLASS:PUBLIC DESCRIPTION:Speaker: MENG-CHIEH (JEREMY) LEE\, Ph.D. Candidate\, Computer S cience\nDepartment\, Carnegie Mellon University\n\nTalk Title: Explainable Mining of Graphs and Time Series: Algorithms\nand Applications\n\nGiven a social network graph\, how can we predict connections between\nusers and determine whether they are based on shared hobbies or common\nfriends? Giv en a database containing molecular graphs\, how can we\ndetermine whether the graphs inhibit HIV replication based on\nsubstructures they frequently share? Similarly\, in time series data\nfrom EEG recording\, how can we i dentify seizures and explain why they\nare considered abnormal? Although r ecent machine learning methods have\nshown improved performance\, many rem ain black-box models\, making\nexplainability challenging. This leads us t o explainable artificial\nintelligence (XAI)\, which offers valuable insig hts through its\nexplanations and is more practical for deployment in real -world\napplications.\n\nIn this thesis\, we focus on developing explainab le machine learning\nmethods tailored for graphs and time series. Each met hod we propose is\neither inherently explainable\, or designed to automati cally provide\ndata analysis or justification for its decisions. In each p art\, we\npresent effective and general algorithms\, and explore a broad r ange of\napplications.\n\nThesis Committee\n\nChristos Faloutsos (Co-chair )\n\nLeman Akoglu (Co-chair)\n\nGeoffrey Gordon\n\nNina Mishra (Amazon)\n\ nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917f553 DTSTART;TZID=America/New_York:20250514T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250514T133000 LOCATION:Rangos Ballroom and Wiegand Gym\, Cohon University Center | Tartan \nPavilion\, Resnick Hall/Legacy Plaza SUMMARY:CMU 2025 Community Picnic! CLASS:PUBLIC DESCRIPTION:Speaker: CMU Community Picnic\n\nThe CMU Community Picnic honor s the vital contributions staff members\nmake to the university. Please jo in us as the CMU community gathers to\nenjoy some good food\, outdoor game s\, raffles and more!\n\nRemember!\n\nArtists Wanted - Get ready to displa y your creativity! The CMU\nCommunity Picnic will once again feature a cur ated art exhibition\, and\nwe invite all CMU community members to submit t heir work.\n\n→ Details on submission guidelines and deadlines will be a nnounced\ncloser to the picnic.\n\n→ Due to limited space\, the organizi ng committee will make the final\nselections for the exhibition.\n\nWe can 't wait to celebrate another successful year with you!\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917f889 DTSTART;TZID=America/New_York:20250513T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250513T150000 URL:https://www.cs.cmu.edu/~pop/seminar/ LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Principles of Programming (PoP) Seminar - Matthew Parkinson CLASS:PUBLIC DESCRIPTION:Speaker: MATTHEW PARKINSON\, Principal Researcher\, Confidentia l\nComputing Group\, Microsoft Research\n\nTalk Title: Dynamic Region Owne rship for Concurrency Safety\n\nThe ways in which the components of a prog ram interact with each other\nin a concurrent setting can be considerably more complex than in a\nsequential setting. The core problem is unrestrict ed shared mutable\nstate. An alternative to unrestricted shared mutable st ate is to\nrestrict the sharing using Ownership. Ownership can turn what w ould\nhave been a race into a deterministic failure that can be explained to\nthe programmer. However\, Ownership has predominantly taken place in\n statically typed languages. \n\nIn this talk\, we explore retrofitting an existing dynamically typed\nprogramming language with an ownership model based on regions. Our\ncore aim is to provide safe concurrency\, that is\, the ownership model\nshould provide deterministic dynamic failures of own ership that can be\nexplained to the programmer. We present a dynamic mode l of ownership\nthat provides ownership of groups objects called regions. We provide\ndynamic enforcement of our region discipline\, which we have\n implemented in a simple interpreter that provides a Python-like syntax\nan d semantics\, and report on our first steps into integrating it into\nan e xisting language\, Python. \n\nThis talk is based on a paper that will ap pear at PLDI'25. \n\n— \n\nMatthew Parkinson works in the Security and Privacy group in Azure\nResearch.  His research focuses on memory safety and concurrency\, in\nparticular\, how to make better programming languag es for Cloud\nInfrastructure programming.  He is the lead maintainer of s nmalloc\, a\nhigh-performance allocator.  Previously\, he worked on exten ding\nseparation logic to reason about both object-oriented and concurrent \nprogramming languages.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef917fca7 DTSTART;TZID=America/New_York:20250512T122000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250512T132000 URL:https://db.cs.cmu.edu/events/dbsp-incremental-computation-on-streams-an d-its-applications-to-databases LOCATION:Gates Hillman 9115 and Zoom SUMMARY:Database Seminar - Mihai Budiu CLASS:PUBLIC DESCRIPTION:Speaker: MIHAI BUDIU\, Chief Scientist\, Feldera.com\n\nTalk Ti tle: DBSP: Incremental Computation on Streams and Its\nApplications to Dat abases\n\nWe describe DBSP\, a framework for incremental computation. Incr emental\ncomputations repeatedly evaluate a function on some input values that\nare \"changing\". The goal of an efficient implementation is to \"re use\"\npreviously computed results. Ideally\, when presented with a new ch ange\nto the input\, an incremental computation should only perform work\n proportional to the size of the changes of the input\, rather than to\nthe size of the entire dataset. \n\nIn databases \"incremental computation\" is known as Incremental View\nMaintenance (IVM)\; IVM has long been a cen tral problem of database\ntheory and practice. \n\nWe define incremental computations as computations on datastreams\,\ni.e.\, sequences of data va lues\, by borrowing ideas from Digital Signal\nProcessing. We then give a general solution to the incremental\ncomputation problem\, including an al gorithm for converting any DBSP\nprogram into an incremental program. Feld era is an early-stage\nstart-up that has built a full-fledged incremental SQL query engine\nbased on a DBSP Rust runtime. \n\nThis work has receive d the 2023 VLDB best paper award\, and the 2024\nACM SIGMOD research highl ights award. \n\n— \n\nMihai Budiu is chief scientist at Feldera.com\, an early-stage startup.\nHe has a Ph.D. in computer science from Carnegie Mellon University. He\nwas previously employed at VMware Research\, Baref oot Networks\, and\nMicrosoft Research. Mihai has published papers on reco nfigurable\nhardware\, computer architecture\, compilers\, security\, dist ributed\nsystems\, big data platforms\, large-scale machine learning\,\npr ogrammable networks and P4\, data visualization\, and databases\; four\nof his papers have received “test of time” awards. He has also\nreceived two technology transfer awards. \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91800f2 DTSTART;TZID=America/New_York:20250512T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250512T123000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Dorian Yao Chan CLASS:PUBLIC DESCRIPTION:Speaker: DORIAN YAO CHAN\, Ph.D. Candidate\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Holographic Display s for Computer Vision\n\nArtificial illumination is ubiquitous in real vis ion systems. By\ncoding extra information from a controlled light source i nto the\nimages captured by a camera\, so-called \"active sensing\" approa ches\nrobustly capture depth\, reflectance and other visual cues crucial t o\ntasks in robotics\, manufacturing\, consumer products and more. However \,\nactive sensors struggle with well-known challenges that limit their\np racticality in modern applications — available power limits range\nand o utdoor performance\, slow speed precludes dynamic scenes and\ndefocused st ructured illumination reduces effective resolution. \n\nTo tackle these c hallenges\, this thesis explores using holographic\ndisplays. Holographic displays have recently seen significant\nattention in the augmented and vi rtual reality (AR/VR) literature. By\nsimply illuminating a spatial-light modulator (SLM) with laser light\,\nsuch devices can simultaneously provid e accommodation cues and\nglasses-free vision correction all in a compact form factor\, key\ncapabilities that are currently missing in modern AR/VR architectures.\n\nIn our work\, we analyze how they can potentially be ad apted as sources\nof active illumination. First\, we show how holographic displays can be\nused to build light redistributive projectors that allow for smarter\nenergy usage in active sensing\, enabling time-of-flight sens ors with\nfar-improved dynamic range. Next\, we demonstrate how this light \nredistribution\, when combined with the underlying fast speed of modern\ nSLMs\, allows for far faster projector systems\, allowing for new types\n of triangulation light curtains. Finally\, we test how the inherent\ncoher ent propagation of holographic systems can be used to program\nmeaningful content at multiple projector depths\, enabling new user\ninterfaces and d epth-sensing methodologies.\n\nOverall\, this defense advances the state-o f-the-art in active sensing\nby demonstrating new ways in which light can be shaped and\nconcentrated via holographic illumination systems. These ab ilities\nunlock vision systems with increased robustness and newfound\ncap abilities. \n\nThesis Committee\n\nMatthew O’Toole (Chair)\n\nIoannis G kioulekas\n\nAswin Sankaranarayanan\n\nMohit Gupta (University of Wisconsi n–Madison)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9180585 DTSTART;TZID=America/New_York:20250511T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250511T160000 URL:https://www.cs.cmu.edu/events/commencement-2025 LOCATION:Carnegie Music Hall\, 4400 Forbes Avenue SUMMARY:SCS Commencement - SCS Undergraduate Diploma Ceremony (S25) CLASS:PUBLIC DESCRIPTION:Speaker: SCS Commencement - SCS Undergraduate Diploma Ceremony\ n\nSCS Undergraduates and Guests - Tickets Required.\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9180838 DTSTART;TZID=America/New_York:20250510T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250510T140000 URL:https://www.cs.cmu.edu/events/commencement-2025 LOCATION:Carnegie Music Hall\, 4400 Forbes Avenue - Oakland SUMMARY:SCS Commencement: SCS Ph.D. Hooding and Diploma Ceremony (S25) CLASS:PUBLIC DESCRIPTION:Speaker: SCS Commencement: SCS Ph.D. Hooding and Diploma Ceremo nyRSVP\nRequired by all graduates.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9180afc DTSTART;TZID=America/New_York:20250509T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250509T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Remote Access - Zoom SUMMARY:AI-SDM Seminar - Giacomo De Nicola CLASS:PUBLIC DESCRIPTION:Speaker: GIACOMO DE NICOLA\, Postdoctoral Research Fellow\, Har vard T.H.\nChan School of Public Health\, Harvard University\n\nTalk Title : Harmful Algal Blooms and Their Impact on Marine Food\nPoisonings in Mada gascar\n\nHarmful algal blooms (HABs) produce toxins that contaminate coas tal\nwaters and aquatic foods\, which can lead to severe and potentially\n deadly intoxications when consumed. Making use of satellite data\, we\nexa mine spatial and temporal patterns of algal blooms around\nMadagascar\, an extremely poor country heavily affected by HABs due to\na combination of increasing nutrient pollution and rising water\ntemperatures. Madagascar's population strongly depends on locally\nsourced seafood to survive\, yet HABs remain understudied and largely\nunmonitored in the region. \n\nWe d evelop a statistical approach drawing from the framework of\ngeneralized a dditive mixed models\, incorporating high-resolution\nhealthcare data from the Madagascar Ministry of Public Health\, to\nexplore the complex associ ation between algal blooms and marine\nfood-related illnesses at the local level. Our findings reveal\ndistinct HAB distribution patterns\, highligh ting high-risk areas and\nseasons. We further demonstrate the link between satellite-detected\nblooms and intoxications\, showcasing remote sensing' s potential for\npublic health applications in settings where resources on the ground\nare limited. \n\n— \n\nGiacomo De Nicola is a postdoctora l fellow at the Harvard T.H. Chan\nSchool of Public Health\, under the joi nt mentorship of Christopher\nGolden and Francesca Dominici. His research broadly seeks to design\,\nimplement and leverage modern statistical tools to address real-world\nproblems\, with a focus on applications in public and planetary health.\nHis current postdoc research is part of the Climate -Smart Public\nHealth project\, where he aims at understanding and measuri ng the\nimpact of climate on public health outcomes in low-income settings . \n\nGiacomo holds a PhD in Statistics from LMU Munich\, an MSc in Econo mic\nand Social Sciences from Bocconi University\, and a BSc in Statistics \nfrom the University of Florence\, where he received the best student\naw ard for graduating top of his class. His research on assessing\nexcess mor tality during crises earned him a special award from the\nFederal Statisti cal Office of Germany. \n\nREGISTER  →  A confirmation with informati on for joining the\nmeeting provided upon registration. \n\nAI-SDM Video Library\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9180fa2 DTSTART;TZID=America/New_York:20250509T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250509T143000 URL:https://www.cs.cmu.edu/events/commencement-2025 LOCATION:Soldiers and Sailors Memorial Hall\, 4141 Fifth Avenue - Oakland SUMMARY:SCS Commencement: SCS Master's Diploma Ceremony (S25) CLASS:PUBLIC DESCRIPTION:Speaker: SCS Commencement: SCS Master's Diploma CeremonyRSVP Re quired\nfor all Graduates.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918124c DTSTART;TZID=America/New_York:20250507T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250507T140000 LOCATION:Wean Hall 4625 SUMMARY:Guest Seminar - Joseph (Yossi) Keshet CLASS:PUBLIC DESCRIPTION:Speaker: JOSEPH (YOSSI) KESHET\, Associate Professor\, Andrew a nd Erna\nViterbi Faculty of Electrical and Computer Engineering\, and Dire ctor\,\nSpeech\, Language\, and Deep Learning Lab\, The Technion\n\nTalk T itle: From Spectrum to Raw Speech: Theoretical and Practical\nAdvances in Diffusion-Based Generation\n\nIn this talk\, I will present two complement ary contributions that push\nthe boundaries of diffusion models from both theoretical and practical\nangles. First\, I will introduce a novel spectr al analysis framework\nthat interprets the inference process of diffusion models through a\nfrequency-domain lens. This allows for a principled desi gn of noise\nschedules tailored to the data’s spectral properties\, repl acing\nheuristic approaches with theoretically grounded strategies. I will \nthen present DiffAR\, an autoregressive diffusion model capable of\ngene rating high-fidelity raw speech waveforms end-to-end. By operating\ndirect ly in the waveform domain and conditioning on overlapping\nframes\, DiffAR achieves coherent\, expressive\, and naturally varied\nspeech generation.   \n\n— \n\nJoseph (Yossi) Keshet received his B.Sc. and M.Sc. degree s in\nElectrical Engineering from Tel Aviv University in 1994 and 2002\,\n respectively. He completed his Ph.D. in Computer Science in 2008 at\nthe S chool of Computer Engineering\, The Hebrew University of\nJerusalem. From 2008 to 2009\, he was a postdoctoral researcher at EPFL\nand the IDIAP Res earch Institute in Switzerland. He then served as a\nResearch Assistant Pr ofessor at TTIC from 2009 to 2012. Between 2013\nand 2022\, he was an Asso ciate Professor in the Department of Computer\nScience at Bar-Ilan Univers ity. Since 2022\, he has been an Associate\nProfessor at the Faculty of El ectrical and Computer Engineering at the\nTechnion. His research interests include speech recognition\, speech\nsynthesis\, and speech analysis.   \n\nMore on the speaker. \n\nFaculty Host:  Bhiksha Ramakrishnan\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9181670 DTSTART;TZID=America/New_York:20250507T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250507T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20250507.html LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Ta-Wei Tu CLASS:PUBLIC DESCRIPTION:Speaker: TA-WEI TU\, Ph.D. Student\, Theory Group\, Computer Sc ience\nDepartment\, Stanford University\n\nTalk Title: Combinatorial Maxim um Flow via Weighted Push-Relabel on\nShortcut Graphs\n\nIn this talk\, I will present a new combinatorial algorithm for maximum\nflow that is based on running the weighted push-relabel algorithm\nintroduced in [BBST'24] o n \"shortcut\" graphs. The use of shortcuts not\nonly improves the running time from n{2+o(1) to Õ(n2) (i.e.\,\nnear-linear in dense graphs)\, but\ , more importantly\, also greatly\nsimplifies both the algorithm and analy sis. Our algorithm is\nrandomized but only because of the use of standard randomized\ncut-matching game. Therefore\, using existing alternatives\, w e also\ngive a deterministic Õ(n2) time algorithm for \"vertex-capacitate d\"\nmax-flow. This is the first near-linear time such algorithm in any\nd ensity regime. \n\nBased on joint work with Aaron Bernstein\, Joakim Blik stad\, Jason Li\,\nand Thatchaphol Saranurak.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91819b8 DTSTART;TZID=America/New_York:20250507T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250507T090000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Doctoral Thesis Proposal - Mingkuan Xu CLASS:PUBLIC DESCRIPTION:Speaker: MINGKUAN XU\, Ph.D. Student\, Computer Science Departm ent \,\nCarnegie Mellon University\n\nTalk Title: Optimization and Simulat ion of Quantum Circuits\n\nOptimizing quantum circuits and simulating them at scale remain\ncritical bottlenecks: manual design of quantum circuit o ptimizations\nis labor-intensive and device-specific\, while simulators st ruggle with\nexponential resource costs. This thesis delivers tools to tac kle these\nchallenges. \n\nFirst\, I introduce Quartz\, a superoptimizer that automates the\ngeneration and verification of circuit transformations for arbitrary\nquantum gate sets. By systematically exploring small circu its and\nemploying an automated theorem prover (Z3)\, Quartz discovers bot h\nexpert-designed and novel optimizations\, outperforming hand-tuned\nopt imizers across various gate sets. \n\nNext\, I present Atlas\, a distribu ted GPU-based simulator that\nhierarchically partitions circuits to exploi t available data\nparallelism while minimizing communication costs\, runni ng over 2×\nfaster than state-of-the-art GPU simulators. Atlas minimizes\ ncommunication overhead via integer linear programming to allocate\n\"near by\" gates to \"nearby\" GPUs and maximizes throughput through\ndynamic pr ogramming for kernel scheduling. \n\nFinally\, I propose an initial forma l verification framework to certify\neach application of transformation-ba sed optimizers like Quartz\,\npaving the way for full correctness guarante es. Together\, these\ncontributions advance automated\, scalable\, and rel iable quantum\ncomputing workflows for emerging devices. \n\nThesis Commi ttee\n\nZhihao Jia (Co-chair)\n\nUmut A. Acar (Co-chair)\n\nRyan O'Donnell \n\nYongshan Ding (Yale University)\n\n \n\nAdditional Information\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9181dfe DTSTART;TZID=America/New_York:20250505T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250505T163000 URL:https://www.ri.cmu.edu/event/video-intelligence-in-the-era-of-multimoda l/ LOCATION:Newell-Simon Hall 3305 SUMMARY:VASC Seminar - Mike Shou CLASS:PUBLIC DESCRIPTION:Speaker: MIKE SHOU\, Assistant Professor / Presidential Young\n Professorship\, Department of Electrical & Computer Engineering\,\nNationa l University of Singapore\n\nTalk Title: Video intelligence in the era of multimodal\n\nThe past few years have witnessed great success in video int elligence\,\nas supercharged by multimodal models. In this talk\, I will s tart with\na brief sharing of our efforts\, in building video-language mod els for\nunderstanding and diffusion models for video generation. Yet\, vi deo\nunderstanding and generation have always been two separate research\n pillars\, despite their strong synergy. This motivates us to develop\nShow -o\, one unified single transformer that can do both multimodal\nunderstan ding and generation. Show-o is the first to unify\nautoregressive and disc rete diffusion modeling\, flexibly supporting a\nwide range of vision-lang uage tasks of any input/output format\,\nincluding visual question-answeri ng\, text-to-image/video generation\,\nand generation of video keyframes w ith captions\, all within one single\n1.3B transformer. Show-o sheds light for building the next-generation\nmultimodal video models. \n\n— \n\n Mike Shou is an Assistant Professor under Presidential Young\nProfessorshi p at National University of Singapore. He was a Research\nScientist at Fac ebook AI in the Bay Area. He obtained his Ph.D. degree\nat Columbia Univer sity with Prof Shih-Fu Chang. His research mainly\nfocuses on video and mu ltimodal. He received the Best Paper Finalist\nat CVPR 2022\, Best Student Paper Nomination at CVPR 2017\, EgoVis\nDistinguished Paper Award 2022/23 . His team won 1st place in the\ninternational challenges including Activi tyNet\, EPIC-Kitchens\, Ego4D.\nHe is a ST Engineering Distinguished Profe ssor and a Fellow of\nNational Research Foundation Singapore. He is on the Forbes 30 Under\n30 Asia list.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91821fd DTSTART;TZID=America/New_York:20250501T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250501T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Crypto Seminar - Isaac Hair CLASS:PUBLIC DESCRIPTION:Speaker: ISAAC HAIR\, Undergraduate Computing Student\, College of\nCreative Studies\, University of California\, Santa Barbara\n\nTalk T itle: Using the Planted Clique Conjecture for Cryptography\n\nWe give a pu blic key encryption scheme that is provably secure against\npoly-size adve rsaries\, assuming n^{(log n)^c} hardness of the standard\nplanted clique conjecture\, for any constant c > 0\, and a relatively\nmild hardness conj ecture about noisy k-LIN over expanders that is not\nknown to imply public -key encryption on its own. Both of our\nconjectures correspond to natural average-case variants of NP-complete\nproblems and have been studied for multiple decades\, with\nunconditional lower bounds supporting them in a v ariety of restricted\nmodels of computation. Our encryption scheme answers an open question\nin a seminal work by Applebaum\, Barak\, and Wigderson [STOC'10]. \n\nThis is a joint work with Riddhi Ghosal (UCLA)\, Aayush Ja in (CMU)\, and\nAmit Sahai (UCLA). \n\nIn Person and Zoom Participation.   See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9182551 DTSTART;TZID=America/New_York:20250501T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250501T160000 URL:https://aco.math.cmu.edu/abs-24-25/may1.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Seth Pettie CLASS:PUBLIC DESCRIPTION:Speaker: SETH PETTIE\, Professor of Computer Science and Engine ering\,\nDepartment of Electrical Engineering and Computer Science\, Unive rsity\nof Michigan\n\nTalk Title: Forbidden 0-1 Patterns and the Pach-Tard os Conjecture\n\nThis talk will survey the extremal theory of pattern-avoi ding 0-1\nmatrices\, and some of their applications in geometry\, combinat orics\,\nand algorithms. If P is a 0-1 matrix\, Ex(P\,n) is the maximum nu mber of\n1s in an n x n 0-1 matrix that does not contain any submatrix tha t\ndominates P. Every 0-1 pattern P can be regarded as the incidence\nmatr ix of a bipartite graph\, in which the two sides of the bipartition\nare o rdered. Thus\, this definition can be seen as a generalization of\nthe Tur an extremal function (for subgraph avoidance). Pattern-avoiding\n0-1 matri ces have been studied since the late 1980s\, and yet the\nprecise relation ship between 0-1 matrices and Turan theory is still\npoorly understood. Fo r many years the foremost open problem has been\nto characterize the extre mal functions of acyclic patterns (those\nwhose graphs correspond to fores ts). In 2005 Pach and Tardos\nconjectured that Ex(P\,n) = O(n polylog(n))\ , for any acyclic P. We give\na simple refutation of the Pach-Tardos conje cture by giving a class of\nacyclic patterns for which Ex(P\,n) > n 2^{sqr t{log n}}. \n\n4:00 pm → Jane Street-sponsored Tea and Cookies in the M ath Lounge \nPlease bring own mug when possible.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91828e3 DTSTART;TZID=America/New_York:20250501T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250501T150000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Justin Raizes CLASS:PUBLIC DESCRIPTION:Speaker: JUSTIN RAIZES\, Ph.D. Candidate\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Quantum Approaches to Verifiable Deletion\n\nIntriguingly\, the laws of physics allow protocols executed by quantum\ncomputers to realize security guarantees that are im possible for\nclassical computers. One of these new possibilities is the a bility to\ntemporarily send ciphertexts to a user\, then later verify that the\nencoded message has been destroyed in an information-theoretic sense .\nBroadbent and Islam (TQC) introduced this notion\, called certified\nde letion\, in the context of encryption. However\, there are many more\nscen arios in which verifiable deletion is useful. For example\, a\nsoftware re ntal company might want to temporarily send their software\nto a user\, th en request that the user destroys the copy at the end of\nthe rental perio d. In this thesis\, we expand the realm of certified\ndeletion to cryptogr aphic primitives beyond just encryption. We define\nand construct the foll owing objects with certified deletion:\n\nObfuscation with Certified Delet ion allows a company to lend a program\nto a user\, allowing them to evalu ate it as they wish. Then\, when the\nuser no longer wish to rent the prog ram\, they can destroy it and prove\nto the issuer that they are no longer able to evaluate the program.\nObfuscation with certified deletion also e nables several new\napplications such as certifiably deletable secret keys .Secret Sharing\nwith Certified Deletion allows a user to distribute share s of a secret\nto several parties. In the event of a data breach\, the use r can\nrequest that the affected party deletes their shares\, rendering th em\nuseless for stealing the secret.Signatures with Certified Deniability\ nallow a signer to endorse a statement in a single message. After the\nrec eiver has verified the signature\, they can destroy it and prove to\nthe s igner that they are no longer able to provide convincing evidence\n- of an y kind - that the signer endorsed this statement. We also show\nhow to con struct the related primitive of NIZKs with certified\ndeniability.Certifie d deniability is a new\, more comprehensive\nparadigm for certified deleti on that rules out additional attacks not\nexplicitly considered by prior d efinitions.\n\nTo build these primitives\, we develop new techniques for v erifying the\ndeletion of information while still allowing access to the i nformation\nunder appropriate conditions.\n\nThesis Committee\n\nVipul Goy al (Chair)\n\nAayush Jain\n\nElaine Shi\n\nGiulio Malavolta (Bocconi Unive rsity)\n\n \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9182dec DTSTART;TZID=America/New_York:20250501T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250501T130000 URL:https://www.ri.cmu.edu/event/when-spatial-computing-meets-accelerated-c omputing/ LOCATION:Newell-Simon Hall 3305 SUMMARY:VASC Seminar - David Chu CLASS:PUBLIC DESCRIPTION:Speaker: DAVID CHU\, Vice President of Spatial Computing and XR \, NVIDIA\n\nTalk Title: When Spatial Computing meets Accelerated Computin g\n\nNVIDIA has been pioneering Accelerated Computing for the past three\n decades\, driving innovations that have transformed society. Among all\npe rsonal computing mediums\, Spatial Computing and Extended Reality\n(XR) st and out as some of the most promising beneficiaries of\naccelerated comput ing. \n\nIn this talk\, we will explore the latest developments and trend s in\nthe XR ecosystem\, highlighting a range of form factors: from augmen ted\nreality on handheld devices and AI glasses\, to fully immersive mixed \nreality experiences in head-mounted displays. We will also delve into\ns ome of the most compelling immersive use-cases. Additionally\, we will\ndi scuss NVIDIA’s contributions at the intersection of XR and AI\,\nillustr ating how AI is being leveraged to mold and enhance XR\nexperiences. By sy nthesizing spatial computing and accelerated\ncomputing\, their dynamic in teraction will shape the future of\ncomputing and society. \n\n— \n\nD avid Chu is VP of Spatial Computing and XR at NVIDIA\, where he is\nbringi ng accelerated computing to spatial computing. Formerly\, he was\nVP of En gineering at Magic Leap\, where he led teams on perception\,\ndeep learnin g and immersion for AR. Prior to that\, David was at Google\nwhere he work ed on AR and VR\, as well as cloud services\, edge\ncomputing and cloud ga ming. Before that\, David was faculty at the\nUniversity of Illinois\, Urb ana-Champaign (UIUC)\, and a staff member at\nMicrosoft Research. David’ s work has received awards such as Best of\nCES\, Best Papers in MobiSys\, and Best Demos in MobiSys and SenSys. He\nhas served as PC Chair for ISMA R and NetGames\, and General Chair of\nHotMobile. His individual work has appeared in such places as CNBC\,\nFast Company\, VentureBeat\, TechCrunch \, PC Magazine\, GameSpot\, Ars\nTechnica\, Slashdot\, The Verge\, Engadge t\, Yahoo and Wired. \n\nFaculty Host:  Srinivasan Seshan\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9183224 DTSTART;TZID=America/New_York:20250501T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250501T130000 LOCATION:McWilliams Classroom\, Gates Hillman 4303 SUMMARY:Doctoral Speaking Skills Talk - Bailey Miller CLASS:PUBLIC DESCRIPTION:Speaker: BAILEY MILLER\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Monte Carlo PDE Simulat ion in Participating Media\n\nThroughout science and engineering\, Monte C arlo methods like walk on\nspheres simulate physical phenomena like therma l conduction on\ncomplex\, large-scale geometric models. Despite their ver satility\,\nthese methods cannot easily handle domains involving participa ting\nmedia that consists of massive quantities of microparticle\ngeometry —such as water droplets in clouds or ions in solution around\nbiological membranes. \n\nIn this talk\, we explore how to generalize Monte Carlo P DE simulation\nto participating media by considering stochastic models of\ nmicroparticle geometry\, which define particles only through aggregate\ns tatistical properties (e.g.\, particle density). Our approach to\nvolumetr ic simulation provides an unbiased and output sensitive means\nof solving PDEs in participating media. We demonstrate how volumetric\nsimulation rel ates to classic volume rendering algorithms and how our\nalgorithm can be leveraged for problems like computing electrostatic\npotential around biol ogical membranes or analyzing atmospheric\nphotochemical systems. \n\nPre sented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91835cf DTSTART;TZID=America/New_York:20250501T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250501T110000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:MSCS Thesis Defense - Charlie Ruan CLASS:PUBLIC DESCRIPTION:Speaker: CHARLIE RUAN\, Master's Student\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Democratizing On-Devi ce LLM Inference with Machine\nLearning Compilers and Web Technologies\n\n Large language models (LLMs) have traditionally relied on cloud-based\ninf erence due to their high computational and memory demands. However\,\nrece nt advances in small LLMs and consumer hardware capabilities have\nmade on -device inference increasingly practical. Among potential\ndeployment targ ets\, the web browser stands out as a uniquely\ncompelling platform: it is universally accessible\, naturally abstracts\nout hardware heterogeneity\ , requires no dependency installation for\nweb applications\, and provides a natural agentic environment for task\nautomation. \n\nWebLLM is a high -performance TypeScript framework that enables LLM\ninference entirely wit hin web browsers. WebLLM compiles LLMs ahead of\ntime using the MLC-LLM an d Apache TVM compiler stack to generate\noptimized WebGPU kernels and a po rtable WebAssembly runtime. The\nsystem exposes a familiar OpenAI-style AP I\, supports efficient GPU\nacceleration\, and integrates seamlessly with browser environments\nusing Web Workers and WebAssembly. To enable structu red generation\,\nwhich is especially challenging for small LLMs\, WebLLM incorporates\nXGrammar\, an efficient grammar-constrained decoding engine\ , allowing\ndevelopers to enforce output formats such as JSON or DSLs with \nnear-zero overhead. Together\, these components demonstrate a path\ntowa rd democratizing LLM access\, making intelligent\, private\, and\nresponsi ve AI experiences universally available through the web. \n\nThesis Commi ttee\n\nTianqi Chen (Chair)\n\nZhihao Jia\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91839ee DTSTART;TZID=America/New_York:20250430T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250430T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20250430.html LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Seth Pettie CLASS:PUBLIC DESCRIPTION:Speaker: SETH PETTIE\, Professor of Computer Science and Engine ering\,\nDepartment of Electrical Engineering and Computer Science\, Unive rsity\nof Michigan\n\nTalk Title: Everything you always wanted to know abo ut Cardinality\nEstimation* (*but were afraid to ask)\n\nThe Cardinality E stimation/Distinct Elements problem is to approximate\nthe number of disti nct elements in a data stream using a small\nprobabilistic data structure called a \"sketch\".  This problem has\nbeen studied for 40 years\, has m any industrial applications\, and is\nfeatured prominently in most courses on Big Data algorithmics.  It is\ntherefore a real puzzle to explain why research on this popular and\nfundamental problem has been unusually slow . This talk presents a\ncomplete history of the Cardinality Estimation pro blem from Flajolet\nand Martin's seminal 1983 paper to the present\, and i ncludes an\naccount of how the research community became fractured\, delay ing many\nnatural developments by decades.  I will present our recent eff orts\nto achieve information-theoretically optimal cardinality sketches\,\ nwhich draws on two notions of \"information\" developed in the 20th\ncent ury: Fisher information (governing optimal point estimation) and\nShannon entropy (governing optimal space/communication). \n\nJoint work with Ding yu Wang.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9183dca DTSTART;TZID=America/New_York:20250430T092000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250430T165000 LOCATION:Gates Hillman 4301 SUMMARY:Meeting of the Minds: SCS Thesis and Poster Presentations S25 CLASS:PUBLIC DESCRIPTION:Speaker: SCS Senior Thesis PresentersTalk Title: Meeting of the Minds\n\n► Morning Session\n\n9:20 AM      →  Janavi Gupta  |  Work in Progress: Passive\nWearable Control Interfaces in Robotic Caregivi ng\n\n   Mentor:   Zackory Erickson \n\n9:30 AM      →  Diya D inesh  | Work in Progress: \nPhysically-Consistent Motion Generation Us ing Diffusion Models for\nAssistive Biomechanics\n\n   Mentor:   Inseu ng Kang\n\n9:40 AM      → Xinfei Cen | How Are Intelligent Tutori ng\nSystems Related to Student Self-Efficacy\n\n   Mentor:   Paulo Car valho\n\n10:10 AM      → Emily Amspoker | A Transformational Game for\nSupporting Computational Empowerment in Responsible AI\n\n   Mento r:   Jessica Hammer\n\n10:30 AM      → Dongkyun Kim  | Explorin g Intra-Object\nComposition Using Scene Graphs and Diffusion Models\n\n    Mentor:  Jun-Yan Zhu\n\n10:50 AM      → Abigail Defranco | St eering with Confidence:\nUncertainty-Aware Classifier-Based Guidance for D iffusion Policies\n\n   Mentor: Andrea Bajcsy\n\n11:10 AM      →  Arthur Jakobsson  | (DORMA) Deformable Object\nRapid Motor Adaptation\n \n   Mentor: Jeffrey Ichnowski\n\n11:30 AM      → Joshua Kim  |  Caramel: Multi-GPU Framework\nfor Encrypted AI\n\n  Mentor:  Dimitrios Skarlatos \n\n► Early Afternoon Session\n\n12:10 PM      → Liji e Yang | TIDALDECODE: Fast and Accurate\nLLM Decoding with Position Persi stent Sparse Attention\n\n   Mentor:  Zhihao Jia\n\n12:30 PM    →  Jerick Shi | Beyond Classical Models: Simulating\nOpinion Dynamics and P olarization Using Large Language Model Agents\n\n   Mentor:  Vincent Co nitzer\n\n12:50 PM   → Rithika Narayan  | Teaching Robot Policies to \nHumans Using Erroneous Examples   \n\n   Mentor:  Henny Admoni \n\ n1:10 PM      → Maxwell Soh | Causal inference identifies\naverage and individual treatment effects for individual cells in\ncase-control sc RNA-seq\n\n   Mentor:   Martin Zhang \n\n1:30 PM      → Andrew Wang | Deep Generative Design of\nEvolutionarily Conserved Celltype-Spec ific Enhancers\n\n   Mentor:  Andreas Pfenning \n\n1:50 PM      → Jessica Yaxuan Liu  | Identifying the impact\nof Alzheimer’s dise ase-associated genetic variants on molecular\npathways via genetic perturb ation of selected enhancers in conditional\nspecific immune cells\n\n   Mentor: Andreas Pfenning\n\n► Late Afternoon Session\n\n2:30 PM      → Siddharth Parekh | Efficiently Emphasizing\nCorrectness during Quest ion Answering Inference\n\n   Mentor: Carolyn Rosé\n\n2:50 PM      → Taekseung Kim | About range retrieval and\nclustering in nearest neig hbors  | \n\n   Mentor: Guy Blelloch\n\n3:10 PM      → Shirley Yu | Diversifying to Verify: Improving\nProgram Verification with Divers e Equivalent Code\n\n   Mentor: Ruben Martins\n\n3:30 PM      → A lly Du | Reconciling model multiplicity for\nimproved conformal predicti on set\n\n   Mentor: Steven Wu\n\n3:50 PM      → Trey Debose | A daptive Lagrangian Fluids Using\nSPH and MLS Techniques\n\n   Mentor:  Minchen Li\n\n4:10 PM      → Kevin You | An Energy-Controllable T ime\nIntegrator for Elastodynamic Contact\n\n   Mentor: Minchen Li\n\n4: 30 PM      → Sophia Qingyang Cao | Possum: A Tail of Dynamic\nFlas h Capacity for Sustainability\n\n   Mentor:  Greg Ganger\n\nMeeting of the Minds - Complete Schedule\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9184388 DTSTART;TZID=America/New_York:20250429T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250429T170000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year Master's Thesis Presentation - Grace Park CLASS:PUBLIC DESCRIPTION:Speaker: GRACE PARK\, Master's Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Action Diversity for Re liable Policy Learning: Assessing\nTreatment Variation in Healthcare\n\nRe cent advances in machine learning for personalized medicine have\ncreated a need to determine when observational healthcare data can\nreliably infor m treatment policies. This thesis examines \"action\ndiversity\" as a crit ical factor for evaluating whether treatment\nvariation in medical dataset s is sufficient for developing dependable\nclinical policies. Through thre e complementary approaches\, we\ninvestigate methods to detect and measure meaningful action diversity\nin healthcare data.\n\nFirst\, we analyze th e MIMIC sepsis dataset using transformer-based\ndynamics models. Our findi ngs reveal that including action information\nprovides minimal improvement in outcome predictions across the entire\ndataset. This suggests limited meaningful treatment diversity when\nanalyzed in aggregate. Second\, in ou r controlled simulation\nexperiments with a one-dimensional GridWorld envi ronment\, we\ndemonstrate that comparing prediction performance between mo dels with\nand without action inputs effectively identifies regions where\ ntreatments meaningfully impact outcomes. Finally\, we present a novel\nin teractive visualization tool that employs t-SNE dimensionality\nreduction and intuitive diversity metrics to help researchers explore\naction divers ity across patient state spaces. This tool helps identify\nsubgroups where treatment policies can be reliably learned. \n\nOur findings demonstrate that dynamics model comparisons can\neffectively identify regions where t reatment policies can be reliably\nlearned\, enabling more targeted and tr ustworthy deployment of machine\nlearning in healthcare. This framework pr ovides researchers with\npractical tools to evaluate data sufficiency befo re deploying\ntreatment recommendation systems\, potentially improving bot h the\nreliability of AI assistance in clinical decision-making and\,\nult imately\, patient outcomes.\n\nThesis Committee\n\nAdam Perer (Chair)\n\nZ achory Erickson\n\nAdditional Information  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91847ef DTSTART;TZID=America/New_York:20250429T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250429T150000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Mingxun Zhou CLASS:PUBLIC DESCRIPTION:Speaker: MINGXUN ZHOU\, Ph.D. Candidate\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Practical Private Info rmation Retrieval and Searching with\nSublinear Cost\n\nIn this thesis\, w e investigate Private Information Retrieval (PIR)\, a\ncryptographic proto col that enables clients to access information from\na database without re vealing their queries to the server. As a\nfundamental building block for privacy-preserving applications\, PIR\nhas been extensively studied in bot h theory and practice for\ndecades. \n\nHowever\, practical implementatio ns have been limited to small-scale\nuse cases due to the linear computati on barrier of PIR\, which requires\nthe server to process the entire datab ase for each query. The seminal\nworks of Beimel\, Ishai\, and Malkin (Cry pto 2000) and Corrigan-Gibbs\nand Kogan (Eurocrypt 2022) introduced Prepro cessing PIR to overcome\nthis barrier. While theoretically efficient\, pre vious\nconstructions remained impractical due to their reliance on expens ive\ncryptographic operations. \n\nTo address this limitation\, we propos e two new PIR schemes: Piano and\nQuarter-PIR. Both achieve sublinear serv er computation and\ncommunication while remaining efficient in practice. T hese\nconstructions transform the practical PIR landscape by providing nea r\nreal-time responses for databases with billions of entries\,\nwhile ma intaining reasonable communication and storage\nrequirements. \n\nFurther more\, we demonstrate the practical utility of our PIR schemes\nthrough an important application – private information searching. We\ndevelop Pacm ann\, a new private approximate nearest neighbor search\nalgorithm that de livers both high search quality and fast response\ntimes for databases wit h hundreds of millions of records. \n\nOur work makes a significant step toward bridging the gap between\ntheory and practice in PIR research. Thes e contributions not only\nadvance the state of the art in PIR designs\, bu t also open new avenues\nfor developing privacy-preserving applications in real-world and\nlarge-scale settings.\n\nThesis Committee\n\nElaine Shi ( Co-Chair)\n\nGiulia Fanti (Co-Chair)\n\nBryan Parno\n\nDavid J. Wu (Univer sity of Texas at Austin)\n\nIn Person and Zoom Participation.  See announ cement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9184cb0 DTSTART;TZID=America/New_York:20250429T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250429T130000 URL:https://www.cs.cmu.edu/~aiseminar/ LOCATION:Newell-Simon 3305 SUMMARY:Joint AI Seminar / Doctoral Speaking Skills Talk - Victor Akinwande CLASS:PUBLIC DESCRIPTION:Speaker: VICTOR AKINWANDE\, Ph.D. Student\, Computer Science De partment\,\nCarnegie Mellon University\n\nTalk Title: Adapting Vision-Lang uage models with Hypernetworks000\n\nSelf-supervised vision-language model s trained with contrastive\nobjectives form the basis of current state-of- the-art methods in AI\nvision tasks. The success of these models is a dire ct consequence of\nthe huge web-scale datasets used to train them\, but th ey require\ncorrespondingly large vision components to properly learn powe rful and\ngeneral representations from such a broad data domain. This pose s a\nchallenge for deploying large vision-language models\, especially in\ nresource-constrained environments. \n\nThis talk presents an alternate v ision-language architecture\, called\nHyperCLIP\, that uses a small image encoder along with a hypernetwork\nthat dynamically adapts image encoder w eights to each new set of text\ninputs. \n\nWith a trained HyperCLIP mode l\, we can generate new zero-shot\ndeployment-friendly image classifiers f or any task with a single\nforward pass through the text encoder and hyper network. HyperCLIP\nincreases the zero-shot accuracy of SigLIP trained mod els with small\nimage encoders by up to 3% on ImageNet and 5% on CIFAR-100 with\nminimal training throughput overhead. \n\nPresented in Partial Ful fillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918508c DTSTART;TZID=America/New_York:20250429T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250429T113000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year Master's Thesis Presentation - Ethan Mackey CLASS:PUBLIC DESCRIPTION:Speaker: ETHAN MACKEY\, Master's Student\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Pinwheels and Polygon s: Symmetric Realizations of\nPolygon-Free Point Placements via SAT\n\nThi s research explores the discovery of symmetrical realizations of\nconvex-h exagon-free point placements on 16 points using satisfiability\nsolving te chniques. The central focus is on identifying point\nconfigurations that e xhibit 3-\, 4-\, and 5-fold rotational symmetry.\nThese 16 point configura tions correspond to the maximal number of\npoints that can be placed in th e Euclidean plane in general position\nwithout forming any convex hexagons \, a central case in the study of\nthe Erdos–Szekeres Conjecture\, a fou ndational problem in\ncombinatorial geometry. \n\nBuilding on previous wo rk in combinatorial geometry and SAT-based\ncombinatorial methods\, this r esearch extends existing Boolean\nsatisfiability encodings by incorporatin g symmetry constraints and\nstructural conditions specific to the hexagon- free problem. Using\nthese ideas\, new conjunctive normal form formulas ar e developed to\nrepresent the search space of symmetric hexagon-free point \nplacements. \n\nTo interpret and visualize solutions\, satisfying assig nments to these\nCNFs are passed through a point realization tool that rec onstructs\ngeometric configurations from orientation triple data. This ena bles\nthe conversion of logical encodings into concrete point placements\n that can be analyzed and compared. Structural analysis of these\nplacement s includes examining the frequency and distribution of\nsmaller convex pol ygons\, such as 4-gons and 5-gons\, to better\nunderstand the local geomet ric implications of hexagon avoidance.\n\nThe resulting symmetric configur ations\, especially those with\nfour-fold and five-fold symmetry\, represe nt some of the first\nstructured\, realizable examples of 16-point hexagon -free sets. These\nfindings contribute new insight into the Erdos–Szeker es Conjecture\nand offer a stepping stone toward understanding larger\ngen eralizations\, such as the existence of 32-point configurations that\navoi d convex 7-gons. \n\nThesis Committee\n\nMarijn Heule (Chair)\n\nRuben Ma rtins\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91854fd DTSTART;TZID=America/New_York:20250428T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250428T163000 LOCATION:Newell-Simon 4305 SUMMARY:VASC Seminar - Hong-Xing “Koven” Yu CLASS:PUBLIC DESCRIPTION:Speaker: HONG-XING \"KOVEN\" YU\, Ph.D. Candidate\, Computer Sc ience\nDepartment\, Stanford University\n\nTalk Title: Generating a Physic al World\n\nGenerating an interactive\, enlivened\, and physical world ena bles a\nwide range of applications in entertainment\, embodied AI\, educat ion\,\nand creative designs. Recent image/video models have shown promise in\nproducing realistic visuals\, yet they operate purely at the pixel\nle vel and lack underlying physical grounding\, leading to failures in\nphysi cal fidelity and user interactivity. In this talk\, I’ll\nintroduce our recent efforts in physical world generation by grounding\npixel models ont o physical models. This methodology inherently\nincorporates physical worl d knowledge about 3D spatial structures and\ndynamics\, simultaneously acq uiring visual realism\, physical fidelity\,\nand user interactivity. I’l l showcase how this methodology is\napplied to enable fast generation of d iverse worlds\, with which users\ncan interact via 3D actions. \n\n— \ n\nHong-Xing “Koven” Yu is a 5th-year PhD candidate at the Computer\nS cience Department of Stanford University\, advised by Prof. Jiajun Wu.\nHi s research interest centers around how AI can understand and\ngenerate the physical world. He is a recipient of the SIGGRAPH Asia\nBest Paper Award\ , the Stanford SoE Fellowship\, the Qualcomm Innovation\nFellowship\, and the Meshy Fellowship\, and a finalist of the NVIDIA\nFellowship\, the Meta Fellowship\, the Jane Street Fellowship\, and the\nRoblox Fellowship.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91858ae DTSTART;TZID=America/New_York:20250428T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250428T170000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Doctoral Thesis Oral - Yi Zhou CLASS:PUBLIC DESCRIPTION:Speaker: YI ZHOU\, Ph.D. Candidate\, Computer Science Departmen t\,\nCarnegie Mellon University\n\nTalk Title: Towards Scalable Automated Program Verification for System\nSoftware\n\nAutomated Program Verificatio n (APV) provides formal guarantees about\nsoftware while promising strong automation in the verification\nprocess. APV has already seen preliminary successes in system software\n(e.g.\, file systems\, network protocols)\, extending beyond academic\nprototypes to industrial applications. However\ , the scalability of APV\nbecomes an issue as we move towards more complex systems\, where\nautomation failures start to show up. Such failures ofte n require\ntedious manual fixes\, breaking the pledge of automation in APV . Worse\nyet\, since program verification is fundamentally undecidable\,\n automation failures are inherently inevitable. \n\nNevertheless\, that do es not mean APV is hopeless beyond small-scale\nsystems. In this thesis\, we organize the discussion around the\ndevelopment stages of APV: (1) crea ting proofs\, (2) reusing proofs\,\n(3) debugging proofs\, and (4) stabili zing proofs. We argue that\,\ndespite the undecidable nature of program ve rification in theory\, we\ncan overcome the scalability challenges that ar ise in practice\, due to\nthe recurrent patterns in APV programming and re asoning. \n\nSpecifically\, we make empirical observations on the common motifs in\nAPV\, and then design formal methods to leverage them for autom ation.\nUsing large-scale verified systems as case studies\, we show this\ ncombination of formal and empirical methods leads to practical\nimprovem ents in APV for system software.   \n\nThesis Committee \n\nBryan Parn o (Chair)\n\nMarijn Heule\n\nRuben Martins\n\nJon Howell (VMware Research / University of Washington)\n\nIn Person and Zoom Participation.  See ann ouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9185cf5 DTSTART;TZID=America/New_York:20250428T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250428T140000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:5th Year Master's Thesis Presentation - Tianwei Owen Li CLASS:PUBLIC DESCRIPTION:Speaker: TIANWEI OWEN LI\, Master's Student\, Computer Science\ nDepartment\, Carnegie Mellon University\n\nTalk Title: Two Simple Algorit hmic Applications of Convex Optimization\n\nMethods and concepts from conv ex optimization have close relations to\ntraditionally discrete algorithms \, and we investigate two separate\ncases of such. In part I\, we leverage a variant of the Mirror-Prox\nalgorithm from Sherman’s 2017 paper on ar ea-convexity and\nmulticommodity flow\, to design a fast Õ(m/ε) algorith m for ε-fair\ncuts\, a special type of approximate st-min-cut that requir es some\nst-flow to 1 − ε saturate all edges across the cut. Such runti me is\nan improvement over the state-of-the-art Õ(m/ε3)\, and the result ing\nalgorithm is much simpler.  In part II\, we design a continuous\nvar iant of the Graham Scan convex hull algorithm that computes the\ntight con vex envelope of degree-n univariate polynomials in O(n3 + n\nlog2 n log b + nb2) with respect to an interval domain\, and updates in\nO(b2) with res pect to a new interval domain\, where 2−b  is the\nrelative precision f or float point arithmetic. Such an algorithm\nrelies on properties of conv ex functions for its proof of correctness\,\nand can be used to construct high quality convex relaxations for\nGeneralized Additive Models (GAM) wit h monotone link functions. \n\nThesis Committee\n\nBarnabás Póczos (Cha ir)\n\nJason Li\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91860c9 DTSTART;TZID=America/New_York:20250428T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250428T123000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Doctoral Speaking Skills Talk - Saranya Vijayakumar CLASS:PUBLIC DESCRIPTION:Speaker: SARANYA VIJAYAKUMAR\, Ph.D. Student\, Computer Science \nDepartment\, Carnegie Mellon University\n\nTalk Title: Grounding Neural Inference with Satisfiability Modulo\nTheories: Leveraging Theory Solvers for Machine Learning\n\nThis presentation describes SMTLayer\, a framework for incorporating\nSatisfiability Modulo Theories (SMT) solvers into deep neural\nnetworks. Unlike previous approaches that approximate solver beha vior\nwith differentiable relaxations\, SMTLayer directly integrates\nnon- differentiable SMT solvers into both forward and backward passes\nof netwo rk training. \n\nSMTLayer enables encoding rich domain knowledge as mathe matical\nformulas within the network architecture. During the forward pass \, the\nsolver uses symbols from prior layers to construct inferences base d on\nthese formulas. In the backward pass\, the solver informs network\nu pdates\, driving it toward representations compatible with the\nsolver's t heory. \n\nThrough experiments on visual arithmetic\, algebraic equation solving\,\nand natural language reasoning tasks\, we demonstrate that mode ls using\nSMTLayer require significantly less training data than conventio nal\nmodels\, show greater robustness to covariate shifts\, and learn\nnat urally interpretable representations.  \n\nPresented in Partial Fulfillm ent of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9186482 DTSTART;TZID=America/New_York:20250428T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250428T110000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Speaking Skills Talk - Kaiyang Zhao CLASS:PUBLIC DESCRIPTION:Speaker: KAIYANG ZHAO\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Contiguitas: The Pursuit of Physical Memory Contiguity in\nDatacenters\n\nThe unabating growth of the memory needs of emerging datacenter\napplications has exacerbated the scalability bottleneck of virtual\nmemory. However\, reducing the excessiv e overhead of address\ntranslation will remain onerous until the physical memory contiguity\npredicament gets resolved. To address this problem\, th is paper\npresents Contiguitas\, a novel redesign of memory management in the\noperating system and hardware that provides ample physical memory\nco ntiguity. We identify that the primary cause of memory fragmentation\nin M eta's datacenters is unmovable allocations scattered across the\naddress s pace that impede large contiguity from being formed. \n\nTo provide ample physical memory contiguity by design\, Contiguitas\nfirst separates regul ar movable allocations from unmovable ones by\nplacing them into two diffe rent continuous regions in physical memory\nand dynamically adjusts the bo undary of the two regions based on\nmemory demand. Drastically reducing un movable allocations is\nchallenging because the majority of unmovable page s cannot be moved\nwith software alone given that access to the page canno t be blocked\nfor a migration to take place. Furthermore\, page migration is\nexpensive as it requires a long downtime to (a) perform TLB shootdowns \nthat scale poorly with the number of victim TLBs\, and (b) copy the\npag e. \n\nTo this end\, Contiguitas eliminates the primary source of unmovab le\nallocations by introducing hardware extensions in the last-level cache \nto enable the transparent and efficient migration of unmovable pages\nev en while the pages remain in use. We build the operating system\ncomponent of Contiguitas into the Linux kernel and run our experiments\nin a produc tion environment at Meta's datacenters. Our results show\nthat Contiguitas 's OS component successfully confines unmovable\nallocations\, drastically reducing unmovable 2MB blocks from an average\nof 31% scattered across th e address space down to 7% confined in the\nunmovable region\, leading to significant performance gains. \n\nSpecifically\, we show that for three major production services\,\nContiguitas achieves end-to-end performance i mprovements of 2-9% for\npartially fragmented servers\, and 7-18% for high ly fragmented servers\,\nwhich account for nearly a quarter of Meta's flee t. We further use\nfull-system simulations to demonstrate the effectivenes s of the\nhardware extensions of Contiguitas. Our evaluation shows that\nC ontiguitas-HW enables the efficient migration of unmovable\nallocations\, scales well with the number of victim TLBs\, and does not\naffect applicat ion performance. \n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918692c DTSTART;TZID=America/New_York:20250425T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250425T130000 LOCATION:Gates Hillman 7101 (Special Time) SUMMARY:Theory Seminar - Tselil Schramm CLASS:PUBLIC DESCRIPTION:Speaker: TSELIL SCHRAMM\, Assistant Professor\, Department of\n Statistics\, Stanford University\n\nTalk Title: Some easy optimization pro blems have the overlap-gap\nproperty\n\nAn optimization problem is said to have the \"overlap-gap property\"\n(OGP) if the near-optimal solutions ar e partitioned into several\nwell-separated clusters. In statistical physic s\, the overlap gap\nproperty is associated with computational intractabil ity for\noptimization. Further\, variants of the OGP imply unconditional l ower\nbounds against local and/or Lipschitz algorithms. In recent years\, the\nOGP has been accepted by some as a good heuristic for predicting\ncom putational intractability\, even beyond these specific unconditional\nlowe r bounds. \n\nIn this talk\, I'll demonstrate that the shortest path prob lem in\nsparse random graphs has the OGP. Because shortest path is\ncomput ationally easy\, this complicates the picture for the overlap-gap\npropert y. \n\nBased on joint work with Shuangping Li.\n\nFaculty Host: Ryan O'Do nnell\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9186c93 DTSTART;TZID=America/New_York:20250425T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250425T130000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Joint AI-SDM Center / S3D Seminar - Tom Griffiths CLASS:PUBLIC DESCRIPTION:Speaker: TOM GRIFFITHS\, Henry R. Luce Professor of Information \nTechnology\, Consciousness and Culture\, Departments of Psychology and\n Computer Science\, Director\, Computational Cognitive Science Lab\,\nPrinc eton University\n\nTalk Title: Using Machine Learning and Psychology to Pr edict and\nUnderstand Human Decisions\n\nMachine learning methods provide increasingly powerful tools for\ngenerating predictions about human behavi or. However\, simply using off\nthe shelf methods to generate predictions potentially misses\nopportunities to benefit from and contribute to the ps ychological\nliterature. \n\nIn this talk I will discuss three ways in wh ich theory and data can\ninteract through machine learning: using theories to pretrain machine\nlearning models\; using theories to constrain machin e learning models\;\nand using unconstrained machine learning models to cr itique\nexplanatory theories. I will illustrate these cases with examples from\nthe study of human decision-making\, discussing risky choice\, moral \njudgments\, behavioral game theory\, and open-ended decision-making\, an d\nalso highlight some recent work using large language models to predict\ nhuman decisions. \n\n— \n\nTom Griffiths is the Henry R. Luce Profess or of Information\nTechnology\, Consciousness and Culture in the Departmen ts of Psychology\nand Computer Science at Princeton University. His resear ch explores\nconnections between human and machine learning\, using ideas from\nstatistics and artificial intelligence to understand how people solv e\nthe challenging computational problems they encounter in everyday\nlife . He has made contributions to the development of Bayesian models\nof cogn ition\, probabilistic machine learning\, nonparametric Bayesian\nstatistic s\, and models of cultural evolution\, and his recent work has\ndemonstrat ed how methods from cognitive science can shed light on\nmodern artificial intelligence systems. \n\nTom completed his PhD in Psychology at Stanfor d University in 2005\,\nand taught at Brown University and the University of California\,\nBerkeley before moving to Princeton. He has received awar ds for his\nresearch from organizations ranging from the American Psycholo gical\nAssociation to the National Academy of Sciences and is a co-author of\nthe book Algorithms to Live By\, introducing ideas from computer\nscie nce and cognitive science to a general audience. \n\nREGISTER→connectin g information will be provided upon registration.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9187107 DTSTART;TZID=America/New_York:20250424T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250424T160000 LOCATION:Gates Hillman 7101 SUMMARY:Doctoral Speaking Skills Talk - Yonghao Zhuang CLASS:PUBLIC DESCRIPTION:Speaker: YONGHAO ZHUANG\, Ph.D. Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Scaling Long Context T raining Data by Long-Distance\nReferrals\n\nTraining large language models for long context understanding faces\nthe challenge of data shortage. Pre vious data engineering approaches\nmechanically concatenate short document s\, which may create many pseudo\nlong documents but raise concerns about data quality. \n\nIn this paper\, we study the core attribute of high qua lity data for\nlong context training\, and provide a data pipeline\, LongP ack\, to scale\nsuch data. We found that long distance referrals\, which o ccur in\nnatural long documents\, are crucial for long-context training.\n However\, simply concatenating short documents does not reliably\ngenerate these relations. We further show that the density of\nlong-distance refer rals\, which is higher in longer documents\, has a\nkey role in training e fficiency\, making previous upsampling methods\nsuboptimal.  \n\nTo enri ch long documents\, we propose LongPack\, a data pipeline that\nconstructs long documents by packing shorter ones based on referral\nrelationships. Specifically\, for web pages\, which are the primary\nsource for language model training\, we found hyper-link a native\nsignal for such a relation. By packing web pages through their\nhyper-link connection\, we can create longer\, high-quality documents. \n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91874d2 DTSTART;TZID=America/New_York:20250424T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250424T160000 URL:https://aco.math.cmu.edu/abs-24-25/apr24.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Vasu Tewari CLASS:PUBLIC DESCRIPTION:Speaker: VASU TEWARI\, Assistant Professor\, Department of Math ematics\,\nMathematical and Computational Sciences\, University of Toronto \nMississauga\n\nTalk Title: Quasisymmetric divided differences and forest polynomials\n\nPostnikov's divided symmetrization\, introduced in the con text of\nvolume polynomials of permutahedra\, possesses a host of remarkab le\n\"positivity\" properties. These turn out to be best understood using a\nfamily of operators we call quasisymmetric divided differences. I will\ nintroduce these operators and then define a basis of the polynomial\nring adapted to these operators in the same way as ordinary divided\ndifferenc es interact with Schubert polynomials. This basis works\nnicely with respe ct to reduction modulo the ideal of positive degree\nquasisymmetric polyno mials. Furthermore the expansion of the Schubert\npolynomials in this basi s is nonnegative—in fact it encodes the\nSchubert class expansions of th e classes of certain toric Richardson\nvarieties whose moment polytopes co me from a cubical subdivision of\nthe permutahedron. I will give a combina torial procedure to compute\nthese Schubert structure constants. \n\nJoin t work with Philippe Nadeau (Lyon) and Hunter Spink (Toronto). \n\n4:00 p m → Jane Street-sponsored tea and cookies in Wean 6220.\n                    \n\nBring you own mug if possible.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918786b DTSTART;TZID=America/New_York:20250423T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250423T160000 LOCATION:Gates Hillman 7101 SUMMARY:5th Year Master's Thesis Presentation - Tongzhou Thomas Liao CLASS:PUBLIC DESCRIPTION:Speaker: TONGZHOU THOMAS LIAO\, Master's Student\, Computer Sci ence\nDepartment\, Carnegie Mellon University\n\nTalk Title: Enhancing GNN s with Encoding\, Rewiring\, and Attention\n\nGraph Neural Networks (GNNs) have become important tools for machine\nlearning on graph-structured dat a. In this paper\, we explore the\nsynergistic combination of graph encodi ng\, graph rewiring\, and graph\nattention\, by introducing Graph Attentio n with Stochastic Structures\n(GRASS)\, a novel GNN architecture. GRASS ut ilizes relative random walk\nprobabilities (RRWP) encoding and a novel dec omposed variant (D-RRWP)\nto efficiently capture structural information. I t rewires the input\ngraph by superimposing a random regular graph to enha nce long-range\ninformation propagation. It also employs a novel additive attention\nmechanism tailored for graph-structured data. Our empirical\nev aluations demonstrate that GRASS achieves state-of-the-art\nperformance on multiple benchmark datasets\, including a 20.3%\nreduction in mean absolu te error on the ZINC dataset. \n\nThesis Committee\n\nBarnabás Póczos ( Chair)\n\nTianqi Chen\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9187bea DTSTART;TZID=America/New_York:20250423T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250423T143000 LOCATION:Gates Hillman 7101 SUMMARY:5th Year Master's Thesis Presentation - Kefan Cao CLASS:PUBLIC DESCRIPTION:Speaker: KEFAN CAO\, Master's Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Computationally Reconstr ucting the Evolution of Cancer\nRisk Evolution\n\nUnderstanding the evolut ion of cancer in its early stages is critical\nto identifying key drivers of cancer progression and developing better\nearly diagnostics or prophyla ctic treatments. Early cancer is\ndifficult to observe\, though\, since it is generally asymptomatic until\nextensive genetic damage has accumulated . \n\nIn this study\, we develop a computational approach to infer how\no nce-healthy cells enter into and become committed to a pathway of\naggress ive cancer. We accomplish this through a strategy of using\ntumor phylogen etics to look backwards in time to earlier stages of\ntumor development co mbined with machine learning to infer how\nprogression risk changes over t hose stages. We apply this paradigm to\npoint mutation data from a set of cohorts from the Cancer Genome Atlas\n(TCGA) to formulate models of how pr ogression risk evolves from the\nearliest stages of tumor growth\, as well as how this evolution varies\nwithin and between cohorts. \n\nThe result s suggest general mechanisms by which risk develops as a\ncell population commits to aggressive cancer\, but with significant\nvariability between c ohorts and individuals. These results imply\nlimits to the potential for e arlier diagnosis and intervention while\nalso providing grounds for hope i n extending these beyond current\npractice. \n\nThesis Committee\n\nRusse ll Schwartz (Chair)\n\nOana Carja\n\n \n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9187fce DTSTART;TZID=America/New_York:20250423T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250423T133000 URL:https://db.cs.cmu.edu/events/real-world-applications-of-gen-ai-and-data bases/ LOCATION:Porter Hall 100 SUMMARY:Database Seminar - Sailesh Krishnamurthy CLASS:PUBLIC DESCRIPTION:Speaker: SAILESH KRISHNAMURTHY\, Vice President of Engineering\ , Google\nCloud\n\nTalk Title: Real-world Applications of Gen AI and Datab ases\n\nIn this talk we will explore the transformative potential of\ninte grating databases and generative AI in enterprise applications. As\nLarge Language Models (LLMs) are being rapidly adopted\, it's clear\nthat they n eed to interact with a plethora of other systems in the\nenterprise and ag entic applications have emerged as the primary\nintegration mechanism. Of these integration targets\, databases are\ncritical\, in order to deliver accurate\, contextually relevant\, and\nuser-friendly experiences. \n\nTh ere is\, however\, a catch - databases have extremely wide surface\nareas\ , and contain highly sensitive and privileged information. Many\nchallenge s\, therefore\, arise in integrating databases with LLMs -\nthese include security\, correctness\, safety\, and reliability. We will\ndescribe a pri ncipled way to reason about how to build agentic\napplications that integr ate database-specific tools that can run the\ngamut from being single-use/ circumscribed to highly flexible. In all\nthese situations\, we argue that security is a first-order problem that\nneeds to be tamed in order to mak e such applications viable. We will\ndescribe \"parameterized secure views \"\, a key innovation that Google\nhas developed in AlloyDB that enables t hese flexible tools that allows\nfor secure natural language querying for databases by building\nsecurity right inside the database. \n\n— \n\nS ailesh Krishnamurthy is Vice President of Engineering at Google Cloud\nwhe re he leads Cloud SQL and Non-Relational Databases as well as a\nprogram t o leverage Gen AI and Google’s Gemini models in various\naspects of data base management. He is an innovator and entrepreneur\nwith 25+ years of ex perience in databases and the cloud. Before\nGoogle\, Sailesh was GM for A urora and RDS MySQL at Amazon preceded by\na stint at Cisco Systems via th e acquisition of Truviso\, a real-time\nstreaming data analytics software company that he co-founded. Sailesh\nis an authority in data management an d is an author of over a dozen\nacademic papers and several issued U.S. pa tents. He earned a Ph.D. in\nComputer Science from UC Berkeley in 2006. \ n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9188471 DTSTART;TZID=America/New_York:20250423T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250423T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20250423.html LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Meghal Gupta CLASS:PUBLIC DESCRIPTION:Speaker: MEGHAL GUPTA\, Ph.D. Student\, Department of Electrica l\nEngineering and Computer Sciences\, University of California\, Berkeley \n\nTalk Title: Stream-Decodable Error-Correcting Codes\n\nIn the standard noisy communication model\, Alice encodes a message\nusing an error-corre cting code and sends it to Bob\, who decodes it\nafter receiving the entir e message and storing it in memory. In this\ntalk\, we'll explore what hap pens when Bob doesn't have enough memory\nto store the whole message and m ust instead decode it bit by bit as it\narrives. We'll define what it mean s for a code to be stream-decodable\nand present nearly matching upper and lower bounds on the code length\nrequired in this setting. \n\nThis is b ased on joint works with Venkat Guruswami\, Mihir Singhal\, and\nRachel Zh ang.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91887c5 DTSTART;TZID=America/New_York:20250422T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250422T160000 LOCATION:Newell-Simon 3001 SUMMARY:5th Year Master's Thesis Presentation - Rohini Banerjee CLASS:PUBLIC DESCRIPTION:Speaker: ROHINI BANERJEE\, Master's Student\, Computer Science\ nDepartment\, Carnegie Mellon University\n\nTalk Title: Uncertainty-Aware AI for Clinical Decision Support\n\nBuilding interpretable-by-design AI mo dels that intuitively\ncommunicate model uncertainty is vital to engenderi ng physician and\npatient trust. We develop uncertainty-guided deep learni ng systems for\ntwo pertinent healthcare settings. Efficient intravascular access in\ntrauma and critical care is a high-stakes intervention affordi ng\nminimal tolerance for error. Autonomous needle insertion systems can\n be useful in austere environments due to the lack of skilled medical\npers onnel. However\, inaccuracies in vessel segmentation modeling can\nresult in vessel damage and hemorrhage. \n\nThe risk can be mitigated via predic tive uncertainty estimation to\nassess model reliability. Thus\, we introd uce MSU-Net\, a novel\nmultistage approach to semantic vessel segmentation in ultrasound\nimages that combines the predictive power of Monte Carlo n etworks and\ndeep ensembles. We demonstrate significant improvements\, 27. 7% over\nthe state-of-the-art\, while enhancing model reliability through a\n20.9% stronger discrimination in epistemic uncertainty between correct\ nand incorrect predictions. \n\nNext\, we investigate the robustness of p redictive modeling in\nquantifying the severity of rash manifestations ass ociated with\nCutaneous Dermatomyositis (CDM)\, a rare and currently incur able\nautoimmune disorder. Given the importance of telemedicine for remote \ndisease monitoring and timely intervention\, we address challenges of\nd ata scarcity and patient diversity by integrating a novel BERT-style\nself -supervised learning (SSL) framework to image-based models.\nPretrained vi a masked image modeling on demographically diverse\nimages\, our model ach ieves over a 40% improvement in fine-tuning\nperformance on high-resolutio n in-clinic hand images from a limited\ncohort of 23 CDM patients. We achi eve 83% accuracy on a held-out\npatient set\, surpassing the clinical benc hmark of 70–75% accuracy. \n\nTo our knowledge\, this is the first work to integrate uncertainty\nestimation into such architectures\, enabling r obustness under\ndistributional shift in skin tone unseen during fine-tuni ng. Our\ncontributions lay the groundwork for developing accurate\,\nstati stically rigorous\, clinically actionable deep learning models.\nFuture wo rk aims to improve the interpretability of models for\nequitable clinical decision support. \n\nThesis Committee\n\nArtur W. Dubrawski (Chair)\n\nL ászló A. Jeni\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9188c75 DTSTART;TZID=America/New_York:20250422T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250422T150000 URL:https://www.cs.cmu.edu/~pop/seminar/ LOCATION:Gates Hillman 8102 SUMMARY:Principles of Programming (PoP) Seminar - Yaron Minsky CLASS:PUBLIC DESCRIPTION:Speaker: YARON MINSKY\, Head of Technology\, Jane Street\n\nTal k Title: The saga of runtime 5\n\nIn December 2022\, after nearly a decade of development\, OCaml 5.0.0\nwas released with OCaml’s first multi-cor e capable runtime. This was\nan exciting milestone\, finally making it pos sible to write\nshared-memory parallel programs in OCaml. The new runtime was designed\nto be easy to adopt: it didn’t disturb OCaml’s FFI\, and \nperformance was meant to be only a couple of percentage points slower\ni n single-core mode. Despite those promising beginnings\, switching to\nrun time-5 was harder than we expected. Indeed\, We’ve only managed to\nswit ch to it this year\, after months and months of research and\nengineering effort to make it work for our use-cases. \n\nThis talk will give an over view of the problems we ran into\, and why\nswitching to runtime-5 was so much harder than we expected. We’ll\nalso discuss what we learned from t he process\, both about how to stage\na complex change like this\, as well as some new things we learned\nabout how to design a GC\, and the improve ments we landed in OCaml as\npart of that work. \n\n— \n\nYaron Minsky received his BA in Mathematics from Princeton and his PhD\nin Computer Sc ience from Cornell focusing on distributed systems. He\njoined Jane Street in 2003\, where he founded the firm's quantitative\nresearch group. He in troduced OCaml\, a statically typed functional\nprogramming language\, to the company and managed the transition to\nusing OCaml for all of its core infrastructure\, turning Jane Street\ninto the world's largest industrial user of the language.  He's been\ninvolved in many different aspects of Jane Street's technology stack\,\nincluding machine learning infrastructur e\, distributed systems design\,\nincremental programming systems\, hardwa re synthesis\,  trading and\nrisk systems\, developer tools\, and user-in terface toolkits. \n\nFaculty Host:  Seth Goldstein\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918909d DTSTART;TZID=America/New_York:20250422T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250422T135000 URL:https://privacy.s3d.cmu.edu/masters/seminar/ LOCATION:Heinz Auditorium\, Hamburg Hall A301 and Zoom SUMMARY:Joint Privacy Seminar / Heinz College Seminar - Alessandro Acquisti CLASS:PUBLIC DESCRIPTION:Speaker: ALESSANDRO ACQUISITI\, Trustees Professor of Informati on\nTechnology and Public Policy\, Co-director\, CMU Center for Behavioral \nDecision Research\, Heinz College\, Carnegie Mellon University\n\nTalk T itle: Twenty years of privacy research at CMU\n\nThis is not a goodbye\, i t’s an arrivederci. In the (more or less)\ntwenty fabulous years I’ve had the privilege of being part of the\nCMU faculty\, much has happened in the field of privacy. \n\nIn this talk\, I will reflect on the evolving privacy landscape through\nthe lens of my own research and that of many es teemed CMU colleagues.\nIt has been an eventful journey — from the quain t days of browser\ncookies in the late 1990s\, to today’s frontiers of n eural privacy. \n\n— \n\nAlessandro Acquisti is the Trustees Professor of Information\nTechnology and Public Policy at the Heinz College\, Carne gie Mellon\nUniversity. His research combines economics\, behavioral resea rch\, and\ndata mining to investigate the role of privacy in a digital soc iety.\nHis studies have promoted the revival of the economics of privacy\, \nadvanced the application of behavioral economics to the understanding\no f consumer privacy valuations and decision-making\, and spearheaded\nthe i nvestigation of privacy and personal disclosures in online social\nnetwork s. \n\nHis studies have won numerous awards and have been published in\nj ournals\, books\, and proceedings across different fields\, including\nSci ence\, Proceedings of the National Academy of Science\, Management\nScienc e\, Journal of Economic Literature\, Marketing Science\, Journal of\nConsu mer Research\, Journal of Personality and Social Psychology\, and\nJournal of Experimental Psychology. Alessandro has testified before\nthe U.S. Sen ate and House committees on issues related to privacy\npolicy and consumer behavior. His findings have been featured in\nnational and international media outlets\, including the Economist\, the\nNew York Times\, the Wall S treet Journal\, the Washington Post\, the\nFinancial Times\, Wired.com\, N PR\, CNN\, and 60 Minutes\; his TED talks on\nprivacy and human behavior h ave been viewed over 1.5 million times\nonline.  Alessandro holds a PhD f rom UC Berkeley\, and Master degrees\nfrom UC Berkeley\, the London School of Economics\, and Trinity College\nDublin.\n\nThis Privacy Seminar is sp onsored by the Masters in Privacy\nEngineering Program and the Heinz Colle ge.\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918950a DTSTART;TZID=America/New_York:20250422T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250422T130000 URL:http://www.cs.cmu.edu/~aiseminar/ LOCATION:Newell-Simon 3305 and Zoom SUMMARY:Artificial Intelligence Seminar - Ameet Talwalkar CLASS:PUBLIC DESCRIPTION:Speaker: AMEET TALWALKAR\, Associate Professor\, Machine Learni ng\nDepartment\, Carnegie Mellon University\, Chief Scientist\, Datadog\n\ nTalk Title: Why AI Needs Specialization\n\nWhile modern AI holds great pr omise\, the gap between its hype and\npractical impact remains substantial . This talk advocates for the\nimportance of specialization to help bridge that gap—urging\nresearchers to tailor problem formulations\, modeling approaches\, data\ncollection\, and evaluation methods to concrete downstr eam tasks. \n\nWe begin by examining the limitations of existing domain-s pecific\nfoundation models–for genomics\, satellite imaging\, and time\n series–that apply techniques from core AI domains such as vision and\nNL P with minimal specialization. We then present recent work from CMU\nand D atadog AI Research that advances specialized approaches on two\ndistinct t asks: autonomously executing complex web tasks and\nproactively detecting or predicting disruptions in production software\nsystems. These efforts h ighlight the critical role of domain-aware\ndesign in moving beyond shiny demos and toward meaningful AI impact. \n\n— \n\nAmeet Talwalkar is an associate professor in the Machine Learning\nDepartment at CMU and Chief Scientist at Datadog. His current research\ninterests include ML for scien ce\, human-AI interaction\, and developing\nspecialized foundation models and agents. He co-founded Determined AI\n(acquired by HPE)\, helped create MLlib in Apache Spark\, co-authored\nthe textbook 'Foundations of Machine Learning\,' and created an\naward-winning edX MOOC on distributed machine learning. He also\nspearheaded the creation of the MLSys conference. \n\ nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9189936 DTSTART;TZID=America/New_York:20250422T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250422T130000 URL:https://db.cs.cmu.edu/events/architecture-of-aerospike-fast-scalable-ge o-replicated-multi-model-database-supporting-strict-serializable-transacti ons LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Database Seminar - Srinivasan \"Sesh\" Seshadri CLASS:PUBLIC DESCRIPTION:Speaker: SRINIVASAN \"SESH\" SESHADRITalk Title: Architecture o f\nAerospike: Fast\, Scalable\, Geo-Replicated\, Multi Model Database\nSup porting Strict Serializable Transactions\n\nAerospike is a fast\, scalable database that supports multiple data\nmodels such as a key value store wi th complex objects\, graphs and\nvectors. Aerospike supports synchronous r eplication within a cluster\nto guarantee linearizability. It also support s asynchronous\nreplication across clusters for disaster recovery with min imal\noverhead to normal transaction processing. Finally\, Aerospike suppo rts\nACID transactions and guarantees strict serializability. \n\nIn this talk\, we will describe the architecture of Aerospike and some\nof the ke y design decisions that have enabled Aerospike to maintain a\nvery high le vel of performance and yet scale to billions of records\nwhile guaranteein g the highest level of transaction correctness while\nsupporting multiple data models.\n\n — \n\nSesh is a veteran technology leader with over t hree decades of\nexperience spanning academia\, startups\, and large compa nies. Sesh is\nan accomplished expert in databases with over 50 academic p ublications\nand more than a dozen patents in this area. He is also a lead ing\nexpert in ML especially as applied to Search and Internet\nAdvertisin g. \n\nSesh has been an Assistant Professor at IIT Bombay\, founded and l ed\nmultiple startups to an exit in the Silicon Valley as well as India\,\ nand finally has been a senior leader at companies such as Google\,\nTarge t\, Yahoo\, and Lucent Technologies. Sesh has a B.Tech\, Computer\nScience from the Indian Institute of Technology\, Madras\, and both an\nM.S. and Ph.D. in Computer Science from the University of\nWisconsin-Madison. \n\n In Person and Zoom Participation. See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9189dbb DTSTART;TZID=America/New_York:20250422T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250422T110000 LOCATION:Gates Hillman 9115 SUMMARY:5th Year Master's Thesis Presentation - Alex Tianyi Xu CLASS:PUBLIC DESCRIPTION:Speaker: ALEX TIANYI XU\, Master's Student\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Automating Real-to- Sim Traffic Scene Generation with Large\nLanguage Models\n\n Simulation-b ased evaluation of autonomous driving (AD) offers a\nscalable and reproduc ible alternative to real-world testing\, yet\ncurrent scenario generation methods often prioritize coverage over\nrealism. This thesis presents an e xploration in enabling open-source\nmodels to automatically generate reali stic traffic scenarios from\nnatural language descriptions of real-world c rashes. I conducted a\nseries of experiments to investigate the effectiven ess of different\ninference-time methods in this domain\, and proposed a f ramework for\nleveraging these approaches to create a dataset that can be used to\nfinetune open-source models with fewer parameters. I found that\n open-source models can effectively learn from synthetic data generated\nby closed-source LLMs in the simulator code generation domain: an\nopen-sour ce model finetuned on this new dataset achieves a 92.5%\nsuccess rate in g enerating syntactically correct scenarios. This work\ndemonstrates the fea sibility of LLM-assisted scenario reconstruction\nat scale and lays the fo undation for open\, realistic\, and automated\nevaluation pipelines for AD algorithms. \n\nThesis Committee\n\nChenyan Xiong (Chair)\n\nReid Simmon s\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918a241 DTSTART;TZID=America/New_York:20250422T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250422T120000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:SCS Faculty Candidate - Vasilis Kontonis CLASS:PUBLIC DESCRIPTION:Speaker: VASILIS KONTONIS\, Postdoctoral Research Fellow\, Inst itute for\nFoundations of Machine Learning\, and\, Department of Computer Science\,\nUniversity of Texas at Austin\n\nTalk Title: Beyond Worst-Case ML\n\nWorst-case theoretical frameworks provide an overly pessimistic view \nof what is computationally feasible in machine learning. In this talk\,\ nI will present new frameworks and algorithmic results that move beyond\nw orst-case assumptions to enable efficient learning in realistic\nsettings. We will examine why fundamental problems are computationally\nintractable in the worst case and how to circumvent these\nbarriers.  \n\nI will fi rst discuss robust classification under label noise\,\nintroducing efficie nt algorithms that challenge long-standing\nimpossibility results while im proving and generalizing prior\nalgorithmic work. Finally\, I will present an application to\nsemi-supervised knowledge distillation\, where our pri ncipled methods\noutperform prior works.\n\n — \n\nVasilis Kontonis ea rned his PhD in Computer Science from the\nUniversity of Wisconsin-Madison \, advised by Professor Christos Tzamos.\nHe is currently a postdoctoral f ellow at the Institute for Foundations\nof Machine Learning (IFML) at the University of Texas at Austin\,\nworking with Professor Adam Klivans. His research focuses on designing\ncomputationally efficient and provably reli able algorithms in machine\nlearning and statistics. \n\nHe has published in top venues in theoretical computer science and\nmachine learning\, inc luding FOCS\, STOC\, COLT\, ICML\, and NeurIPS. His\nwork has been recogni zed with awards\, including the Best Paper Award\nat the Conference on Lea rning Theory (COLT) 2024. \n\nIn Person and Zoom Participation.  See ann ouncement. \n\n→ Attendance at this talk is restricted to members of th e SCS\ncommunity and relevant CMU stakeholders.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918a66a DTSTART;TZID=America/New_York:20250421T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250421T173000 URL:https://db.cs.cmu.edu/events/sql-death-gel-replacing-sql-and-improving- on-the-relational-database-model LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Michael Sullivan CLASS:PUBLIC DESCRIPTION:Speaker: MICHAEL SULLIVAN\, Lead Compiler Engineer\, Gel\n\nTal k Title: Gel: Replacing* SQL and Improving on the Relational\nDatabase Mod el\n\nGel (formerly EdgeDB) is a new database built around an evolution of \nthe relational model that we call \"graph-relational\".  In the\ngraph- relational model\, data is represented as strongly typed objects\ncontaini ng set-valued scalar properties and links to other\nobjects. Missing valu es are represented in the language as empty sets\n(no NULL!)\, and have co nsistent semantics. The query language\, EdgeQL\,\nsupports convenient fet ching and modification of nested data. EdgeQL\nqueries are composable: que ries can be used without restriction as\nsubexpressions and bound to varia bles. \n\nGel is open source and is implemented on top of PostgreSQL. The \ncompiler maps Gel's graph-relational schemas to Postgres schemas and\nco mpiles EdgeQL queries into Postgres's flavor of SQL. Gel also\, of\ncourse \, now also supports querying and modifying the database using\nSQL. \n\n — \n\nMichael (Sully) Sullivan is the lead compiler engineer at Gel.\nP reviously\, he worked at Dropbox on the mypy typechecker for Python\nand w as a principal developer of the mypyc compiler for typed Python.\nHe compl eted his Ph.D at Carnegie Mellon University in 2017\, advised\nby Karl Cra ry\, writing a thesis on the design and implementation of a\nnew language memory model for low level concurrency. \n\nThis talk is part of the SQL or Death? Seminar Series\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918aa7f DTSTART;TZID=America/New_York:20250421T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250421T150000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Mehrabian Collaborative Innovation Center 1313 (Special Time /\nLo cation) SUMMARY:Crypto Seminar - Zeyu Liu CLASS:PUBLIC DESCRIPTION:Speaker: ZEYU LIU\, Ph.D. Student\, Applied Cryptography Labora tory\,\nComputer Science Department\, Yale University\n\nTalk Title: Snake -eye Resistant PKE from LWE for Oblivious Message\nRetrieval and Robust En cryption\n\nIn this work\, we introduce snake-eye resistance\, a new secur ity\nproperty for public-key encryption (PKE). This property ensures that a\nciphertext—potentially adversarially generated—cannot decrypt to\nt he same plaintext under two different secret keys. Snake-eye\nresistance i s particularly useful for (1) preventing spamming attacks\nin oblivious me ssage retrieval (OMR) and (2) enabling efficient robust\nencryption scheme s. We first analyze the snake-eye resistance of\nlattice-based PKE schemes . \n\nOur study reveals that while Regev05 and PVW08 satisfy this propert y\,\nmore efficient\, state-of-the-art schemes like Crystals-Kyber do not. \nTo bridge this gap\, we propose LWEmongrass\, a new lattice-based PKE\ns cheme that is provably snake-eye resistant under the standard LWE\nassumpt ion while significantly improving efficiency over Regev05 and\nPVW08. Appl ying LWEmongrass to OMR\, we achieve a 12× speedup over\nexisting spammin g-attack-resistant OMR schemes (conjectured in LT22\nand proven in this wo rk)\, while maintaining provable security under\nthe LWE assumption. \n\n Additionally\, we establish that snake-eye resistance implies\nrobustness\ , yielding the first robust lattice-based PKE scheme that\navoids the inef ficiencies of the KEM-DEM paradigm. As a contribution\nof independent inte rest\, we introduce two LWE variants with side\ninformation\, which serve as key building blocks in our security proofs\nand enable reductions from standard LWE for relevant parameter\nsettings.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918ae7a DTSTART;TZID=America/New_York:20250421T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250421T150000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Mai Elkady CLASS:PUBLIC DESCRIPTION:Speaker: MAI ELKADY\, Ph.D. Candidate\, Department of Computer Science\,\nPurdue University\n\nTalk Title: An Introduction to Graph Neura l Networks\n\nGraphs are a natural representation of complex systems\, fro m social\nnetworks to molecular structures. Working directly on graphs all ows us\nto capture relational information that would be difficult to encod e\notherwise. Graph Neural Networks (GNN) are a class of models that\nexte nds deep learning to non-Euclidean data\, making learning directly\non gra phs feasible. In this talk\, I will discuss the motivation behind\nusing g raphs in machine learning\, then walk through the key building\nblocks of GNNs—message passing\, aggregation\, and update functions. I\nwill prese nt examples of recent GNN architectures and highlight common\nGNN tasks in cluding node classification\, link prediction\, and graph\nclassification.  \n\n— \n\nMai Elkady is a Ph.D. candidate at Purdue University specia lizing in\nmachine learning with a focus on discrete generative models\, a nd graph\nlearning. During her Ph.D. Mai had an extensive teaching experie nce\,\nserving as a teaching assistant for multiple undergraduate courses and\nas an instructor of record for a programming course through Purdue’ s\nGraduate Teaching Fellowship program. Her commitment to education\nearn ed her the Purdue Graduate Teaching Award in Spring 2020. In\naddition to her academic work\, Mai interned at Microsoft and Block\nInc.\, contributi ng to both applied and research projects in Natural\nLanguage Processing ( NLP)\, heterogeneous graph learning\, and scalable\nGraph Neural Network a rchitectures. \n\nJoint Machine Learning Department and Computer Science Department\n\nIn Person and Zoom Participation.  See announcement. \n\n → Attendance at this talk is restricted to members of the SCS\ncommunity and relevant CMU stakeholders.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918b2b5 DTSTART;TZID=America/New_York:20250421T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250421T130000 URL:https://cmubiglab.github.io/lunch/ LOCATION:Group Viewing Newell-Simon Hall 4513 and Zoom SUMMARY:Accessibility Lunch - Bryan F. Shaw CLASS:PUBLIC DESCRIPTION:Speaker: BRYAN F. SHAW\, Professor\, Department of Chemistry an d\nBiochemistry\, Baylor University\n\nTalk Title: And Science For All: Ma king Science Accessible to People\nwith Blindness and Low Vision\n\nThis s eminar will present a handful of interventions to make science\naccessible to children and adults with blindness and low vision. These\ntechnologic interventions utilize universal designs that might help\nall people learn and do science more easily. Key interventions that\nwill be presented incl ude: use of light-scattering tactile graphics\n(“lithophanes”) as univ ersal 2D graphics\; oral somatosensory\nperception of 3D models\; and the use of 3D printed tools and machine\nvision to make benchtop techniques (e .g.\, SDS-PAGE\, TLC) accessible to\npeople with diverse physical and visu al abilities. Funding for this\nresearch is provided by ongoing support fr om the National Institutes\nof Health (Institute of General Medical Scienc es and National Eye\nInstitute)\, the National Science Foundation\, and th e Robert A. Welch\nFoundation. \n\n— \n\nBryan F. Shaw is a Professor in the Department of Chemistry and\nBiochemistry at Baylor University. His research is at the interface of\nspecial education and molecular science. Half of the work in his\nresearch group aims to make science (chemistry a nd biochemistry in\nparticular)\, accessible to people with blindness and low vision. The\nother half involves basic research in bioinorganic and bi ophysical\nchemistry and applied research in pediatric ophthalmology. \n\ nIn Person Group Viewing and Zoom.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918b694 DTSTART;TZID=America/New_York:20250419T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250419T200000 URL:https://www.cs.cmu.edu/scs-career-center/career-center-events LOCATION:Swartz Center for Entrepreneurship\, 3rd Floor\, Tepper School of\ nBusiness SUMMARY:ScottySPARK 2025 CLASS:PUBLIC DESCRIPTION:Speaker: Hosted by ScottyLabsTalk Title: ScottySpark\n\nScottyL abs is excited to see you at ScottySpark\, our upcoming project\nexpo. Sto p by the Swartz Center for Entrepreneurship anytime between\n5:00 pm and 8:00 pm to get an early look at the latest products being\nbuilt by CMU s tudents. Your peers are working on all kinds of\napplications\, from a cen tralized database of every research\nopportunity across campus to an CMU-s pecific AI-assistant.  \n\nSee the list of projects. \n\nIf you have a project you’d like to share\, sign up for a slot in\nthe expo by Wedne sday\, April 16. Viewing projects is open-entry for\nall. If you plan on s topping by\, we’d appreciate it if you’d let\nus know via the interes t form. \n\nWe hope to see you there!    \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918ba39 DTSTART;TZID=America/New_York:20250418T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250418T150000 URL:https://www.cs.cmu.edu/~pop/seminar/ LOCATION:Gates Hillman 8102 SUMMARY:Principals of Programming (PoP) Seminar - Liron Cohen CLASS:PUBLIC DESCRIPTION:Speaker: LIRON COHEN\, Assistant Professor\, Department of Comp uter\nScience\, Ben-Gurion University\n\nTalk Title: Realizability Meets E ffects: Advancing Computational\nFoundations\n\nThe Curry-Howard correspon dence and its semantic embodiment in the\nform of realizability\, establis h a foundational connection between\nlogic\, programming\, and type system s. However\, these fundamental\nframeworks fall short of addressing modern computational paradigms\nthat rely on effects such as state\, concurrency \, and probabilistic\nbehavior. \n\nThis talk explores effectful realizab ility\, a new approach that\nextends realizability to encompass a wide ran ge of computational\neffects. We will introduce (at least) two new algebra ic structures:\nEvidenced Frames\, which provide an abstract structure for building\neffectful realizability models\, and Monadic Combinatory Algebr as\, a\nmonadic-based extension of traditional partial combinatory algebra s.\nWe will see how these extensions not only have a critical impact on\nt he resulting theory\, allowing novel computational interpretations of\nkey principles\, but also offer a robust framework for verifying modern\ncomp utational systems. \n\n— \n\nLiron Cohen is an assistant professor of Computer Science at\nBen-Gurion University who uses formal methods to deve lop frameworks\nthat support formal reasoning about programs and proofs\, aiming to\nimprove reliable software and formalized mathematics. Her resea rch\nfocuses on expanding notions of computation and the logical principle s\ngoverning the properties of systems\, as well as developing formal\ntec hniques that facilitate program verification and mechanized\nmathematics. She was a Fulbright postdoctoral researcher at Cornell\nUniversity\, and s he received her PhD and MSc from Tel Aviv\nUniversity. \n\nFaculty Host:   Robert Harper\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918be61 DTSTART;TZID=America/New_York:20250418T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250418T140000 URL:https://www.cmu.edu/flame/events/index.html LOCATION:Tepper Building 1403 and Zoom SUMMARY:CMU FLAME Center Seminar - Nicholas Carlini CLASS:PUBLIC DESCRIPTION:Speaker: NICOLAS CARLINI\, Research Scientist\, Anthropic\n\nTa lk Title: Computer Security in the Age of Large Language Models\n\nIn ten years Large Language Models (LLMs) have grown from something\nthat can som etimes spell some words correctly\, to something that can\nsolve PhD level math problems and write code at the level of\ncompetitive programmers. Wh at does this mean for security? \n\nIn this talk I first discuss one atte mpt at studying to what extent\nLLMs are capable of performing challenging research-level security\ntasks. As it turns out\, they're not very good a t this yet. So then I\nconsider a second question: what could LLMs do to c hange security\ntoday? I argue there are many domains where even the capab ilities of\ncurrent models would more than suffice to fundamentally alter the\neconomics of how attacks are performed and monetized. Finally\, I\nco nclude with some thoughts looking towards the future. \n\n— \n\nNichol as Carlini is a research scientist at Anthropic studying what\nbad things you could do with\, or do to\, language models. Broadly his\nwork is at th e intersection of machine learning and computer security\,\nand has receiv ed best paper awards from USENIX Security\, ICML\, and\nIEEE S&P. He recei ved his PhD from UC Berkeley under David Wagner. \n\nIn Person and Zoom P articipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918c24b DTSTART;TZID=America/New_York:20250418T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250418T130000 URL:https://ece.cmu.edu/news-and-events/seminars.htm LOCATION:Panther Hollow Conference Room 4105\, Mehrabian Collaborative\nInn ovation Center SUMMARY:ECE Graduate Seminar - Matthew C. Stamm CLASS:PUBLIC DESCRIPTION:Speaker: MATTHEW STAMM\, Associate Professor\, Electrical and C omputer\nEngineering Department\, Drexel University\n\nTalk Title: Multime dia Forensics: Using AI to Protect Against Fake\nMedia Threats\n\nRecent a dvances in photo editing software and generative AI have made\nit easier t han ever to create visually realistic fake images and\nvideos. While these technologies hold tremendous potential for\ncreative and commercial appli cations\, they also pose growing risks\nwhen misused by malicious actors. From orchestrated disinformation\ncampaigns to identity theft and fraud\, the threats from increasingly\nconvincing synthetic or manipulated media c ontinue to raise serious\nconcerns throughout government\, industry\, and society at large. \n\nIn this talk\, I will introduce several deep learni ng-based multimedia\nforensic systems developed in the Multimedia and Info rmation Security\nLab (MISL) at Drexel University that combat the escalati ng threat of\nfake media. These systems are able to identify AI-generated images and\nvideos even when they are produced by generative systems that were not\nencountered during training.  \n\nThis enables detection syste ms to remain effective in the\nfast-evolving generative AI landscape\, whe re new generative systems\ncontinue to emerge at a rapid pace. I will also discuss novel\ntechniques to identify edited and manipulated images and v ideos\ncreated by both AI-based and traditional editing systems\, as well as\nexamine emerging challenges faced when creating authentication systems \nto combat the latest generation of fake media threats. \n\n— \n\nDr. Matthew C. Stamm is an Associate Professor in the Electrical and\nCompute r Engineering Department at Drexel University\, where he leads\nthe Multim edia and Information Security Lab (MISL). He and his lab\ndevelop AI techn iques to detect fake media\, including deepfake videos\,\nAI generated ima ges and photoshopped forgeries. \n\nDr. Stamm is the recipient of several awards for his research\nincluding Drexel University’s 2023 Provost’s Award for Outstanding\nMid-Career Scholarly Achievement\, the University of Maryland’s 2023\nAlumni Excellence Award for Research\, and an NSF CA REER Award in\n2013.  In 2021\, he was named one of Popular Science Magaz ine’s\nBrilliant 10 - its annual list of the “10 most innovative\nup-a nd-coming minds in science.”  Dr. Stamm served as an elected\nmember of the IEEE Signal Processing Society’s Information Forensics\nand Securit y Technical Committee from 2018-2020\, and as General Chair\nof the 2017 A CM Workshop on Information Hiding and Multimedia\nSecurity. \n\nIn additi on to his academic work\, he has helped authenticate media\nused in nation al news stories for the New York Times\, Reuters\, AFP\,\nPolitiFact\, and several other organizations.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918c718 DTSTART;TZID=America/New_York:20250418T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250418T123000 LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI-SDM — Student Brainstorming Session - Frieda Chen\, Mandy Dai\ ,\nSamuel Dinesh CLASS:PUBLIC DESCRIPTION:Talk Title: AI-SDM — Student Brainstorming Session\n\nThe AI Institute for Societal Decision Making (AI-SDM) —which brings\ntogether AI and social sciences researchers to develop human-centric\nAI for societ al good — hosts student-led brainstorming discussion\nsessions. Our goal is to foster interdisciplinary collaboration and\ngenerate ideas on how A I can help solve societal problems\,\nparticularly from an angle of decisi on-making. These sessions are\nintended for graduate and undergraduate stu dents to connect\, share\nideas\, and collaborate on AI-related projects.\ n\nThis week\, we will host members of a team at CMU that is developing a \ndigital simulation of disaster response management in collaboration\nwit h the American Red Cross.  Frieda Chen (Master's student\,\nIntegrated In novation Institute)\, Mandy Dai (Master's student\,\nElectrical & Computer Engineering)\, and Samuel Dinesh (Master's\nstudent\, Information Network ing Institute) are part of the team\nworking on a tool that will not only enhance training for disaster\nresponse volunteers but also serve as a cu tting-edge research tool for\nexploring human decision-making and teaming with AI agents. Please\njoin us to learn more about this impactful project . \n\nRSVP → to anticipate a headcount for lunch In Person and Zoom\nPa rticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918cac2 DTSTART;TZID=America/New_York:20250418T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250418T123000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Doctoral Thesis Proposal - Madhusudhan Reddy Pittu CLASS:PUBLIC DESCRIPTION:Speaker: MADHUSUDHAN REDDY PITTU\, Ph.D. Student\, Computer Sci ence\nDepartment\, Carnegie Mellon University\n\nTalk Title: Fairness\, Di versity\, Explainability\, and Robustness for\nAlgorithmic Decision-making \n\nFairness\, diversity\, explainability\, and robustness are key challen ges\nin computational decision-making\, impacting machine learning\, resou rce\nallocation\, and data analysis. Balancing these principles with\neffi ciency presents significant computational and structural\nchallenges. This proposal investigates algorithmic approaches for\ndiverse selection\, fai r allocation\, interpretable clustering\,\nconstrained subspace approximat ion\, and comparison-based optimization.\nTogether\, these directions cont ribute to more equitable\,\nrepresentative\, interpretable\, and robust al gorithmic decision-making\nunder structural and informational constraints.  \n\nThesis Committee\n\nDavid Woodruff (Chair)\n\nAnupam Gupta\n\nPrasad Tetali\n\nMohit Singh (Georgia Institute of Technology)\n\nOla Svensson ( EPFL)\n\nAdditional Information\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918ce7b DTSTART;TZID=America/New_York:20250418T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250418T113000 LOCATION:Gates Hillman 8102 SUMMARY:Theory Seminar - Amit Rajaraman CLASS:PUBLIC DESCRIPTION:Speaker: AMIT RAJARAMAN\, Ph.D. Student\, Theory Group \, Depar tment of\nElectrical Engineering and Computer Science\, Massachusetts Inst itute\nof Technology\n\nTalk Title: Weak Poincaré Inequalities and Mixing from Non-Worst-Case\nInitializations\n\nThere has been a recent surge of powerful tools to show rapid mixing\nof Markov chains\, via functional ine qualities such as Poincaré\ninequalities. In many situations\, Markov cha ins fail to mix rapidly\nfrom a worst-case initialization\, yet are expect ed to approximately\nsample from a random initialization. Under such condi tions\, a\nPoincaré inequality does not hold\, necessitating new tools to prove\nsampling guarantees. \n\nWe develop a framework to analyze such i nitializations\, based on\nestablishing so-called \"weak Poincaré inequal ities.\" As an\napplication\, we prove that \"simulated annealing\" sample s from the\nGibbs measure of a spherical spin glass for inverse temperatur es up to\na natural threshold\, matching recent algorithms based on algori thmic\nstochastic localization. In this talk\, we will focus on an applica tion\nof our techniques to sampling from mixtures of log-concave\ndistribu tions using data-based initializations. \n\nBased on joint work with Bric e Huang (MIT)\, Sidhanth Mohanty (MIT)\,\nand David X Wu (Berkeley)  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918d240 DTSTART;TZID=America/New_York:20250417T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250417T160000 LOCATION:Gates Hillman 7101 SUMMARY:Doctoral Speaking Skills Talk - Daiyaan Arfeen CLASS:PUBLIC DESCRIPTION:Speaker: DAIYAAN ARFEEN\, Ph.D. Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Nonuniform-Tensor-Para llelism: Mitigating GPU failure\nimpact for Scaled-up LLM Training\n\nLLM training is scaled up to 10Ks of GPUs by a mix of data-(DP) and\nmodel-par allel (MP) execution. Critical to achieving efficiency is\ntensor-parallel (TP\; a form of MP) execution within tightly-coupled\nsubsets of GPUs\, r eferred to as a scale-up domain\, and the larger the\nscale-up domain the better the performance. New datacenter\narchitectures are emerging with mo re GPUs able to be tightly-coupled\nin a scale-up domain\, such as moving from 8 GPUs to 72 GPUs connected\nvia NVLink. \n\nUnfortunately\, larger scale-up domains increase the blast-radius of\nfailures\, with a failure o f single GPU potentially impacting TP\nexecution on the full scale-up doma in\, which can degrade overall LLM\ntraining throughput dramatically. With as few as 0.1% of GPUs being in\na failed state\, a high TP-degree job ca n experience nearly 10%\nreduction in LLM training throughput. We propose\ nnonuniform-tensor-parallelism (NTP) to mitigate this amplified impact\nof GPU failures. In NTP\, a DP replica that experiences GPU failures\noperat es at a reduced TP degree\, contributing throughput equal to the\npercenta ge of still-functional GPUs. \n\nWe also propose a rack-design with impro ved electrical and thermal\ncapabilities in order to sustain power-boostin g of scale-up domains\nthat have experienced failures\; combined with NTP\ , this can allow the\nDP replica with the reduced TP degree (i.e.\, with f ailed GPUs) to keep\nup with the others\, thereby achieving near-zero thro ughput loss for\nlarge-scale LLM training. \n\nPresented in Partial Fulfi llment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918d673 DTSTART;TZID=America/New_York:20250417T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250417T160000 URL:https://aco.math.cmu.edu/abs-24-25/apr17.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Nitya Mani CLASS:PUBLIC DESCRIPTION:Speaker: NITYA MANI\, Ph.D. Student\, Department of Mathematics \,\nMassachusetts Institute of Technology\n\nTalk Title: Strong spatial m ixing for colorings on trees and its\nalgorithmic applications\n\nStrong s patial mixing (SSM) is an important and widely studied\nquantitative notio n of \"correlation decay\" for a variety of natural\ndistributions arising in statistical physics and theoretical computer\nscience. One of the most widely studied such distributions is the\nuniform distribution over prope r q-colorings on a maximum degree Δ\ngraph\, μ. Provided that q ≥ Δ+1 \, any such graph has at least one\nproper q-coloring that can be identif ied via e.g. a greedy search. It\nis a longstanding folklore conjecture th at whenever q satisfies this\nminimal inequality\, μ exhibits SSM and the correlations between the\ncolors of vertices in a sample from μ decay ex ponentially fast in the\ngraph distance between the vertices. However\, ev en the specialization\nof this basic question to bounded-degree trees has remained wide open\,\nhighlighting how much there still is to learn about random colorings.\n\nWe essentially resolve the SSM conjecture on trees\, holds for random\nq-colorings on trees of maximum degree Δ whenever q≥ Δ+3. In the\nalgorithmic direction\, we use SSM on trees to establish opt imal mixing\nof the Glauber dynamics (a classical Markov chain) for sampli ng\nq-colorings on graphs of maximum degree Δ and girth \\girth whenever\ nq≥Δ+3. We discuss failures of this style of reduction to extend\nmore generally to all graphs. \n\nIncludes joint work with Zongchen Chen\, Kui kui Liu\, and Ankur Moitra\nand with Kuikui Liu and Francisco Pernice.\n\n 4:00 pm → Jane Street sponsored Tea and Cookies in the Math Lounge.\nBri ng your own mug if possible.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918da88 DTSTART;TZID=America/New_York:20250417T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250417T153000 LOCATION:Gordon Bell Conference Room\, Gates Hillman 5117 SUMMARY:Doctoral Thesis Proposal - Shawn Chen CLASS:PUBLIC DESCRIPTION:Speaker: SHAWN SHUOSHUO CHEN\, Ph.D. Student\, Computer Science \nDepartment\, Carnegie Mellon University\n\nTalk Title: Reshaping Data Ce nter Networks with Reconfigurability\n\nData center networks are fundament al to cloud computing—they tightly\ncouple compute and storage with high bandwidth and low latency. The\ndemand for data center network bandwidth is continuously growing\,\ndriven by the proliferation of data-intensive a pplications like AI/ML\nand video streaming. However\, electrical packet s witches struggle to\ndeliver the total bandwidth required by the growing d emands because a\n“plateauing” Moore’s Law limits I/O density and hi gh-speed\nmemory capacity. Moreover\, the sheer scale of modern data cente r\nnetworks makes electrical packet-switched networks increasingly\nexpens ive and power-hungry. Reconfigurable optical switching\ntechnology is a pr omising alternative\, offering the potential for\nhigher bandwidth\, reduc ed energy consumption\, and runtime\nreconfigurability. Reconfigurable dat a center networks (RDCNs) combine\nthe benefits of both optical and packet switches to accommodate\ndiverse traffic patterns and enhance network per formance. \n\nThis thesis addresses the limitations of current network de signs in\nRDCNs by revisiting underlying assumptions and redesigning core\ nnetwork components\, focusing on transport\, traffic engineering\, and\nt opology. First\, we present Time-division TCP (TDTCP)\, a new transport\np rotocol that adapts to the fluctuating bandwidth and latency in\ndemand-ob livious RDCNs by maintaining independent network states for\neach time-div ision multiplexed path. Second\, we tackle traffic\nengineering in demand- aware RDCNs with approaches that help implement\ncomplex traffic engineeri ng solutions in switches with minimum\nprecision loss. Third\, we propose a flexible machine learning job\nscheduling mechanism for reconfigurable c lusters based on torus\ntopologies\, ensuring optimal job performance whil e mitigating resource\nfragmentation. Together\, these innovations aim to unlock the full\npotential of RDCNs\, achieving higher performance\, cost- efficiency\, and\nscalability for future data center workloads. \n\nThesi s Committee\n\nSrinivasan Seshan (Chair)\n\nPeter Steenkiste\n\nTim Dettme rs\n\nMinlan Yu (Harvard University)\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918dee6 DTSTART;TZID=America/New_York:20250417T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250417T133000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Speaking Skills Talk - Andy Zou CLASS:PUBLIC DESCRIPTION:Speaker: ANDY ZHOU\, Ph.D Student\, Computer Science Department \,\nCarnegie Mellon University\n\nTalk Title: Progress and Challenges in A I Red Teaming\n\nThis presentation demonstrates how red teaming uncovers c ritical\nvulnerabilities in AI agents that challenge assumptions about saf e\ndeployment.  The talk discusses the risks of integrating AI into\nreal -world applications and dives into practical safeguards to enhance\nresili ence and ensure dependable deployment in high-risk settings. \n\nPresente d in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918e1e3 DTSTART;TZID=America/New_York:20250416T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250416T163000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:15-422/642 Machine Learning and Systems: Guest Lecture - Tim Dettme rs CLASS:PUBLIC DESCRIPTION:Speaker: TIM DETTMERS\, Research Scientist\, Allen Institute fo r AI\,\nand\, Assistant Professor\, Machine Learning Department\, Carnegie Mellon\nUniversity\n\nTalk Title: Efficient Foundation Models via Quantiz ation\n\nThe ever-increasing scale of foundation models\, such as ChatGPT and\nAlphaFold\, has revolutionized AI and science more generally. However \,\nincreasing scale also steadily raises computational barriers\, blockin g\nalmost everyone from studying\, adapting\, or otherwise using these\nmo dels for anything beyond static API queries. \n\nIn this talk\, I will pr esent research that significantly lowers these\nbarriers for a wide range of use cases\, including quantized inference\nalgorithms that are used to make predictions after training and\nfine-tuning approaches — such as QL oRA — that adapt a trained\nmodel to new data. I will also talk about th e different requirements\nwhen models are deployed for private use vs for company use and how\nthis altered the effectiveness of quantization algori thms.\n\n — Tim Dettmers is a Research Scientist at the Allen Institute for\nAI and an Assistant Professor at Carnegie Mellon University. His wor k\nfocuses on making foundation models\, such as ChatGPT\, accessible to\n researchers and practitioners by reducing their resource requirements.\nHi s main focus is to develop high-quality agent systems that are\nopen-sourc e and can be run on consumer hardware\, such as laptops. His\nresearch won oral\, spotlight\, and best paper awards at conferences\nsuch as ICLR and NeurIPS and was awarded the Block Award and Madrona\nPrize. He created th e bitsandbytes open-source library for efficient\nfoundation models\, whic h is growing at 2.2 million installations per\nmonth\, and for which he re ceived Google Open Source and PyTorch\nFoundation awards. \n\nFaculty Hos t: Tianqi Chen\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918e5d1 DTSTART;TZID=America/New_York:20250416T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250416T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20250416.html LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Gary Hoppenworth CLASS:PUBLIC DESCRIPTION:Speaker: GARY HOPPENWORTH\, Ph.D. Student in Computer Science\, Computer\nScience and Engineering\, University of Michigan\n\nTalk Title: Covering Approximate Shortest Paths with DAGs\n\nWe define and study anal ogs of probabilistic tree embeddings and tree\ncovers for directed graphs. We define the notion of a DAG cover of a\ngeneral directed graph G: a sma ll collection D1\,… Dg of DAGs so that\nfor all pairs of vertices s\, t\ , some DAG Di provides low distortion\nfor \\dist(s\,t)\; i.e. \\ distG(s\ , t) ≤ min{i ∈ [g] \\ distDi(s\, t)\n≤ α ·  \\ distG(s\, t)\, wh ere α is the distortion.     \n\nAs a trivial upper bound\, there is a DAG cover with n DAGs and α = 1\nby taking the shortest-paths tree from each vertex. When each DAG is\nrestricted to be a subgraph of G\, there is a simple matching lower\nbound (via a directed cycle) that n DAGs are nec essary\, even to\npreserve reachability. Thus\, we allow the DAGs to inclu de a limited\nnumber of additional edges not from the original graph. When n2\nadditional edges are allowed\, there is a simple upper bound of two\n DAGs and α = 1. Our first result is an almost-matching lower bound\nthat even for n2-o(1) additional edges\, at least n1-o(1) DAGs are\nneeded\, ev en to preserve reachability. However\, the story is different\nwhen the nu mber of additional edges is Õ(m)\, a natural setting where\nthe sparsity of the DAG collection asymptotically matches that of the\noriginal graph. Our main upper bound is that there is a near-linear\ntime algorithm to con struct a DAG cover with Õ(m) additional edges\,\npolylogarithmic distorti on\, and only O(\\log n) DAGs. This is similar\nto known results for undir ected graphs: the well-known FRT\nprobabilistic tree embedding implies a t ree cover where both the\nnumber of trees and the distortion are logarithm ic. \n\nOur algorithm also extends to a certain probabilistic embedding\n guarantee. We complement our upper bound with a lower bound showing\nthat achieving a DAG cover with no distortion and Õ(m) additional\nedges requi res a polynomial number of DAGs. Finally\, we point out that\nour DAG cove r upper bounds have implications for a variety of\ncombinatorial and algor ithmic problems related to approximate shortest\npaths in directed graphs.   \n\nThis talk is based on joint work with Sepehr Assadi and Nicole Wei n.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918ea64 DTSTART;TZID=America/New_York:20250416T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250416T120000 LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:SCS Faculty Candidate - Gabriel Poesia CLASS:PUBLIC DESCRIPTION:Speaker: GABRIEL POESIA\, Ph.D. Student\, Computer Science Depa rtment\,\nStanford University\n\nTalk Title: Learning Formal Reasoning\n\n Formal systems\, such as type theories\, provide general foundations for\n representing mathematics and computation\, with increasing adoption in\nth e formalization of research-level mathematical results\, as well as\nfor i mplementing and verifying critical real-world software. However\,\ntheir f lexibility comes at a cost: most key problems in these systems\,\nlike fin ding proofs\, are computationally undecidable. Nonetheless\,\nhumans routi nely solve novel mathematical problems\, write new programs\nand prove the m correct. Crucially\, we leverage our ability to learn\,\ndeveloping incr easingly better heuristics and abstractions for our\nparticular domains of interest as we gain experience. We do so even\nwithout specific goals oth er than exploring and making interesting\ndiscoveries. \n\nIn this talk\, I’ll present my research addressing fundamental\nchallenges arising in learning formal reasoning. We’ll aim at\nbuilding systems that self-impr ove by spending compute without\nrequiring human examples\, much like Alph aZero was able to achieve for\nchallenging games. First\, I’ll show how reinforcement learning and\nabstraction learning combined enable an agent to master sections from\nthe Khan Academy algebra curriculum\, and even re construct the\nhuman-designed curriculum using its learned abstractions\, despite\nseeing problems in a random order. Then\, I’ll present my work on\nopen-ended learning for theorem proving\, where an agent starts only\n from axioms and learns from self-generated conjectures\, bootstrapping\nit s ability to prove human-written theorems despite only training on\nproofs it found by itself. Along the way\, I’ll present methods for\ninterfaci ng symbolic and neural systems\, with applications to program\ngeneration and verification\, and discuss standing challenges in\ndeveloping self-imp roving reasoning machines. \n\n— \n\nGabriel Poesia is a PhD student a t Stanford University in the\nComputation and Cognition Lab. His research is centered on learning\nformal reasoning\, interfacing dependent type the ory\, language models\,\nreinforcement learning and intrinsically motivate d learning\, with work\nat the intersection of all these areas recognized with an Oral\npresentation at NeurIPS 2024. His research has been supporte d by the\nStanford Interdisciplinary Graduate Fellowship. \n\nFaculty Hos ts: Marijn Heule\, Pradeep Ravikumar Joint Computer Science\nDepartment a nd Machine Learning Department\n\nIn Person and Zoom Participation.  See announcement. \n\n→ Attendance at this talk is restricted to members of the SCS\ncommunity and relevant CMU stakeholders.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918ef40 DTSTART;TZID=America/New_York:20250415T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250415T183000 URL:https://www.cs.cmu.edu/~ai-in-action/ LOCATION:Rashid Auditorium\, Gates Hillman 4401 and Livestream SUMMARY:AI in Action Distinguished Lecture - Sudha Rao CLASS:PUBLIC DESCRIPTION:Speaker: SUDHA RAO\, Principal Researcher\, Natural Language Pr ocessing\nGroup\, Microsoft Research\n\nAn Seminar Series for AI in Indust ry and Government hosted by the\nCarnegie Mellon School of Computer Scienc e Undergraduate AI Program\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918f201 DTSTART;TZID=America/New_York:20250415T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250415T153000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Runtian Zhai CLASS:PUBLIC DESCRIPTION:Speaker: RUNTIAN ZHAI\, Ph.D. Candidate\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Contextures: The Mecha nism of Representation Learning\n\nThis thesis establishes the contexture theory to mathematically\ncharacterize the mechanism of representation lea rning\, also known as\npretraining. Despite the remarkable empirical succe ss of foundation\nmodels\, it is not very clear what representations they learn\, and why\nthese representations are useful for various disparate do wnstream\ntasks. A scientific understanding of representation learning is\ ncritical\, especially at this point when scaling up the model size is\npr oducing diminishing returns\, and designing new pretraining methods\nis im perative for further progress. Prior work treated different\nrepresentatio n learning methods quite differently\, whereas the\ncontexture theory prov ides a unified framework for delineating the\nrepresentations these method s learn. \n\nThe central argument is that a representation is learned fro m the\nassociation between the input X and a context variable A. We prove\ nthat if an encoder captures the maximum information of this\nassociation\ , in which case we say that the encoder learns the\ncontexture\, then it w ill be optimal on the class of tasks that are\ncompatible with the context . We also show that a context is the most\nuseful when the association bet ween X and A is neither too strong nor\ntoo weak. The important implicatio n of the contexture theory is that\nincreasing the model size alone will a chieve diminishing returns\, and\nfurther advancements require better cont exts. We demonstrate that lots\nof existing pretraining objectives can lea rn the contexture\, including\nsupervised learning\, self-supervised learn ing\, generative models\, etc.\nBased on that\, we introduce two general o bjectives---SVME and KISE\,\nfor learning the contexture. We also show how to mix multiple contexts\ntogether\, which is an effortless way to create better contexts from\nexisting ones. Then\, we prove statistical learning bounds for\nrepresentation learning\, and extend the framework to spectra lly\ntransformed kernel regression for semi-supervised learning. Finally\, \nwe discuss the effect of the data distribution shift from pretraining\nt o the downstream task.  \n\nThesis Committee\n\nPradeep Ravikumar (Co-Ch air)\n\nZico Kolter (Co-Chair)\n\nAndrej Risteski\n\nYuandong Tian (Meta)\ n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918f676 DTSTART;TZID=America/New_York:20250415T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250415T135000 URL:https://privacy.cs.cmu.edu/masters/seminar/index.html LOCATION:Hamburg Hall 1002 and Zoom SUMMARY:Privacy Seminar - Marco Paes CLASS:PUBLIC DESCRIPTION:Speaker: MARCO PAES\, Researcher\, CERT Division\, Software Eng ineering\nInstitute\, Carnegie Mellon University\n\nTalk Title: Trustworth y Surveillance? Navigating Enterprise Risk and\nOrganizational Privacy\n\n The intersection of privacy and insider threat mitigation presents a\ncomp lex challenge for organizations navigating the demands of\nsecurity\, ethi cs\, and regulatory compliance. As enterprises\nincreasingly adopt advance d monitoring technologies—including\nAI-driven surveillance and data ana lytics—the tensions between risk\nmanagement and the protection of civil liberties grow more pronounced.\nThis seminar will explore the multifacet ed relationship between\ninsider threats\, privacy preservation\, and emer ging technologies\,\nwhile also examining the broader policy and human dim ensions of the\nissue. \n\nCentral to the discussion is the balance betwe en effective enterprise\nrisk management and the ethical handling of emplo yee data. The talk\nwill address the spectrum of insider threats\, from ne gligent to\nmalicious actors\, and their implications for organizational s ecurity.\nIt will also consider the human element of privacy perceptions w ithin\nenterprise environments\, drawing on survey-based research to highl ight\ntensions between security practices and individual rights. \n\nA cr itical examination of governance frameworks will explore how\nadministrati ve data can be leveraged for threat analysis while\nrespecting privacy bou ndaries. The discussion will extend to the role\nof policy in shaping resp onsible monitoring practices\, including an\nanalysis of privacy legislati on such as GDPR and CCPA\, as well as\nproposed policy alternatives to rec oncile security needs with civil\nliberties. Finally\, the presentation wi ll evaluate the promise and\nlimitations of privacy-enhancing technologies \, including advances in\nAI and encryption\, in addressing insider threat s while providing\ncertain privacy guarantees. By synthesizing technical\, legal\, and\nmanagerial perspectives\, this presentation aims to move tow ards\naccountable and transparent enterprise risk management. \n\n— \n \nMarco Paes is a Researcher in the CERT Division of Carnegie Mellon\nUniv ersity's Software Engineering Institute\, where he specializes in\nHuman-C entered Risk Management in Enterprise Security\, and the\nincorporation of Artificial Intelligence into security and risk\nmanagement. Previously\, he was a Security and Privacy Architect at\nMITRE\, where he worked on the team developing the PANOPTIC Privacy\nThreat Model. He has instructed cou rses at George Mason University\,\nPurdue University\, and the Software En gineering Institute related to\nHuman Factors\, Machine Learning\, and Inf ormation Security.\n\nSpring 2025 Privacy Seminars are Sponsored by the Ma sters in Privacy\nEngineering Program and the Carnegie Bosch Institute \n \nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918fb66 DTSTART;TZID=America/New_York:20250414T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250414T173000 URL:https://db.cs.cmu.edu/events/sql-death-mariadbs-new-query-optimizer LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Ulf Michael Widenius CLASS:PUBLIC DESCRIPTION:Speaker: MICHAEL WIDENIUS\, Chief Technology Officer\, MariaDB\ nCorporation AB\n\nTalk Title: MariaDB's New Query Optimizer: A Multi-tool That Does Some\nThings Differently\n\nMariaDB's query optimizer stems fro m MySQL's original implementation.\nIt didn't come from a textbook\, inste ad\, it grew organically to meet\nthe demands of the workloads we were tar geting. This talk will discuss\nsome of the uncommon choices we've made in MariaDB's new optimizer and\nwhat we've got (and lost) as a result. \n\n — \n\nUlf Michael Widenius (often called Monty) is the main author of t he\noriginal version of the open source MySQL database\, a founding member \nof the MySQL AB company and CTO of the MariaDB Corporation AB.\nAddition ally\, he is a founder and general partner at venture capital\nfirm OpenOc ean. This talk is part of the SQL or Death? Seminar Series\n\nZoom Partici pation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef918fee3 DTSTART;TZID=America/New_York:20250414T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250414T130000 URL:https://www.cylab.cmu.edu/events/2025/04/14-seminar-neel.html LOCATION:Panther Hollow Room 4105\, Mehrabian Collaborative Innovation Cent er SUMMARY:CyLab Seminar - Seth Neel CLASS:PUBLIC DESCRIPTION:Speaker: SETH NEEL\, Research Scientist\, Google Research\n\nTa lk Title: Machine Unlearning is as \"easy\" as Data Attribution\n\nMachine unlearning—efficiently removing the effect of a small\n\"forget set\" o f training data on a trained machine learning\nmodel—has attracted signi ficant research interest due to potential\napplications in privacy\, model editing\, safety training and more.\nDespite this\, recent work shows tha t existing techniques are\nineffective for non-convex models\; remnants of the removed points can\nstill be reliably detected after \"unlearning.\"  \n\nIn this talk\, we'll first briefly review the earlier work on\ngradi ent-based machine unlearning in the convex setting that motivates\nmany of these more recent algorithms\, building intuition for why these\nmethods may fail with non-convexity. We will then introduce a new\nmachine unlearn ing technique that exhibits strong empirical\nperformance when removing da ta from neural networks\, and is\nwell-motivated theoretically. Our meta-a lgorithm\, which we call Data\nModel Matching (DMM)\, leverages recent adv ances in data attribution to\npredict the output of the model if it were r e-trained on all but the\nforget set points\, and then fine-tunes the mode l to match these\npredicted outputs. \n\nWe show that in a simple convex setting DMM converges faster than\nprior unlearning algorithms\, and in no n-convex settings (ResNets\ntrained on CIFAR and ImageNet)\, DMM achieves superior empirical\nunlearning performance. An added benefit of DMM is tha t it is a\nmeta-algorithm\, meaning that future advances in data attributi on will\ntranslate directly into better unlearning algorithms\, pointing t o a\nclear direction for future progress in this field. \n\n— \n\nSeth Neel is currently a Research Scientist at Google Research focused\non val ue of data\, synthetic data\, and data privacy questions as they\npertain to training generative models. He joined Google from Harvard\,\nwhere he w as an Assistant Professor at Harvard Business School and\nfaculty affiliat e in Computer Science. He received his Ph.D. in 2020\nfrom the University of Pennsylvania\, under the supervision of Aaron\nRoth and Michael Kearns.  \n\nFaculty Host:  Steven Wu\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919037e DTSTART;TZID=America/New_York:20250414T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250414T130000 URL:https://cmubiglab.github.io/lunch/ SUMMARY:Accessibility Lunch Seminar - Eshed Ohn-Bar CLASS:PUBLIC DESCRIPTION:Speaker: ESHED OHN-BAR\, Assistant Professor\, Department of El ectrical\nand Computer Engineering\, Boston University\n\nTalk Title: Adap tive Accessibility Systems\n\nDespite decades of progress in AI\, making A I systems usable and\nhelpful in real-world situations remains challenging . This gap becomes\nespecially clear when designing systems for people wit h disabilities\,\nwhose diverse needs and behaviors often confound existin g models and\nremain a (suboptimal) afterthought. In this talk\, I will de scribe our\nefforts to build adaptive AI systems that safely and efficient ly\nassist individuals with visual impairments in complex navigation\ntask s. \n\nAt the core of our approach are AI models that can leverage divers e\nsupervision and realistic interactions to learn multimodal models for\n how people perceive the environment\, respond to instructions\, and\ndynam ically move through 3D spaces. Beyond addressing fundamental\nquestions in interactive AI\, e.g.\, for effectively modeling diverse\ndata\, the rese arch will pave the way toward a foundational framework\nfor building scala ble systems that can seamlessly adapt to varying\nusers\, environments\, a nd applications—whether in transportation\,\nhealthcare\, and education.  \n\n— \n\nEshed Ohn-Bar is an Assistant Professor in the Department o f\nElectrical and Computer Engineering at Boston University (BU). His\nres earch lies at the intersection of machine intelligence\, computer\nvision\ , and interactive systems\, with the goal of developing scalable\nAI frame works that enhance quality of life—from self-driving\nvehicles to wearab le technologies for individuals with disabilities.\nHis collaborations wit h colleagues have been recognized with numerous\nawards\, including the BU College of Engineering Early Career\nExcellence in Research Award\, best PhD dissertation award from the\nIEEE Intelligent Transportation Systems S ociety\, and best paper awards\n(the CVPR AMFG Workshop\, Web for All Conf erence\, and finalist at\nICPR). His team was also a semifinalist in the 2 022 DoT Inclusive\nDesign Challenge. Prior to joining BU\, he was a Humbol dt Fellow at the\nMax Planck Institute and a Postdoctoral Researcher at Ca rnegie Mellon\nUniversity. Eshed received a BS degree in Mathematics from UC Los\nAngeles in 2010\, MEd from UC Los Angeles in 2011\, and the PhD de gree\nin Electrical Engineering from UC San Diego in 2017. \n\nZoom Parti cipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919079c DTSTART;TZID=America/New_York:20250411T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250411T133000 URL:https://safety21.cmu.edu/whats-happening/ LOCATION:Scott Hall 6142 and Livestream SUMMARY:Smart Safety Connection Seminar - Christoph Mertz CLASS:PUBLIC DESCRIPTION:Speaker: CHRISTOPH MERTZ\, Principal Project Scientist\, The Ro botics\nInstitute\, Carnegie Mellon University\,\n\nTalk Title: The Danger ous Long Tail of Transportation: Construction\nZones\n\nNavigating work zo nes autonomously is challenging: They contain unique\nobjects\, signs have different meanings\, and different rules apply\ncompared to normal roads. There is also a lack of open datasets. Our\nnew ROADWork dataset addresse s this gap\, and with it we are able to\nimprove state-of-the-art models a cross perception and navigation\ntasks. In this talk we will show how our new dataset is a step forward\nin solving the autonomous driving problem i n construction zones. \n\nSafety21 is a collaboration between several ins titutions\, spanning\nmany disciplines. \n\nREGISTRATION  →  Requeste d in advance (lunch provided for in\nperson attendees)\n                             \n\n → Due to space constraints\, there will be a limit for in-person\nattendance with a virtual option\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9190af1 DTSTART;TZID=America/New_York:20250411T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250411T140000 URL:https://www.cmu.edu/flame/events/index.html LOCATION:Tepper Building 1403 and Zoom SUMMARY:CMU Flame Center Seminar - Jacob Springer CLASS:PUBLIC DESCRIPTION:Speaker: JACOB SPRING\, Ph.D. Student\, Machine Learning Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Overtrained Language Mod els Are Harder to Fine-Tune\n\nLarge language models are pre-trained on ev er-growing token budgets\nunder the assumption that better pre-training pe rformance translates\nto improved downstream models. In this work\, we cha llenge this\nassumption and show that extended pre-training can make model s harder\nto fine-tune\, leading to degraded final performance. We term th is\nphenomenon catastrophic overtraining. For example\, the\ninstruction-t uned OLMo-1B model pre-trained on 3T tokens leads to over\n2% worse perfor mance on multiple standard LLM benchmarks than its 2.3T\ntoken counterpart . Through controlled experiments and theoretical\nanalysis\, we show that catastrophic overtraining arises from a\nsystematic increase in the broad sensitivity of pre-trained parameters\nto modifications\, including but no t limited to fine-tuning. Our\nfindings call for a critical reassessment o f pre-training design that\nconsiders the downstream adaptability of the m odel.  \n\nPaper Reference\n\n — \n\nJacob Springer is a third-year PhD student in the Machine Learning\nDepartment at CMU\, advised by Aditi Raghunathan. His research broadly\nfocuses on the science of foundation mo dels\, emphasizing pre-training\,\nfine-tuning\, and optimization. His cur rent research investigates\nfactors affecting the adaptability of language models to new\ntasks—through fine-tuning or prompting—across all stag es of the\nmodel lifecycle\, from pre-training to inference. \n\nIn Perso n and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9190ecc DTSTART;TZID=America/New_York:20250411T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250411T130000 LOCATION:Traffic21 Classroom and Zoom SUMMARY:Doctoral Thesis Oral Defense - Magdalen Dobson Manohar CLASS:PUBLIC DESCRIPTION:Speaker: MAGDALEN DOBSON MANOHAR\, Ph.D. Candidate\, Computer S cience\nDepartment\, Carnegie Mellon University\n\nTalk Title: New Techniq ues for Parallelism and Concurrency in Nearest\nNeighbor Search\n\nNearest neighbor search in both high and low dimensions is an\nimportant problem in the field of computer science and beyond. In this\nthesis\, we extend t he capabilities of nearest neighbor search\nalgorithms to meet modern dema nds\, including support for billion-scale\nindices\, parallelism and concu rrency on machines with hundreds of\ncores\, efficient updates\, and exten sions to related geometric problems\nsuch as range search. We progress tow ards this goal by introducing the\nzd-tree\, a data structure for low-dime nsional nearest neighbor search\nwith provable guarantees on the work and span of search\, build\, and\nupdate\, a scalable parallel build\, and the ability to perform\nbatch-dynamic updates in parallel.\n\nBuilding on the zd-tree\, we also present the CLEANN-Tree (for\nConcurrent Linearizable E fficient Augmented Nearest Neighbor Search\nTree)\, a generalization of th e zd-tree which supports concurrent\nqueries and updates utilizing version ed pointers and lock-free locks.\nIn high dimensions\, we introduce techni ques to make existing nearest\nneighbor search algorithms lock-free\, dete rministic\, and scalable to\nbillion-size datasets. We apply these techniq ues to four existing\ngraph-based nearest neighbor search algorithms in a library called\nParlayANN. Building off our work in ParlayANN\, we extend the search\nalgorithms for graph-based nearest neighbor indices to the rel ated but\nrelatively under-studied problem of range search in high dimensi ons\,\nmaking significant gains over a naive baseline. \n\nThesis Committ ee\n\nGuy E. Blelloch (Chair)\n\nPhillip B. Gibbons\n\nAndrew Pavlo\n\nHar sha Vardhan Simhadri (Microsoft Azure)\n\nMatthijs Douze (Meta AI Research )\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919132d DTSTART;TZID=America/New_York:20250410T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250410T183000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 and Zoom SUMMARY:Hans J. Berliner Lecture in Artificial Intelligence - Peter Stone CLASS:PUBLIC DESCRIPTION:Speaker: PETER STONE\, Truchard Foundation Chair in Computer Sc ience\,\nAssociate Chair of the Computer Science Department\, University\n Distinguished Teaching Professor\, and Director\, Texas Robotics\,\nUniver sity of Texas at Austin and\, Chief Scientist\, Sony AI\n\nTalk Title: Out racing Champion Gran Turismo Drivers with Deep\nReinforcement Learning\n\n Many potential applications of artificial intelligence involve making\nrea l-time decisions in physical systems while interacting with humans.\nAutom obile racing represents an extreme example of these conditions\;\ndrivers must execute complex tactical manoeuvres to pass or block\nopponents while operating their vehicles at their traction limits.\nRacing simulations\, such as the PlayStation game Gran Turismo\,\nfaithfully reproduce the non- linear control challenges of real race\ncars while also encapsulating the complex multi-agent interactions. \n\nThis talk describes how we trained agents for Gran Turismo that can\ncompete with the world's best e-sports d rivers. We combine\nstate-of-the-art\, model-free\, deep reinforcement lea rning algorithms\nwith mixed-scenario training to learn an integrated cont rol policy\nthat combines exceptional speed with impressive tactics. In ad dition\,\nwe construct a reward function that enables the agent to be\ncom petitive while adhering to racing's important\, but under-specified\,\nspo rtsmanship rules. We demonstrate the capabilities of our agent\,\nGran Tur ismo Sophy\, by winning a head-to-head competition against four\nof the wo rld's best Gran Turismo drivers.  A version is now available\nto everyone with a Playstation 5 and has since become one of the\nlargest commercial deployments of a reinforcement learning agent. \n\n— \n\nDr. Peter Sto ne holds the Truchard Foundation Chair in Computer\nScience at the Univers ity of Texas at Austin. He is Associate Chair of\nthe Computer Science Dep artment\, as well as Director of Texas\nRobotics. In 2013 he was awarded t he University of Texas System\nRegents' Outstanding Teaching Award and in 2014 he was inducted into\nthe UT Austin Academy of Distinguished Teachers \, earning him the title\nof University Distinguished Teaching Professor. Professor Stone's\nresearch interests in Artificial Intelligence include m achine learning\n(especially reinforcement learning)\, multiagent systems\ , and robotics.\nProfessor Stone received his Ph.D in Computer Science in 1998 from\nCarnegie Mellon University. From 1999 to 2002 he was a Senior\n Technical Staff Member in the Artificial Intelligence Principles\nResearch Department at AT&T Labs - Research. He is an Alfred P. Sloan\nResearch Fe llow\, Guggenheim Fellow\, AAAI Fellow\, IEEE Fellow\, AAAS\nFellow\, ACM Fellow\, Fulbright Scholar\, and 2004 ONR Young\nInvestigator. In 2007 he received the prestigious IJCAI Computers and\nThought Award\, given biannu ally to the top AI researcher under the age\nof 35\, and in 2016 he was aw arded the ACM/SIGAI Autonomous Agents\nResearch Award. Professor Stone co- founded Cogitai\, Inc.\, a startup\ncompany focused on continual learning\ , in 2015\, and currently serves\nas Chief Scientist of Sony AI. \n\nFacu lty Host:  Vincent Conitzer \n\n— About the Lecture: \n\nThe Hans J. Berliner Lecture in Artificial Intelligence has been\nestablished in tribu te to Hans J. Berliner\, (CS'74)  in recognition\nof the significant and critical accomplishments as faculty\,\nresearcher\, advisor\, and exemplar y colleague and friend to many.  \n\nThis endowed lecture is presented b y the Computer Science Department\,\nin conjunction with the SCS Distingui shed Lecture Series\, and will let\nus reflect on Hans' contributions and all they enabled. Learn more.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91918df DTSTART;TZID=America/New_York:20250410T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250410T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - Miranda Christ CLASS:PUBLIC DESCRIPTION:Speaker: MIRANDA CHRIST\, Ph.D. Student\, Columbia University\n \nTalk Title: Ideal Pseudorandom Error-Correcting Codes with\nApplications to Watermarking Generative AI\n\nMotivated by the growing need to identif y AI-generated content\, we\n([CG24]) introduced a powerful new framework for generative AI\nwatermarking. This framework leverages a new cryptograp hic primitive\ncalled a pseudorandom error-correcting code (PRC). A PRC is an\nerror-correcting code with the property that any polynomial number of \ncodewords are pseudorandom to any efficient adversary. We construct\nPRC s from standard cryptographic assumptions\, and in this talk I will\ngive an overview of our construction from subexponential LPN. Since\nthe introd uction of PRCs\, there has been a flurry of exciting works\nstrengthening their properties and implementing them in practice. I\nwill highlight new work with my collaborators ([AAC+25]) in which we\ndefine and construct a notion of an ideal PRC\, with stronger\nrobustness and pseudorandomness mo tivated by applications. Our proof\nof security uses tools from the analys is of Boolean functions. \n\nThis is based on works with Sam Gunn\, Omar Alrabiah\, Prabhanjan\nAnanth\, and Yevgeniy Dodis: [CG24] \, [AAC+25] \n \n— Miranda Christ is a computer science PhD student at Columbia\nUniver sity\, advised by Tal Malkin and Mihalis Yannakakis.  She is a\nmember of the Theory Group and the Crypto Lab. Her research is\ngenerally on theore tical cryptography\, and recently has focused on the\nintersection of cryp tography and machine learning. \n\nIn Person and Zoom Participation.  Se e announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9191cda DTSTART;TZID=America/New_York:20250410T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250410T150000 LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:SCS Faculty - Media Relations Refresher CLASS:PUBLIC DESCRIPTION:Talk Title: Enhancing Reinforcement Learning with Error-Prone L anguage\nModelsTo help SCS faculty more effectively talk about their rese arch\nwith journalists and feel more comfortable doing so\, the SCS Market ing\nand Communications team invites you to a refresher session on engagin g\nwith the media.  In this session\, you will learn about the media\nrel ations resources available in SCS and University Communications and\nMarke ting and pick up useful tips for interviews\, including how to\nrespond to difficult questions.\nQuestions welcome!In Person and Zoom Participation .  See\nannouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9191ff5 DTSTART;TZID=America/New_York:20250410T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250410T140000 URL:https://www.cylab.cmu.edu/events/ LOCATION:Newell-Simon 3305 SUMMARY:Special CyLab Seminar - Mihai Christodorescu CLASS:PUBLIC DESCRIPTION:Speaker: MIHAI CHRISTODORESCU\, Research Scientist\, Google\n\n Talk Title: Can We Use Language Models for Security Reviews Yet?\n\nLarge Language Models (LLMs) have shown success in text generation\nand\, increa singly\, in code generation and related tasks. We\ninvestigate applying LL Ms to code comprehension\, specifically for\nsecurity and privacy code rev iews. While existing models trained on\nclean code repositories (like GitH ub and Stack Overflow) perform well\nat summarizing and explaining standar d code\, enabling AI assistance in\ncode reviews\, they falter when faced with the unusual or adversarial\ncode typical in security analysis. Using counterfactual testing\, we\nevaluated LLM understanding of programming co ncepts and found\nsignificant gaps\, particularly concerning data flow and control flow.\nTo address this\, we developed a framework that automatica lly generates\na synthetic dataset of mutated code. This reward-guided app roach\nsystematically creates diverse and realistically obfuscated samples .\nWe demonstrate that incorporating this synthetic data into training\nsi gnificantly improves code ML model robustness and performance on\nobfuscat ed code\, paving the way for more reliable AI tools in\nsecurity-critical domains. \n\n— \n\nDr. Mihai Christodorescu is a Research Scientist at Google\, where he\nfocuses on software security and privacy\, especially for the mobile\ndomain. His research interests are in fundamental approach es to\ncomputer security and privacy problems by combining methods from\nm ultiple domains\, from programming languages\, to machine learning\,\nbeha vioral modeling\, and formal methods. Most recently\, he focused on\ntrans lating progress in user authentication to software service\nauthentication and on designing cryptographic techniques to allow\nusers to disclose the ir personal data in flexible ways. He received\nhis Ph.D. in Computer Scie nces from the University of\nWisconsin–Madison in 2007. Dr. Christodores cu holds 25 patents and\nhas published more than 35 papers in several inte rnational conferences\nand journals\, including the IEEE Symposium on Secu rity and Privacy\n(S&P)\, the ACM Conference on Computer and Communication s Security\n(CCS)\, the USENIX Security Symposium\, the Annual Computer Se curity\nApplications Conference (ACSAC)\, and many more. Faculty Host:  L imin\nJia\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91924c5 DTSTART;TZID=America/New_York:20250410T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250410T130000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:JPMorgan Chase Quantum Engineering Research Center Seminar CLASS:PUBLIC DESCRIPTION:Speaker: RUSLAN SHAYDULIN\, Head\, Quantum Engineering Research \, Global\nTechnology Applied Research Venter \, JPMorgan Chase\,\n\nTalk Title: Certified randomness using a trapped-ion quantum processor\n\nAltho ugh quantum computers can perform a wide range of practically\nimportant t asks beyond the abilities of classical computers\, realizing\nthis potenti al remains a challenge.  An example is to use an\nuntrusted remote device to generate random bits that can be certified\nto contain a certain amoun t of entropy. Certified randomnesshas many\napplications but is impossible to achieve solely by classical\ncomputation. \n\nHere we demonstrate the generation of certifiably random bits using\nthe 56-qubit Quantinuum H2-1 trapped-ion quantum computeraccessed over\nthe Internet. Our protocol lev erages the classical hardness of recent\nrandom circuit sampling demonstra tions: a client generates quantum\n‘challenge’ circuits using a small randomness seed\, sends them to\nan untrusted quantum server to execute an d verifies the results of the\nserver. We analyze the security of our prot ocol against a restricted\nclass of realistic near-term adversaries. \n\n Using classical verification with measured combined sustained\nperformance of 1.1 × 1018 floating-point operations per second across\nmultiple  su percomputers\, we certify 71\,313 bits of entropy under\nthis restricted a dversary and additional assumptions. Our results\ndemonstrate a step towar ds the practical applicability of present-day\nquantum computers. Referenc e Paper \n\n— \n\nRuslan Shaydulin is Head of Quantum Engineering Rese arch at the Global\nTechnology Applied Research center at JPMorgan Chase\, where he and his\nteam focus on practical aspects of evaluating quantum a lgorithmic\nspeedups and realizing them on hardware. Areas of responsibili ty of\nRuslan’s team include numerical benchmarking of quantum algorithm s\,\ncompilation and execution on quantum hardware\, compilation to\nfault -tolerant architectures and error correction. Prior to joining\nJPMorgan C hase\, Ruslan was a Maria Goeppert Mayer fellow at Argonne\nNational Labor atory.  \n\nFaculty Host: Umut Acar  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9192918 DTSTART;TZID=America/New_York:20250410T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250410T120000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Mengzhou Xia CLASS:PUBLIC DESCRIPTION:Speaker: MENGZHOU XIA\, Ph.D. Student\, Department of Computer Science\,\nPrinceton University\n\nTalk Title: Advancing the Pareto Fronti er of Open Language Models\n\nLarge language models (LLMs) have reshaped A I by enabling\nbreakthroughs in language understanding\, reasoning\, and d iverse\napplications. However\, their massive computational demands and th e\nproprietary nature of leading models hinder broad accessibility and\ncu stomization. My work addresses these challenges by optimizing the\nuse of existing compute\, data\, and models to push the Pareto frontier\nfor LLM training.  In doing so\, it not only produces stronger\nlanguage models b ut also offers universal approaches that support\neffective customization and advance our scientific understanding of\nmodel training. \n\nFirst\, I will discuss how structured pruning can be leveraged to\npre-train compa ct\, high-performing models at a fraction of the usual\npre-training cost\ , demonstrating its effectiveness in pushing the\nPareto frontier for gene ral-purpose pre-training. Next\, I will turn to\nthe post-training phase t o explore the critical role of data in\nshaping model behavior\, presentin g principled data optimization\ntechniques that enhance models’ capabili ties\, safety\, and\ntransparency—showing that “less is more” when i t comes to\nconstructing effective training datasets. Finally\, I will int roduce\nnovel post-training approaches that more effectively align languag e\nmodels with desired behaviors and objectives. By revealing gaps in the\ nreasoning abilities of even proprietary models\, I outline future\ndirect ions for building AI systems with enhanced reasoning\ncapabilities—focus ing on broader data synthesis through agentic\nprocesses and enabling adva nced applications. \n\n— \n\nMengzhou Xia is a final-year PhD student at Princeton University\,\nadvised by Prof. Danqi Chen. She develops effic ient training methods\nfor compact\, capable\, and open language models by optimizing the use\nof compute\, data\, and existing models\, enabling ea sy and effective\ncustomization. Her open-source models are widely used in the\ncommunity. Mengzhou is a recipient of the 2024 Apple Scholars in AI/ ML\nPhD Fellowship\, and the 2022 Bloomberg Data Science PhD Fellowship.\n She has also been awarded as a 2024 EECS Rising Star at MIT.\n\nFaculty Ho sts: Max Simchowitz\, Chris Donahue \n\nJoint Machine Learning Department and Computer Science Department In\nPerson and Zoom Participation.  See announcement. → Attendance at\nthis talk is restricted to members of the SCS community and relevant\nCMU stakeholders.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9192dde DTSTART;TZID=America/New_York:20250409T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250409T165000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:15-422/642 Machine Learning and Systems: Guest Lecture - Zihao Ye CLASS:PUBLIC DESCRIPTION:Speaker: ZIHAO YE\, Ph.D. Student in Computer Science\, Paul G. Allen\nSchool of Computer Science and Engineering\, University of Washing ton\n\nTalk Title: FlashInfer: Efficient and Customizable Kernel Generatio n\nfor LLM Inference Serving\n\nTransformers\, driven by attention mechani sms\, form the foundation of\nlarge language models (LLMs). As these model s scale up\, efficient GPU\nattention kernels become essential for high-th roughput and low-latency\ninference. Diverse LLM applications demand flexi ble and\nhigh-performance attention solutions. We present FlashInfer: a\nc ustomizable and efficient attention engine for LLM serving.\nFlashInfer ta ckles KV-cache storage heterogeneity using block-sparse\nformat and compos able formats to optimize memory access and reduce\nredundancy. It also off ers a customizable attention template\, enabling\nadaptation to various se ttings through Just-In-Time (JIT) compilation.\nAdditionally\, FlashInfer' s load-balanced scheduling algorithm adjusts\nto dynamism of user requests while maintaining compatibility with\nCUDAGraph which requires static con figuration. FlashInfer have been\nintegrated into leading LLM serving fram eworks like SGLang\, vLLM and\nMLC-Engine. Comprehensive kernel-level and end-to-end evaluations\ndemonstrate FlashInfer's ability to significantly boost kernel\nperformance across diverse inference scenarios: compared to\ nstate-of-the-art LLM serving solutions\, FlashInfer achieve 29-69%\ninter -token-latency reduction compared to compiler backends for LLM\nserving be nchmark\, 28-30% latency reduction for long-context\ninference\, and 13-17 % speedup for LLM serving with parallel\ngeneration. \n\n— \n\nZihao Y e is a fifth year PhD student at University of Washington\,\nadvised by Lu is Ceze and Tianqi Chen. Zihao joined the catalyst group\nat CMU starting from 2025 as a visiting researcher\, and his research\ninterest focuses o n machine learning systems. Faculty Host: Tianqi\nChen\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9193275 DTSTART;TZID=America/New_York:20250409T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250409T140000 LOCATION:Gordon Bell Conference Room\, Gates Hillman 5117 SUMMARY:Doctoral Thesis Oral Defense - Eric Mark Sturzinger CLASS:PUBLIC DESCRIPTION:Speaker: ERIC MARK STURZINGER\, Ph.D. Candidate\, Computer Scie nce\nDepartment\, Carnegie Mellon University\n\nTalk Title: Survival-Criti cal Machine Learning\n\nAutonomous systems must be able to survive in adve rsarial or hostile\nenvironments where threats evolve and morph.  Under c onditions in\nwhich a class of adversarial agents is novel but rare\, thes e systems\nmust rapidly learn and adapt.  We introduce Survival-Critical Machine\nLearning (SCML)\, a new ML paradigm that defines how autonomous s ystems\nthat rely on machine learning can negotiate such adversarial\nenvi ronments.  Inspired by the ability of a biological entity's\nimmune syste m to develop defenses against new viruses\, SCML systems\nleverage the wor kflow of Live Learning to iteratively improve ML\nmodels for threat detect ion.\n\nBeyond the conceptualization of SCML\, the main contributions of t his\ndissertation are an analytical model\, a prototype implementation\, a nd\nexperimental results of the SCML design tradeoff space.  We evaluate\ nthe impact on survivability of the various design parameters and\ndemonst rate the intimate relationship between SCML and Live\nLearning.  Notably\ , we evaluate the impact of the availability of\nfinite countermeasures (C Ms)\, the CM deployment threshold\, the number\nof deployed systems\, and the average threat arrival rate\, among\nothers\, on the probability of su rvival of a given mission duration. \nAdditionally\, we model SCML as a M arkov Decision Process (MDP) to\ndemonstrate how it can be analyzed within existing\, well-understood ML\nframeworks such as MDPs and Reinforcement Learning (RL).\n\nOur experimental results confirm that learning can indee d improve\nsurvivability in an SCML system.  It further shows that the CM \ndeployment threshold and the number of available CMs have a\nsignificant impact on survivability.  Allowing flexibility in the CM\ndeployment thr eshold during the mission enhances such survivability\nunder most conditio ns.  Similarly\, Live Learning improves the\nprobability of mission succe ss by increasing the likelihood of\naccurately classifying actual threats (true positives) and decreasing\nthe likelihood of wasting CMs on non-thre ats (false positives).  By\ndefining an SCML MDP\, we also show how an SC ML system can optimally\nadjust its CM deployment threshold as a function of state\, defined by\nthe number of remaining CMs and the time until miss ion completion.\n\nThesis Committee\n\nMahadev Satyanarayanan (Chair)\n\nP admanabhan Pillai\n\nJeff Schneider\n\nRashmi Vinayak\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9193732 DTSTART;TZID=America/New_York:20250409T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250409T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20250409.html LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Hanna Komlos CLASS:PUBLIC DESCRIPTION:Speaker: HANNA KOMLOS\, Ph.D. Student in Theoretical Computer S cience\,\nComputer Science and Engineering\, Tandom School of Engineering\ , New\nYork University\n\nTalk Title: Online List Labeling: Near optimalit y through hiding and\nexploiting the history of operations This talk focus es on the\nlist-labeling problem\, a fundamental data structural primitive with a\nlong history of both theoretical study and practical application. List\nlabeling captures the basic task of storing a dynamically changing set\nof up to N elements in sorted order in an array of size M=Θ(N). The\ ngoal is to support online insertions and deletions while moving\nelements around in the array as little as possible. Despite over four\ndecades of study\, there has until recently been a longstanding gap\nbetween the uppe r and lower bounds for list labeling (O(log^2(N)) and\nΩ(log(N)) element moves per operation\, respectively). \nTwo recent breakthroughs have clos ed this gap to within\npolylogarithmic factors. In a paper at FOCS 2022\, we give a randomized\nalgorithm which uses the traditional data structural security property\nof history independence to hide information from an ob livious\nadversarial inserter in order to achieve an improved O(log^1.5(N) )\nworst-case expected cost. In a FOCS 2024 paper\, we build a randomized\ npredictor of future operations from a workload’s history that\nperforms well on all inputs in expectation. Combining this with a\nprediction-augm ented algorithm\, we achieve a further improved\nworst-case upper bound of Õ(log(N)). To the best of our knowledge\,\nthis is the first example of an application of the\nalgorithms-with-predictions framework that uses an internal predictor\nto achieve a better worst-case upper bound.\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9193af3 DTSTART;TZID=America/New_York:20250408T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250408T130000 LOCATION:Gates Hillman 6115 SUMMARY:Doctoral Thesis Proposal - Wan Shen Lim CLASS:PUBLIC DESCRIPTION:Speaker: WAN SHEN LIM\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Database Gyms: Towards A utonomous Database Tuning\n\nDatabase management systems (DBMSs) are the f oundation of modern\ndata-intensive applications. But as more features are developed to\nsupport new workloads\, they become increasingly complex an d difficult\nto configure. Decades of research on autonomous DBMS configur ation\nhave largely produced advisory tools that still rely on human\nexpe rtise for their deployment into database tuning pipelines. Using\nthese to ols involves a multi-step process where a human operator (1)\ndetermines a n optimization objective\, (2) selects a suitable tool to\nimprove the obj ective\, (3) sets up and configures the DBMS to run a\nparticular workload \, (4) runs the workload to collect telemetry\, (5)\nuses the collected te lemetry to calibrate the tool\, and (6) operates\nthe tool to obtain recom mendations\, which the operator must then\nreview and apply. Because of th e ad-hoc nature of these pipelines\,\nthey require significant human effor t to set up\, extend\, and deploy.\nMoreover\, these tools are difficult t o compose and swap.\n\nThis proposal presents the database gym\, an integr ated framework that\nsystematizes and automates the DBMS configuration pip eline. The gym\neliminates repetition in the setup and operation of such p ipelines by\nproviding a set of reusable\, interoperable\, and interchange able\ncomponents that simplify the development and integration of ML-drive n\nDBMS configuration tools. It leverages its complete control over the\nt uning process to enable optimizations that require end-to-end\nknowledge. First\, it eliminates step-level repetition by skipping over\nredundant co mputation during telemetry collection to reduce the\nlatency of the tuning pipeline. Next\, it eliminates pipeline-level\nrepetition by reusing past experience to improve tool performance. For\nexample\, it adapts models o f DBMS behavior to account for how operator\nsemantics differ across DBMS versions. \n\nWe propose to extend our preliminary work by developing a t ool for\nDBMS upgrades that uses version-aware models to predict performan ce\nimprovements and regressions\, addressing another database\nadministra tion task with significant human involvement. Lastly\, we\nwill leverage r ecent advances in agentic artificial intelligence to\norchestrate tools on behalf of a human operator. These efforts will\ntransform the database gy m from a platform for developing and\ndeploying DBMS configuration tools i nto an autonomous database\nadministrator for production environments. \n \nThesis Committee\n\nAndrew Pavlo (Chair)\n\nJignesh Patel\n\nDavid Ander sen\n\nLin Ma (University of Michigan)\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9193fb6 DTSTART;TZID=America/New_York:20250408T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250408T170000 URL:https://scsdean.cs.cmu.edu/foundersday/index.html LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Founders Day (2025) CLASS:PUBLIC DESCRIPTION:Since 2012\, SCS Founders Day  is our special day to honor tho se\nfaculty and staff members who through their work best exemplify the\ni deals made so important to us by Allen Newell\, Hebert A. Simon and\nAlan J. Perlis. At SCS Founders Day\, we present the Allen Newell Award\nfor Re search Excellence\, the Herbert A. Simon Award for Teaching\nExcellence in Computer Science\, the Alan J. Perlis Award for\nImagination in Computer Science\, and staff recognition awards. We also\ncelebrate and pay tribute to faculty and staff who have been\nrecognized for their work and contrib utions beyond SCS.   \n\nJoin us for this special day. It's let's us co me together as a\ncommunity and reflect on our achievements and the future ahead for the\nSchool of Computer Science.  \n\nAgenda includes:\n\nInt roduction of new FacultyFaculty recognition awardsStaff recognition\naward s\n\nProgram Website\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9194347 DTSTART;TZID=America/New_York:20250407T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250407T173000 URL:https://db.cs.cmu.edu/events/sql-death-oxql-oximeter-query-language/ LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Benjamin Naecker CLASS:PUBLIC DESCRIPTION:Speaker: BEN NAECKER\, Software Engineer\, Oxide Computer Compa ny\n\nTalk Title: OxQL: Oximeter Query Language\n\nOxide Computer Company builds private cloud computers–co-designing\nhardware and software that works together. Our choice to own so much\nof the design was to enable sys temic control and observability: we\ncollect data from the smallest hardwa re sensor to the distributed\ncontrol plane software. \n\nThis talk cover s the Oximeter Query Language (OxQL)\, the\ndomain-specific language used to query and analyze this telemetry\ndata\, both for Oxide engineers and a lso external customers. The choice\nto build a custom query language in Ox QL was not taken lightly –\ngiven the choice of SQL or death\, we natura lly prefer SQL! Our query\nsystem did in fact initially expose a SQL inter face\, but issues around\nperformance\, efficiency\, and expressiveness ul timately led us to\nreconsider. OxQL is the result. The language includes analysis methods\ntailored to timeseries data\; an expressive\, pipe-based syntax making\nboth queries and the language itself easy to modify\; and a clear\ninterface boundary that lets Oxide incrementally improve the lang uage\nas needs evolve. \n\nThis talk will explore the underlying data mod el\, syntax\, semantics\,\nand query engine of OxQL. We also present some of the criteria we used\nfor evaluating the “choice” of SQL or death\, which may help others\nfacing the same question to understand when a spec ial-purpose DSL is\nappropriate. \n\n— \n\nBenjamin Naecker is a softw are engineer at Oxide Computer Company\,\nwhere he works primarily on the distributed control plane\, telemetry\nsubsystem\, and networking. In the past\, he has worked on\nhigh-performance scientific computing to improve the quality and\nreduce the duration of MRI scans for preventative medicin e. He\nreceived his PhD in computational neuroscience from Stanford\nUnive rsity\, where he built large-scale realtime data recording and\nvisualizat ion software for neuroscientific experiments. \n\nThis talk is part of th e SQL or Death? Seminar Series\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91947ca DTSTART;TZID=America/New_York:20250407T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250407T150000 LOCATION:Newell-Simon 3305 SUMMARY:Machine Learning Seminar - Ke Li CLASS:PUBLIC DESCRIPTION:Speaker: KE LI\, Assistant Professor and Visual Computing Chair \, School\nof Computing Science \, Simon Fraser University\n\nTalk Title: The Devil Is in the Gaps: How (Not) To Interpolate Between\nDiscrete Point s\n\nWhat do diffusion models/flow matching\, Gaussian splatting and\neffi cient transformer architectures have in common? Under the hood\,\nthey all turn a discrete set of points into a function defined\neverywhere. In the case of diffusion models/flow matching\, the points\nare training data po ints\, and the function is the probability density.\nIn the case of Gaussi an splatting\, the points are splat centres and\nthe function is the volum e density. In the case of efficient\ntransformers\, the points are keys us ed by attention and the function\nis the mapping from query to attention w eights.  It turns out that\nhow gaps between points are filled in is crit ical.\n\nIn this talk\, I will show how seemingly innocent choices made in \npopular techniques give rise to profound consequences. Such choices\nmak e diffusion models/flow matching data-hungry and slow to sample\nfrom\, Ga ussian splats hard to move and edit\, and hashing-based\nefficient transfo rmers error-prone. To address these issues\, I will\ngive an overview of t hree methods my lab developed\, Implicit Maximum\nLikelihood Estimation (I MLE)\, Proximity Attention Point Rendering\n(PAPR) and IceFormer\, and sho w applications in few-shot image\nsynthesis\, trajectory prediction\, visu omotor policy learning\, novel\nview synthesis\, 3D shape and albedo editi ng\, scene interpolation and\nlanguage modelling.  \n\n— \n\nKe Li is an Assistant Professor and Visual Computing Chair in the\nSchool of Compu ting Science at Simon Fraser University (SFU)\, where he\ndirects the APEX Lab. Ke’s current research interests are on\ngenerative models\, neural rendering\, efficient transformers and\nreinforcement learning. Ke was pr eviously a Research Scientist at\nGoogle and a Member of the Institute for Advanced Study (IAS) in\nPrinceton\, and received his Ph.D. from Universi ty of California\,\nBerkeley and B.Sc. from the University of Toronto.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9194c16 DTSTART;TZID=America/New_York:20250405T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250405T193000 URL:https://www.givecampus.com/schools/CarnegieMellonUniversity/events/steh lik-leaves-his-mark-celebration LOCATION:Simmons Auditorium\, Tepper Building SUMMARY:Celebrating Mark Stehlik @ Spring Carnival 2025 CLASS:PUBLIC DESCRIPTION:Mark Stehlik\,  University Teaching Professor\, Assistant Dean for\nOutreach\, Director of the CSD Undergraduate Program and Co-founder\ nof CMU CS Academy\, will be honored by the School of Computer Science.\n Alumni \, please join us in recognizing Mark and his incredible\n42-year career at CMU. CMU launched its undergraduate computer science\ndegree pro gram in 1989\, graduating its first cohort of 75 undergrads\nin 1992. Mark Stehlik advised them all. While countless things have\nchanged since then \, one factor has remained constant: Mark.   \n\nFor 35 years\, he has   advised students\, helping them navigate the\noccasionally rough waters between high school and college graduation.\nClose to 4\,000 advisees (in Pittsburgh and in Doha\, Qatar) have\nbenefited from his dedicated guidan ce and support. And by this May\,\nhe'll celebrate the graduation of his f ourth advisee whose parent(s)\ncalled Mark a mentor\, advisor and friend.   Mark has also taught\nthousands of students in a variety of 100- and 20 0-level computer\nscience courses since he became faculty in 1982. \n\nDr inks\, food and camaraderie to be provided.\n\nREGISTER → by 31 March 2 025\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9194fd5 DTSTART;TZID=America/New_York:20250404T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250404T200000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:The Nineteenth Annual SIGBOVIK Conference (take 2) CLASS:PUBLIC DESCRIPTION:Speaker: Calling all Students\, Faculty\, Staff\, and Sentient AIsThe\nNineteenth Annual SIGBOVIK Conference will take place at Carnegie\ nMellon University! The Special Interest Group on Harry Q. Bovik\n(SIGBOVI K) is a multidisciplinary conference dedicated to lesser-known\nresearch a reas neglected by mainstream conferences\, such as:\nComplexity practice H ighlight syntaxing Multi-armed philanthropists\nConnotational semantics Sc ience computers Data of unusual size And\nmany more!We hope to see you all there for a celebration of the\ngreatest scientific minds around.\nThe fi rst-extended submission deadline is March 29. Everyone is\nencouraged to s ubmit\, regardless of affiliation\, discipline\,\nrationality\, or goodwil l. Authors of accepted papers are encouraged\n(but not required) to give a five minute presentation of their work.\nFor those not in the know  —   SIGBOVIK is an evening of\ntongue-in-cheek academic presentations. If w e have a goal\, it's to\npoke fun at our fields and provide a venue for si lly ideas with\n(often\, but not always) serious executions. SIGBOVIK has both\npublished proceedings and live talks\, and everyone is welcome to\np articipate. All subjects are welcome\, although our audience is\nprimarily made up of computer scientists. The best way to get a feel\nfor SIGBOVIK is to look at our past proceedings.\nFurther disinformation will eventuall y be available at our website\n(yeah\, under construction)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919538b DTSTART;TZID=America/New_York:20250404T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250404T133000 URL:https://safety21.cmu.edu/events/ LOCATION:Scott Hall 6002 and Remote Access SUMMARY:Smart Safety Connection Seminar - Rahul Mangharam & Shinpei Kato CLASS:PUBLIC DESCRIPTION:Speaker: RAHUL MANGHARAM and SHINPEI KATOTalk Title: Reimaginin g\nIntelligent Vehicles: The Art & Technology of Open Source\n\nSpeakers\n \nRahul Mangharam — Professor\, Department of Electrical & Systems\nEngi neering\, University of PennsylvaniaShinpei Kato — Founder and\nChief Ex ecutive Officer\, TIER IV\n\nAutonomous control and decision systems are f orming the basis for\nsignificant pieces of our nation’s critical infras tructure. In\nparticular\, Autonomous Vehicles (AVs) present direct and ur gent\nsafety-critical challenges. About 10 years ago\, the self-driving ca r\nindustry made rosy predictions on widespread adoption of the\ntechnolog y by 2020\; however\, this has not come to pass due\, in large\npart\, to concerns about safety and robustness of the technology while\noperating in complex scenarios and the underestimation of the true\ndifficulty of fiel ding robust perception and control systems in the\nreal world. Much of tod ay’s research on AVs is limited to Big Tech\nexperimentation on expensiv e commercial vehicles that require large\nteams with diverse skills and po wer-hungry platforms. \n\nThis talk presents Autoware\, an open-source so ftware project that\nprovides full-stack capabilities of autonomous drivin g\, alongside the\nlessons learned on building intelligent vehicles. Autow are is deployed\nin commercially-successful robotaxis in Tokyo\, buses in Istanbul and\nMichigan\, cargo vehicles for logistics and Indy Autonomous Racing\nvehicles. The Autoware Foundation has over 80+ industry and academ ic\nconsortium partners who collaboratively develop the AV software and\nh ardware stack. We will walk through what it takes to make open-source\nsof tware successful for future mobility. \n\nREGISTER →  required.\n\n→   Due to space constraints\, there will be a limit for in-person\nattenda nce with a virtual option.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91957c9 DTSTART;TZID=America/New_York:20250404T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250404T140000 LOCATION:Tepper Building 1403 SUMMARY:CMU FLAME Center Seminar - Hao Zhu CLASS:PUBLIC DESCRIPTION:Speaker: HAO ZHU\, Postdoctoral Researcher\, Computer Science\n Department\, Stanford University\n\nTalk Title: Ushering AI Agents to an O pen Social World\n\nUnlike frontier AI models trained on static datasets\, humans learn\nthrough dynamic interactions with other people and the worl d. This\nfundamental difference in learning methodology not only makes lan guage\nagents less sample-efficient than humans but also introduces\nsigni ficant risks when these agents are deployed to interact with real\nhumans in the real world. Building agents that can efficiently learn\nthrough int eraction with other agents\, humans and the world is a\nchallenging proble m. In this presentation\, I will outline three\nfoundational approaches we 've developed to address this challenge:\n\nLearning through exploration o n the internet (NNetNav-live) — We\ndeploy an open-ended agent (without explicit task instructions) to\nexplore the web\, gather experience and re troactively label and train\non the data.Learning from human normative dec ision-making (EgoNormia)\n— We explore methods for agents to observe and internalize social\nnorms in physical interactions through crowd-sourced annotation with\ncontext perturbation.Learning to build metrics from human feedback\n(AutoLibra\, in prep) — We present a framework for automatica lly\nbuilding behavior evaluation metric systems that help both humans\nun derstand agent performance\, and agents improve the policy based on\nhuman feedback.\n\nThese complementary approaches offer a path toward creating AI agents\nthat can more effectively learn\, adapt\, and integrate into ou r open\nsocial world. \n\n— \n\nHao Zhu is a postdoctoral researcher i n the Computer Science\nDepartment at Stanford University. He finished his PhD last summer\nfrom CMU LTI. He is interested in AI agents\, human-agen t interaction\,\nrobotics and embodied AI\, and what AI agents tell us abo ut human\nsocial and embodied cognition. \n\nPizza Lunch Provided\n\nIn P erson and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9195c5a DTSTART;TZID=America/New_York:20250404T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250404T130000 URL:https://www.cylab.cmu.edu/events/2025/04/04-seminar-yang.html LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom (new date) SUMMARY:CyLab Seminar — Yaoquing Yang CLASS:PUBLIC DESCRIPTION:Speaker: YAOQUING YANG\, Assistant Professor\, Department of Co mputer\nScience\, Dartmouth College\n\nTalk Title: Weight Matrix Diagnosti cs for Improved Neural Network\nTraining and Compression\n\nThis talk will introduce several useful metrics derived from studying\nthe heavy-tail ph enomenon in neural network weight matrices. I will\nbegin by motivating th ese metrics through random matrix theory and\ndiscussing their connection to recent studies on feature learning. I\nwill then demonstrate how these metrics can be applied to various\nneural network applications\, including layer-wise pruning of large\nlanguage models\, tuning training and fine-t uning learning rates\,\ntraining scientific machine learning models with l imited data\,\nadjusting the architectural hyperparameters of LoRA network s\, and\nmodel selection on Hugging Face Transformers without access to\nt raining or testing data. In a recent study\, we show that these\ntheory-dr iven metrics can be scaled to prune large language models\nwith up to 65 b illion parameters\, outperforming some of the latest\npruning methods. \n \n— \n\nMy current research focuses on diagnosing and mitigating failur es in\nmachine learning models. For example\, I analyze shape and geometri c\nfeatures in high-dimensional spaces\, such as loss landscapes\, weight\ nmatrix spectral densities\, and decision boundaries\, to provide\nactiona ble insights for addressing common failure modes in these\nmodels. I also apply these techniques to applications such as 3D point\nclouds and graphs . My research draws inspiration from statistical\nlearning and information theory. \n\nFaculty Host:  Pulkit Grover \n\nRescheduled from 31 March 2025.\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9196092 DTSTART;TZID=America/New_York:20250403T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250403T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - Adam O'Neill CLASS:PUBLIC DESCRIPTION:Speaker: ADAM O'NEILL\, Assistant Professor\, Manning College o f\nInformation and Computer Sciences\, University of Massachusetts Amherst \n\nTalk Title: On The Tight Security of (Threshold) Schnorr Signatures\n\ nWe show that the widely-used Schnorr signature scheme meets\nexistential unforgeability under chosen-message attack (EUF-CMA) in\nthe random oracle model (ROM) if the circular discrete-logarithm (CDL)\nassumption holds in the underlying group. CDL is a new\,\nnon-interactive and falsifiable var iant of the discrete-log assumption\nthat we introduce. Our reduction is c ompletely tight\, meaning the\nconstructed adversary against CDL has essen tially the same running\ntime and success probability as the assumed forge r. This serves to\njustify the size of the underlying group used in practi ce. \n\nTo our knowledge\, we are the first to exhibit such a reduction.\ nIndeed\, prior results required interactive and non-falsifiable\nassumpti ons (Bellare and Dai\, INDOCRYPT 2020) or additional idealized\nmodels lik e the algebraic group model (Fuchsbauer et al.\, EUROCRYPT\n2020). We then extend our result to threshold Schnorr signatures. In\nparticular\, we sh ow that Sparkle+ (Crites et al.\, CRYPTO 2023) is\ntightly secure under st atic corruptions assuming CDL. Finally\, we\njustify CDL by showing it hol ds in two carefully-chosen idealized\nmodels that idealize different aspec ts of the assumption. I\n\nn Person and Zoom Participation.  See announce ment.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9196466 DTSTART;TZID=America/New_York:20250403T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250403T180000 URL:https://www.cmu.edu/dietrich/statistics-datascience/events/index.html LOCATION:Poster Hall A35 SUMMARY:Morris H. DeGroot Memorial Lecture - Eric Tchetgen Tchetgen CLASS:PUBLIC DESCRIPTION:Speaker: ERIC J. TCHETGEN TCHETGEN\, University Professor\, Pro fessor of\nBiostatistics in Biostatistics and Epidemiology\, Professor of\ nStatistics and Data Science\, The Wharton School\, University of\nPennsyl vania\n\nTalk Title: On Identification in the Binary Instrumental Variable \nModel: Introducing the NATE and Beyond\n\nWe revisit the identification problem in the canonical binary\ninstrumental variable model. Our work rev eals new conditions for the\nclassical Wald ratio estimand to be endowed w ith a nonparametric\ncausal interpretation. Specifically\, we describe a s traightforward set\nof conditions under which the Wald Ratio point identif ies the Nudge\nAverage Treatment Effect (NATE)\, defined as the average ca usal effect\nfor the subgroup of units whose treatment can be manipulated by the\ninstrument\, a sub-group referred to as Nudge-able. Crucially\, th e\nNudge-able may include both compliers and defiers therefore obviating\n the need for the standard no-defier condition known to identify the\nLocal Average Treatment Effect (LATE). \n\nOur key identification condition fo r the NATE is that any variability\nof the treatment effect induced by a h idden counfounder must be is\nuncorrelated with corresponding variability in the share of compliers\namong the Nudge-able. An important and easily i nterpretable sufficient\ncondition for this assumption is that\, although a priori unrestricted\,\nthe share of compliers within the subgroup of Nud ge-able units is\nbalanced across strata of the unmeasured confounders. Im portantly\,\nmonotonicity is recovered as a degenerate case where the nudg e-able\nare all compliers\, thus ruling out the existence of defiers\, in which\ncase the NATE matches the LATE. \n\nCrucially\, the Wald ratio ret ains a nonparametric causal\ninterpretation as the NATE under the proposed identification\ncondition\, even when monotonicity does not hold\,therefo re providing a\ncausal interpretation under weaker conditions than previou sly\navailable. Various generalizations of the results will be given\,\nin cluding new straightforward conditions for identification of the\naverage treatment effect for the treated by a generalized Wald ratio\nestimand\, t ogether with newquasi-IV identification results with an\nimperfect instrum ent which violates the exclusion restriction\nassumption. \n\n— \n\nEr ic Tchetgen Tchetgen’s primary area of interest is in\nsemi-parametric e fficiency theory with application to causal\ninference\, missing data prob lems\, statistical genetics\, and mixed\nmodel theory. In general\, he wor ks on the development of statistical\nand epidemiologic methods that make efficient useof the information in\ndata collected by scientific investiga tors\, while avoiding unnecessary\nassumptions about the underlying data g enerating mechanism. Dr.\nTchetgen Tchetgen received his PhD from Harvard University. He is a\nco-winner of the 2022 Rousseeuw Prize for Statistics and was awarded\nthe Myrto Lefkopoulou Distinguished Lectureship in 2020.  \n\nThe lecture is funded in part by the friends and colleagues of Morri s\nH. DeGroot\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9196955 DTSTART;TZID=America/New_York:20250403T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250403T160000 URL:https://aco.math.cmu.edu/abs-24-25/apr3.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Eric Katz CLASS:PUBLIC DESCRIPTION:Speaker: ERIC KATZ\, Associate Professor\, Department of Mathem atics\,\nOhio State University\n\nTalk Title: Invariants of Lattice Polyto pes and Matroids\n\nWe will discuss invariants of lattice polytopes and th eir subdivisions\narising from Ehrhart and Hodge theory and introduce thei r matroid\ntheoretic analogues which are enriched versions of the characte ristic\nand Tutte polynomials. \n\n4:00 pm →  Jane Street sponsors tea and cookies\, Math Lounge\n(please bring your own mug if possible).\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9196c53 DTSTART;TZID=America/New_York:20250403T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250403T133000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:Teruko Yata Memorial Lecture in Robotics - Katerina Fragkiadaki CLASS:PUBLIC DESCRIPTION:Speaker: KATERINA FRAGKIADAKI\, JPMorgan Chase Associate Profes sor of\nComputer Science\, Machine Learning Department\, Carnegie Mellon\n University\n\nTalk Title: Learning World Simulators from Data\n\nModern fo undational models have achieved superhuman performance in\nmany logic and mathematical  reasoning tasks by learning to think\nstep by step.  Howe ver\, their ability to understand videos\, and\,\nconsequently\, control e mbodied agents\, lags behind. They often make\nmistakes in recognizing sim ple activities\, and often hallucinate\nwhen  generating videos. This rai ses a fundamental question: What is\nthe equivalent of thinking step-by-st ep for visual recognition and\nprediction?\n\nIn this talk\, we argue that step-by-step visual reasoning has much to\ndo with inverting a physics si mulator\, that is\, mapping raw video\npixels back to a structured\, 3D-li ke neural representation of the\nworld. This  involves inferring 3D neura l representations of \nobjects\, parts\, their 3D motion and appearance t rajectories\,\nestimating camera movements and 3D scene structure and phys ics\nproperties.  We will discuss methods to automatically extract such 3 D\nneural representations from images and videos using generative model\np riors and end-to-end feed-forward models. We will present methods\nthat in ject such knowledge of camera motion and 3D scene structure in\nmodern VLM s and show it improves their ability to ground language and\ncontrol robot manipulators.\n\nHow can we scale up annotations for such simulator inver sion? We will\ndiscuss methods that use generative models  of language an d vision to\nautomate development of 3D simulations in physics engines.\nA dditionally\, we will discuss our efforts in developing faster and\nmore g eneral physics engines. The integration of physics engines with\ngenerativ e models aims to automate the replication of real physical\nenvironments w ithin the physics simulator\, enabling more accurate and\nscalable world s imulation data for sim-to-real learning of 3D\nperception and action. We b elieve such real-to-sim and sim-to-real\nlearning paradigms are very hopef ul for developing robots that can see\nand think accurately\, step-by-step .\n\n—\n\nKaterina Fragkiadaki is the JPMorgan Chase Associate Professor in the\nMachine Learning Department in Carnegie Mellon University. She\nr eceived her undergraduate diploma from Electrical and Computer\nEngineerin g in the National Technical University of Athens. She\nreceived her Ph.D. from University of Pennsylvania and was a\npostdoctoral fellow in  UC Ber keley and Google research after that.\nHer work focuses on combining forms of common sense reasoning\, such as\nspatial understanding and 3D scene u nderstanding\, with deep visuomotor\nlearning. The goal of her work is to enable few-shot learning and\ncontinual learning for perception\, action a nd language grounding. Her\ngroup develops methods for computer vision for mobile agents\, 2D and\n3D visual parsing\, 2D-to-3D perception\, vision- language grounding\,\nlearning of object dynamics\, navigation and manipul ation policies.\nPioneering innovations of her group's research include 2D -to-3D\ngeometry-aware neural networks for 3D understanding from 2D video\ nstreams\, analogy-forming networks for memory-augmented few-shot visual\n parsing\, and language-grounding in 2D and 3D scenes with bottom-up and\nt op-down attention. Her work has been awarded with a best Ph.D. thesis\nawa rd\, an NSF  CAREER award\, AFOSR Young Investigator award\, a DARPA\nYou ng Investigator award\, Google\, TRI\, Amazon\, UPMC and Sony faculty\nres earch awards. She is a program chair for ICLR 2024. \n\n— \n\nAbout th e Lecture: The Yata Memorial Lecture in Robotics is part of\nthe School of Computer Science Distinguished Lecture Series.\nTeruko Yata was a pos tdoctoral fellow in the Robotics Institute from\n2000 until her untimely d eath in 2002. After graduating from the\nUniversity of Tsukuba\, working u nder the guidance of Prof. Yuta\, she\ncame to the United States. At Carne gie Mellon\, she served as a\npost-doctoral fellow in the Robotics Institu te for three years\, under\nChuck Thorpe. Teruko’s accomplishments in th e field of ultrasonic\nsensing were highly regarded and won her the Best S tudent Paper Award\nat the International Conference on Robotics and Automa tion in 1999. It\nwas frequently noted\, and we always remember\, that “ the quality of\nher work was exceeded only by her kindness and thoughtfuln ess as a\nfriend.” Join us in paying tribute to our extraordinary collea gue\nand friend through this most unique and always exciting lecture.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9197287 DTSTART;TZID=America/New_York:20250402T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250402T130000 LOCATION:Mehrabian Collaborative Innovation Center 1301 SUMMARY:Doctoral Speaking Skills Talk - Pratap Singh CLASS:PUBLIC DESCRIPTION:Speaker: PRATAP SINGH\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Owl: Bringing Verified C ryptographic Protocols to Practice\n\nCryptographic security protocols\, s uch as TLS or WireGuard\, form the\nfoundation of a secure Internet\, but vulnerabilities are discovered in\nthem with alarming frequency. Formal ve rification promises a\nfoundational solution to this problem\, with signif icant prior work on\nmechanizing cryptographic proofs of high-level protoc ol designs\, and\nmore recent work on verifying particular implementations of single\nprotocols against higher-level specifications. \n\nIn this ta lk\, I will discuss our work on Owl\, a verifier and secure\ncompiler for cryptographic protocols. Owl's verifier uses a novel\ncombination of infor mation-flow and refinement types to prove security\nin the computational m odel\, enabling greater modularity and automation\nthan prior computationa l tools. Owl's compiler translates well-typed\nprotocols into performant\, interoperable\, side-channel resistant Rust\nlibraries that are automatic ally formally verified to be correct. With\nOwl\, developers can define an d prove security for their protocols in\nan intuitive\, high-level languag e\, and obtain for free a drop-in\nexecutable implementation with formal g uarantees of correctness and\nsecurity. \n\nWe evaluate Owl using a range of case studies. We verify the core\nlogic of 14 protocols\, including SS H and Kerberos. We additionally\ndevelop two large-scale\, interoperable c ase studies for WireGuard and\nHybrid Public-Key Encryption (HPKE)\, yield ing verified implementations\nthat interoperate with\, and match the perfo rmance of\, existing\nindustrial baselines on end-to-end benchmarks. \n\n Presented as part of the CyLab Student Seminar Series\n\nPresented in Part ial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91976a1 DTSTART;TZID=America/New_York:20250401T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250401T150000 LOCATION:Newell-SImon 4305 and Zoom SUMMARY:SCS Teaching Faculty Candidate - Dimitrios Myrisiotis CLASS:PUBLIC DESCRIPTION:Speaker: DIMITRIOS MYRISIOTIS\, Research Fellow\, French Nation al Centre\nfor Scientific Research (CNRS)\, Campus for Research Excellence and\nTechnological Enterprise (CREATE)\n\nTalk Title: LeetCode #793: A St ory of Two Algorithms\n\nIn this teaching demo\, we will explore two algor ithms for solving\nLeetCode problem #793. \n\n— \n\nDimitrios Myrisiot is is a Research Fellow at CNRS@CREATE LTD.\n(Singapore). He is mainly wor king on computational complexity theory\nand the foundations of machine le arning. Faculty Host: Iliano\nCervesato Joint Computer Science Department / CMU-Qatar In Person and\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91979c2 DTSTART;TZID=America/New_York:20250401T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250401T130000 URL:https://db.cs.cmu.edu/events/evolving-transactions-in-oracle-23ai-new-a dvances-in-concurrency-and-consistency LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Kisht Kumar and Akshay S. Kulkarni CLASS:PUBLIC DESCRIPTION:Speaker: KISHY KUMAR and AKSHAY S. KULKARNI\, KISHY - Director\ ,\nDatabase Transactions Group\, AKSHAY - Manager\, Data and Transactions\ nGroup\, Oracle\n\nTalk Title: Evolving Transactions in Oracle 23ai: New A dvances in\nConcurrency and Consistency\n\nOracle Database 23ai introduces groundbreaking enhancements that\nredefine transaction processing\, ensur ing greater concurrency\,\nflexibility\, and efficiency. This talk will ex plore four key features\nthat impact database transactions: Logical Replic ation of JSON\nRelational Duality Views\, Lock-Free Reservations\, Session less\nTransactions\, and Value-based Concurrency Control. Learn how we\nlo gically and consistently replicate data stored in JSON Relational\nDuality Views to heterogenous target data stores. The Lock-free\nReservations fea ture boosts concurrency of hot data\, enhancing system\nefficiency. Sessio nless Transactions enables lightweight and efficient\ntransactions in the database\, reducing complexity in building modern\napplications. Value-bas ed Concurrency Control introduces a novel\nconcurrency control approach th at increases parallelism by reducing\nthe locking duration. Join us as we dive into these innovations and\ntheir impact on transaction processing in Oracle Database 23ai. \n\n— \n\nThe two speakers:\n\nKishy Kumar is a Director at the Oracle Database Transactions Group.\nHe holds a Master’ s in Electrical and Computer Engineering with a\nfocus on Systems Software from Carnegie Mellon and Bachelor’s in\nElectronics and Communications Engineering from Thapar University\,\nIndia. Since graduating in 2013\, he has been with Oracle working in\ndifferent layers of the database and Exa data engineered system and is\ndeeply invested in advancing databases and storage systems. In his\nfree time\, he volunteers for CMU’s Alumni Asso ciation by organizing\nevents\, mentors tech startups at accelerators\, an d likes to spend time\nwith family\, work out and play basketball. \n\nAk shay S Kulkarni is a manager at Oracle in the Data and Transactions\nGroup . He graduated from Carnegie Mellon University with a dual\nMaster’s in Electrical & Computer Engineering and Technology\nInnovation Management\, and joined Oracle in 2017. At Oracle\, he works\non projects related to tr ansaction processing and flashback\ntechnologies. In his free time\, he li kes to read\, swim\, and play\nbadminton. \n\nZoom Participation.  See a nnouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9197e42 DTSTART;TZID=America/New_York:20250331T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250331T173000 URL:https://db.cs.cmu.edu/events/sql-death-starrocks-query-optimizer LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Kaisen Kang CLASS:PUBLIC DESCRIPTION:Speaker: KAISEN KANG\, Query Team Leader\, StarRocks\n\nTalk Ti tle: StarRocks Query Optimizer\n\nStarRocks is a C++ analytical database t hat handles large-scale\,\nhigh-concurrency\, low-latency OLAP queries. In this presentation\,\nKaisen Kang\, Tech Lead of the StarRocks Query Team\ , will provide an\nin-depth look at the StarRocks optimizer\, the key opti mizations\, and\ndesign choices. The talk will focus on two key areas:\n\n Key Optimizations – A deep dive into three representative\noptimizations : Multi-left join colocate optimization\, Partitioned\nMaterialized Views auto union rewrite\, and Low Cardinality global\ndictionary optimization\, showcasing how they enhance query\nperformance. Cost Estimation Challenge s & Improvements – Lessons\nfrom real-world production use cases and the enhancements made to\naddress them\, including Auto Analyze\, Query Feedb ack\, Adaptive\nExecution\, Predicate Columns\, and SQL Plan Manager.\n\nF inally\, Kaisen will share insights and experiences from building the\nopt imizer from the ground up over the past four years\, reflecting on\nkey de sign decisions and challenges. \n\nThis talk is part of the SQL or Death? Seminar Series\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91981f0 DTSTART;TZID=America/New_York:20250331T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250331T160000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Mihir Kiran Bala CLASS:PUBLIC DESCRIPTION:Speaker: MIHIR KIRAN BALA\, Ph.D. Candidate\, Computer Science\ nDepartment\, Carnegie Mellon University\n\nTalk Title: Towards Fully-Auto nomous Ultra-Light Drones\n\nAutonomous drones have emerged as an exciting new technology which\ncould revolutionize infrastructure inspection\, mil itary\nreconnaissance\, and police surveillance. However\, the vast majori ty of\ntoday’s platforms are heavy\, costly\, and difficult to operate. This\nrestricts them from use in many mission settings\, such as in densel y\npopulated environments\, where government regulation forbids autonomous \noperation of heavy drones near people. Much of this weight comes from\nt he onboard compute resources required for these drones to run the\ncritica l computer vision algorithms that provide situational\nawareness. \n\nIn this thesis oral\, I show how autonomy can be induced on lightweight\ndron es using edge computing\, offloading high compute jobs to a\nnetwork-proxi mal server. I demonstrate how this technique can lead to\nautonomous aircr aft that fly much closer to the FAAs regulatory limits\nat acceptable perf ormance cost. I also reveal a new operating system\ndesigned to unify the disparate landscape of drones under a single\,\neasy-to-program API. I sho w how this can be leveraged to create\nheterogeneous collaborative drone s warms on commercial off-the-shelf\nhardware. \n\nThesis Committee\n\nMaha dev Satyanarayanan (Chair)\n\nDavid O’Hallaron\n\nJeff Schneider\n\nPadm anabhan Pillai\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91985d5 DTSTART;TZID=America/New_York:20250331T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250331T140000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Asher James Trockman CLASS:PUBLIC DESCRIPTION:Speaker: ASHER JAMES TROCKMAN\, Ph.D. Candidate\, Computer Scie nce\nDepartment\, Carnegie Mellon University\n\nTalk Title: Mimetic Initia lization for Deep Neural Networks\n\nWhile neural network weights are typi cally initialized randomly from\nunivariate distributions\, pre-trained we ights often have\nvisually-discernible multivariate structure. We propose a technique\ncalled \"mimetic initialization\" that aims to replicate su ch\nstructures when initializing convolutional networks (CNNs)\,\nTransfor mers\, and State Space Models (SSMs). For CNNs\, we handcraft a\nclass of multivariate Gaussian distributions to initialize filters for\ndepthwise c onvolutional layers\; for Transformers\, we initialize the\nquery and key weights for self-attention layers such that their\nproduct approximates th e identity\; and for SSMs\, we initialize layers\nto approximate simple li near attention. Mimetic initialization\nsubstantially reduces training t ime and increases final accuracy on\nvarious common small-scale benchmarks .  \n\nOur technique enables us to almost close the gap between untraine d and\npre-trained Vision Transformers on small datasets like CIFAR-10\,\n achieving up to a 6% gain in accuracy through initialization alone.\nFor c onvolutional networks like ConvMixer and ConvNeXt\, we observe\nimprovemen ts in accuracy and reductions in training time\, even when\nconvolutional filters are frozen (untrained) after initialization. For\nSSMs\, mimetic  initialization substantially improves generalization\nabilities on synth etic language tasks like copying and associative\nrecall. Overall\, our fi ndings suggest that the benefits of\npre-training can be separated into tw o components: serving as a good\ninitialization and storing transferable k nowledge\, with the former\nbeing simple enough to (at least partially) ca pture by hand in\nclosed-form.  \n\nThesis Committee\n\nZico Kolter (Cha ir)\n\nAlbert Gu\n\nAditi Raghunathan\n\nSébastien Bubeck (OpenAI)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9198a53 DTSTART;TZID=America/New_York:20250327T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250327T180000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:Hank Suz-Chi Wan Distinguished Lecture: An SCS Distinguished Lectur e CLASS:PUBLIC DESCRIPTION:Speaker: JAIME TEEVAN\, Chief Scientist and Technical Fellow\, Microsoft\n\nTalk Title: The Future of Work is Conversations That Matter\n \nDecades of research into how computing can support work have taught us\n that work is inherently purposeful\, persistent\, and collaborative. But\n now that people can communicate with computers via natural language\,\nwe must re-examine what we know about each of these fundamental\nelements. \ n\nThis talk will explore how AI can transition from merely executing\ntas ks to unlocking human potential\; from storing knowledge in\ndocuments to sustained knowledge exchanges\; and from individual\nefforts to coordinate d collective intelligence. While AI has already\ndemonstrated its ability to enhance workplace efficiency\, early\ninsights from its use suggest tha t the most profound impact will come\nfrom how it transforms what we do. T he future of work will be shaped\nby our ability to engage in meaningful a nd interesting conversations\,\nfundamentally changing how we interact wit h technology and each\nother. \n\n— \n\nJaime Teevan is Chief Scientis t and Technical Fellow at Microsoft\,\nwhere she is responsible for drivi ng research-backed innovation in\nthe company's core products.  Jaime is world-renowned for her\nresearch into productivity and personalized search \, and was\nrecognized by TIME as one of the top 100 people to play a n\ninstrumental role in AI development and societal advancement. She led\ nthe creation of M365 Copilot by integrating AI into Microsoft\nproducts\, invented the first personalized search algorithm used by\nBing\, and coor dinated the company's hybrid work research during the\npandemic. Previous ly she was Technical Advisor to Microsoft CEO Satya\nNadella. Jaime is an ACM Fellow and a member of the SIGIR and CHI\nAcademies. She holds a Ph.D. in AI from MIT and a B.S. from Yale\,\nand is an Affiliate Professor at the University of Washington.   \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9198ea7 DTSTART;TZID=America/New_York:20250327T143000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250327T170000 URL:https://aco.math.cmu.edu/seminar.html LOCATION:Wean Hall 8220 (special time) SUMMARY:ACO Seminar - Marcus Michelen and Shayan Oveis Gharan CLASS:PUBLIC DESCRIPTION:Speaker: MARCUS MICHELEN and SHAYAN OVEIS GHARANTalk Title: Two Talks\n\nThere will be two back-to-back talks during this week's seminar. \n\n2:30 pm\n\n► Marcus Michelen     \n\nAssistant Professor\, Depar tment of Mathematics\, Statistics\, and\nComputer Science\, University of Illinois Chicago      \n\n— New lower bounds for sphere packings an d independent sets via\nrandomness \n\nWe construct new lower bounds for sphere packings in high dimensions\nand for independent sets in graphs wit h not-too-large co-degrees. For\ndimension d\, this achieves a sphere pack ing of density (1 + o(1)) d\nlog d / 2(d+1). In general dimension this pro vides the first\nasymptotically growing improvement for sphere packing low er bounds\nsince Rogers' bound of c*d/2d in 1947. The proof amounts to a r andom\n(very dense) discretization together with a new theorem on\nconstru cting independent sets on graphs with not-too-large co-degree.\nBoth steps will be discussed and no knowledge of sphere packings will\nbe assumed or required. Central to the analysis is a nibble method. \n\nThis is based on joint work with Marcelo Campos\, Matthew Jenssen\nandJulian Sahasrabudh e.\n\n4:00 pm\n\n► Shayan Oveis Gharan      \n\nLazowska Professor of Computer Science & Engineering Computer Science\nand Engineering\, Univ ersity of Washington      \n\n— C-Lorentzian Polynomials\, Trickled own Thms\, Mixing Time and Log\nConcavity \n\nCompletely Log Concave\, a. k.a.\, Lorentzian polynomials\, were\ndiscovered a few years ago where the y were used to relate seemingly\ndistant areas of Math and CS such as geom etry of polynomials\, Hodge\ntheory for combinatorial geometries\, theory of high dimensional\nexpanders and mixing time of Markov chains. Consequen tly\, they lead to\na resolution of several long-standing open problems on matroids such\nas the Mason's log-concavity conjecture and the Mihail-Vaz irani\nconjecture on the expansion of the bases exchange graph.\nUnfortuna tely\, this family of polynomials are limited as their support\ncorrespond s to bases of a matroid or more generally vertices of a\ngeneralized permu tahedra. I will explain a generalization of\nLorentzian polynomial to conv ex cones in the positive orthant\, called\nC-Lorentzian polynomials\, and use them to study combinatorial objects\nsuch as distributive\, modular\, or geometric lattices and the\ncorresponding sampling\, and log-concavity problems. Enroute we will\nalso discuss new local-to-global theorems for h igh-dimensional\nexpanders called Trickledown theorems. \n\nBased on a jo int work with Jonathan Leake and Kasper Lindberg. 3:30 pm\n- Jane Street-s ponsored tea and cookies in the Math Lounge. Please\nbring your own mug if possible.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91993fb DTSTART;TZID=America/New_York:20250327T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250327T120000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:SCS Faculty Candidate - Rishabh Iyer CLASS:PUBLIC DESCRIPTION:Speaker: RISHABH IYER\, Postdoctoral Researcher\, Department of \nElectrical Engineering and Computer Sciences\, University of\nCalifornia \, Berkeley\n\nTalk Title: Performance Interfaces for Systems Software and Hardware\n\nSemantic interfaces—such as code documentation and\nspecifi cations—provide simple\, abstract descriptions of a system’s\nfunction ality\, enabling engineers to reason about and use the\nsystem’s functio nality without having to understand the\nimplementation. In contrast\, the re exist no equivalent interfaces for\nsystem performance\, despite perfor mance having become a first-class\ncitizen in system design. Popular appro aches to thinking about\nperformance\, like benchmarking\, profiling\, and deriving performance\nenvelopes\, provide incomplete visibility\, leading to frequent hiccups\nand meltdowns in production when the workload or run time environment\nchanges in unpredicted ways. \n\nIn this talk\, I will introduce the notion of a performance interface\nand describe two techniqu es that derive simple\, abstract performance\ninterfaces for systems softw are and hardware\, respectively. First\,\nCFAR\, which derives interfaces that enable precise reasoning about how\nsystems code uses the CPU cache. Then LTC\, which derives interfaces\nthat enable engineers to reason about the performance of hardware\naccelerators. The improved performance visib ility provided by CFAR and\nLTC has tangible benefits: for instance\, we u sed the CFAR-derived\ninterfaces to identify several cache-inefficient acc ess patterns and\nperformance bugs (including in the Linux kernel's TCP st ack) and the\nLTC-derived interfaces to speed up compilation for ML accele rators by\n5-41x. \n\n— \n\nRishabh Iyer is a postdoctoral researcher at UC Berkeley\, working with\nSylvia Ratnasamy and Scott Shenker. He rece ived his PhD from EPFL\nunder the supervision of George Candea and Katerin a Argyraki\, and his\nbachelor's degree from IIT Bombay.  Rishabh's resea rch focuses on\ndeveloping techniques that enable developers to reason abo ut the\nexpected performance of their systems before they are deployed in\ nproduction. His dissertation introduced the notion of performance\ninterf aces and was awarded the ACM SIGOPS Dennis M. Ritchie Award\, the\nEurosys Roger Needham PhD Award\, and the Dimitris N. Chorafas Award. \n\nFacult y Host:  Justine Sherry \n\nIn Person and Zoom Participation. \n\n→ A ttendance at this talk is restricted to members of the SCS\ncommunity and relevant CMU stakeholders. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91998e4 DTSTART;TZID=America/New_York:20250326T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250326T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Mirabel Reid CLASS:PUBLIC DESCRIPTION:Speaker: MIRABEL REID\, Ph.D. Student in Computer Science\, Sch ool of\nComputer Science\, Georgia Institute of Technology\n\nTalk Title: On k-Winners-Take-All as a Model of Neuron Communication\n\nUnderstanding how low-level neuron dynamics give rise to high-level\nemergent capabiliti es is one of the key open questions in\nneuroscience. In this talk\, I wi ll investigate discrete-time\,\nweighted digraph models of the connectome\ , which not only provide\ninsight into biological neural networks\, but al so form the backbone of\nmodern artificial neural networks. In particular\ , I focus on recurrent\nneural networks employing the k-winners-take-all (k-WTA) function as\na nonlinear gating function\, a mechanism which model s firing rate\nregulation via inhibition. The process proceeds as follows: at each\ntime step\, a subset of k neurons are assigned the value 1 (sa id to\nbe active at t). At the next time step\, the active set consists of\nthe k neurons with the largest weighted degree to the previous\nact ive set. By exploring the computational properties arising from\nthis mode l\, this talk aims to offer insights into the emergence of\nstructures in the connectome. \n\nAbout the Speaker\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9199c9a DTSTART;TZID=America/New_York:20250326T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250326T113000 LOCATION:Mehrabian Collaborative Innovation Center 2201 SUMMARY:Doctoral Speaking Skills Talk - Zhengyao Lin CLASS:PUBLIC DESCRIPTION:Speaker: ZHENGYAO LIN\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: FlowCert: Translation Va lidation for Asynchronous Dataflow\nvia Dynamic Fractional Permissions\n\n Coarse-grained reconfigurable arrays (CGRAs) have gained attention in\nrec ent years due to their promising power efficiency compared to\ntraditional von Neumann architectures. \n\nTo program these architectures using ordi nary languages such as C\, a\ndataflow compiler must transform the origina l sequential\, imperative\nprogram into an equivalent dataflow graph\, com posed of dataflow\noperators running in parallel.This transformation is ch allenging since\nthe asynchronous nature of dataflow graphs allows out-of- order\nexecution of operators\, leading to behaviors not present in the\no riginal imperative programs. \n\nIn this talk\, we address this challenge by developing a translation\nvalidation technique for dataflow compilers to ensure that the\ndataflow program has the same behavior as the original imperative\nprogram on all possible inputs and schedules of execution.We apply\nthis method to a state-of-the-art dataflow compiler targeting the\n RipTide CGRA architecture.\n\nOur tool uncovers 8 compiler bugs where the compiler outputs incorrect\ndataflow graphs\, including a data race that i s otherwise hard to\ndiscover via testing. After repairing these bugs\, ou r tool verifies\nthe correct compilation of all programs in the RipTide be nchmark\nsuite. \n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919a079 DTSTART;TZID=America/New_York:20250325T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250325T140000 LOCATION:Newell-SImon 3305 SUMMARY:Doctoral Speaking Skills Talk - Honghao Lin CLASS:PUBLIC DESCRIPTION:Speaker: HONGHAO LIN\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: A Strong Separation for A dversarially Robust L_0\nEstimation for Linear Sketches\nMathJax.Hub.Confi g({\ntex2jax: {\ninlineMath: [ ['$'\,'$']\, [\"\\\\(\"\,\"\\\\)\"] ]\,\npr ocessEscapes: true\n}\n})\;\n\nThe majority of streaming problems are defi ned and analyzed in a\nstatic setting\, where the data stream is any worst -case sequence of\ninsertions and deletions that is fixed in advance. Howe ver\, many\nreal-world applications require a more flexible model\, where an\nadaptive adversary may select future stream elements after observing\n the previous outputs of the algorithm. Over the last few years\, there\nha s been increased interest in proving lower bounds for natural\nproblems in the adaptive streaming model.\n\nIn this talk\, we will give the first kn own adaptive attack against\nlinear sketches for the well-studied $\\ell_0 $-estimation problem over\nturnstile\, integer streams. For any linear str eaming algorithm 𝒜\nthat uses sketching matrix $\\mathbf{A}\\in \\mathb b{Z}^{r \\times n}$\nwhere $n$ is the size of the universe\, this attack m akes\n$\\tilde{O}(r^8)$ queries and succeeds with high constant probabilit y\nin breaking the sketch. We will also give an adaptive attack against\nl inear sketches for the $\\ell_0$-estimation problem over finite fields\n$\ \mathbb{F}_p$\, which requires a smaller number of $\\tilde{O}(r^3)$\nquer ies. \n\nFinally\, we will provide an adaptive attack over $\\mathbb{R}^n $\nagainst linear sketches $\\mathbf{A} \\in \\mathbb{R}^{r \\times n}$ fo r\n$\\ell_0$-estimation\, in the setting where $\\mathbf{A}$ has all nonze ro\nsubdeterminants at least $\\frac{1}{\\text{poly}(r)}$. Our results\npr ovide an exponential improvement over the previous number of queries\nknow n to break an $\\ell_0$-estimation sketch.\n\nPresented in Partial Fulfill ment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919a492 DTSTART;TZID=America/New_York:20250325T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250325T130000 URL:https://www.cs.cmu.edu/~aiseminar/ LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:AI Seminar - Andy Zou CLASS:PUBLIC DESCRIPTION:Speaker: ANDY ZOU\, Ph.D. Student\, Computer Science Department \,\nCarnegie Mellon University\, and\, Chief Technology Office and\nCo-fou nder\, Gray Swan AI\n\nTalk Title: Red Teaming AI Agents in-the-wild\n\nTh is presentation demonstrates how red teaming uncovers critical\nvulnerabil ities in AI agents that challenge assumptions about safe\ndeployment. The talk discusses the risks of integrating AI into\nreal-world applications a nd recommends practical safeguards to enhance\nresilience and ensure depen dable deployment in high-risk settings. \n\n— \n\nAndy Zou is a PhD st udent at CMU. He is the CTO and cofounder at Gray\nSwan AI and a cofounder of Center for AI Safety. He works in AI Safety\nand Security. \n\nIn Per son and Zoom Participation.  See announcement.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919a7db DTSTART;TZID=America/New_York:20250325T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250325T120000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Teaching Faculty Candidate - Bilal Taha CLASS:PUBLIC DESCRIPTION:Speaker: BILAL TAHA\, Postdoctoral Fellow\, Department of Elect rical and\nComputer Engineering\, and\, Lecturer\, Department of Computer Science\,\nUniversity of Toronto\n\nTalk Title: Iterative Gradient-Based A lgorithms for Optimization in ML\n\nThis teaching demo focuses on Iterativ e Gradient-Based Algorithms for\nOptimization in Machine Learning (ML). Th e lecture will introduce\nstudents to the fundamental concepts of gradient descent and ascent\,\ndetailing their roles in optimizing ML models. Stud ents will gain an\nunderstanding of the mathematical foundations behind it erative\ngradient methods\, including the significance of gradients\, lear ning\nrates\, and the convergence process. The session will also explore\n different types of gradient descent approaches—Batch Gradient\nDescent\, Stochastic Gradient Descent\, and Mini-Batch Gradient\nDescent—and high light how advanced variants like Momentum can\naddress the limitations inh erent in the standard methods. By the end\nof the lecture\, students will be equipped with the knowledge to\neffectively explain and differentiate b etween these optimization\ntechniques and their applications in machine le arning. \n\n— \n\nBilal Taha is currently a postdoctoral fellow in the Department of\nElectrical and Computer Engineering (ECE) and a lecturer i n the\nDepartment of Computer Science at the University of Toronto. He ear ned\nhis Ph.D. in 2023 from the ECE department at the University of\nToron to. Previously\, he served as a Postgraduate Researcher at the\nVector Ins titute\, a Research Intern at the ICUBE Laboratory\,\nUniversity of Strasb ourg\, and at Amazon Alexa AI. He has received\nseveral awards\, including the NSERC Postdoctoral Fellowship\,\nInnovators Under 35 by MIT Technolog y Review Arabia\, and the DiDi\nChuxing Award for Biometrics. His research interests include machine\nlearning\, wearables\, and biometrics. Faculty Hosts:  Dilsun Kaynar\,\nReid Simmons Joint Computer Science Department and CMU Qatar\n\nIn Person and Zoom Participation. \n\n→ Attendance at this talk is restricted to members of the SCS\ncommunity and relevant CMU stakeholders.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919ac2b DTSTART;TZID=America/New_York:20250324T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250324T173000 URL:https://db.cs.cmu.edu/events/sql-death-prql-pipelined-relational-query- language LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Tobias Brandt CLASS:PUBLIC DESCRIPTION:Speaker: TOBIAS BRANDTTalk Title: PRQL: Pipelined Relational Qu ery\nLanguageThis talk is part of the SQL or Death? Seminar SeriesZoom\nPa rticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919aed3 DTSTART;TZID=America/New_York:20250324T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250324T163000 URL:https://www.ri.cmu.edu/event/autoregressive-models-foundations-and-open -questions/ LOCATION:Remote Access - Zoom SUMMARY:VASC Seminar - Kaiming He CLASS:PUBLIC DESCRIPTION:Speaker: KAIMING HE\, Associate Professor\, Department of Elect rical\nEngineering and Computer Science\, Massachusetts Institute of\nTech nology\n\nTalk Title: Autoregressive Models: Foundations and Open Question s\n\nThe success of Autoregressive (AR) models in language today is so\ntr emendous that their scope has\, in turn\, been largely narrowed to\nspecif ic instantiations. In this talk\, we will revisit the foundations\nof clas sical AR models\, discussing essential concepts that may have\nbeen overlo oked in modern practice. We will then introduce our recent\nresearch on br oadening the scope of modern AR models in the context of\nimage generation \, challenging the common beliefs about how AR models\ncan be built. We wi ll also discuss open questions and potential\ndirections for future resear ch.   \n\n— \n\nKaiming He is an Associate Professor in the Departmen t of EECS at MIT\nwhich he joined in Feb 2024. Before that\, he was a rese arch scientist\nin industrial labs including Facebook AI Research (FAIR\, 2016-2024)\nand Microsoft Research (MSR\, 2011-2016). His research covers a wide\nrange of topics in Computer Vision and Machine Learning. His work has\nbeen recognized by numerous prestigious awards in the community\,\nin cluding the PAMI Young Researcher Award 2018 and multiple Best Paper\nAwar ds at top-tier conferences such as CVPR\, ICCV\, and ECCV. \n\nZoom Parti cipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919b270 DTSTART;TZID=America/New_York:20250324T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250324T150000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Mohammad Salameh CLASS:PUBLIC DESCRIPTION:Speaker: MOHAMMAD SALAMEH\, Principal Researcher and Team Lead\ , Huawei\nTechnologies Canada\n\nTalk Title: Teaching Demo on \"Logistic R egression\"\n\nLogistic Regression is a supervised machine learning algori thm for\nclassification problems\, where it predicts the probability of a\ ncertain observation falling into a particular class. In this demo\nlectur e\, we will explain how logistic Regression models the\nprobability of bin ary outcomes. We will address the sigmoid functions\nand their properties\ , derive the log-loss function\, and utilize\ngradient descent to minimize it. Additionally\, we'll cover the case of\nmultinomial logistic regressi on where more than 2 classes are\nconsidered for classification. \n\n— Dr. Mohammad Salameh is a Principal Researcher and Team Lead at\nHuawei Te chnologies Canada\, specializing in optimizing deep learning\nmodels for d evices with limited compute resources. His expertise spans\n software-har dware co-design\, neural architecture search\, computer\nvision\, natural language processing (NLP) and autonomous car driving.\nPrior to his role a t Huawei\, Dr. Salameh was a Postdoctoral Researcher\nat Carnegie Mellon U niversity in Qatar\, working on the Multi-Arabic\nDialect Applications and Resources project. He holds a Ph.D. in\nComputing Science from the Univer sity of Alberta\, where his research\nfocused on Statistical Machine Trans lation from English to Arabic\, and\nsentiment analysis. \n\nHe has activ ely contributed to the research community\, co-organizing\nSemEval shared tasks on Determining Sentiment Intensity and Affects in\nTweets. His resea rch has been published in top-tier AI and machine\nlearning conferences\, including NeurIPS\, CVPR\, ICLR\, AAAI\, ACL\, NAACL\,\nand IEEE. \n\nFac ulty Hosts:  Dilsun Kaynar\, Patrick Virtue \n\nJoint Computer Science D epartment with CMU-Qatar\n\nIn Person and Zoom Participation.  See announ cement.\n\n-> Attendance at this talk is restricted to members of the SCS\ ncommunity and relevant CMU stakeholders.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919b6ab DTSTART;TZID=America/New_York:20250321T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250321T163000 LOCATION:Frick Fine Arts Building/Library\, Room 125\, 650 Schenley Drive SUMMARY:Pittsburgh Mathematical Horizons Distinguished Lecture - James A.\n Maynard CLASS:PUBLIC DESCRIPTION:Speaker: JAMES A. MAYNARD\, Professor of Number Theory\, The\nM athematical Institute\, University of Oxford\n\nTalk Title: Patterns in th e Primes\n\nJoin in for an inspiring lecture on the mysteries of prime num bers\,\npresented by James A. Maynard\, a Fields Medal laureate and Fellow of\nthe Royal Society.  Prime numbers hold secrets that have fascinated\ nmathematicians for centuries.  For instance\, how often do two prime\nnu mber differ by exactly 2?  This simple yet famous question remains\nunsol ved\, despite over a hundred years of dedicated effort. \nRecognized for his groundbreaking contributions to number theory\,\nMaynard will explore these enduring puzzles\, revealing how patterns in\nprime numbers connect not only to pure mathematics but also to\nreal-world applications\, such a s securing our online communications. \n\nAbstract  — How often do two prime numbers differ by exactly 2?\nThis simple (and famous) question is unsolved despite mathematicians\nworking on it for over 100 years. Moreove r\, basic questions like this\nabout prime numbers turn out to lie at the heart of many important\nreal-world questions. For example\, it turns out that the security of\ninternet communication (which we rely on whenever we buy something\nonline) is related to basic questions about patterns in pr ime numbers\nwhich we don't know how to solve! I'll talk about why prime n umbers\nare so special to mathematicians\, how important questions in both pure\nmathematics and the real world turn out to be connected to primes\, and\nhow we are making progress towards solving these famous problems whi ch\nhave been studied for hundreds of years. \n\n— \n\nThe Pittsburgh Mathematical Horizons lecture series was made possible\nby a generous dona tion of the Benter Foundation. The lectures are\nhosted by the University of Pittsburgh.  The lecture is  open and\naccessible to the science-inte rested public of all ages\, this\npresentation promises to illuminate the elegance of mathematics and\nits profound impact on our understanding of t he world.  \n\nReception will follow the presentation.\n\nPoster with QR code\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919baec DTSTART;TZID=America/New_York:20250321T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250321T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Gates Hlilman 8102 and Zoom SUMMARY:AI-SDM Seminar - Terri Adams CLASS:PUBLIC DESCRIPTION:Speaker: TERRI ADAMS\, Professor of Sociology and Criminology\, \, Deputy\nDirector of the NOAA Cooperative Science Center in Atmospheric \nSciences & Meteorology\,\, and Associate Dean\, Office of External\nAffa irs\, Howard University\n\nTalk Title: The Role of Uncertainty in Decision -Making During High\nConsequential Events\n\nWith the increased risk of na tural disasters\, efforts are underway to\nimprove all aspects of emergenc y response. Regardless of the type of\ndisaster\, whether health or weathe r-related\, there are high degrees of\nuncertainty associated with the unf olding of each event. As emergency\nmanagers\, first responders\, and the public seek to maximize their\nresponse to disasters\, it is essential to understand the impacts of\nuncertainty on decision-making. This seminar wi ll explore the impact\nof uncertainty on decision-makers during high conse quence events. \n\n— \n\nTerri Adams\, Ph.D.\, is a Professor in the D epartment of Sociology and\nCriminology at Howard University. She currentl y serves as the\nAssociate Dean of the Office of External Affairs\, respon sible for the\nresearch portfolio at the Graduate School. Dr. Adams also s erves as\nthe Deputy Director of the NOAA Cooperative Science Center in\nA tmospheric Sciences and Meteorology (NCAS-M) at Howard University and\nDir ector of the Howard University Initiative on Public Opinion. Her\nresearch takes an interdisciplinary approach to examine issues with\ntheoretical a nd practical implications. Dr. Adams’ specific research\ninterests inclu de emergency management\, behavioral responses to severe\nweather and clim ate\, and the impact of trauma and disasters on\nindividuals and organizat ions. Her work centers on individuals' and\norganizations' decision-making processes in the face of crisis events.\nShe is the author of Policing Du ring in Natural Disasters: Stress\,\nResilience\, and the Challenges of Em ergency Management\, which takes a\ncritical review of the challenges face d by first responders before\,\nduring\, and after natural disasters. \n\ nREGISTER  → confirmation email containing information about joining\nt he meeting forwarded upon registration.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919bf95 DTSTART;TZID=America/New_York:20250321T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250321T133000 URL:https://safety21.cmu.edu/category/seminars/ LOCATION:Scott Hall 6142 and Remote Access SUMMARY:Smart Safety Connection Seminar - Multiple Panelists CLASS:PUBLIC DESCRIPTION:Speaker: Multiple PanelistsTalk Title: You Can’t Put a Value on\nHuman Life” but Economists Do: How Ethics and Economics Intersect in \nTransportation\n\nHigh-profile transportation incidents reveal the criti cal interplay\nbetween policy\, economics\, and physical realities. In thi s seminar\,\nexperts from different fields will share quick\, five minute case\nstudies about the trade-offs they face in their research and areas o f\nexpertise. Guided by a moderator\, panelists will engage with the\naudi ence to find common themes\, as well as identify issues that\nrequire furt her exploration. The intended outcome of this panel is to\nhelp make trans portation more visible and interesting to students. \n\nPanelists\n\nKare n Clay — Teresa and H. John Heinz III Professor of Economics and\nPublic Policy\, Carnegie Mellon UniversityAlex London — K&L Gates\nProfessor o f Ethics and Computational Technologies\, Carnegie Mellon\nUniversityHenry Posner III — Chairman\, Railroad Development\nCorporationAlbert Presto — Research Professor\, Mechanical\nEngineering\, Carnegie Mellon Univers ity\n\nPanel Moderator\n\nKaren Lightman — Executive Director\, Safety21 UTC\, Carnegie Mellon\nUniversity\n\n⇒  Due to space constraints\, the re will be a 30 person limit for\nin-person attendance with a virtual opti on.       \n\nAdvance registration is required.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919c339 DTSTART;TZID=America/New_York:20250321T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250321T140000 URL:https://www.cmu.edu/flame/events/index.html LOCATION:Tepper Builing 1403 SUMMARY:CMU FLAME Center Seminar - Neehar Peri CLASS:PUBLIC DESCRIPTION:Speaker: NEEHAR PERI\, Ph.D. Student\, Robotics Institute \, Ca rnegie\nMellon University\n\nTalk Title: Towards Foundation Models for 3D Perception\n\nFoundation models trained on internet-scale multi-modal data achieves\nremarkable zero-shot accuracy across a variety of 2D perception tasks.\nHowever\, extending such foundation models for 3D perception rema ins\nchallenging due to a lack of diverse large-scale 3D training data. To \nbridge this gap\, we leverage 2D foundation models to generate 3D\npseud o-labels\, allowing us to turn any 2D dataset into a 3D dataset!\nWe demon strate the effectiveness of our shelf-supervised approach on\npopular Auto nomous Vehicle (AV) benchmarks across several 3D\nperception tasks\, achie ving competitive performance with\nfully-supervised baselines. Lastly\, we highlight applications beyond\nAVs and examine the limitations of adaptin g off-the-shelf 2D\nfoundation models for 3D perception in-the-wild. \n\n — \n\nNeehar Peri is a fourth year Ph.D. student in the Robotics Instit ute\n(RI) at CMU\, advised by Prof. Deva Ramanan. He is broadly interested \nin computer vision and machine learning\, with applications to robot\npe rception and action. His current research focuses on leveraging 2D\npriors to build foundation models for 3D perception. \n\nIn Person and Zoom Par ticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919c6e3 DTSTART;TZID=America/New_York:20250320T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250320T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - Seyoon Ragavan CLASS:PUBLIC DESCRIPTION:Speaker: SEYOON RAGAVAN\, Ph.D. Student\, Theory of Computation \,\nComputer Science and Artificial Intelligence Laboratory\, Massachusett s\nInstitute of Technology\n\nTalk Title: Cloning Games\, Black Holes\, an d Cryptography\n\nWe revisit the notion of unclonable encryption introduce d by Broadbent\nand Lord (TQC 2020)\, where a classical message comprising n bits is\nencrypted into a quantum ciphertext state comprising n qubits. \nInformally\, an adversary should not be able to distribute this\nciphert ext state between two isolated parties such that both parties\ncan decrypt if the secret key is later revealed to them. Our main\ncontributions are twofold:\n\nWe propose the first candidate unclonable encryption scheme wh ich is\nplausibly multi-copy secure: if the adversary receives t identical \nciphertext states\, they cannot enable t+1 isolated parties to\nsimultan eously decrypt. Previous candidates\, based on BB84 states and\ncoset stat es\, are multi-copy insecure for some t = poly(n).  \nAdditionally\, we provide evidence towards this scheme’s multi-copy\nsecurity by proving s ecurity whenever t\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919caa6 DTSTART;TZID=America/New_York:20250320T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250320T163000 URL:https://brain.andrew.cmu.edu/seminar LOCATION:Baker Hall 341A and Zoom SUMMARY:brAIn Seminar - Xuan Ma CLASS:PUBLIC DESCRIPTION:Speaker: XUAN MA\, Research Assistant Professor\, Department of \nNeuroscience\, Northwestern University\n\nTalk Title: Decoding motor con trol from a nonlinear manifold spanning\nnaturalistic behaviors\n\nRecent advances in computational neuroscience suggest that neural\nactivity for m otor control is organized within low-dimensional\nmanifolds\, capturing th e key patterns of neural covariation underlying\ndiverse motor behaviors. This framework enables the projection of\ncomplex\, high-dimensional neura l recordings into more interpretable\nsubspaces. However\, our understandi ng of these neural manifolds\nremains limited\, primarily because most stu dies have focused on a few\nhighly constrained and stereotyped motor tasks \, presumably capturing\nonly isolated fragments of a broader manifold str ucture. \n\nIn this talk\, I will present our work based on neural record ings from\nmonkeys performing tasks in both a conventional laboratory sett ing and\na less constrained\, in-cage environment. Although the tasks were \nnominally similar\, involving various forms of grasp\, they were\nperfor med in quite different contexts. In the cage\, the monkey must\nmaintain a quadrupedal stance\, generating more complex proprioceptive\ninputs from all four limbs\, than occurs while seated in the primate\nchair. We found that the latent representations of neural activity in\nlow-dimensional spa ce formed distinct clusters within the neural\nmanifold\, which we identif ied using unsupervised algorithms. By\ncomputing a separate linear decoder for each cluster\, we achieved\nsignificantly more accurate EMG predictio ns from M1 activity than was\npossible with a single global linear decoder trained on data from all\nbehaviors. This piecewise linear approach even slightly outperformed a\nglobal LSTM decoder. However\, the piecewise deco der\, like the global\nlinear and LSTM decoders\, failed to predict EMG ac curately for\nheld-out tasks. \n\nWe also tried to predict data for point s that fell between clusters by\ncombining the outputs of multiple decoder s based on the geometric\nrelationships between the clusters\; these effor ts failed as well. We\nanalyzed pairwise Euclidean and geodesic distances between neural\nactivity samples across different tasks and during hour-lo ng\ncontinuous recordings. The discrepancy between these distance measures \nwas relatively small for in-lab tasks but significantly larger for\nin-c age tasks and even more pronounced in continuous recordings that\nincluded spontaneous behaviors. We hope that by further studying the\ntopology of the curved manifolds and the geometric relationships\nbetween the clusters using more intuitive computational tools\, we may\nbe able to build decod ers through interpolation across the curved\nmanifold\, at least for novel tasks having activity that falls within\nthe convex hull of the training data.    \n\n— \n\nXuan Ma is a Research Assistant Professor at Nort hwestern University.\nHe earned his Ph.D. from Huazhong University of Scie nce and\nTechnology\, Wuhan\, China\, where he worked with Prof. Jiping He to\nestablish a nonhuman primate lab from the ground up. In 2018\, he mov ed\nto the U.S. and joined Prof. Lee Miller’s lab at Northwestern\nUnive rsity as a postdoctoral researcher\, investigating how neuronal\nactivity in the brain coordinates hand movements to advance\nbrain-controlled funct ional electrical stimulation (FES)\nbrain-computer interfaces (BCIs). His research focuses on\nunderstanding the neural mechanisms underlying motor control as well\nas developing advanced computational methods to enhance B CI\nperformance\, making them more robust to drifts and adaptable to\nvari ations in task demands. \n\nIn Person and Zoom Participation.  See annou ncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919d021 DTSTART;TZID=America/New_York:20250320T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250320T150000 URL:https://www.ri.cmu.edu/event/the-new-era-of-video-generation/ LOCATION:Newell-Simon 4305 SUMMARY:VASC Seminar - Dr. Rong Yan CLASS:PUBLIC DESCRIPTION:Speaker: RONG YAN\, Chief Technology Officer\, HeyGen\n\nTalk T itle: The New Era of Video GEneration\n\nTraditional video production is s low\, expensive\, and requires\nspecialized skills. Founded by CMU alumni\ , HeyGen is an AI-native\nvideo platform designed to revolutionize the vid eo creation process by\nmaking visual storytelling accessible to all. We'v e successfully grown\nto more than 20M users\, and tens of millions revenu e in less than one\nyear\, with recognition as the #1 Fastest Growing Soft ware Product by\nG2 in 2025.\n\nIn this talk\, I will talk about how HeyG en leverages cutting-edge AI\nto enable users to create\, localize\, perso nalize\, and interact with\nvideos with our state-of-the-art human-centric  video engine. In\nparticular\, I will cover our key user cases including avatar videos\,\nvideo translation\, interactive avatar and a number of\n AI innovations for video generation. I will share more on our\nin-house research areas around video generation as well as real-world\nchallenges i n building AI applications.  \n\n—\n\nDr. Rong Yan is currently the CT O for HeyGen which is an innovative\nAI-driven video platform that allows users to create videos with\nAI-generated avatars and voices. Our mission is to make visual\nstorytelling accessible to everyone. Before joining Hey Gen\, Rong was a\nVP of Engineering in Hubspot Inc. responsible for their Data products\nincluding data platform\, automation\, data integration and reporting.\nBefore Hubspot\, he was the Senior Director of Product Engine ering in\nSnapchat\, Director of Data in Square\, and Eng Manager of Ads R anking\nin Facebook. Dr. Yan received his M.Sc. (2004) and Ph.D. (2006)\n degree from Carnegie Mellon University's School of Computer Science.\nHe h as received 2 Best Paper runner-Up awards\, published more than 60\npapers \, and co-chaired more than 10 conferences / workshops in the\nrelated dom ains.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919d422 DTSTART;TZID=America/New_York:20250320T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250320T143000 URL:https://www.cmu.edu/math/news-events/calendar.html#event=77840144\;inst ance=20250320133000?popup=1&lang=en-US LOCATION:Hamburg Hall A301 SUMMARY:Joint PITT-CMU Math Colloquium - James A. Maynard CLASS:PUBLIC DESCRIPTION:Speaker: JAMES A. MAYNARD\, Professor of Number Theory\, The\nM athematical Institute\, University of Oxford\n\nPlease join James A. Mayna rd\, a Fields Medal laureate and Fellow of\nthe Royal Society\, who will v isit CMU\, as part of a Joint CMU-PITT\nMathematics Colloquium. \n\nThe t alk is geared toward a wide audience and is open to all.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919d732 DTSTART;TZID=America/New_York:20250320T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250320T140000 LOCATION:Gates Hillman 8102 SUMMARY:Doctoral Speaking Skills Talk - Benjamin Stoler CLASS:PUBLIC DESCRIPTION:Speaker: BENJAMIN STOLER\, Ph.D. Student\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Towards Robust Autono mous Driving via Enhanced Data\nUtilization\n\nAutonomous vehicles must na vigate among humans in a safe and\nsocially-compliant manner. Current appr oaches for building and\nevaluating such systems rely on data-driven techn iques\; however\, a\ngeneralization gap emerges\, as methods trained in th ese traditional\nparadigms are unable to cope with unexpected real-world s cenarios.\nTherefore\, we aim to develop improved evaluation settings and\ nmethodologies to increase and assess robustness in autonomous driving\nag ainst these challenges.\n\nFor robustness against out-of-distribution\, sa fety-relevant scenarios\,\nwe create a hierarchical characterization metho d which leverages\ncounterfactual probes to find hidden safety-relevant sc enarios in\nlarge datasets. We then address the induced generalization gap by\nincorporating the characterizations into downstream trajectory\npredi ction models' inductive biases. Next\, for robustness against\nadversarial \, safety-critical scenarios\, we develop a reactive\,\nskill-based advers ary policy which leverages a learned\, multi-faceted\ncriticality objectiv e to perturb existing scenarios. We then train ego\npolicies in a closed-l oop manner against these generated scenarios\,\ndemonstrating improved dow nstream ego performance.\n\nThis talk concludes by outlining future direct ions to further advance\nreliability and safety in autonomous driving. \n \nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919db06 DTSTART;TZID=America/New_York:20250320T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250321T200000 LOCATION:Gates Hillman SUMMARY:Admitted PhD Student 2025 Open House CLASS:PUBLIC DESCRIPTION:Talk Title: CSD Admitted PhD Student Open House\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919dd9d DTSTART;TZID=America/New_York:20250319T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250319T160000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Adithya Abraham Philip CLASS:PUBLIC DESCRIPTION:Speaker: ADITHYA ABRAHAM PHILIP\, Ph.D. Candidate\, Computer Sc ience\nDepartment\, Carnegie Mellon University\n\nTalk Title: Accurately P arameterizing Internet Performance Testing for\nRealistic Evaluations\n\nT he performance of Internet services — be it file download\ncompletion ti mes\, video quality\, or lag-free video conferencing — is\nheavily influ enced by network parameters. These include the bottleneck\nbandwidth\, pac ket loss\, network delays\, and how fairly the bottleneck\nlink is shared with other services. However\, current techniques to\nevaluate service per formance display three major issues: (a) testing\npredominantly in setting s representing the \"edge\" of the Internet\, and\nnot the core\; (b) an o veremphasis on the role of Congestion Control\nAlgorithms (CCAs) in determ ining application performance\; (c) testing\nin settings that do not neces sarily reflect where congestion occurs on\nthe Internet today. The goal of this thesis is to improve the state of\nthe art in testing for a more rea listic evaluation of Internet service\nperformance. We achieve this by cha nging measurement methodology to\ntest in more diverse network conditions\ , evaluate deployed Internet\nservices as opposed to just their underlying CCAs\, and identify more\nrealistic network parameters for evaluations.  \n\nWe first examine the changes in CCA behavior when evaluated in\nsettin gs representing the core of the Internet as opposed to the edge.\nWe find that the change to core Internet speeds and flow counts\ndramatically alte rs fairness outcomes\, and challenges long-standing\nassumptions about CCA behavior. This highlights the need to run\nInternet evaluations in more d iverse settings. \n\nWe then build Prudentia\, an Internet fairness watch dog\, to understand\nhow fairly two Internet services can share a bottlene ck link. In\naddition to discovering extreme unfairness on the Internet to day\, we\ngain key insights into improving current testing methodology — (a)\nThe most and least fair services both use variants of the same CCA\, \nhighlighting the need to test services in addition to CCAs\; (b)\nnetwor k settings can drastically affect even service-level fairness\noutcomes\, necessitating their careful selection. \n\nIn the final part of this thes is\, we leverage end-to-end measurements\nfrom a leading video-streaming s ervice to identify the prevalent\nnetwork conditions experienced by its us ers. Based on these\nmeasurements\, we recommend guidelines for parameteri zing future\nInternet evaluations so that their results are more relevant and\nreliable indicators of real-world CCA and service performance. \n\nT hesis Committee\n\nJustine Sherry (Chair)\n\nSrinivasan Seshan\n\nTheophil us A. Benson\n\nRenata Teixeira (Netflix)\n\nIn Person and Zoom Participat ion.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919e275 DTSTART;TZID=America/New_York:20250319T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250319T120000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:Teaching Track Faculty Candidate - Tanvi Bajpai CLASS:PUBLIC DESCRIPTION:Speaker: TANVI BAJPAI\, Ph.D. Candidate in Computer Science\, S iebel\nSchool of Computing and Data Science\, University of Illinois\nUrba na-Champaign\n\nTalk Title: Fundamentals of Functions\n\nFunctions are a f oundational concept in mathematics and computer\nscience\, forming the bas is for understanding transformations\,\nmappings\, and relationships betwe en sets. In this lecture\, I will\nintroduce the fundamental properties of functions—injectivity\,\nsurjectivity\, and bijectivity—and explore t heir significance in both\ntheoretical and applied contexts. Emphasizing b oth formal definitions\nand intuition\, I will highlight key connections t hat deepen\nunderstanding and make abstract concepts more accessible. \n\ n— \n\nTanvi Bajpai is a Ph.D. candidate in Computer Science at the\nUn iversity of Illinois Urbana-Champaign\, advised by Dr. Chandra\nChekuri an d Dr. Eshwar Chandrasekharan. Her research spans theory\,\nHCI\, and their intersection\, where she focuses on designing\noptimization and evaluatio n metrics that incorporate human-centered\nconsiderations for algorithmic decision-making systems. She is\npassionate about teaching and mentorship\ , particularly in making\ntheoretical concepts more intuitive and accessib le to students.\nDrawing from her experience across both theoretical and a pplied\ndomains\, she emphasizes building strong foundational reasoning sk ills\nwhile fostering an inclusive and engaging learning environment. \n\ nShe has taught and advised students at various levels\, mentoring\nunderg raduate researchers and course assistants to help them grow both\ntechnica lly and pedagogically. Through her teaching\, she aims to equip\nstudents with the confidence and problem-solving skills necessary to\ntackle comple x ideas both in and beyond the classroom. \n\nFaculty Host:  Anil Ada In Person and Zoom Participation.  See\nannouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919e6ca DTSTART;TZID=America/New_York:20250318T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250318T150000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Tianyu Li CLASS:PUBLIC DESCRIPTION:Speaker: TIANYU LI\, Ph.D. Student\, Computer Science and Artif icial\nIntelligence Laboratory\, Massachusetts Institute of Technology\n\n Talk Title: Building Novel Abstractions for a Declarative Cloud\n\nAs the cloud evolves in capability\, it has also become increasingly\ncomplex and difficult to program. New abstractions are necessary to\nensure next-gene ration cloud applications are correct\, simple\, and\nefficient. \n\nIn t his talk\, I will first describe Resilient Composition\, a new\nabstractio n that ensures fault-tolerance in applications composed from\nindependent\ , distributed components. The key insight is to rely on\natomic\, fault-to lerant “steps” that span component operations and\nmessages. I will pr esent DARQ\, an efficient execution engine for such\nsteps\, and Distribut ed Speculative Execution\, a transparent\noptimization that dramatically r educes overhead of Resilient\nComposition. \n\nI will also briefly discus s BRAD\, an abstraction that virtualizes\ncloud data management and enable s automatic data infrastructure\ndesign. Together\, these solutions repres ent an important first step\ntowards a more declarative cloud\, where stro ng primitives separate\nuser applications from their underlying infrastruc ture\, paving a way\nfor simplicity and efficiency through automation. \n \n— \n\nTianyu Li is a final-year PhD student at MIT\, advised by Profe ssor Sam\nMadden. His research interests span distributed systems and data base\nsystems\, with an emphasis on building new abstractions for the mode rn\ncloud. Tianyu currently focuses on building better fault-tolerance for \nthe cloud through the novel abstraction of Resilient Composition\, and\n has worked closely with industry leaders like Microsoft to deploy\nthese s olutions. Before MIT\, Tianyu obtained his MS and BS from CMU. \n\nFacult y Host:  Jignesh Patel In Person and Zoom Participation.  See\nannounce ment.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919eaf1 DTSTART;TZID=America/New_York:20250318T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250318T130000 LOCATION:Newel-Simon 3305 and Zoom SUMMARY:Joint AI Seminar / Doctoral Speaking Skills Talk - Alexander Goldbe rg CLASS:PUBLIC DESCRIPTION:Speaker: ALEXANDER GOLDBERG\, Ph.D. Student\, Computer Science\ nDepartment\, Carnegie Mellon University\n\nTalk Title: Balancing Transpar ency and Reviewer Anonymity in\nScientific Peer Review\n\nScientific peer review plays a crucial role in maintaining the\ncredibility of scholarly w ork. Concerns about its integrity and\nquality have motivated increasing t ransparency into the peer review\nprocess. In this talk\, I will examine m echanisms for sharing more data\nabout peer review while preserving review er anonymity. Our analysis of\npeer review highlights more general privacy -utility trade-offs that\narise in sharing time series and graph-structure d data. \n\nFirst\, I will discuss “open peer review” systems (like\n OpenReview.net)\, where reviews are publicly posted while keeping\nreviewe r identities anonymous. We show that temporal batching patterns\nin review activity create realistic risks of reviewer\nde-anonymization. We then pr opose a privacy-preserving mechanism that\n adds random delays to review postings and prove that this mechanism\noptimally balances privacy risks a nd system latency under a formal\nprivacy framework. \n\nNext\, I will ex plore the challenge of sharing private data about\nreviewer-paper assignme nts to facilitate the evaluation of fraud\ndetection algorithms. I will sh ow that reporting even a single\naccuracy statistic on this dataset can en able a malicious third party\nto de-anonymize reviewers. Then\, I will con sider the use of techniques\nfrom differential privacy to mitigate these r isks and outline\npotential directions for improving privacy-preserving da ta sharing in\npeer review. \n\nPresented as part of the Artificial Intel ligence Seminar Series \n\nPresented in Partial Fulfillment of the CSD Sp eaking Skills\nRequirement \n\nIn Person and Zoom Participation.  See an nouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919eef2 DTSTART;TZID=America/New_York:20250317T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250317T173000 URL:https://db.cs.cmu.edu/events/sql-death-malloy-a-modern-open-source-lang uage-for-analyzing-transforming-and-modeling-data SUMMARY:Database Seminar - Lloyd Tabb CLASS:PUBLIC DESCRIPTION:Speaker: LLOYD TABB\, Founder/Former Chief Technology Officer\, Looker\,\nCo-creator of Malloy\, Meta\n\nTalk Title: Malloy: A Modern Ope n Source Language for Analyzing\,\nTransforming\, and Modeling Data\n\nIn software we express our ideas through tools. In data\, those tools\nthink in rectangles. From spreadsheets to the data warehouses\, to do\nany analy tical calculation\, you must first go through a rectangle..\nForcing data through a rectangle shapes the way we solve problems (for\nexample\, dimen sional fact tables\, OLAP Cubes). But really\, most Data\nisn’t rectangu lar. Most data exists in hierarchies (orders\, items\,\nproducts\, users). Most query results are better returned as a\nhierarchy (category\, brand\ , product). Can we escape the rectangle?\nMalloy is a new experimental dat a programming language that\, among\nother things\, breaks the rectangle p aradigm and several other long\nheld misconceptions in the way we analyze data. \n\n— \n\nLloyd Tabb spent the last 30 years revolutionizing how the world uses\nthe internet and\, by extension\, data. Originally a data base &\nlanguages architect at Borland\, Lloyd founded Looker\, which Goog le\nacquired in 2019. Lloyd's work at Looker helped define the Modern Data \nStack. At Google\, Lloyd continues to pursue his passion for data\, and\ nlove of programming languages through his current project\, Malloy —\na new open source experimental programming language that aims to\nreinvent how programmers describe data relationships and\ntransformations. Before L ooker\, Lloyd was one of the internet\npioneers\, working at Netscape duri ng the browser wars as the Principal\nEngineer on Navigator Gold\, the fir st HTML WYSIWYG editor. During his\nfive-year tenure at Netscape\, he was the technical lead on several\nreleases of Netscape's flagship product\, t he Netscape Web Browser.\nLloyd was also an early founder and heavily invo lved in defining\nMozilla.org. In 2003\, Lloyd was the CTO of LiveOps\, on e of the very\nfirst gig-economy companies. Lloyd also co-founded Readyfor ce and\nLuminate. \n\nThis talk is part of the SQL or Death? Seminar Seri es\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919f2e2 DTSTART;TZID=America/New_York:20250317T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250317T150000 LOCATION:Newell-Simon 3305 and Zoom SUMMARY:SCS Faculty Candidate - Allen Liu CLASS:PUBLIC DESCRIPTION:Speaker: ALLEN LIU\, Ph.D. Student in Computer Science\, Electr ical\nEngineering and Computer Science Department\, Massachusetts Institut e\nof Technology\n\nTalk Title: Learning Theoretic Foundations for Modern (Data) Science\n\nIn this talk\, I will explain how fundamental problems i n computational\nlearning theory are at the heart of modern problems in ma chine\nlearning and scientific applications and how algorithmic insights i n\nmathematically tractable models can inspire new solutions in a wide\nva riety of domains. I will explore two directions. \n\nFirst\, I will explo re algorithmic foundations for model stealing of\nlanguage models.  Model stealing\, where a learner tries to recover an\nunknown model through que ry access\, is a critical problem in machine\nlearning. Here\, I will aim to build a theoretical foundation for\ndesigning model stealing algorithms .\n\nSecond\, I will introduce Hamiltonian learning\, a central computatio nal\ntask towards understanding and benchmarking quantum systems.  I will \nhighlight how the lens of learning theory plays a key role in\nidentifyi ng and circumventing previous barriers and allows us to give\nefficient al gorithms in settings that were previously conjectured to\nbe intractable.  \n\n— \n\nAllen Liu is currently a fifth-year graduate student in EEC S at MIT\,\nadvised by Ankur Moitra. His research is in learning theory\, broadly\ndefined\, encompassing classical learning theory and statistics\, as\nwell as problems in modern machine learning and scientific\napplicati ons such as quantum information.\n\nAllen is the recipient of a Hertz Fell owship and a Citadel GQS\nFellowship. His work has been awarded Best Stude nt Paper at QIP in\n2024 and featured in popular science media including Q uanta Magazine's\nBiggest Breakthroughs in Computer Science for 2024. \n\ nFaculty Hosts:  Ryan O'Donnell (CSD)\, Nina Balcan (MLD) \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919f709 DTSTART;TZID=America/New_York:20250317T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250317T103000 LOCATION:Traffic21 Classroom\, Gates Hillman 6121 SUMMARY:Doctoral Speaking Skills Talk - Ziyue Qiu CLASS:PUBLIC DESCRIPTION:Speaker: ZIYUE QIU\, Ph.D. Student\, Computer Science Departmen t\,\nCarnegie Mellon University\n\nTalk Title: Moirai: Optimizing Placemen t of Data and Compute in Hybrid\nClouds\n\nHybrid cloud deployment of larg e-scale data analytics requires careful\npartitioning of the data and jobs between on-premise and cloud sites\nto avoid massive networking costs. Mo irai is a new framework that\nanalyzes job logs\, including which data eac h job accessed\, to\ndetermine which data should go on each side and which should be\nreplicated. It also provides the job scheduler with table loca tion and\naccess-size prediction information\, so it can determine the bes t\nlocation to execute each new job to minimize inter-site data\nfetching.   \n\nMoirai’s optimization scales to huge data corpuses and minimizes \ndollar costs\, by exploiting recurring job templates to identify data\ni nter-dependencies\, per-job read volumes\, and ignore dependencies for\nli ghtly-used data to reduce optimizer complexity. Simulations driven\nby a 9 -month trace of CorpX’s Presto cluster (84M queries\, 24EB\ndata-read vo lume) show that Moirai can reduce dollar costs for an\non-premise/in-cloud hybrid deployment by ?95% relative to the\nstate-of-the-art partitioning approach and over 99.5% relative to\nother public approaches. The savings come from 97–99.8% reduction in\ncloud egress\, up to 99% reduction in r eplication\, and 85–97%\nreduction in on-premises uplink requirements.  \n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequireme nt\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919faae DTSTART;TZID=America/New_York:20250314T143000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250314T153000 URL:https://www.ri.cmu.edu/event/ri-seminar-2-2-2-2-2/ LOCATION:Tepper Building 1403 SUMMARY:Robotics Seminar - Ken Goldberg CLASS:PUBLIC DESCRIPTION:Speaker: KEN GOLDBERG\, Distinguished Professor\, Industrial En gineering\nand Operations Research\, William S. Floyd Distinguished Profes sor of\nEngineering\, Electrical Engineering & Computer Sciences\, Univers ity of\nCalifornia\, Berkeley\n\nTalk Title: Is Data All You Need?: Large Robot Action Models and Good\nOld Fashioned Engineering\n\nEnthusiasm has been skyrocketing for humanoids based on recent\nadvances in \"end-to-end\ " large robot action models. Initial results\nare promising\, and several collaborative efforts are underway to\ncollect the needed demonstration da ta. But is data really all you\nneed?\n\nAlthough end-to-end Large Vision\ , Language\, Action (VLA) Models have\npotential to generalize and reliabl y solve all problems in robotics\,\ninitial results have been mixed.  It seems likely that the size of\nthe VLA state space and dearth of available demonstration data\,\ncombined with challenges in getting models to gener alize beyond the\ntraining distribution and the inherent challenges in int erpreting and\ndebugging large models\, will make it difficult for pure en d-to-end\nsystems to provide the kind of robot performance that investors expect\nin the near future. \n\nIn this presentation\, I share my concern s about current trends in\nrobotics\, including task definition\, data col lection\, and experimental\nevaluation.  I propose that to reach expected performance levels\, we\nwill need \"Good Old Fashioned Engineering (GOFE )\" – modularity\,\nalgorithms\, and metrics.   \n\nI'll present MANIP \, a modular systems architecture that can integrate\nlearning with well-e stablished procedural algorithmic primitives such\nas Inverse Kinematics\, Kalman Filters\, RANSAC outlier rejection\, PID\nmodules\, etc. I’ll sh ow how we are using MANIP to improve\nperformance on robot manipulation ta sks such as grasping\, cable\nuntangling\, surgical suturing\, motion plan ning\, and bagging\, and\npropose open directions for research.\n\n— \n \nKen Goldberg is President of the Robot Learning Foundation (organizer\no f the annual Conference on Robot Learning) and William S. Floyd\nDistingui shed Chair of Engineering at UC Berkeley and Chief Scientist\nof Ambi Robo tics and Jacobi Robotics. Ken leads research in robotics\nand automation: grasping\, manipulation\, and learning for applications\nin warehouses\, i ndustry\, homes\, agriculture\, and robot-assisted\nsurgery.  He is Profe ssor of IEOR with appointments in EECS and Art\nPractice.  Ken is Chair o f the Berkeley AI Research (BAIR) Steering\nCommittee (60 faculty) and is co-founder and Editor-in-Chief emeritus\nof the IEEE Transactions on Autom ation Science and Engineering\n(T-ASE).  He has published ten US patents\ , over 400 refereed papers\,\nand presented over 600 invited lectures to a cademic and corporate\naudiences.  Ken earn his PhD in Computer Science a t Carnegie Mellon\nUniversity\, his BSEE in Electrical Engineering\, the U niversity of\nPennsylvania\, and a BSE in Economics from UPenn-Wharton.   \n\nAdditional Information\n\nRI seminars are recorded and can be access ed after the seminar.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef919ffad DTSTART;TZID=America/New_York:20250314T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250314T123000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI-SDM - Student Brainstorming Session - Jianyu Xu CLASS:PUBLIC DESCRIPTION:Speaker: JAINYU XU\, Postdoctoral Researcher\, Machine Learning \nDepartment\, Carnegie Mellon University\n\nThe AI Institute for Societal Decision Making (AI-SDM) — which\nbrings together AI and social science s researchers to develop\nhuman-centric AI for societal good — hosts stu dent-led brainstorming\ndiscussion sessions. Our goal is to foster interdi sciplinary\ncollaboration and generate ideas on how AI can help solve soci etal\nproblems\, particularly from an angle of decision-making. \n\nThese sessions are intended for graduate and undergraduate students to\nconnect \, share ideas\, and collaborate on AI-related projects. \n\nThis week\, Jianyu Xu will present his work on machine learning for\nmaternal health . This work contributes to the Public Health use-case\nthrust within AI-SD M. You won't want to miss the opportunity to learn\nabout this exciting pr oject. \n\nRSVP  → for headcounts In Person and Zoom Participation.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a0302 DTSTART;TZID=America/New_York:20250314T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250314T110000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Speaking Skills Talk - Meng-Chieh Lee CLASS:PUBLIC DESCRIPTION:Speaker: MENG-CHIEH LEE\, Ph.D. Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Accurate\, Robust\, an d Interpretable Graph Mining\n\nHow can we solve semi-supervised node clas sification in various graphs\npossibly with noisy features and structures? Graph neural networks\n(GNNs) have succeeded in many graph mining tasks\, but their\ngeneralizability to various graph scenarios is limited due to the\ndifficulty of training\, hyperparameter tuning\, and the selection of a\nmodel itself.   In this talk\, I will present a carefully-designed\ns imple model SlimG for solving semi-supervised node classification. It\nexh ibits the following desirable properties: accurate\, robust\, fast\,\nscal able\, and interpretable. \n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a067e DTSTART;TZID=America/New_York:20250313T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250313T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room and Zoom SUMMARY:Crypto Seminar - Hadas Zeilberger CLASS:PUBLIC DESCRIPTION:Speaker: HADAS ZEILBERGER\, Ph.D. Student\, Department of Compu ter\nScience\, Yale University\n\nTalk Title: Khatam: Reducing the Communi cation in Multilinear\,\nCode-Based SNARKs\n\nTwo successful techniques ha ve recently emerged in the construction of\nSNARKs with extremely fast pro vers\; The use of multilinear (instead of\nunivariate) polynomial commitme nt schemes (PCS) and the construction\nof PCS from error-correcting codes. Recently\, BaseFold (Crypto 2024)\nintroduced a family of  PCS that comb ine these two techniques\,\nthereby achieving a better tradeoff between pr over and verifier costs\nthan the state of the art. Despite its impressive overall efficiency\,\nBaseFold suffered from larger proof sizes than its univariate\ncounterparts\, due to unproven claims about linear codes\, whi ch were\nnot relevant in the univariate setting. This work closes this gap by\nproving a new fact about linear codes – that if  if πL\, πR ar e\ntwo vectors in 𝔽n and if πL + r πR is close to C\,\nthen π L\, πR and (πL + r πR) all agree with codewords at\npositions indexe d by the same set S\, except with negligible\nprobability over r ← 𝔽. Our result holds as long as |S| > (1 -\nΔC + ε)1/3\, and with fa ilure probability smaller than 2/ε2|𝔽|\n. Furthermore\, our results e xtend to any finite field and any linear\ncode. \n\nIn Person and Zoom Pa rticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a0a27 DTSTART;TZID=America/New_York:20250313T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250313T163000 URL:https://brain.andrew.cmu.edu/seminar LOCATION:Baker Hall 340A and Zoom SUMMARY:brAIn Seminar - Jonathan Pillow CLASS:PUBLIC DESCRIPTION:Speaker: JONATHAN PILLOW\, Professor\, Princeton Neuroscience I nstitute\,\nPrinceton University\n\nTalk Title: Disentangling the Roles of Distinct Cell Classes with\nCell-Type Dynamical Systems\n\nLatent dynamic al systems have been widely used to characterize the\ndynamics of neural p opulation activity in the brain. However\, these\nmodels typically ignore the fact that the brain contains multiple cell\ntypes\, which limits their ability to capture the functional roles of\ndistinct cell classes or pred ict the effects of cell-specific\nperturbations. To overcome these limitat ions\, we introduce the\n“cell-type dynamical systems” (CTDS) model\, which extends latent\nlinear dynamical systems to contain distinct latent variables for each\ncell class\, with appropriate sign constraints on the interactions\nbetween them.\n\nIn this talk\, I will describe the CTDS mod el and show that fitting in\nthe noiseless case can be reduced to non-nega tive matrix\nfactorization.  I will then show an application of a multi-r egion\nmodel CTDS to simultaneous recordings from rat frontal orienting\nf ields (FOF) and anterior dorsal striatum (ADS) during an auditory\ndecisio n-making task.  Remarkably\, the model — fit only to\nunperturbed neura l activity — predicts the time-dependent effects of\ndifferent optogenet ic perturbations on behavior\, specifically in FOF\,\nADS\, and FOF-to-ADS axon terminals. I will close by discussing the\nfuture directions and oth er applications for biologically-constrained\ndynamical models of neural a ctivity and behavior. \n\n— \n\nJonathan Pillow completed his undergra duate education at the\nUniversity of Arizona in Tucson\, where he studied mathematics and\nphilosophy. He received a Ph.D. in neuroscience from New York\nUniversity in 2005\, and was postdoctoral fellow at the Gatsby\nCom putational Neuroscience Unit at University College London. In 2009\,\nhe b ecame an assistant professor at the University of Texas at Austin\,\nand i n 2014 Jonathan moved to Princeton University to join the\nPrinceton Neuro science Institute\, Psychology department\, and Center\nfor Statistics & M achine Learning. Jonathan's current research sits at\nthe border between n euroscience and statistical machine learning\, and\nfocuses on computation al and statistical methods for understanding how\nlarge populations of neu rons transmit and process information. \nAdditional Information. \n\nIn Person Group Viewing and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a0ecb DTSTART;TZID=America/New_York:20250313T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250313T160000 URL:https://aco.math.cmu.edu/abs-24-25/mar13.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Luis Ferroni CLASS:PUBLIC DESCRIPTION:Speaker: LUIS FERRONI\, Postdoctoral Researcher\, Institute for Advanced\nStudy\, Princeton\n\nTalk Title: The polytope of all matroids\n \nIn this talk we introduce a polytope that encodes all matroids of a\nfix ed size and rank. The polytope is constructed using the\nrepresentation of each matroid as a combination of Schubert matroids\nwithin the valuative group. Linear maps in the polytope's ambient\nspace correspond to valuativ e invariants on matroids. Each matroid is\nrepresented as a lattice point\ , but only certain special matroids\ncorrespond to the vertices of the pol ytope\, giving rise to a new\nnotion of \"extremality.\" \n\nWe argue tha t in cases where a conjecture in matroid theory posits the\npositivity of a specific invariant\, extremal matroids should be\nexamined first. For in stance\, we will explain why the counterexamples\nto the Ehrhart positivit y conjecture by De Loera\, Haws\, and Köppe\, as\nwell as the counterexam ples to the Merino-Welsh conjecture on Tutte\npolynomials\, both correspon d to vertices of these polytopes. In\ngeneral\, if the conjecture involves a valuative invariant\, extremal\nmatroids are the correct candidates to be counterexamples.\n\nFurthermore\, the framework of these polytopes allo ws us to address a\nwell-known folklore question in matroid theory. We sho w the existence\nof a valuative invariant that serves as a test for repres entability.\nIn other words\, we show that there are valuative invariants that are\nnon-negative for all realizable matroids but fail to remain\nnon -negative in general. \n\nThis is joint work with Alex Fink. \n\n4:00 pm - Jane Street-sponsored tea and cookies in the math lounge\n(please bring your own mug if you have one).\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a12f7 DTSTART;TZID=America/New_York:20250313T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250313T130000 URL:https://db.cs.cmu.edu/events/redesigning-blockchains-ssd-optimized-veri fiable-databases-and-beyond-daniel-lin-kit-wong LOCATION:Newell-Simon 3305 SUMMARY:Joint SDI / Database Seminar - Daniel Wong CLASS:PUBLIC DESCRIPTION:Speaker: DANIEL LIN-KIT WONG\, Research Engineer\, LayerZero La bs\nResearch\n\nTalk Title: Redesigning Blockchains: SSD-optimized Verifia ble\nDatabases and Beyond\n\nBlockchains are decentralized ledgers that re place trusted central\nauthorities with verifiable distributed consensus. This\ndecentralization has computers’\, but there are huge opportunities \nfor architectural optimization across the entire blockchain software\nst ack. \n\nWe begin this talk by outlining the scaling challenges from a sy stems\nresearcher’s perspective\, and discussing the bottlenecks faced i n\ncomputation\, storage\, and network bandwidth. We then discuss how we\n optimized the blockchain storage layer using our novel verifiable\ndatabas e\, the Quick Merkle Database (QMDB). QMDB addresses\nlongstanding state m anagement bottlenecks with a streamlined design\nintegrating Merkle tree s torage and key-value stores. QMDB’s\nflash-friendly append-only design e nables zero disk IO Merklelization\nwith a minimal DRAM footprint. \n\nWe demonstrate that QMDB scales linearly with IOPS on both consumer\nand ent erprise hardware\, reaching up to 2.3 million state updates per\nsecond. Q MDB achieves throughput up to 6x higher than RocksDB\, and up\nto 8x over NOMT\, the previous state-of-the-art verifiable database. \n\nOpen-source d code and preprint \n\nQMDB is the first in a series of planned papers f rom LayerZero Labs\nthat intend to challenge the assumption that high perf ormance\nnecessarily requires centralization. \n\n— \n\nDaniel Wong is a Research Engineer in LayerZero Labs Research\, a newly\nformed lab focu sed on scaling blockchain performance\, and the youngest\nconsortium membe r of the CMU Parallel Data Lab. Daniel completed his\nPhD thesis on ML for flash caching at the PDL in 2024. Like his\nadvisor Greg Ganger\, he has broad interests across systems research\nincluding caching\, distributed s torage\, cloud computing\, ML systems\nand security. He received his BA in Computer Science from the\nUniversity of Cambridge. In his spare time\, D aniel enjoys Singaporean\nfood and making sushi.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a174f DTSTART;TZID=America/New_York:20250313T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250313T120000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Ewin Tang CLASS:PUBLIC DESCRIPTION:Speaker: EWIN TANG\, Postdoctoral Fellow\, Miller Institute\, U niversity\nof California\, Berkeley\n\nTalk Title: What are quantum comput ers good for?\n\nQuantum computers—computers which exploit quantum mecha nics—are\npoised to reshape the landscape of computation. But understand ing when\n'quantum' can help speed up algorithmic tasks is tricky\, partic ularly\nfor those tasks which have the greatest potential for real-world\n impact. \n\nIn this talk\, I will survey my work in quantum algorithms to \nunderstand where quantum computers will be useful. I will argue that\nth is research can shed light\, not only on quantum computation\, but\nalso o n its proposed applications. \n\n— \n\nEwin Tang is a second-year post doctoral fellow at the Miller Institute\nat UC Berkeley\, broadly interest ed in quantum algorithms and quantum\nsystems. She received her PhD at the University of Washington\, advised\nby James Lee\, where she worked on in vestigating quantum speedups for\nmachine learning. Her work has been awar ded plenary talks and best\nstudent paper at QIP\; been featured in Quanta magazine\; and in 2019\nshe was named Forbes 30 Under 30 in science. \n\ nFaculty Host:  Aayush Jain \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a1b0b DTSTART;TZID=America/New_York:20250312T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250312T150000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Thomas Ristenpart CLASS:PUBLIC DESCRIPTION:Speaker: TOM RISTENPART\, Professor\, Cornell Tech\, Department of\nComputer Science\, Cornell University\n\nTalk Title: Computer Securit y in Known-Adversary Threat Models\n\nComputer security is traditionally a bout the protection of digital\nsystems from adversaries such as criminals or governments. In this\ntalk\, I will explore what I call known-adversar y threat models\, in\nwhich the adversary is a member of the victim's soci al circles—an\nintimate partner\, family member\, or other close acquain tance—and who\nseeks to cause harm to the victim\, not just their techno logy.\nKnown-adversary threat models are a widespread and increasingly sev ere\nproblem\, and their study opens up a new frontier for computer securi ty\nresearch and practice.\n\nThis perspective has emerged from my researc h and advocacy in the\ncontext of intimate partner violence (IPV). Our res earch details the\nmultifaceted and damaging ways in which abusers exploit technology to\nharass\, impersonate\, threaten\, monitor\, intimidate\, a nd otherwise harm\ntheir partner via spyware\, social media\, tracking too ls\, account\ncompromise\, and more. To make progress\, I take an advocate -scientist\napproach that blends ongoing evidence-based design and deliver y of\nclinics that directly assist survivors suffering tech abuse\; system ic\nadvocacy for new laws and policies\; and basic research into how to\nd esign technology to improve its security in the face of\nknown-adversary t hreats. The latter requires understanding and\nresolving complex tensions between known-adversary and traditional\nthreat models\, with result being principled reworkings of some of our\nmost fundamental security tools to increase security and safety for\nall users. \n\nThe talk will include di scussion of physical\, sexual\, and emotional\nviolence. \n\n— \n\nTho mas Ristenpart is a Professor at Cornell Tech and a member of the\nCompute r Science department at Cornell University. Before joining\nCornell Tech i n May\, 2015\, he spent four and a half years as an\nAssistant Professor a t the University of Wisconsin-Madison. He\ncompleted his PhD at UC San Die go in 2010.  His research spans a wide\nrange of computer security topics \, with recent focuses including\ndigital privacy and safety in interperso nal abuse\, anti-abuse\nmitigations for encrypted messaging systems\, impr ovements to\nauthentication mechanisms including passwords\, and topics in applied\nand theoretical cryptography. His work is routinely featured in the\nmedia and has been recognized by numerous distinguished paper awards\ ,\ntwo ACM CCS test-of-time awards\, a USENIX Security test-of-time award\ ,\nan Advocate of New York City award\, an NSF CAREER Award\, and a Sloan\ nResearch Fellowship. \n\nFaculty Hosts:  Lujo Bauer and Srini Seshan \ n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a1fe8 DTSTART;TZID=America/New_York:20250312T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250312T130000 LOCATION:Newell-Simon 3305 SUMMARY:Doctoral Speaking Skills Talk - Long Pham CLASS:PUBLIC DESCRIPTION:Speaker: LONG PHAM\, Ph.D. Student\, Computer Science Departmen t\,\nCarnegie Mellon University\n\nTalk Title: Hybrid Resource-Bound Analy sis of Programs\n\nResource-bound analysis aims to infer symbolic bounds o f worst-case\nresource usage (e.g.\, running time\, memory\, and energy) o f programs as\nfunctions of program inputs. Resource analysis has numerous \napplications\, including job scheduling in cloud computing and\npreventi on of side-channel attacks. Two approaches to resource\nanalysis are stati c analysis and data-driven analysis. Static analysis\nreasons about progra ms' behaviors by examining their source code.\nData-driven analysis first runs programs on many inputs to collect\ncost measurements and then statis tically analyzes this dataset.\n\nStatic and data-driven resource analyses have complementary strengths\nand weaknesses. Static analysis is sound: a ny cost bound successfully\ninferred is guaranteed to be a valid bound. Ho wever\, static analysis\nis incomplete: every analysis technique has a pro gram that it cannot\nhandle. On the other hand\, data-driven analysis can infer a candidate\ncost bound for any program. However\, it does not guara ntee soundness\nof inference results.\n\nIn this talk\, I describe my rece nt work (PLDI 2024) on hybrid resource\nanalysis\, which integrates static and data-driven analyses. Given a\nprogram\, the user first specifies whi ch analysis techniques are\napplied to which code fragments. The two analy ses are then performed\non their respective code fragments. Finally\, thei r inference results\nare combined into an overall cost bound of the progra m. Hybrid\nresource analysis retains the strengths of constituent analyses while\nmitigating their respective weaknesses.\n\nI start by introducing background on type-based resource analysis\nAutomatic Amortized Resource A nalysis (AARA). It automates the\npotential method from amortized analysis . I next introduce our first\ncontribution: Bayesian data-driven resource analysis. Finally\, I\npresent Hybrid AARA\, which integrates AARA and Bay esian data-driven\nanalysis via a novel interface between linear programmi ng and sample\nalgorithms. \n\nJoint work with Feras Saad and Jan Hoffman n.\n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequire ment\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a246e DTSTART;TZID=America/New_York:20250312T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250312T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20250312.html LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Yaowei Long CLASS:PUBLIC DESCRIPTION:Speaker: YAOWEI LONG\, Ph.D. Student\, Computer Science and Eng ineering\nDivision\, University of Michigan\n\nTalk Title: Connectivity La beling Schemes for Edge Faults via Expander\nHierarchies\n\nWe consider th e problem of assigning short labels to the vertices and\nedges of a graph G so that given any query   |F| ≤ f\, we can\ndetermine whether s and t are still connected in G — F\, given only\nthe labels of F ⊂ {s\,t}. This problem has been considered when F ⊂\nE (edge faults) and F ⊂ V (vertex faults)\, where correctness is\nguaranteed with high probability ( w.h.p.) or deterministically. In\nthis talk\, I will introduce a new deter ministic labeling scheme for\nedge faults that uses Õ(√ f¯ )-bit label s\, which can be\nconstructed in polynomial time. This improves on Dory an d Parter's\n[PODC 2021] existential bound of O(f log n) (requiring exponen tial\ntime to compute) and the efficient Õ(f2)-bit scheme of Izumi\, Emek \,\nWadayama\, and Masuzawa [PODC 2023]. Our construction uses an improved \nedge-expander hierarchy and a distributed coding technique based on\nRee d-Solomon codes. \n\nBased on joint work  with Seth Pettie and Thatchaph ol Saranurak.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a27eb DTSTART;TZID=America/New_York:20250311T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250311T200000 LOCATION:McConomy Auditorium\, Cohon University Center and Livestream SUMMARY:2025 3MT Championship CLASS:PUBLIC DESCRIPTION:Attend the 2025 3MT Championship \n\nThe Three Minute Thesis ( 3MT) is an internationally recognized\ncompetition that challenges Ph.D. s tudents to present a compelling\noration on their thesis and its significa nce in just three minutes\, in\nlanguage that anyone can understand. In th e 2025 3MT Championship\, CMU\nPh.D. students representing a variety of re search areas take to the\nstage to discuss their thesis for a chance to wi n cash prizes. CMU\nleadership and alumni judge the competition and the au dience has two\nopportunities to vote for their favorites:\n\nPeople's Cho ice Award  → selected by the in-person audience\; \nandAlumni Choice A ward → selected by the virtual audience\n\n3MT is not an exercise in tri vializing or ‘dumbing-down’ research\nbut requires students to consoli date their ideas\, crystallize their\nresearch discoveries and capture the imagination of their audience.\nThe 2025 3MT Championship will feature th e 10 preliminary round\nwinners as they compete for first\, second and thi rd place\, as well as\na live audience vote for People’s Choice and a li ve virtual vote for\nthe Alumni Award. \n\nREGISTRATION Required \n\n→   Event registration closes March 11 at 4:00 pm ET Sponsored and\nhosted by the CMU Libraries\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a2ba7 DTSTART;TZID=America/New_York:20250311T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250311T173000 LOCATION:Newell-Simon 3305 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Ananya Ashish Joshi CLASS:PUBLIC DESCRIPTION:Speaker: ANANYA ASHISH JOSHI\, Ph.D. Candidate\, Computer Scien ce\nDepartment\, Carnegie Mellon University\n\nTalk Title: Event Monitorin g in Modern Public Health Data Streams\n\nGrowing volumes of public health -related data render standard\ntechniques for syndromic surveillance (desi gned for smaller data\nvolumes) obsolete. My thesis presents a practical a pproach for experts\nto monitor large-scale aggregate public health data. These novel big\ndata monitoring methods identify data corresponding to qu ality issues\nor changes in disease dynamics and are simple\, scalable\,\n generalizable\, and shown to be accurate in real-world settings based\non human-labeled data. When paired with custom user interfaces\, these\nmetho ds have led to a 53-fold increase in monitoring efficiency for\ndata exper ts at the Delphi Group at Carnegie Mellon University.\nExperts can now det ect over 200 noteworthy data issues from 15 million\nnew data points each week. The output of this thesis' monitoring\napproach can directly support public health surveillance\, especially\nat the state or national level\, and increase the utility of public\nhealth data modernization efforts for data-driven decision-making.  \n\nThesis Committee\n\nRoni Rosenfeld (C o-Chair)\n\nBryan Wilder (Co-Chair)\n\nRayid Ghani\n\nMatthew Biggerstaff (Centers for Disease Control and Prevention)\n\nIn Person and Zoom Partici pation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a2f60 DTSTART;TZID=America/New_York:20250311T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250311T143000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Thesis Proposal - Caspar Oesterheld CLASS:PUBLIC DESCRIPTION:Speaker: CASPAR OESTERHELD\, Ph.D. Student\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: New foundational id eas for cooperative AI\n\nMy doctoral research addresses two fundamental o bstacles to beneficial\noutcomes from strategic interactions between multi ple parties:\nstrategic incentives against cooperation (as in the Prisoner 's\nDilemma) and the multiplicity of solutions (sometimes called the\nequi librium selection problem). As AI systems are increasingly\ninvolved in co nsequential decision making processes on behalf of human\nprincipals\, und erstanding how to achieve desirable outcomes in\nmulti-agent AI settings b ecomes critical. My research leverages unique\nfeatures of AI systems — including their transparency\,\nreproducibility\, and malleability — to develop novel game-theoretic\napproaches that enable better\, more coopera tive outcomes.      \n\nThree primary research directions form the cor e of this dissertation.\nFirst\, the concept of safe Pareto improvements  provides a rigorous\nframework for improving outcomes without resolving equilibrium\nselection problems. Unlike traditional solution concepts\, sa fe Pareto\nimprovements make qualitative assumptions about pairs of games rather\nthan individual games. This sometimes allows us to prefer playing one\ngame over another\, without any judgment about how each of the\nindiv idual games is played. Second\, the concept of program\nequilibrium expl ores how the use of mutually transparent\ndecision-making algorithms can a llow for cooperation. Third\, my\nresearch on so-called Newcomb-like dec ision problems takes\ninspiration from philosophical branches of decision theory. I\ninvestigate how cooperation can be achieved when different part ies\ndeploy similar AI systems.\n\nCurrent and planned work extends these directions through several\nprojects\, including: connecting program equil ibrium with mediated\nequilibrium\; exploring sequential program/mediated equilibrium-type\nsettings\; investigating the relationship between self-l ocating beliefs\nand decision theory\; developing theoretical foundations for safe\nPareto improvements\, as well as analyzing safe Pareto improveme nts in\na new setting. I've also started to implement some of these\ntheor etical ideas in language models to test their practical\napplicability. \ n\nThesis Committee\n\nVincent Conitzer (Chair)\n\nTuomas Sandholm\n\nFei Fang\n\nStuart Russell (University of California\, Berkeley)\n\nBen Levins tein (University of Illinois at Urbana-Champaign)\n\nAdditional Informatio n\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a3431 DTSTART;TZID=America/New_York:20250311T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250312T171000 URL:https://sites.google.com/andrew.cmu.edu/cbi-resai/schedule LOCATION:Bosch Spark Room\, Scott Hall 5201 SUMMARY:Carnegie Bosch Institute Symposium on Responsible & Sustainable AI CLASS:PUBLIC DESCRIPTION:Talk Title: Carnegie Bosch Institute Symposium on Responsible & \nSustainable AISave the date and REGISTER for this two-day summit\nfeatur ing keynote speakers from Carnegie Mellon University\, Penn State\nUnivers ity and Bosch\, who will share cutting-edge research and\nindustry insight s on responsible and sustainable AI.\nCBI Fellows will also present the un ique work they have had the\nopportunity to do as part of the fellowship p rogram\, covering topics\nfrom privacy\, security\, and inclusiveness in A I systems to\ntechnologies for environmental management.\nKeynote speakers include Lorrie Cranor\, Director and Bosch\nDistinguished Professor in Se curity and Privacy Technologies\, CyLab\nSecurity and Privacy Institute\; FORE Systems University Professor\,\nComputer Science and Engineering & Pu blic Policy\, Carnegie Mellon\nUniversity Valerie Karplus\, Professor\, De partment of Engineering and\nPublic Policy\; Associate Director of the Wil ton E. Scott Institute for\nEnergy Innovation\, Carnegie Mellon University Zico Kolter\, Professor\nand Director\, Machine Learning Department\, Car negie Mellon\nUniversity\;  Chief Expert\, BoschGlobal\; Board Member at OpenAI\;Chief\nTechnical Advisor\, GraySwanAI Jonatas Soares dos Santos. T echnology &\nInnovation Manager\, Bosch Aarti Singh\, Professor\, Machine Learning\nDepartment\, Carnegie Mellon University\; Director\, AI Institut e for\nSocietal Decision Making (NSF) Param Singh\, Carnegie Bosch Profess or\nof Business Technologies and Marketing\; Associate Dean for Research\, \nTepper School of Business\, Carnegie Mellon University Emma Strubell\,\n Raj Reddy Assistant Professor\, Language Technologies Institute\,\nCarnegi e Mellon University\;\nVisiting Scientist\, Allen Institute for Artificial Intelligence Shomir\nWilson\, Associate Professor\, College of Informatio n Sciences and\nTechnology\, Pennylsvania State UniversitySchedule and Add itional\nInformation\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a37fd DTSTART;TZID=America/New_York:20250310T203000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250311T133000 URL:https://www.cmu.edu/ethics-ai/events/generative-ai-conference-2025.html LOCATION:Simmons Auditorium\, First Floor\, Tepper Building SUMMARY:K&L Gates – Carnegie Mellon University Conference in Ethics and\n Computational Technologies CLASS:PUBLIC DESCRIPTION:Speaker: Generative AI: Ethics and Governance\n\nJoin for insig htful discussions on the intersection of ethics and\nartificial intelligen ce. \n\nGenerative AI is poised to create massive economic and societal\n impacts across a wide range of domains. This conference seeks to\nillumina te the new ethical considerations and societal implications of\nthe techno logy and elucidate the pros and cons of existing approaches\nto the govern ance of the technology to ensure safe\, responsible\, and\nethical use. Wa tch a panel discussion on the ethics of Generative AI\nfrom last year’s conference. \n\nKeynote Speakers  (learn more about all speakers/panelis ts)\n\nD.J. Patil\, General Partner\, GreatPoint VenturesNatasha Crampton\ , Vice\nPresident and Chief Responsible AI Officer\, Microsoft Corporation \n\nSessions include:\n\nKeynote speeches by prominent voices from industr y and\ngovernmentFireside chats with keynote speakersExpert panels on both \nsector-agnostic and sector-specific governance of generative AI\n\nREGIS TER → Registration Request by 3 March 2025 \n\nAbout the Conference:    The K&L Gates – CMU Conference in Ethics\nand Computational Technolog ies brings together thought leaders from\nacademia\, industry\, government and civil society to address the\nemerging ethical challenges and societa l implications of generative AI\nand examine the effectiveness of various governance mechanisms in\npromoting the safe\, responsible\, and ethical d evelopment and\ndeployment of this transformative technology to benefit so ciety and\nhumanity. \n\nConference Agenda\n\nMonday\, March 10 \n\n8:30 a.m - 4:30 p.m \n\nDay 1 of the conference will bring together experts f rom government\,\nindustry\, academia and civil society to discuss the sec tor-agnostic\nchallenges and opportunities of Generative AI governance. \ n\n►  Expert Panel 1: Ethics of Generative AI: New Issues and\nChalleng es       \n\n— This panel will explore the new ethical concerns th at arise with\nthe use of GenAI (compared to conventional predictive AI). Such issues\ninclude copyright\, widening power disparities\, and more.   \n\n►   Expert Panel 2: Governmental Policies on Generative AI\n        \n\n— This panel will explore existing governmental policies a nd\nguidance to ensure the ethical and responsible use of GenAI.  \n\n ►   Expert Panel 3: Organizational Governance of Generative AI\n       \n\n — This panel will explore internal organizational governance\ nstructures in major AI firms and how they map to international/federal\nr equirements.  \n\nTuesday\, March 11\n\n8:30 a.m - 1:30 p.m \n\nDay 2 o f the conference focuses on sector-specific governance of GenAI\nand evalu ating its impact on climate\, democracy\, labor and\ninnovation.  \n\n ►   Expert Panel 4: Governing the Impacts of GenAI Across Sectors\nand Domains        \n\n— This panel brings together experts to discu ss the impacts of\nGenerative AI in education\, employment\, medicine\, an d the environment\nand effective domain-specific governance mechanisms for mitigating\nharms and maximizing benefits.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a3e6c DTSTART;TZID=America/New_York:20250310T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250310T173000 URL:https://db.cs.cmu.edu/events/sql-death-pipe-syntax-in-sql LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Jeff Shute CLASS:PUBLIC DESCRIPTION:Speaker: JEFF SHUTE\, Google\n\nTalk Title: Pipe Syntax in SQL: SQL for the 21st Century\n\nSQL has been extremely successful as the de f acto standard language\nfor working with data. Virtually all mainstream da tabase-like systems\nuse SQL as their primary query language. But SQL is a n old language\nwith significant design problems\, making it difficult to learn\,\ndifficult to use\, and difficult to extend. Many have observed th ese\nchallenges with SQL\, and proposed solutions involving new languages. \nNew language adoption is a significant obstacle for users\, and none of\ nthe potential replacements have been successful enough to displace\nSQL.  \n\nWe've taken a different approach - solving SQL’s problems by\nexte nding SQL. Inspired by patterns that work well in other modern\ndata langu ages\, we added piped data flow syntax to SQL. The results\nare transforma tive - SQL becomes a flexible language that’s easier\nto learn\, use and extend. \n\nWe've seen great results from early adopters in F1 and BigQu ery\,\nshowing this syntax is compelling for users\, and showing the benef its\nof improving SQL from within\, allowing users to adopt new language\n features incrementally. \n\n— \n\nJeff Shute does data storage\, query \, and processing at Google\,\nincluding leading SQL language design and i mplementation in GoogleSQL\,\nthe SQL implementation underneath all of Goo gle's SQL-based products.\nThis talk is part of the \n\nSQL or Death? Sem inar Series\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a4271 DTSTART;TZID=America/New_York:20250310T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250310T150000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Charles Yuan CLASS:PUBLIC DESCRIPTION:Speaker: CHARLES YUAN\, Ph.D. Candidate\, Computer Science & Ar tificial\nIntelligence Laboratory\, Massachusetts Institute of Technology\ n\nTalk Title: Building the Tools to Program a Quantum Computer\n\nBringin g the promise of quantum computation into reality requires not\nonly build ing a quantum computer but also correctly programming it to\nrun a quantum algorithm. To obtain asymptotic advantage over classical\nalgorithms\, qu antum algorithms rely on the ability of data in quantum\nsuperposition to exhibit phenomena such as interference and\nentanglement. In turn\, an imp lementation of the algorithm as a program\nmust correctly orchestrate thes e phenomena in the states of qubits.\nOtherwise\, the algorithm would yiel d incorrect outputs or lose its\ncomputational advantage. Given a quantum algorithm\, what are the\nchallenges and costs to realizing it as a progra m that can run on a\nphysical quantum computer? \n\nIn this talk\, I answ er this question by showing how basic programming\nabstractions upon which many quantum algorithms rely – such as data\nstructures and control flo w – can fail to work correctly or\nefficiently on a quantum computer. I then show how we can leverage\ninsights from programming languages to re-i nvent the software stack of\nabstractions\, libraries\, and compilers to m eet the demands of quantum\nalgorithms. This approach holds out a promise of expressive and\nefficient tools to program a quantum computer and pract ically realize\nits computational advantage. \n\n— \n\nCharles Yuan is a Ph.D. candidate at MIT CSAIL working with Prof.\nMichael Carbin. His cu rrent research examines the challenges of\nprogramming quantum computers a nd other emerging models of\ncomputation. His work has appeared in the ACM SIGPLAN POPL\, OOPSLA\,\nand PLDI conferences and has been recognized wit h the SIGPLAN\nDistinguished Artifact Award and the CQE-LPS Doc Bedard Fel lowship. \n\nFaculty Host:  Robert Harper \n\nIn Person and Zoom Partic ipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a469f DTSTART;TZID=America/New_York:20250307T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250307T103000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Doctoral Speaking Skills Talk - Nicole Feng CLASS:PUBLIC DESCRIPTION:Speaker: NICOLE FENG\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Robust Algorithms for Win ding Numbers and Signed Distance\non General Domains\n\nAlgorithms for geo metric problems enable modeling and simulation of\nthe world around us — but they often break down on real-world\ngeometric data comprising noisy\ , incomplete\, or inaccurate\nobservations or representations of shapes. T hese imperfections\, along\nwith the ever-increasing heterogenity of geome tric data\, require\nalgorithms for robust geometry processing: versatile algorithms that\nwork reliably across varying degrees of quality of their input. I'll\ndemonstrate that the requisite robustness and versatility can be\nachieved by processing smooth\, globally-defined functions encoding t he\ngeometry of interest. Moreover\, we can unlock further control over\ng eometry and topology by processing higher-order derivatives of\nfunctions. As examples\, I'll present algorithms for curve and surface\nreconstructi on\, and signed distance computation. \n\nPresented in Partial Fulfillmen t of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a4a2e DTSTART;TZID=America/New_York:20250306T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250307T123000 URL:https://www.cmu.edu/block-center/news-events/block-events/cmu-mit-event -page.html LOCATION:Hamburg Hall SUMMARY:AI: Who Wins and Who Loses? CLASS:PUBLIC DESCRIPTION:Speaker: Hosted by: CMU Block Center for Technology and Society and\nMIT CSAIL FutureTech\n\nWe are pleased to invite you to our upcoming conference titled AI: Who\nWins and Who Loses? - hosted by Carnegie Mello n University’s Block\nCenter for Technology and Society and MIT’s CSAI L FutureTech\, with\nfunding from the NSF Directorate of Technology\, Inno vation\, and\nPartnerships (TIPs).  This interdisciplinary event will tak e place\n6-7 March 2025 on the CMU campus. \n\nAs artificial intelligence (AI) continues to reshape industries\,\nbusiness leaders\, policymakers\, and researchers must grapple with\ncritical questions about its economic diffusion\, workforce\nimplications\, and productivity potential. This wor kshop will bring\ntogether leading experts to explore:\n\nEmpirical method s for assessing AI adoption in business and\nindustryAI’s impact on labo r markets—automation\, augmentation\, and\nnew work creationThe reorgani zation of production and business\nprocesses due to AI integrationStrategi es for policymakers\, business\nleaders\, and researchers to navigate AI ’s economic challenges and\nopportunities\n\nThe workshop will feature e ngaging discussions\, data-driven insights\,\nand collaborative problem-so lving aimed at advancing AI research and\npolicy. \n\nPlease REGISTER  t o confirm your attendance.\n\n → Space is limited.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a4ddd DTSTART;TZID=America/New_York:20250303T000000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250308T000000 SUMMARY:Spring Break 2025 CLASS:PUBLIC DESCRIPTION:CMU Business Office will be open\, but there are no classes.\nN ormal schedules resume Monday\, March 10.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a5028 DTSTART;TZID=America/New_York:20250228T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250228T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Newell Simon 4305 and Zoom SUMMARY:AI-SDM Seminar - Ananya Rao CLASS:PUBLIC DESCRIPTION:Speaker: ANANYA RAO\, Ph.D. Student\, Robotics Institute\, Carn egie\nMellon University\n\nTalk Title: Heterogeneous Multi-Agent Coordinat ion for Disaster\nResponse\n\nUsing robots in disaster response scenarios can improve situational\nawareness and search effectiveness. Heterogeneous robots\, each with\ndifferent sensing and motion capabilities\, can be co ordinated to\ngather information more efficiently than a single robot or a \nhomogeneous team. This talk will cover two key aspects of\nheterogeneous multi-agent coordination for disaster response. First\,\nwe present metho ds of using the spectral decomposition of an\ninformation prior to distrib uting robotic agents with diverse skill\nsets to different sub-parts of th e overall search and coverage\nproblem\, such that each robot is leveragin g its unique capabilities to\nbest contribute to the team's goal. Second\, we present methods for\nautomating the process of classifying bases of op erations for disaster\nresponse and optimally selecting start locations fo r robots with\nheterogeneous capabilities. Experiments conducted on both\n synthetically generated and real-world maps demonstrate the efficacy\nof t he presented methods. \n\n— \n\nAnanya Rao is a PhD candidate at Carne gie Mellon University’s\nRobotics Institute\, advised by David Wettergre en and Howie Choset. Her\nwork focuses on multi-agent exploration of extre me environments\,\nparticularly on planning for search and rescue and plan etary\nexploration. She has a B.S. in Computer Science and a M.S. in Robot ics\nfrom Carnegie Mellon University. Her work has been published at\nseve ral top-tier robotics conferences and received the Best Paper\nAward at th e International Symposium on Distributed Autonomous Robotic\nSystems (DARS ) in 2021. REGISTER\n\nIn Person and Zoom Participation.  See announcemen t.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a5412 DTSTART;TZID=America/New_York:20250227T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250227T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Gates Hillman 8215 and Zoom SUMMARY:Crypto Seminar - Jonathan Shafer CLASS:PUBLIC DESCRIPTION:Speaker: JONATHAN SHAFER\, Postdoctoral Associate\, Computer Sc ience &\nArtificial Intelligence Laboratory\, Massachusetts Institute of\n Technology\n\nTalk Title: Oblivious Defense in ML Models: Backdoor Removal without\nDetection\n\nAs society grows more reliant on machine learning\, ensuring the\nsecurity of machine learning systems against sophisticated attacks\nbecomes a pressing concern. A recent result of Goldwasser\, Kim\, \nVaikuntanathan\, and Zamir (FOCS 2022) shows that an adversary can\nplan t undetectable backdoors in machine learning models\, allowing the\nadvers ary to covertly control the model’s behavior. Backdoors can be\nplanted in such a way that the backdoored machine learning model is\ncomputational ly indistinguishable from an honest model without\nbackdoors. \n\nIn this talk\, we present strategies for defending against backdoors in\nML model s\, even if they are undetectable. The key observation is that\nit is some times possible to provably mitigate or even remove backdoors\nwithout need ing to detect them\, using techniques inspired by the\nnotion of random se lf-reducibility. This depends on properties of the\nground-truth labels (c hosen by nature)\, and not of the proposed ML\nmodel (which may be chosen by an attacker). \n\nWe give formal definitions for secure backdoor mitig ation\, and proceed\nto show two types of results. First\, we show a “gl obal mitigation”\ntechnique\, which removes all backdoors from a machine learning model\nunder the assumption that the ground-truth labels are clo se to a\nFourier-heavy function. Second\, we consider distributions where the\nground-truth labels are close to a linear or polynomial function in\n R^n. Here\, we show “local mitigation” techniques\, which remove\nback doors with high probability for every input of interest\, and are\ncomputa tionally cheaper than global mitigation. All of our\nconstructions are bla ck-box\, so our techniques work without needing\naccess to the model’s r epresentation (i.e.\, its code or\nparameters). \n\nTo appear in STOC 202 5. Joint work with Shafi Goldwasser\, Neekon Vafa\,\nand Vinod Vaikuntanat han.\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a5842 DTSTART;TZID=America/New_York:20250227T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250227T160000 URL:https://aco.math.cmu.edu/abs-24-25/feb27.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Swee Hong Chang CLASS:PUBLIC DESCRIPTION:Speaker: SWEE HONG CHANG\, Assistant Professor\, Department of\ nMathematics\, Rutgers University\n\nTalk Title: Complexity of log-concave inequalities in matroids\n\nA sequence of nonnegative real numbers a1\, a 2\,…\, a_n\, is log-concave\nif a 2/1 i/2 ≥ ai-1}a_{i+1} for all i ran ging from 2 to n-1.\nExamples of log-concave inequalities range from inequ alities that are\nreadily provable\, such as the binomial coefficients ai = (n/i)\, to\nintricate inequalities that have taken decades to resolve\, such as the\nnumber of independent sets a in a matroid M with i elements ( otherwise\nknown as the first Mason's conjecture\; and was resolved by Jun e Huh in\n2010s in a remarkable breakthrough). It is then natural to ask i f it\ncan be shown that the latter type of inequalities is intrinsically\n more challenging than the former. In this talk\, we provide a rigorous\nfr amework to answer this type of questions\, by employing a combination\nof combinatorics\, complexity theory\, and geometry. \n\nThis is a joint wor k with Igor Pak.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a5bb5 DTSTART;TZID=America/New_York:20250227T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250227T130000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Alex Ozdemir CLASS:PUBLIC DESCRIPTION:Speaker: ALEX OZDEMIR\, Ph.D. Student\, Computer Science Depart ment\,\nStanford University\n\nTalk Title: Security and Privacy through Pr ogrammable Cryptography\,\nCompilers\, and Verification\n\nHow can we buil d large applications with strong security and privacy?\nHow can we balance complexity of the application's logic with the\ncomplexity of the cryptog raphic tools needed to achieve security? \n\nIn this talk\, I will discus s an answer to this question: the\nprogrammable cryptography stack. A prog rammable cryptosystem allows a\ndeveloper to build applications that secur ely compute over encrypted\,\nsecret\, or distributed data. The programmab le cryptography stack\nallows developers to express these secure applicati ons using high\nlevel abstractions. I will discuss three of my contributio ns to the\nprogrammable cryptography stack in the domains of cryptography\ ,\ncompilers\, and formal verification. \n\n— \n\nAlex Ozdemir is a Ph D student at Stanford advised by Dan Boneh and\nClark Barrett. His work co mbines ideas from cryptography\, compilers\,\nand verification. He and his collaborators have published 18 papers at\ntop-tier venues in computer se curity\, verification\, and distributed\nsystems\, and they won the Best T ool Paper award at TACAS 2022. \n\nFaculty Hosts: Wenting Zheng\, Fraser Brown \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a5f7a DTSTART;TZID=America/New_York:20250227T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250227T120000 LOCATION:Gates Hillman 6121 SUMMARY:Doctoral Speaking Skills Talk - Isabel Suizo CLASS:PUBLIC DESCRIPTION:Speaker: ISABEL SUIZO\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: A First Look at Starlink ’s Impact on Internet Access\nEquity\n\nWhile the digital divide still p ersists across social\, economic\, and\ngeographic boundaries\, low-earth orbiting (LEO) satellite networks\nshow great promise in bridging the Inte rnet connectivity gap. These\nnetworks feature a dense constellation of sa tellites that orbit at low\naltitudes\, allowing users to bypass many chal lenges of terrestrial\nconnectivity by transmitting packets directly to sp ace. As a result\,\nLEO network companies like Starlink\, Project Kuiper ( Amazon)\, and\nOneWeb frequently advertise their ability to bridge the con nectivity\ngap. \n\nIn this talk\, we introduce a measurement study with the goal of\nunderstanding how well Starlink\, the most widely deployed LE O network\,\nis actually able to improve Internet access in low income and remote\nregions. Using a combination of data from speed tests\, tracerout es\,\nand custom simulations\, we show that Starlink\, like similar attemp ts\nto reach low income and remote regions\, is still an incomplete\nsolut ion as currently deployed. However\, our findings regarding client\nto edg e server mapping and inter-satellite link deployment suggest\nthat there a re clear opportunities for improvements in these\nregions.  \n\nPresente d in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a6314 DTSTART;TZID=America/New_York:20250416T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250416T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting - April 2025 CLASS:PUBLIC DESCRIPTION:Speaker: CSD Faculty Meeting\n\nSee email announcement for deta ils.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a65ae DTSTART;TZID=America/New_York:20250212T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250212T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting - February 2025 CLASS:PUBLIC DESCRIPTION:See email announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a6836 DTSTART;TZID=America/New_York:20250226T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250226T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20250226.html LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Amir Azarmehr CLASS:PUBLIC DESCRIPTION:Speaker: AMIR AZARMEHR\, Ph.D. Student in Computer Science\, Kh oury\nCollege of Computer Sciences\, Northeastern University\n\nTalk Title : Stochastic Matching via In-n-Out Local Computation\nAlgorithms\n\nWe pro ve that a polynomial degree (in inverse of realization\nprobability) subgr aph can obtain a (1-epsilon)-approximation for\nstochastic matching\, impr oving over the previous quasi-polynomial\ndegree bound of [Behnezhad\, Der akhshan\, Hajiaghayi STOC'20]. Beyond\nits quantitative improvement\, a ke y conceptual contribution of our\nwork is to connect local computation alg orithms (LCAs) to the\nstochastic matching problem for the first time. Whi le prior work on\nLCAs mainly focuses on their out-queries (the number of vertices\nprobed to produce the output of a given vertex)\, our analysis a lso\nbounds the in-queries (the number of vertices that probe a given\nver tex). We prove that the outputs of LCAs with bounded in- and\nout-queries (in-n-out LCAs for short) have limited correlation\, a\nproperty that our analysis crucially relies on and might find\napplications beyond stochasti c matchings. \n\nThis talk is based on a joint work with Soheil Behnezhad \, Alma\nGhafari\, and Ronitt Rubinfeld\, to appear in STOC'25\, .\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a6b92 DTSTART;TZID=America/New_York:20250225T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250225T130000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Doctoral Speaking Skills Talk - Sam Arch CLASS:PUBLIC DESCRIPTION:Speaker: SAM ARCH\, Ph.D. Student\, Computer Science Department \,\nCarnegie Mellon University\n\nTalk Title: The Key to Effective UDF Opt imization: Before Inlining\,\nFirst Perform Outlining\n\nUser-defined func tions (UDFs) are procedural functions (written in\nPL/SQL\, Python) callab le from SQL queries. However\, queries with UDFs\nare notoriously hard to optimize as they mix declarative SQL and\nprocedural UDF code. In SQL Serv er 2019\, Microsoft shipped UDF\ninlining\, a feature eliminating all UDF code by translating entire\nUDFs to SQL. However\, inlining often leads to sub-optimal performance.\nWe propose a new optimization technique\, UDF o utlining\, to\nintentionally hide pieces of a UDF from the query optimizer \, resulting\nin significantly simpler and faster queries.  \n\nPresente d as part of the Database Group Lunch Talks \n\nPresented in Partial Fulf illment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a6eea DTSTART;TZID=America/New_York:20250225T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250225T130000 URL:http://www.cs.cmu.edu/~aiseminar/ LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:AI Seminar - Samy Bengio CLASS:PUBLIC DESCRIPTION:Speaker: SAMY BENGIO\, Senior Director\, AI and Machine Learnin g\nResearch\, Apple\n\nTalk Title: How far can transformers reason? the gl obality barrier and\ninductive scratchpad\n\nCan Transformers predict new syllogisms by composing established\nones?  More generally\, what type of targets can be learned by such\nmodels from scratch? Recent works show th at Transformers can be\nTuring-complete in terms of expressivity\, but thi s does not address\nthe learnability objective. \n\nThis presentation put s forward the notion of 'globality degree' to\ncapture when weak learning is efficiently achievable by regular\nTransformers\, where the latter meas ures the least number of tokens\nrequired in addition to the tokens histog ram to correlate nontrivially\nwith the target. As shown experimentally an d theoretically under\nadditional assumptions\, distributions with high gl obality cannot be\nlearned efficiently. \n\nIn particular\, syllogisms ca nnot be composed on long chains.\nFurthermore\, we show that (i) an agnost ic scratchpad cannot help to\nbreak the globality barrier\, (ii) an educat ed scratchpad can help if\nit breaks the globality barrier at each step\, (iii) a notion of\n'inductive scratchpad' can both break the globality bar rier and\nimprove the out-of-distribution generalization\, e.g.\, generali zing to\nalmost double input size for some arithmetic tasks. \n\n— \n\ nSamy Bengio (PhD in computer science\, University of Montreal\, 1993) is\ na senior director of machine learning research at Apple since 2021 and\na n adjunct professor at EPFL since 2024. Before that\, he was a\ndistinguis hed scientist at Google Research since 2007 where he was\nheading part of the Google Brain team\, and at IDIAP in the early 2000s\nwhere he co-wrote the well-known open-source Torch machine learning\nlibrary. \n\nHis rese arch interests span many areas of machine learning such as\ndeep architect ures\, representation learning\, vision and language\nprocessing and more recently\, reasoning. He is action editor of the\nJournal of Machine Learn ing Research and on the board of the NeurIPS\nfoundation. \n\nHe was on t he editorial board of the Machine Learning Journal\, has\nbeen program cha ir (2017) and general chair (2018) of NeurIPS\, program\nchair of ICLR (20 15\, 2016)\, general chair of BayLearn (2012-2015)\,\nMLMI (2004-2006)\, a s well as NNSP (2002)\, and on the program committee\nof several internati onal conferences such as NeurIPS\, ICML\, ICLR\, ECML\nand IJCAI. \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a73ac DTSTART;TZID=America/New_York:20250224T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250224T173000 URL:https://db.cs.cmu.edu/events/sql-death-apache-pinot-query-optimizer LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Yash Mayya\, Gonzalo Ortiz CLASS:PUBLIC DESCRIPTION:Speaker: YASH MAYYA\, GONZALO ORTIZ\, Pinot\n\nTalk Title: Apac he Pinot Query OptimizerThis talk is part of the SQL\nor Death? Seminar Se riesZoom Participation. See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a766b DTSTART;TZID=America/New_York:20250224T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250224T163000 URL:https://www.ri.cmu.edu/event/discovering-and-erasing-undesired-concepts / LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Niv Cohen CLASS:PUBLIC DESCRIPTION:Speaker: NIV COHEN\, Postdoctoral Researcher\, Computer Science and\nEngineering\, Tandon School of Engineering\, New York University\n\n Talk Title: Discovering and Erasing Undesired Concepts\n\nThe rapid growth of generative models allows an ever-increasing\nvariety of capabilities. Yet\, these models may also produce undesired\ncontent such as unsafe or m isleading images\, private information\, or\ncopyrighted material. \n\nIn this talk\, I will discuss practical methods to prevent undesired\ngenera tions. First\, I will show how the challenge of avoiding\nundesired genera tions manifested itself in a simple Capture-the-Flag\nLLM setting\, where even our top defense strategy was breached. Next\, I\nwill demonstrate a s imilar vulnerability in state-of-the-art concept\nerasure methods for Text -to-Image models. Finally\, I will distinguish\nbetween erasure through Gu idance-Based Avoidance and Destruction-Based\nRemoval methods. I will disc uss the trade-offs of each approach and\ntheir behavior in various setting s. \n\n—\n\nNiv Cohen is a postdoctoral researcher at New York Universi ty hosted\nby Prof. Chinmay Hegde. He received a BSc in mathematics with p hysics\nas part of the Technion Excellence Program. He received his PhD in \ncomputer science from the Hebrew University of Jerusalem\, advised by\nP rof. Yedid Hoshen. Niv was awarded the Israeli data science\nscholarship f or outstanding postdoctoral fellows (VATAT). He is\ninterested in anomaly detection\, representation learning\, and AI\nsafety for Vision & Language models.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a7a38 DTSTART;TZID=America/New_York:20250224T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250224T130000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Speaking Skills Talk - Yiran Lei CLASS:PUBLIC DESCRIPTION:Speaker: YIRAN LEI\, Ph.D. Student\, Computer Science Departmen t\,\nCarnegie Mellon University\n\nTalk Title: FLASH: Fast All-to-All Comm unication in GPU Clusters\n\nScheduling All-to-All communications efficien tly is fundamental to\nminimizing job completion times in distributed syst ems. Incast and\nstraggler flows can slow down All-to-All transfers\; and GPU clusters\nbring additional straggler challenges due to highly heteroge neous link\ncapacities between technologies like NVLink and Ethernet. Exis ting\nschedulers all suffer high overheads relative to theoretically optim al\ntransfers. Classical\, simple scheduling algorithms such as SpreadOut\ nfail to minimize transfer completion times\; modern optimization-based\ns chedulers such as TACCL achieve better completion times but with\ncomputat ion times that can be orders of magnitude longer than the\ntransfer itself . \n\nThis paper presents FLASH\, which schedules near-optimal All-to-All \ntransfers with a simple\, polynomial time algorithm. We prove that\, so\ nlong as intra-server networks are significantly faster than\ninter-server networks\, FLASH approaches near-optimal transfer\ncompletion times. We i mplement FLASH and demonstrate that its\ncomputational overheads are negli gible\, yet it achieves transfer\ncompletion times that are comparable to state-of-the-art solver-based\nschedulers. \n\nPresented in Partial Fulfi llment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a7de7 DTSTART;TZID=America/New_York:20250224T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250224T130000 URL:https://cmubiglab.github.io/lunch/ LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Accessibility Lunch Seminar - Kayo Yin CLASS:PUBLIC DESCRIPTION:Speaker: KAYO YIN\, Ph.D. Student\, Berkeley AI Research (BAIR) \,\nDepartment of Electrical Engineering and Computer Sciences \,\nUniver sity of California\, Berkeley\n\nTalk Title: AI for Signed Languages: Chal lenges and Opportunities\n\nSigned languages are complex natural languages that rely on multiple\narticulators moving simultaneously in 3D space. De spite recent\nadvances in AI and language modeling\, developing technology that can\ntruly understand and work with signed languages remains challen ging.\nThis talk will introduce the linguistic properties of American Sign \nLanguage (ASL) that underscore the complexity of signed language\nmodeli ng\, and discuss current progress in signed language technology.\nThen\, I will dive deeper into our recent work on developing AI-powered\ntools to make STEM education more accessible to deaf and\nhard-of-hearing students. Finally\, I will conclude with open\nchallenges and future directions for signed language AI research. \n\n— \n\nKayo Yin is a PhD student at U C Berkeley advised by Jacob Steinhardt\nand Dan Klein. She currently works on LLM interpretability and NLP for\nsigned languages. Before that\, she was a Master’s student at\nCarnegie Mellon University advised by Graham Neubig\, and completed her\nundergraduate studies at École Polytechnique. Her research has been\nrecognized by the ACL 2023 Best Resource Paper awa rd\, EMNLP 2022 Best\nPaper Honorable Mention award\, ACL 2021 Best Theme Paper award\, a\nSiebel Scholarship\, the Thomas Clarkson medal\, and the Vitalik Buterin\nFellowship on AI Safety. \n\nIn Person and Zoom Particip ation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a81d5 DTSTART;TZID=America/New_York:20250220T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250220T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Gates Hillman 8215 and Zoom SUMMARY:Crypto Seminar - Yanyi Liu CLASS:PUBLIC DESCRIPTION:Speaker: YANYI LIU\, Ph.D. Student in Computer Science\, Cornel l Tech\,\nCornell University\n\nTalk Title: White-Box Learning and Public- Key Encryption\n\nWe consider a generalization of the Learning With Error problem\,\nreferred to as the white-box learning problem: You are given th e code\nof a sampler that with high probability produces samples of the fo rm\ny\,f(y) + ϵ where ϵ is small\, and f is computable in polynomial-siz e\,\nand the computational task consists of outputting a polynomial-size\n circuit C that with probability\, say\, 1/3 over a new sample y'\naccordin g to the same distributions\, approximates f(y') (i.e.\,\n|C(y')-f(y')| is small). This problem can be thought of as a\ngeneralization of the Learni ng with Error Problem (LWE) from linear\nfunctions f to polynomial-size co mputable functions. \n\nWe demonstrate that worst-case hardness of the wh ite-box learning\nproblem\, conditioned on the instances satisfying a noti on of\ncomputational shallowness (a concept from the study of Kolmogorov\n complexity) not only suffices to get public-key encryption\, but is\nalso necessary\; as such\, this yields the first problem whose\nworst-case hard ness characterizes the existence of public-key\nencryption. Additionally\, our results highlight to what extent LWE\n``overshoots\" the task of publ ic-key encryption. We complement these\nresults by noting that worst-case hardness of the same problem\, but\nrestricting the learner to only get bl ack-box access to the sampler\,\ncharacterizes one-way functions. \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a8b68 DTSTART;TZID=America/New_York:20250220T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250220T160000 URL:https://aco.math.cmu.edu/abs-24-25/feb20.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Siamak Yassemi CLASS:PUBLIC DESCRIPTION:Speaker: SIAMAK YASSEMI\, Shelly Visiting Professor\, Departmen t of\nMathematical Sciences\, Carnegie Mellon University\n\nTalk Title: f- vectors and h-vectors of Simplicial Complexes: A\nConjecture by Kalai\n\nI n this talk\, we revisit the concepts of f-vectors and h-vectors for\nsimp licial complexes\, with a focus on Kalai’s conjecture concerning\nthe f- vectors of flag simplicial complexes. We then review key results\nand deve lopments related to this conjecture. \n\n4:00 pm - Jane Street-sponsored tea and cookies in Wean 6220 (please\nbring your own mug)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a8f43 DTSTART;TZID=America/New_York:20250219T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250219T130000 URL:https://www.cs.cmu.edu/~theorylunch LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Louis Golowich CLASS:PUBLIC DESCRIPTION:Speaker: LOUIS GOLOWICHTalk Title: Improved Fault-Tolerant\nNon -Clifford Gates (Or: How to Multiply Quantumly)\n\nA principal challenge i n realizing the potential of quantum computing\nlies in our ability to per form computations fault-tolerantly\, in the\npresence of the noise inheren t to quantum devices. Non-Clifford\nquantum gates\, which are analogous to the classical multiplication\n(i.e. AND) gate\, are particularly difficul t to implement\nfault-tolerantly. We show how to perform such gates with s ignificantly\nlower asymptotic overhead than was achievable with prior\nte chniques.  \n\nFor this purpose\, we present new constructions of quantu m\nerror-correcting codes supporting transversal non-Clifford gates\,\nmea ning that the desired logical gates can be executed by applying a\nconstan t-depth physical circuit to the code state. In particular\, we\npresent th e first asymptotically good such codes\, as well as the first\nsuch codes with low-weight stabilizers that have almost linear\ndimension and polynom ial distance. Our constructions are based on a\ncombination of algebraic a nd topological techniques.  No prior\nknowledge of quantum computing will be assumed. \n\nBased on joint works with Venkatesan Guruswami and Ting- Chun Lin\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a9327 DTSTART;TZID=America/New_York:20250218T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250218T130000 URL:http://www.cs.cmu.edu/~aiseminar/ LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Artificial Intelligence Seminar - Keegan Harris CLASS:PUBLIC DESCRIPTION:Speaker: KEEGAN HARRIS\, Ph.D. Student \, Machine Learning Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Should You Use Your La rge Language Model to Explore or\nExploit?\n\nIn-context (supervised) lear ning is the ability of an LLM to perform\nnew prediction tasks by conditio ning on examples provided in the\nprompt\, without any updates to internal model parameters. Although\nsupervised learning is an important capabilit y\, many applications\ndemand the use of ML models for downstream decision making. Thus\,\nin-context reinforcement learning (ICRL) is a natural nex t frontier. \n\nIn this talk\, we investigate the extent to which contemp orary LLMs can\nsolve ICRL tasks. We begin by deploying LLMs as agents in simple\nmulti-armed bandit environments\, specifying the environment\ndesc ription and interaction history entirely in-context. We experiment\nwith s everal frontier models and find that they do not engage in\nrobust decisio n making behavior without substantial task-specific\nmitigations. Motivate d by this observation\, we then use LLMs to\nexplore and exploit in silos in various (contextual) bandit tasks. We\nfind that while the current gene ration of LLMs often struggle to\nexploit\, in-context mitigations may be used to improve performance on\nsmall-scale tasks. On the other hand\, we find that LLMs do help at\nexploring large action spaces with inherent sem antics\, by suggesting\nsuitable candidates to explore. \n\nThis talk is based on joint work with Alex Slivkins\, Akshay\nKrishnamurthy\, Dylan Fos ter\, and Cyril Zhang. \n\n— \n\nKeegan Harris is a final-year Machine Learning PhD candidate at CMU\,\nwhere he is advised by Nina Balcan and S teven Wu\, and does research on\nmachine learning for decision making. He has been recognized as a\nRising Star in Data Science and his research is supported by an NDSEG\nFellowship. He is also the head editor of the ML@CM U blog. Previously\,\nKeegan spent two summers as an intern at Microsoft R esearch and\ngraduated from Penn State with BS degrees in Computer Science and\nPhysics. \n\nIn Person and Zoom Participation.  See announcement.\ n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a9794 DTSTART;TZID=America/New_York:20250217T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250217T183000 URL:https://www.cmu.edu/cmist/news-events/index.html LOCATION:Grand Room\, Posner Hall 340 SUMMARY:CMIST: Scientists and Strategists Series - Jonas Sandbrink CLASS:PUBLIC DESCRIPTION:Speaker: JONAS SANDBRINK\, Workstream Lead \, UK AI Safety Inst itute\n\nTalk Title: How Can We Manage the Misuse Potential of Increasingl y\nPowerful Technologies?\n\nJoin CMIST for a discussion with Jonas Sandbr ink on how humanity is\ndeveloping increasingly impactful technologies and scientific\ninsights. Whether it's large language models\, the ability to engineer\nviruses for delivering genetic material\, or a detailed underst anding\nof potential pandemic pathogens\, these advancements promise\nsign ificant societal benefits. However\, they also pose risks of\nunprecedente d harm if misused. Dr. Sandbrink will explore how the\ndynamics of misuse risks and benefits scale across different types of\ntechnologies and insig hts. He will discuss how scientists\, technology\ndevelopers\, and policym akers can consider the unique properties of\nvarious technologies to maxim ize their positive impact. \n\nThis event will explore how\, by fostering risk-sensitive innovation\,\nwe can help unlock the tremendous benefits o f new technologies while\nsafeguarding humanity from potential downsides.  \n\n— \n\nJonas Sandbrink is a Workstream Lead at the UK AI Safety In stitute.\nPreviously\, Jonas was a researcher at the University of Oxford\ , where\nhis work focused on the implications of synthetic biology and\nar tificial intelligence for biological security risks. He has served\nas a B iosecurity Advisor to the UK Cabinet Office and Google DeepMind.\nHe has w orked as a consultant for the Nuclear Threat Initiative\, and\nhas complet ed fellowships with the Johns Hopkins Center for Health\nSecurity\, the Co uncil on Strategic Risks\, and the Arms Control\nNegotiation Academy. Jona s holds a Bachelor in Medical Sciences from\nthe University of Oxford and a Ph.D. (DPhil) from Nuffield Department\nof Medicine\, University of Oxfo rd. \n\nREGISTER\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91a9c06 DTSTART;TZID=America/New_York:20250217T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250217T173000 URL:https://db.cs.cmu.edu/events/sql-death-towards-sanity-in-query-language s LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Victor Leis\, Thomas Neumann CLASS:PUBLIC DESCRIPTION:Speaker: VICTOR LEIS\, THOMAS NEUMANN\, School of Computation\, \nInformation and Technology\, Technical University of Munich (TUM_\n\nTal k Title: Towards Sanity in Query Languages\n\nThe relational model has sto od the test of time is the foundation of\nmost database systems. But let's be honest -- its success is not\nbecause of SQL\, but in spite of it. SQL 's syntax is arcane\,\ninconsistent\, and bears little resemblance to the actual execution\nsemantics of queries. Worse yet\, SQL is not even a true standard --\nevery system implements its own incompatible dialect\, creat ing a\nfractured ecosystem where portability and interoperability are\naft erthoughts. This lack of standardization slows down innovation\,\nforcing database developers to reinvent the wheel instead of pushing\nthe field fo rward. In this talk\, we argue that we need a formally\ndefined Sane Inter mediate Representation (SaneIR) — a\nwell-specified\, interoperable abst raction for relational query\nexecution. SaneIR would serve as a lingua fr anca between database\nsystems\, enabling compatibility\, portability\, an d accelerated progress\nin the database world. It's time to stop treating SQL as the immovable\nfoundation of relational databases and start designi ng a future where\ndatabase systems can truly evolve. \n\n— \n\nViktor Leis and Thomas Neumann are two of the best database\nresearchers in the world. They are German. \n\nThis talk is part of the SQL or Death? Semina r Series\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91aa03b DTSTART;TZID=America/New_York:20250217T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250217T163000 LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Jia-Bin Huang CLASS:PUBLIC DESCRIPTION:Speaker: JIA-BIN HUANG\, Capital One Associate Professor\, Depa rtment of\nComputer Science\, University of Maryland College Park\n\nTalk Title: Controllable Visual Imagination\n\nGenerative models have empowered human creators to visualize their\nimaginations without artistic skills a nd labor. A prominent example is\nlarge-scale text-to-image generation mod els. However\, these models\noften are difficult to control and do not res pect 3D perspective\ngeometry and temporal consistency of videos. In this talk\, I will\nshowcase several of our recent efforts to improve controll ability for\nimage\, video\, 3D generation\, and editing. Specifically\, I will talk\nabout how we improve semantic control for 2D image generation s\,\ngenerate realistic textures from reference images for 3D objects\, an d\nsynthesize novel views\, lighting\, and weather for 3D scenes. \n\n—  \n\nJia-Bin Huang is a Capital One-endowed Associate Professor of Comput er\nScience at the University of Maryland College Park. Before coming to\ nUMD\, Huang was a research scientist at Meta Reality Labs. Before that\,\ nhe was an Assistant Professor of Electrical and Computer Engineering\nat Virginia Tech. Huang received his Ph.D. from the University of\nIllinois\, Urbana-Champaign (UIUC) in 2016. His research interests\ninclude 3D compu ter vision\, generative models\, and computational\nphotography. Huang is the recipient of the Thomas & Margaret Huang\nAward\, NSF CRII award\, fa culty award from Samsung\, Google\, 3M\,\nQualcomm\, and a Google Research Scholar Award. \n\nFaculty Host: Fernando De La Torre Frade\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91aa426 DTSTART;TZID=America/New_York:20250217T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250217T120000 URL:https://www.cylab.cmu.edu/events/2025/02/17-seminar-pagliarini.html LOCATION:Panther Hollow Conference Room 4105\, Mehrabian Collaborative\nInn ovation Center SUMMARY:CyLab Seminar - Samuel Pagliarini CLASS:PUBLIC DESCRIPTION:Speaker: SAM PAGLIARINI\, Special Professor\, Department of Ele ctrical\nand Computer Engineering\, Carnegie Mellon University\n\nTalk Tit le: REPQC: Reverse Engineering and Backdooring Hardware\nAccelerators for Post-quantum Cryptography\n\nSignificant research efforts have been dedica ted to designing\ncryptographic algorithms that are quantum-resistant. The motivation is\nclear: robust quantum computers\, once available\, will re nder current\ncryptographic standards vulnerable. Thus\, we need new Post- Quantum\nCryptography (PQC) algorithms\, and\, due to the inherent complex ity of\nsuch algorithms\, there is also a demand to accelerate them in\nha rdware. In this talk\, we show that PQC hardware accelerators can be\nback doored by two different adversaries located in the chip supply\nchain. We propose REPQC\, a  sophisticated reverse engineering\nalgorithm that can be employed to confidently identify hashing\noperations (i.e.\, Keccak) wi thin the PQC accelerator - the location of\nwhich serves as an anchor for finding secret information to be leaked.\nFrom there\, an adversary can pr oceed to mount specific attacks that\nmake a quantum-resistant accelerator vulnerable to classic hardware\nsecurity threats. \n\n— \n\nSamuel Pa gliarini is a Special Professor of Electrical and Computer\nEngineering at Carnegie Mellon University in Pittsburgh\, PA. His\nresearch interests in clude Application Specific Integrated Circuits\,\nComputer Aided Design\, and Hardware Security. Prof. Pagliarini holds\naPhD from Telecom ParisTech \, Paris\, France. He also holds a Master’s\ndegree and a Bachelor’s d egree from Universidade Federal do Rio\nGrande do Sul\, Porto Alegre\, Bra zil. From 2019 to 2023\, Prof.\nPagliarini held a tenured professor positi on at Tallinn University of\nTechnology where he led the Centre for Hardwa re Security. His research\non hardware security and chip design has been f unded by several\nresearch agencies in the US and in Europe\, including th e NSF\, DARPA\,\nIARPA\, and the European Commission. \n\nIn Person and Z oom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91aa87b DTSTART;TZID=America/New_York:20250214T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250214T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Newell Simon 4305 and Zoom SUMMARY:AI-SDM Seminar - Valerie Chen CLASS:PUBLIC DESCRIPTION:Speaker: VALERIE CHEN\, Ph.D. Student\, Machine Learning Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Effective Human-AI Inter actions in an Evolving AI\nLandscapeThe evolving AI landscape means that m odels have the\npotential to support human decision-makers and users in a wide range\nof fields and applications. And yet\, human-AI interactions in many\ndeployment contexts still face practical challenges. In this talk\, we\ndiscuss how to design and evaluate these interactions in two common\n paradigms. The first half focuses on predictive decision-making\, and\nthe role of explanations in facilitating appropriate reliance on AI\nrecommen dations. The second half explores more complex settings where\nAI are “c opilots”\, highlighting gaps in existing evaluations and\nproposing a ne w in-the-wild framework. We conclude by discussing\ncommon themes across p aradigms and lessons learned for building\neffective human-AI interactions .\n\n—\n\nValerie Chen is a Machine Learning Ph.D. student at CMU. Her r esearch\naims to improve human-AI interactions by leveraging insights from \npractical user studies to design new interactive systems. Her research\n sits at the intersection of ML\, NLP\, and HCI. Valerie is a recipient\nof the NSF Graduate Research Fellowship\, a former intern at MSR’s\nFairne ss\, Accountability\, Transparency & Ethics in AI group\, and was\nselecte d as a Rising Star in Data Science.\n\nREGISTER  → upon registration\, confirmation information on joining\nthe meeting will be provided\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91aac58 DTSTART;TZID=America/New_York:20250214T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250214T160000 LOCATION:Tepper Building 4219 SUMMARY:Operations Research Seminar - Rad Niazadeh CLASS:PUBLIC DESCRIPTION:Speaker: RAD NIAZADEH\, Assistant Professor of Operations Manag ement\,\nand Asness Junior Faculty Fellow\, Booth School of Business\, The \nUniversity of Chicago\n\nTalk Title: Batching and Optimal Multi-stage Bi partite Allocations:\nTheory and Practice\n\n In many real-time matching applications in online\nmarketplaces—such as matching riders to drivers in ride-sharing or\nallocating video ads to YouTube users—the platform a llows (or\ninherently has) some latency to batch demand and improve effici ency in\nnon-stationary environments. Motivated by these scenarios\, I\nin vestigate the optimal trade-off between batching and inefficiency in\ndesi gning online allocations under non-stationary arrivals. In\nparticular\, I consider a variant of the classic adversarial online\nbipartite allocatio n family of problems where demand arrives\nstage-wise in batches rather th an fully online. Our main result is an\noptimal-competitive (fractional) m atching algorithm\, improving the\nclassic competitive ratio known for the online variants [Mehta et al.\,\n2007\; Aggarwal et al.\, 2011]. \n\nTim e permitting\, I will briefly discuss a generalization to multistage\nconf iguration allocations and its applications to video display\nadvertising a nd AdX.   Our main technique is to develop algorithmic\ntools that vary the trade-off between “greediness” and\n“hedging” across stages. W e rely on a particular family of\nconvex-programming-based matchings that distribute demand in a\nspecifically balanced way among supply across diff erent stages\, while\ncarefully adjusting the resulting matching. More pre cisely\, we\nidentify a unique sequence of polynomials with decreasing deg rees to\nserve as strictly concave regularizers for the basic linear progr am at\neach stage. At each stage\, our multi-stage algorithm solves the\nc orresponding convex program to obtain the regularized optimal\nmatching. B y providing a structural decomposition (tied to these\nconvex programs) of the subgraph at each stage and recursively\nconnecting the regularizers\, we develop a new multi-stage primal-dual\nframework to analyze the compet itive ratio. For the extension to\nmulti-stage configuration allocation\, we propose a new idea of\nregularizing separately at different price level s and analyze the\nresulting price-level convex programming-based multi-st age algorithm\nwith another primal-dual argument. The talk is based on two papers:\n\nPaper 1  ( Joint with Yiding Feng\, Management Science 2024)P aper 2 \n( Joint with Yiding Feng and Amin Saberi\, Operations Research 2 023)\n\nRad Niazadeh is an Assistant Professor of Operations Management an d\nAsness Junior Faculty Fellow at The University of Chicago Booth School\ nof Business. He is also part of the faculty at Toyota Technological\nInst itute of Chicago (TTIC) by a courtesy appointment. Prior to\njoining Chica go Booth\, he was a visiting researcher at Google Research\nNYC and a post doctoral fellow at Stanford University\, Computer Science\nDepartment. He finished his PhD in Computer Science (minored in\nApplied Mathematics) at Cornell University. Rad primarily studies the\ntheory and applications of online algorithms\, online learning\, and\nmechanism design. His research aims to design market algorithms and\nmechanisms for real-time operations of online marketplaces and\ne-commerce platforms\, as well as operations o f governmental agencies\nand non-profit organizations. \n\nRad has receiv ed several awards for his research\, including the\nINFORMS Auctions and M arket Design Rothkopf Junior Researcher Paper\nAward (2021\, 2024: first p lace\, 2023:second place)\, the INFORMS\nRevenue Management and Pricing Di ssertation Award (honorable mention)\,\nIJCAI 2024 Distinguished Paper Awa rd\, MSOM Best Student Paper (first\nplace)\, Service Science Best Student Paper (third place)\, and the\nGoogle PhD Fellowship (in Market Algorithm s). \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ab20f DTSTART;TZID=America/New_York:20250214T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250214T140000 URL:https://www.cmu.edu/flame/events/index.html LOCATION:Tepper Building 1403 SUMMARY:CMU FLAME Center Seminar - Sachin Goyal & Christina Baek CLASS:PUBLIC DESCRIPTION:Speaker: SACHIN GOYAL\, CHRISTINA BAEK\, Ph.D. Students\, Machi ne\nLearning Department\, Carnegie Mellon University\n\nTalk Title: Contex t-Parametric Inversion: Why Instruction Finetuning\nCan Worsen Context Rel iance\n\nA standard practice when using large language models is for users to\nsupplement their instruction with an input context containing new\nin formation for the model to process. However\, models struggle to\nreliably follow the input context\, especially when it conflicts with\ntheir param etric knowledge from pretraining. In-principle\, one would\nexpect models to adapt to the user context better after instruction\nfinetuning\, partic ularly when handling knowledge conflicts. However\,\nwe observe a surprisi ng failure mode: during instruction tuning\, the\ncontext reliance under k nowledge conflicts initially increases as\nexpected\, but then gradually d ecreases as instruction finetuning\nprogresses. This happens while the per formance on standard benchmarks\nkeeps on increasing far after this drop. We call this phenomenon\ncontext-parametric inversion and observe it acros s multiple general\npurpose instruction tuning datasets such as TULU\, Alp aca and\nUltrachat\, across different model families like Llama\, Mistral\ , and\nPythia. We perform various controlled studies and theoretical analy sis\nto show that context-parametric inversion occurs due to examples in\n the instruction finetuning data where the input context provides\ninformat ion that aligns with model's parametric knowledge. Our\nanalysis suggests some natural mitigation strategies with limited but\ninsightful gains\, an d serves as a useful starting point in addressing\nthis deficiency in inst ruction finetuning. \n\nPaper \n\n— \n\nSachin Goyal is a fourth yea r PhD student in the Machine Learning\nDepartment (MLD) at CMU\, advised b y Prof. Zico Kolter. His current\nresearch focus includes robust training and fine-tuning of foundation\nmodels. \n\nChristina Baek is a fourth-y ear PhD student in the Machine Learning\nDepartment (MLD) at CMU\, advised by Zico Kolter and Aditi Raghunathan.\nShe is broadly interested in ML sa fety and understanding deep learning\nthrough scientific methods. Her rese arch focuses on understanding the\nout-of-distribution robustness and long -tail behaviors of models. She\nhas worked on strategies for model assessm ent under real-world shifts\nwith limited labeled data. Lately\, she has b een interested in ensuring\nthe safety of agentic systems through theory-g uided insights for model\nfailures and how they snowball across training\, inference\, and\ninteraction. \n\nIn Person and Zoom Participation.  Se e announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ab76a DTSTART;TZID=America/New_York:20250214T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250214T113000 LOCATION:Gates Hillman 8102 SUMMARY:Doctoral Speaking Skills - Madhusudan Reddy Pittu CLASS:PUBLIC DESCRIPTION:Speaker: MADHUSUDHAN REDDY PITTU\, Ph.D. Student\, Computer Sci ence\nDepartment\, Carnegie Mellon University\n\nTalk Title: Matroid Inter section and Determinant Maximization\nMathJax.Hub.Config({\ntex2jax: {\nin lineMath: [ ['$'\,'$']\, [\"\\\\(\"\,\"\\\\)\"] ]\,\nprocessEscapes: true\ n}\n})\;\n\nDeterminant maximization provides an elegant generalization of \nproblems in many areas\, including convex geometry\, statistics\, machin e\nlearning\, fair allocation of goods\, and network design.In an instance \nof the determinant maximization problem\, we are given a collection of\n vectors $v_1\,\\ldots\, v_n \\in \\mathbb{R}^d$\, and the goal is to pick a\nsubset $S\\subseteq [n]$ of given vectors to maximize the determinant\n of the matrix $\\sum_{i \\in S} v_iv_i^\\top$\, where the picked set of\nv ectors $S$ must satisfy some combinatorial constraint such as\ncardinality constraint ($|S| \\leq k$) or matroid constraint ($S$ is a\nbasis of a ma troid defined on $[n]$).\n\nWe give efficient deterministic combinatorial algorithms for the\ndeterminant maximization problem under a matroid const raint that\nachieves $O(r^{O(r)})$-approximation for any matroid of rank $ r\\leq d$\nand $O(d^O(d))$-approximation for any matroid of rank $r\\geq d $. The\nalgorithm for the $r\\leq d$ case relies on the geometric\ninterpr etation of the determinant whereas the algorithm for the $r\\geq\nd$ case relies on the algebraic properties of the determinant and the\nproperties of a convex programming relaxation introduced by Madan et\nal. (FOCS '20). In both cases\, we use matroid intersection as a local\nsearch tool to it eratively improve a solution by finding an\nalternating negative cycle in an appropriately defined exchange graph\ndefined by the matroids.\n\nJoint work with Adam Brown\, Aditi Laddha\, Mohit Singh\, and Prasad\nTetali.\n \nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91abbb2 DTSTART;TZID=America/New_York:20250213T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250213T160000 URL:https://aco.math.cmu.edu/abs-24-25/feb13.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Dhruv Mybayi CLASS:PUBLIC DESCRIPTION:Speaker: DHRUV MUBAYI\, Professor\, Department of Mathematics\, \nStatistics\, and Computer Science\, University of Illinois Chicago\n\nTa lk Title: New results on the Erdos Rogers functionWhat is the\nmaximum siz e of a triangle-free subgraph that every n vertex K4–free\ngraph is guar anteed to contain? This problem was posed by Hajnal\,\nErdos and Rogers in the 1960s as a way to generalize classical graph\nRamsey numbers. We prov e almost optimal results using recent\nconstructions in Ramsey theory. We also consider the problem where we\nreplace triangle and K4 by arbitrary g raphs H and G and discover\nseveral interesting new phenomena.\nThis is jo int work with Jacques Verstraete.4:00 pm → Jane\nStreet-sponsored tea an d cookies in Wean 6220 (bring your own mug).\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91abee7 DTSTART;TZID=America/New_York:20250212T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250212T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Rachel Zhang CLASS:PUBLIC DESCRIPTION:Speaker: RACHEL ZHANG\, Ph.D. Student in Computer Science\, Ele ctrical\nEngineering and Computer Science Department\, Massachusetts Insti tute\nof Technology\n\nTalk Title: Explicit Vertex Expanders Beyond the Sp ectral BarrierWe\ngive the first explicit constructions of vertex expander s that pass\nthe spectral barrier. Previously\, the strongest known explic it vertex\nexpanders were those given by d-regular Ramanujan graphs\, whos e\nspectral properties imply that every small set S of vertices has at\nle ast 0.5d|S| distinct neighbors. However\, it is possible to construct\nRam anujan graphs containing a small set S that has no more than\n0.5d|S| dist inct neighbors. In fact\, no explicit construction was\nknown to beat the 0.5 barrier. In this talk\, I will discuss how we\nconstruct vertex expand ers for which every small set expands by a\nfactor of 0.6d. In fact\, our construction satisfies an even stronger\nproperty: small sets actually hav e 0.6d|S| unique neighbors.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ac278 DTSTART;TZID=America/New_York:20250211T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250211T183000 URL:https://www.cmu.edu/cmist/news-events/index.html LOCATION:Grand Room\, Posner Hall 340 SUMMARY:CMIST Scientists & Strategists - Adam Segal CLASS:PUBLIC DESCRIPTION:Speaker: ADAM SEGAL\, Ira A. Lipman Chair in Emerging Technolog ies and\nNational Security\, and\, Director\, Digital and Cyberspace Polic y\nProgram\, Council on Foreign Relations\n\nTalk Title: How Tech Firms Sh ape Geopolitics\n\nThe United States has increasingly with worked with and depended on\nprivately owned companies to spur innovation and advance for eign\npolicy interests. Starting in the 1990s\, with the help of governmen t\nfunding of basic R&D and a hands-off approach to regulation\,\ntechnolo gy companies commercialized the internet\, created the\nsmartphone\, and b uild global social media platforms. The U.S. tech\nsector dominated its co mpetitors in other countries\, supplied the\ntechnology needed for defense modernization\, and promoted American\npower. This relationship is now be ing remade through shifts in\ndomestic politics and geopolitical competiti on\, especially with China.\nWhat does this mean for US national security and foreign policy\ninterests? \n\n— \n\nAdam Segal is the Ira A. Lipm an chair in emerging technologies and\nnational security and director of t he Digital and Cyberspace Policy\nprogram at the Council on Foreign Relati ons (CFR). An expert on\nsecurity issues\, technology development\, and Ch inese domestic and\nforeign policy\, Segal was the project director for th e CFR-sponsored\nIndependent Task Force reports Confronting Reality in Cyb erspace\,\nInnovation and National Security\, Defending an Open\, Global\, Secure\,\nand Resilient Internet\, and Chinese Military Power. His book T he\nHacked World Order: How Nations Fight\, Trade\, Maneuver\, and Manipul ate\nin the Digital Age (PublicAffairs\, 2016) describes the increasingly\ ncontentious geopolitics of cyberspace. His work has appeared in the\nFina ncial Times\, the New York Times\, Foreign Policy\, the Wall Street\nJourn al\, and Foreign Affairs\, among others. From April 2023 to June\n2024\, S egal was a senior advisor in the State Department's Bureau of\nCyberspace and Digital Policy\, where he led the development of the\nUnited States In ternational Cyberspace and Digital Policy. Segal has a\nBA and PhD in gove rnment from Cornell University\, and an MA in\ninternational relations fro m the Fletcher School of Law and Diplomacy\,\nTufts University \n\nREGIST ER\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ac70d DTSTART;TZID=America/New_York:20250210T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250210T173000 URL:https://db.cs.cmu.edu/events/sql-death-larry-ellison-was-right-kinda-ty pescript-stored-procedures-for-the-modern-age/ LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - James Cowling CLASS:PUBLIC DESCRIPTION:Speaker: JAME COWLING\, Chief Technology Officer and Co-founder \, Convex\n\nTalk Title: Larry Ellison was Right (kinda)! TypeScript Store d\nProcedures for the Modern Age\n\nNo one uses SQL to write business logi c. It's written in a programming\nlanguage with libraries\, tests\, type s afety\, and expressive syntax.\nTraditionally this was the domain of a bac kend team\, who’d try to\nbuild enough functionality to keep the fronten d team happy without\nbreaking the database. This model hasn't kept up wit h the needs of\nfull stack developers though\, so they've turned to platfo rms that\nexpose the database directly to code running on client devices. This\nintroduces a host of new problems\, like request waterfalls\, row-le vel\nsecurity\, and difficulty implementing complex transactional logic.\n Like all good database problems it turns out this was already solved\nin t he 80s\, as stored procedures written in PL/SQL. No application\ndeveloper wants to use a niche language that tightly couples business\nlogic with t he database though. But what if we just ran stored\nprocedures written in TypeScript\, that are integrated directly with\napplication developer flow s? It turns out it works\, really well.\n\n—\n\nJames Cowling is CTO and cofounder of Convex\, the open-source reactive\ndatabase for app develope rs. James was Senior Principal Engineer at\nDropbox and tech lead on proje cts including their multi-exabyte\ngeo-distributed storage system\, multi- homing Dropbox infra\, and\nbuilding database systems that handle millions of queries per second.\nJames received his PhD at MIT specializing in lar ge-scale distributed\ntransactions and consensus protocols. Rumor has it h e has a cover band\nwith his cofounders. \n\nThis talk is part of the SQL or Death? Seminar\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91acb21 DTSTART;TZID=America/New_York:20250210T143000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250210T160000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Thesis Proposal - Jeff Xu CLASS:PUBLIC DESCRIPTION:Speaker: JEFF XU\, Ph.D. Student\, Computer Science Department\ , Carnegie\nMellon University\n\nTalk Title: Spectral Techniques for Avera ge-Case Complexity and Beyond\n\nIn recent years\, algorithmic challenges across diverse areas including\nstatistical physics\, machine learning and cryptography have centered\naround statistical inference problems\, i.e.\ , computational problems\nwith average-case inputs. For many of these prob lems\, the best-known\nefficient algorithms are often suboptimal\, giving rise to information\nvs. computation gaps\, discrepancies between what is theoretically\npossible given the amount of information and what can be at tained via\nefficient algorithms. One fundamental question is how we can p rovide\nrigorous evidence of hardness to show that such gaps are\ninsurmou ntable for efficient computation. \n\nWe propose to provide rigorous evid ence via the lens of the\nSum-of-Squares (SoS) algorithms\, a hierarchy of semidefinite\nprogrammings. Unlike several other popular models in the av erage-case\nsetting (eg. low-degree polynomials/statistical-query/ overlap -gap).\nSum-of-Squares algorithms are known to capture various optimal\nal gorithms in both the average and worst-case setting\, and therefore\nprovi de one of the strongest form of hardness. \n\nIn this talk\, I will illus trate that average-case SoS hardness usually\nboils down to the study of s pectral techniques\, specifically the study\nof random matrices with corre lated input\, and how a refined\nunderstanding of such matrices gives rise to the resolution of key\naverage-case complexity questions including spa rse graph\nindependent-set\, densest-k subgraph\, and coloring. Finally\, I will\ntalk about some exciting challenges in the future along this\ndire ction\, and specifically how they boil down again to\nsimple-to-state\, se lf-contained questions about random matrices\, and\nexplore the applicatio ns beyond average-case hardness.\n\nThesis Committee\n\nPravesh K. Kothari (Chair)\n\nAayush Jain\n\nRyan O’Donnell\n\nMadhur Tulsiani (Toyota Tec hnical Institute at Chicago / University of\nChicago)\n\nAdditional Inform ation\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91acfae DTSTART;TZID=America/New_York:20250210T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250210T150000 LOCATION:Mehrabian Collaborative Innovation Center 2214 and Zoom SUMMARY:Robots + Privacy Reading Group - Sarah Scheffler CLASS:PUBLIC DESCRIPTION:Speaker: SARAH SCHEFFLER\, Assistant Professor\, Department of\ nEngineering and Public Policy\, and Software and Societal Systems\nDepart ment\, Carnegie Mellon University\n\nThis is an opportunity to learn more and to better understand what the\ncurrent state of the art\, methods\, an d approaches are to privacy (and\nsecurity) as specifically relevant to ro botics.\nTo kick things off....\nSarah will offer the general practice for how this reading group will\nwork going forward and will also present on: Rueben et al: Themes and\nResearch Directions in Privacy-Sensitive Roboti cs (the results of a\n2018 workshop that tried to identify areas for futur e research in\nprivacy and robotics) Aouedi et al\, 2024: A Survey on Inte lligent\nInternet of Things: Applications\, Security\, Privacy\, and Futur e\nDirections(or at least the privacy parts of this -- so we can see how\n things have progressed).Subscribe to the reading group mailing listIn\nPer son and Zoom Participation.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ad31c DTSTART;TZID=America/New_York:20250206T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250206T163000 URL:https://brain.andrew.cmu.edu/seminar LOCATION:Group Viewing Baker Hall 340A and Zoom SUMMARY:brAIn Seminar - SueYeon Chung CLASS:PUBLIC DESCRIPTION:Speaker: SUEYEON CHUNG\, Assistant Professor\, Center for Neura l\nScience\, New York University\, and\, Project Leader\, Center for\nComp utational Neuroscience\, Flatiron Institute\, Simons Foundation\n\nTalk Ti tle: Towards a Multi-Scale Understanding of Biological and\nArtificial Neu ral Networks\n\nRecent breakthroughs in experimental neuroscience and mach ine learning\nhave opened new frontiers in understanding the computational \nprinciples governing neural circuits and artificial neural networks\n(AN Ns). Both biological and artificial systems exhibit an astonishing\ndegree of orchestrated information processing capabilities across\nmultiple scal es - from the microscopic responses of individual neurons\nto the emergent macroscopic phenomena of cognition and task functions.\nAt the mesoscopic scale\, the structures of neuron population\nactivities manifest themselv es as neural representations. Neural\ncomputation can be viewed as a serie s of transformations of these\nrepresentations through various processing stages of the brain. The\nprimary focus of my lab's research is to develop theories of neural\nrepresentations that describe the principles of neura l coding and\,\nimportantly\, capture the complex structure of real data f rom both\nbiological and artificial systems. \n\nIn this talk\, I will pr esent three related approaches that leverage\ntechniques from statistical physics\, machine learning\, and geometry to\nstudy the multi-scale nature of neural computation. First\, I will\nintroduce new statistical mechanic al theories that connect geometric\nstructures that arise from neural resp onses (i.e.\, neural manifolds)\nto the efficiency of neural representatio ns in implementing a task.\nSecond\, I will employ these theories to analy ze how these\nrepresentations evolve across scales\, shaped by the propert ies of\nsingle neurons and the transformations across distinct brain regio ns.\nFinally\, I will demonstrate how insights from the theories of neural \nrepresentations can elucidate why certain ANN models better predict\nneu ral data\, facilitating model comparison and selection. \n\n— \n\nSueY eon Chung is an Assistant Professor in the Center for Neural\nScience at N YU\, with a joint appointment in the Center for\nComputational Neuroscienc e at the Flatiron Institute\, an internal\nresearch division of the Simons Foundation. She is also an affiliated\nfaculty member at the Center for D ata Science and Cognition &\nPerception Program at NYU. Prior to joining N YU\, she was a\nPostdoctoral Fellow in the Center for Theoretical Neurosci ence at\nColumbia University\, and BCS Fellow in Computation at MIT. Befor e\nthat\, she received a Ph.D. in applied physics at Harvard University\,\ nand a B.A. in mathematics and physics at Cornell University. She\nreceive d the Klingenstein-Simons Fellowship Award in Neuroscience in\n2023\, and the Sloan Research Fellowship in 2024. Her main research\ninterests lie at the intersection between statistical physics\,\nneuroscience and machine learning\, with a particular focus on\nunderstanding and interpreting neur al computation in biological and\nartificial neural networks by employing methods from neural network\ntheory\, statistical physics\, and high-dimen sional statistics. \n\nGroup Viewing and Zoom Participation. See announce ment.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ad81b DTSTART;TZID=America/New_York:20250206T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250206T160000 URL:https://aco.math.cmu.edu/abs-24-25/feb6.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Ilya Shkredov CLASS:PUBLIC DESCRIPTION:Speaker: ILYA SHKREDOV\, Professor of Mathematics \, Department of\nMathematics\, Purdue University\n\nTalk Title: Some applications of t he higher energy method to\ndistribution irregularities\n\nWe review recen t results obtained by the method of higher sumsets and\nhigher energies. I n particular\, we discuss two applications:\nirregularities in the distrib ution of the difference set and\nirregularities in the large Fourier coeff icients of sets with small\nsumsets. \n\n4:00 pm - Jane Street-sponsored tea and cookies in the math lounge\n(bring your own mug if possible)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91add99 DTSTART;TZID=America/New_York:20250205T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250205T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20250205.html LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Shivam Nadimpalli CLASS:PUBLIC DESCRIPTION:Speaker: SHIVAM NADIMPALLI\, Instructor in Applied Mathematics\ ,\nDepartment of Mathematics\, Massachusetts Institute of Technology\n\nTa lk Title: Gaussian Polytope ApproximationWe study the\napproximability of high-dimensional convex sets by intersections of\nhalfspaces\, where the a pproximation quality is measured with respect\nto the standard Gaussian di stribution and the complexity of an\napproximation is the number of halfsp aces used.\n\nWe establish a range of upper and lower bounds both for gene ral convex\nsets and for specific natural convex sets that are of particul ar\ninterest. We rely on techniques from many different areas\, including\ nclassical results from convex geometry\, Cramér-type bounds from\nprobab ility theory\, and—perhaps surprisingly—a range of topics\nfrom comput ational complexity theory\, including computational learning\ntheory\, unc onditional pseudorandomness\, and the study of influences\nand noise sensi tivity in the analysis of Boolean functions.\n\nBased on joint work  with Anindya De and Rocco Servedio.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ae1b1 DTSTART;TZID=America/New_York:20250204T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250204T203000 URL:https://www.givecampus.com/schools/CarnegieMellonUniversity/events/scs- nyc-alumni-mixer-at-haven-rooftop LOCATION:Haven Rooftop\, 132 W 47th Street\, New York\, NY 10036 SUMMARY:SCS Alumni Mixer in New York City - January 2025 CLASS:PUBLIC DESCRIPTION:Talk Title: NYC - SCS Alumni Mixer\n\nJoin us for our SCS Alumn i Mixer in Midtown! The event will include\nhors d’oeuvres and drinks by Haven Rooftop NYC. SCS legend\, Mark\nStehlik\, will be in attendance. Do n’t miss the opportunity to catch\nup with old friends and meet new ones ! \n\nREGISTRATION  \n\n→ Opens on January 7\, 2025\n\n→ Registrati ons close January 30\, 2025 at 11:59 pm\n\nThis event is ONLY for SCS alum ni.  Attendees must be 21+ to attend\nper the venue. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ae4d0 DTSTART;TZID=America/New_York:20250204T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250204T135000 URL:https://privacy.s3d.cmu.edu/masters/seminar/index.html LOCATION:Hamburg Hall 1002 and Zoom SUMMARY:Privacy Seminar - Michael Feffer CLASS:PUBLIC DESCRIPTION:Speaker: MICHAEL FEFFER\, Ph.D. Student\, Ph.D. Program in Soci etal\nComputing\, Software and Societal Systems Department\, Carnegie Mell on\nUniversity\n\nTalk Title: Red-Teaming for Generative AI: Silver Bullet or Security\nTheater?\n\nIn response to rising concerns surrounding the s afety\, security\, and\ntrustworthiness of Generative AI (GenAI) models\, practitioners and\nregulators alike have pointed to AI red-teaming as a ke y component of\ntheir strategies for identifying and mitigating these risk s. However\,\ndespite AI red-teaming’s central role in policy discussion s and\ncorporate messaging\, significant questions remain about what preci sely\nit means\, what role it can play in regulation\, and how it relates to\nconventional red-teaming practices as originally conceived in the\nfie ld of cybersecurity. \n\nIn this work\, we identify recent cases of red-t eaming activities in\nthe AI industry and conduct an extensive survey of r elevant research\nliterature to characterize the scope\, structure\, and c riteria for AI\nred-teaming practices. Our analysis reveals that prior met hods and\npractices of AI red-teaming diverge along several axes\, includi ng the\npurpose of the activity (which is often vague)\, the artifact unde r\nevaluation\, the setting in which the activity is conducted (e.g.\,\nac tors\, resources\, and methods)\, and the resulting decisions it\ninforms (e.g.\, reporting\, disclosure\, and mitigation). \n\nIn light of our fin dings\, we argue that while red-teaming may be a\nvaluable big-tent idea f or characterizing GenAI harm mitigations\, and\nthat industry may effectiv ely apply red-teaming and other strategies\nbehind closed doors to safegua rd AI\, gestures towards red-teaming\n(based on public definitions) as a p anacea for every possible risk\nverge on security theater. To move toward a more robust toolbox of\nevaluations for generative AI\, we synthesize ou r recommendations into\na question bank meant to guide and scaffold future AI red-teaming\npractices. \n\n— \n\nMichael Feffer is a fourth-year Societal Computing PhD student at\nCarnegie Mellon University (CMU). He st udies interactions between AI\nand society\, including algorithmic fairnes s\, participatory ML\, and\ngenerative AI model evaluation. He aims to dev elop frameworks whereby\neveryday people impacted by ML models can influen ce model\ndevelopment. \n\nIn Person and Zoom Participation.  See announ cement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ae984 DTSTART;TZID=America/New_York:20250204T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250204T130000 URL:http://www.cs.cmu.edu/~aiseminar/ LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:AI Seminar - Yuchen Li CLASS:PUBLIC DESCRIPTION:Speaker: YUCHEN LI\, Ph.D. Student\, Machine Learning Departmen t\,\nCarnegie Mellon University\n\nTalk Title: Towards Mathematical Unders tanding of Modern Language\nModels\n\nTo mathematically reason about how n eural networks learn languages\,\nour methodology involves three major com ponents: (1) mathematically\ncharacterizing key structures in language dat a distributions\, (2)\ntheoretically proving how neural networks capture s uch structures\nthrough self-supervision during pre-training\, and (3) con ducting\ncontrolled experiments using synthetic data. In this talk\, I wil l\nsurvey a few applications of this methodology: understanding\nTransform ers training dynamics via the lens of topic models\, and\nproving pitfalls in common Transformer interpretability heuristics via\nthe lens of a form al language (the Dyck grammar). These results\nillustrate some promises an d challenges for this methodology.\nFinally\, I will share some thoughts on key open questions. \n\n— \n\nYuchen Li is a Ph.D. student in the M achine Learning Department at\nCarnegie Mellon University\, advised by Pro fessor Andrej Risteski. \nYuchen's research interest is in improving the mathematical\nunderstanding of language models (training dynamics\, effici ent\nsampling\, mechanistic interpretability). References\n\nYuchen Li\, Y uanzhi Li\, and Andrej Risteski. How Do Transformers Learn\nTopic Structur e: Towards a Mechanistic Understanding. ICML\n2023. https://arxiv.org/abs /2303.04245Kaiyue Wen\, Yuchen Li\, Bingbin\nLiu\, Andrej Risteski. Transf ormers are uninterpretable with myopic\nmethods: a case study with bounded Dyck grammars. NeurIPS 2023.\n\nIn Person and Zoom Participation.  See a nnouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91aee0f DTSTART;TZID=America/New_York:20250203T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250203T160000 URL:https://www.cs.cmu.edu/~pop/seminar/2025-02-03-Albarghouthi/ LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:PoP Seminar - Aws Albarghouthi CLASS:PUBLIC DESCRIPTION:Speaker: AWS ALBARGHOUTHI\, Associate Professor\, Computer Scie nces\,\nSchool of Computer\, Data & Information Sciences\, University of\n Wisconsin-Madison\n\nTalk Title: Synthesizing Quantum-Circuit Compilers\n\ nThe promise of quantum computing has tantalized researchers for\ndecades\ , and recent breakthroughs in physical implementations have\nbrought this technology closer to reality. However\, the quantum\ncomputing landscape r emains highly dynamic: competing physical\nsubstrates\, fault tolerance sc hemes\, and architectures continue to\nemerge with no clear frontrunner. T his diversity creates a significant\nbottleneck in the compilation pipelin e – developing and maintaining\nseparate compilers for each new device o r experimental setup is both\ntime-consuming and error-prone.\n\nIn this t alk\, I will present an alternative approach: automatically\nsynthesizing device-specific quantum circuit compilers. This\nsynthesis-based methodolo gy enables rapid iteration while maintaining\ncorrectness guarantees. I wi ll focus on the optimizer component\, which\nreduces circuit size to minim ize quantum computation errors. I will\ndemonstrate how automatically synt hesized optimizers can achieve\nsuperior performance compared to sophistic ated hand-crafted\nalternatives.\n\nThis talk is based on joint work with Amanda Xu\, Abtin Molavi\, and\nSwamit Tannu.\n\n—\n\nAws Albarghouthi i s an associate professor at the University of\nWisconsin-Madison. He studi es the problems of automated synthesis and\nverification of programs. He r eceived his PhD from the University of\nToronto in 2015. He has received a number of best-paper awards for his\nwork (at FSE\, UIST\, and FAST)\, a CACM Research Highlight for his PLDI\n2020 paper\, an NSF CAREER award\, a Google Faculty Research Award\, and\nmultiple Facebook Research Awards.\n Faculty Host:  Feras Saad\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91af25c DTSTART;TZID=America/New_York:20250203T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250203T130000 URL:https://www.cylab.cmu.edu/events/2025/02/03-seminar-ippolito.html LOCATION:Panther Hollow Conference Room\, Mehrabian Collaborative Innovatio n\nCenter 4105\, and Zoom SUMMARY:CyLab Seminar - Daphne Ippolito CLASS:PUBLIC DESCRIPTION:Speaker: DAPHNE IPPOLITO\, Assistant Professor\, Language Techn ologies\nInstitute\, Carnegie Mellon University\, and Senior Research Scie ntist\,\nGoogle Deepmind\n\nTalk Title: Troubles with Training Data for La rge Language Models\n\nModern large language models (LLMs) derive their ca pabilities from the\ndata used to train their underlying neural networks. While this data\nis the source of LLMs’ strength\, it also creates falli bilities.\nThough the companies releasing LLMs aim to hide their training data\nfrom users\, we demonstrate how it is surprisingly difficult to keep \nmalicious\, or even typical users\, from accessing long strings of text\ nthat LLMs have memorized from the source data. Furthermore\, most\ntraini ng data is derived from large-scale crawls of the Internet. We\ninvestigat e whether by poisoning portions of the Internet\, an\nadversary can insert backdoors or otherwise change the behavior of the\nLLMs trained on this d ata. \n\n— \n\nDaphne Ippolito is an assistant professor at the Langua ge Technologies\nInstitute at Carnegie Mellon University and a senior rese arch\nscientist at Google Deepmind. Among other topics\, she studies priva cy\nand security issues around language generation systems\, strategies fo r\nbetter evaluation of language models\, and customizability of language\ nmodels for different real-world applications. In Person and Zoom\nPartici pation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91af62c DTSTART;TZID=America/New_York:20250131T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250131T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Newell-Simon Hall 4305 and Zoom SUMMARY:AI-SDM - Seminar Series - Ozgur Eris\, Christina Liaghati\, Ben Wel lner CLASS:PUBLIC DESCRIPTION:Speaker: OZGUR ERIS\, CHRISTINA LIAGHATI\, BEN WELLNER\, MITRE\ n\nTalk Title: AI Assurance for Public Good\n\n►Ozgur Eris is the Managi ng Director of the AI and Autonomy\nInnovation Center\, where he leads ove r 300 AI engineers and scientists\ndedicated to advancing AI for public go od. His research focuses on\naugmenting human cognition with AI to tackle sociotechnical\nchallenges\, including course of action planning\, clinica l decision\nsupport\, AI assurance\, IoT-enabled situation awareness\, and \ncollaboration systems. Dr. Eris founded MITRE’s AI Assurance and\nDisc overy Lab\, which aims to identify and mitigate critical\, yet\npoorly und erstood\, assurance risks in AI-enabled systems. Before\njoining MITRE\, h e was a tenured Associate Professor in the Product\nInnovation Management department at Delft University of Technology in\nThe Netherlands. Prior to Delft\, he was a founding faculty member at\nFranklin W. Olin College of Engineering in Massachusetts\, where he\nco-created a novel engineering de sign program. Earlier in his career\,\nhe served as the Associate Director of the Center for Design Research\nat Stanford University. Dr. Eris earne d his B.S. in Mechanical\nEngineering\, cum laude\, from the University of Washington\, and both\nhis M.S. and Ph.D. in Mechanical Engineering with a focus on\nengineering design from Stanford University.\n\n►Christina L iaghati is the Manager of the Trustworthy & Secure AI\nDepartment and MITR E ATLAS Lead. Working across a collaborative global\ncommunity of industry \, government\, and academia\, Dr. Liaghati leads\nMITRE’s Trustworthy & Secure AI Department and MITRE ATLAS\, where\nshe passionately drives res earch and developments in trustworthy and\nsecure AI for everyone working to leverage AI-enabled systems. Leading\nher department of 50+ scientist a nd engineers and serving the\ncommunity with the not-for-profit\, objectiv e\, MITRE perspective\, she\nis dedicated to working together to create an d openly share actionable\ntools\, capabilities\, data\, and frameworks fo r trustworthy and secure\nAI like ATLAS\, an ATT&CK-style framework of the threats and\nvulnerabilities of AI-enabled systems.\n\nAs Dr. Liaghati ha s worked across the community to improve the common\nunderstanding of AI s ecurity concerns\, her work quickly started\noverlapping with broader AI a ssurance concerns\, which includes AI\nequitability\, interpretability\, r eliability\, robustness\, safety\, and\nneeds for privacy enhancement. As a result of this expansion beyond AI\nsecurity into more of these elements of trustworthy AI and AI\nassurance\, her current focus under ATLAS and a cross the international\ncommunity is to build a protected mechanism for i ncreased knowledge\nand incident sharing across government and industry in both AI\nsecurity and the broader areas of AI assurance. Dr. Liaghati als o\nchairs the NATO Science and Technology Organization Research Task\nGrou p on the AI Assurance and Security\, focused on fostering an\nenduring col laborative community of NATO organizations and industry\npartners\, levera ging the Science and Technology Organization to shape\nfuture interoperabl e capability developments in AI security and\nassurance.\n\n► Ben Wellne r is Chief Scientist for Generative AI within MITRE’s\nArtificial Intell igence and Innovation Center\, based on Bedford\, MA.\nHe is the principal investigator for LLM-SIREN\, an internal research\neffort focused on deve loping novel large language model-based\ncapabilities and providing robust methods for assuring them. Dr.\nWellner has led numerous efforts at MITRE addressing research\nchallenges in natural language processing\, healthca re analytics and\naviation safety and has published over 25 peer-reviewed journal and\nconference papers\, in areas of machine learning\, human lang uage\ntechnology as well as clinical- and bio-informatics. These various\n research efforts have also led to the development of multiple\nopen-source software artifacts that have seen widespread use within\nMITRE and beyond in areas of topic modeling and information\nextraction. He also serves as Adjunct Lecturer at Brandeis University\nwhere he teaches graduate level courses in computational linguistics\nand machine learning. Dr. Wellner ho lds a B.S. in Mathematics and\nComputer Sciences from the University of Wi sconsin – Madison\, an\nM.Sc. in Advanced Computing from Imperial Colleg e – London\, and a\nPh.D. in Computer Science from Brandeis University.\ n\nREGISTER  → confirmation email will provide details on joining the\n meeting.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91afc2b DTSTART;TZID=America/New_York:20250130T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250130T163000 URL:https://brain.andrew.cmu.edu/seminar LOCATION:Group Viewing\, Baker Hall 340A and Zoom SUMMARY:brAIn Seminar - Scott Linderman CLASS:PUBLIC DESCRIPTION:Speaker: SCOTT LINDERMAN\, Assistant Professor of Statistics\, and\,\nInstitute Scholar\, Wu Tsai Neurosciences Institute \, Stanford\nUn iversity\n\nTalk Title: State Space Models for Biological and Artificial\n Intelligence\n\nNew recording technologies are revolutionizing neuroscienc e\, allowing\nus to measure the spiking activity of large numbers of neuro ns in\nfreely behaving animals. These technologies offer exciting\nopportu nities to link brain activity to behavioral output\, but they\nalso pose s tatistical challenges. Neural and behavioral data are\nnoisy\, high-dimens ional time series with nonlinear dynamics and\nsubstantial variability acr oss subjects. \n\nI will present our work on state space models (SSMs) to tackle these\nchallenges. The key idea is that high-dimensional measureme nts often\nreflect the evolution of underlying latent states\, whose dynam ics may\nshed light on neural computation. For example\, we have used SSMs to\nstudy how attractor dynamics in the hypothalamus encode persistent\ni nternal states during social interaction\, and to connect stereotyped\nmov ements to moment-to-moment fluctuations in brain activity. \n\nThere has been a resurgence of interest in SSMs within the machine\nlearning communi ty as well\, and SSMs now form the backbone of several\nstate-of-the-art m odels for sequential data. I will present recent\nwork from my lab that fo cuses on novel models and efficient algorithms\nfor sequential data\, with applications in neuroscience and beyond.\nTogether\, these projects highl ight the central role of state space\nmodels in our studies of both biolog ical and artificial\nintelligence. \n\n— \n\nScott Linderman PhD is an Assistant Professor at Stanford University\nin the Statistics Department and the Wu Tsai Neurosciences\nInstitute.  His research focuses on machin e learning\, computational\nneuroscience\, and the general question of how computational and\nstatistical methods can help to decipher neural comput ation. His work\ncombines novel methodological development in the areas of state space\nmodels\, deep generative models\, point processes\, and appr oximate\nBayesian inference with applied statistical analyses of large-sca le\nneural and behavioral data.  \n\nPreviously\, he was a postdoctoral fellow with David Blei and Liam\nPaninski at Columbia University and a gra duate student at Harvard\nUniversity with Ryan Adams. His work has been re cognized with a Savage\nAward from the International Society for Bayesian Analysis\, an AISTATS\nBest Paper Award\, and Fellowships from the McKnigh t\, Sloan\, and Simons\nFoundations. \n\nIn Person/Group Viewing and Zoom Participation.  See announcement.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b017c DTSTART;TZID=America/New_York:20250130T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250130T160000 URL:https://aco.math.cmu.edu/abs-24-25/jan30.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Caroline Klivans CLASS:PUBLIC DESCRIPTION:Speaker: CAROLINE KLIVANS\, Professor of Applied Mathematics\, Division\nof Applied Mathematics\, Brown University\n\nTalk Title: The Arb oricity Polynomial\n\nI will introduce a new matroid (graph) invariant: Th e Arboricity\nPolynomial. Arboricity is a numerical invariant first introd uced by\nNash-Williams\, Tutte and Edmonds. It captures the minimum number of\nindependent sets (forests) needed to decompose the ground set of a\nm atroid (edges of a graph). The arboricity polynomial enumerates the\nnumbe r of such decompositions. We examine this counting function in\nterms of s cheduling\, Ehrhart theory\, quasisymmetric functions\, matroid\npolytopes and the permutohedral fan. \n\n4:00 pm  → Jane Street sponsored Tea a nd Cookies\, Wean 6220\n                → bring your own mug if you have one\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b04ae DTSTART;TZID=America/New_York:20250130T143000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250130T160000 LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:SCS Teaching Track Faculty Candidate - Elijah Rivera CLASS:PUBLIC DESCRIPTION:Speaker: ELIJAH RIVERA\, Ph.D. Candidate\, Department of Comput er\nScience\, Brown University\n\nTalk Title: Freedom from the Numbers\, b y the Numbers\, for the Numbers:\nAn Introduction to Hashmaps\n\nAdapted l ecture from the course “Program Design with Data Structures\nand Algorit hms\" at Brown University. This talk assumes familiarity\nwith list-like a nd array-like data structures. Arrays offer efficient\nstorage and lookup operations for ordered or numerically labeled data\,\nbut real-world data often doesn't lend itself to numeric organization.\nDoes this mean we miss out on the benefits of arrays? We don't have\nto! \n\nIn this lecture\, we’ll design a powerful new data structure—the\nhashmap—that allows us to use non-numeric keys with the same ease as\nnumeric indices. We’ll dive into the design challenges and\ntrade-offs involved in transforming keys into array indices\, and by\nthe end we’ll achieve a data structure that maintains constant time\naccess to data\, no matter what kinds of la bels our information\ncarries. \n\n— \n\nElijah Rivera is a PhD candid ate in CS at Brown University\, where he\nis serving as Teaching Fellow fo r \"Program Design with Data Structures\nand Algorithms.\" He is advised b y Prof. Shriram Krishnamurthi and\nProf. Kathi Fisler. His expertise and h is heart are both in computing\neducation\, but his mind is also easily di stracted by interesting\nproblems from many other fields of CS. Prior to h is time at Brown\, he\nreceived both his S.B. and M.Eng. from MIT\, where he worked on several\nresearch projects in the fields of programming langu ages\, program\nsynthesis\, and formal verification. S3D Joint with CSD\n\ nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b089a DTSTART;TZID=America/New_York:20250129T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250129T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Gates Hillman 8102 SUMMARY:Joint Theory Lunch Seminar / Speaking Skills Talk - Jingxun Liang CLASS:PUBLIC DESCRIPTION:Speaker: JINGXUN LIANG\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Tight Cell-Probe Lower Bounds for Dynamic Succinct\nDictionaries\n\nThe dictionary problem involv es maintaining a set  S ⊂ [U] of n\nkeys under insertions\, deletions\, and membership queries focusing on\nefficiency in both time and space. Th is fundamental problem has been\nstudied extensively for six decades since 1953. Recently\, a series of\nworks has established the optimal time-spac e trade-off for\ndictionaries. In this talk\, I will describe the tight lo wer bound\nresult that any dictionary with an operational time t just incu r a\nredundant Ω(log(t) n) bits per key in space. \n\nPresented as part of the Theory Lunch Seminar \n\nPresented in Partial Fulfillment of the C SD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b0bbd DTSTART;TZID=America/New_York:20250128T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250128T180000 LOCATION:TBA SUMMARY:SCS Distinguished Industry Lecture - Mark Russinovich CLASS:PUBLIC DESCRIPTION:Speaker: MARK RUSSINOVICH\, CTO\, Deputy CISO\, and Technical F ellow\,\nMicrosoft Azure\n\nMark Russinovich is CTO\, Deputy CISO\, and Te chnical Fellow for\nMicrosoft Azure\, Microsoft’s global enterprise-grad e cloud platform.\nA widely recognized expert in distributed systems\, ope rating systems\nand cybersecurity\, Mark earned a Ph.D. in computer engine ering from\nCarnegie Mellon University. He later co-founded Winternals Sof tware\,\njoining Microsoft in 2006 when the company was acquired. Mark is a\npopular speaker at industry conferences such as Microsoft Ignite\,\nMic rosoft Build\, and RSA Conference. He has authored several\nnonfiction and fiction books\, including the Microsoft Press Windows\nInternals book ser ies\, Troubleshooting with the Sysinternals Tools\, as\nwell as fictional cyber security thrillers Zero Day\, Trojan Horse and\nRogue Code.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b0eda DTSTART;TZID=America/New_York:20250127T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250127T130000 URL:https://www.cylab.cmu.edu/events/2025/01/27-seminar-ram-shankar-siva-ku mar.html LOCATION:Danforth Conference Room\, 2nd Floor\, Cohon University Center and Zoom SUMMARY:CyLab Seminar - Ram Shankar Siva Kumar CLASS:PUBLIC DESCRIPTION:Speaker: RAM SHANKAR SIVA KUMAR\, Data Cowboy\, Microsoft\n\nTa lk Title: Ignore previous instructions - Attacks on AI systems and\nwhat t o do about it\n\nThis talk will provide an overview of the art and science of attacking\nAI systems and its societal implications. We will walk thro ugh the\nevolution of adversarial examples to the now famous\, jailbreaks\ , to\nask why it is so darn difficult to secure AI systems from adversarie s.\nWe will look with a critical eye at the explosion of AI Safety\nInstit utes\, especially US AI Safety Institute\, and their efforts to\nassuage t his problem. \n\n— \n\nRam Shankar Siva Kumar is a Data Cowboy working on the intersection of\nmachine learning and security. At Microsoft\, he founded the AI Red\nTeam\, bringing together an interdisciplinary group of researchers and\nengineers to proactively attack AI systems and find fail ures. His\nrecent book on attacking AI systems\, Not With a Bug\, has been called\n“Essential Reading” by Microsoft’s Chief Technology Officer and\nreceived wide praise from industry leaders at DeepMind\, OpenAI as w ell\nas policy makers and academia. He is donating his proceeds of the boo k\nroyalty to Black In AI. His work on AI and Security has appeared in\nin dustry conferences like RSA\, BlackHat\, Defcon\, BlueHat\, DerbyCon\,\nMI RCon\, Infiltrate\, academic workshops at NeurIPS\, ICLR\, ICML\, IEEE\nS& P\, ACM - CCS. \n\nHis work has been covered by Bloomberg\, VentureBeat\, Wired\, and\nGeekwire. He founded the Adversarial ML Threat Matrix\, an A TT&CK style\nframework enumerating threats to machine learning. His work o n\nadversarial machine learning appeared notably in the National Security\ nCommission on Artificial Intelligence (NSCAI) Final report presented\nto the United States Congress and the President. He is currently Tech\nPolicy Fellow at UC Berkeley and an affiliate at the Berkman Klein\nCenter for I nternet and Society at Harvard University\, where he is\nbroadly investiga ting two questions: How do we assess the safety of ML\nsystems? What are t he policy and legal ramifications of AI\, in the\ncontext of security? He is also Technical Advisory Board Member at the\nUniversity of Washington.  \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b136a DTSTART;TZID=America/New_York:20250126T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250126T130000 URL:https://privacy.cs.cmu.edu/cmu-privacy-day/index.html LOCATION:Carnegie Library of Pittsburgh - Main (Oakland) SUMMARY:Privacy Day 2025 CLASS:PUBLIC DESCRIPTION:Speaker: An event celebrating International Data Privacy Day\n\ nJoin in on for our Privacy Day 2025 event at the Carnegie Library of\nPit tsburgh (CLP).  \n\nFor the first time\, CMU is partnering with CLP to c elebrate\nInternational Data Privacy Day by bringing practical advice on\n protecting privacy online to the general public\, off the university's\nca mpus. \n\nThe event is open to the public\, and no registration is requir ed. Data\nPrivacy Day is an international effort to empower and educate pe ople\nto protect their privacy and control their digital footprint. \n\nM ore information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b16a0 DTSTART;TZID=America/New_York:20250124T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250124T133000 URL:https://safety21.cmu.edu/events/smart-safety-connection-seminar-series/ LOCATION:Scott Hall 6142 and Remote SUMMARY:Smart Safety Connection Seminar - Nat Beuse CLASS:PUBLIC DESCRIPTION:Speaker: NAT BEUSE\, Chief Safety Officer\, Aurora\n\nTalk Titl e: Safety Management System (SMS)Supporting Safe Autonomous\nFreight Opera tions with SMS:  Aurora is a leader in the development\nof automated vehi cle technology and services for freight movement in\nthe country.  We ha ve been developing our product for a number of\nyears and are on the cusp of deploying our first driverless product. \nAurora has been utilizing a Safety Management System to support the\ndevelopment of our technology an d our operations.  This talk will\nfocus on how we started\, SMS in actio n\, and highlight various industry\nactivities.REGISTER\n→Due to space c onstraints\, there will be a 30 person limit for\nin-person attendance wit h a virtual option\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b19c1 DTSTART;TZID=America/New_York:20250123T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250123T130000 URL:https://pdl.cmu.edu/SDI/2025/012325.html LOCATION:Panther Hollow Conference Room 4105\, Mehrabian Collaborative\nInn ovation Center SUMMARY:SDI Seminar - Jeff Mogul CLASS:PUBLIC DESCRIPTION:Speaker: JEFF MOGUL\, Principal Software Engineer\, Network\nIn frastructure\, Google\n\nTalk Title: Thinking About Availability in Large Service\nInfrastructures\, and Especially in Cloud Networks\n\nThis talk c overs several papers from HotOS 2017 and 2019\, plus some\nadditional mate rial. We increasingly depend on the availability of\nonline services\, who se availability depends in complex ways on the\navailability of a complex underlying set of invisible infrastructure\nservices. Most software engine ers lack useful frameworks to create and\nevaluate designs for individual services that support end-to-end\navailability in these infrastructures\, especially given cost\,\nperformance\, and other constraints on viable com mercial services. Even\ngiven the extensive research literature on techniq ues for replicated\nstate machines and other fault-tolerance mechanisms\, we found little\nhelp in this literature for addressing infrastructure-wid e\navailability. \n\nThe first part of the talk argues that\, in many but not all ways\, one\ncan think about availability with the mindset that we have learned to\nuse for security\, and discusses some general techniques that appear\nuseful for implementing and operating high-availability\ninf rastructures. The second part does a deeper dive into the specific\nproble m of defining SLOs for cloud networks. The third part revisits\nthe questi on of SLO definition\, and suggests that the problem needs to\nbe reframed for cloud-computing providers\, to incorporate some\n\"statistical thinki ng\" and expectations about customer behavior\, not\njust about provider b ehavior. \n\n— \n\nJeff Mogul works on fast\, cheap\, reliable\, and f lexible infrastructure\nfor Google. Until 2013\, he was Fellow at HP Labs\ , doing research\nprimarily on computer networks and operating systems iss ues for\nenterprise and cloud computer systems\; previously\, he worked at the\nDEC/Compaq Western Research Lab. He received his PhD from Stanford i n\n1986\, an MS from Stanford in 1980\, and an SB from MIT in 1979. He is\ nan ACM Fellow. Jeff is the author or co-author of several Internet\nStand ards\; he contributed extensively to the HTTP/1.1 specification.\nHe was a n associate editor of Internetworking: Research and\nExperience\, and has been the chair or co-chair of a variety of\nconferences and workshops\, in cluding SIGCOMM\, OSDI\, NSDI\, USENIX\,\nHotOS\, and ANCS. \n\nFaculty H ost:  Justin Sherry\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b1e7d DTSTART;TZID=America/New_York:20250123T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250123T160000 URL:https://www.cmu.edu/ideas-social-cybersecurity/events/ai_disinfo_summit .html SUMMARY:AI & Disinformation Summit CLASS:PUBLIC DESCRIPTION:Speaker: Presented by the CMU Center for Informed Democracy & S ocial -\ncybersecurity\n\nThe CMU Center for Informed Democracy & Social - cybersecurity\n(IDeaS) is hosting a summit on AI & Disinformation\, wi th invited\nfaculty and experts throughout the US. The summit is being ru n\nby Dr. Kathleen M. Carley\, professor in the computer science\ndepartm ent and Director for the Center for Computational Analysis of\nSocial and Organizational Systems.\n\nThe goal of the summit is time to take stock o f the state of the\nfield\, identify gaps\, and share insights to foster t he next decade of\nresearch and practice in this area. The events includes series of\npanels and guided general discussions. The outcome of this sum mit will\nbe a brief on the state of the art in the area\, and a discussio n of\nidentified research gaps.\n\nREGISTER for virtual participation.  N o registration is needed for\nin-person participants. \n\n     ⇒  Panels will be livestreamed during the event but Q&A\nwill not be availabl e for those attending virtually. \n\nSchedule\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b21df DTSTART;TZID=America/New_York:20250122T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250122T130000 LOCATION:Gates Hillman 8102 SUMMARY:Joint Theory Lunch Seminar / Speaking Skills Talk - Noah Singer CLASS:PUBLIC DESCRIPTION:Speaker: NOAH SINGER\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Cosystolic expansion of C hevalley coset complex HDXs\n\nCosystolic expansion is a generalization of vertex expansion in graphs\nto \"higher dimensions\" (i.e.\, to simplicia l complexes). Sparse\,\nefficiently-constructible cosystolic expanders hav e powered recent\nbreakthroughs in quantum coding and local testability\, but these\nconstructions are hard to come by. In this talk\, I will presen t recent\njoint work with Ryan O'Donnell which analyzes the cosystolic exp ansion\nof a certain algebraically-defined complex\, called a \"B-type Che valley\ncoset complex\"\, constructed by O’Donnell & Pratt. Our analysis \nbuilds on a recent\, simpler analysis of a related \"A-type complex\" du e\nto Kaufman & Oppenheim. These complexes have the advantage that\,\nunli ke earlier known complexes due to Lubotzky\, Samuels\, & Vishne\,\nanalyzi ng their expansion does NOT require any deep mathematical\ntools. \n\nI w ill define all relevant notions\, so no prior knowledge of expansion\nor a lgebra is required.\n\nPresented as part of the Theory Lunch Seminar\n\nPr esented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b257a DTSTART;TZID=America/New_York:20250121T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250121T150000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Proposal - Joshua Nathaniel Williams CLASS:PUBLIC DESCRIPTION:Speaker: JOSHUA NATHANIEL WILLIAMS\, Ph.D. Student\, Computer S cience\nDepartment\, Carnegie Mellon University\n\nTalk Title: Understandi ng Generative Image Modeling Through Discrete\nCounterfactual Prompt Optim ization\n\nIt is well understood that generative models can exhibit\nrepre sentational biases across various social groups. Without explicit\nconditi oning\, the vast majority of airplane pilots will be male and\nthe vast ma jority of bank tellers will be female. While much of the\nexisting work fo cuses on hypothesis-driven investigations into these\nbiases\, in this pro posed thesis\, we focus on discovery of new and\nunexpected patterns in ho w these models represent people. We show\nthrough the lens of counterfactu al explainability – processes that\nfind minimal input changes that alte r a model's output – a framework\nfor hypothesis generation in order to reveal surprising patterns in\nhow text-to-image models align textual inpu t with socially meaningful\ngroup memberships.\n\nWe establish a formal\, mathematical distinction between counterfactual\nexplanations and adversar ial examples\, which enables the development\nof novel distance metrics an d optimization strategies for identifying\nhuman-readable counterfactual p rompts. Our approach employs discrete\noptimization techniques\, including an adapter for computing gradients\nacross embeddings from diverse models . Furthermore\, we evaluate\ndiscrete prompt optimization methods and demo nstrate that additional\nregularization is crucial for generating coherent and human-like\nprompts. This work lays the foundation for more nuanced u nderstanding\nof representational biases in generative models and offers t ools for\ntheir systematic exploration.\n\nThesis Committee\n\nZico Kolter (Chair)\n\nAditi Raghunathan\n\nHoda Heidari\n\nSarah Laszlo (Visa)\n\n  \n\nAdditional Information\n\nIn Person and Zoom Participation.  See anno uncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b29af DTSTART;TZID=America/New_York:20250508T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250508T170000 URL:https://csd.cmu.edu/academics/doctoral-resources/dsr-schedule LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Student Review (DSR) S25 - PL/Security/Systems CLASS:PUBLIC DESCRIPTION:Speaker: CSD Doctoral Student Review (DSR)Talk Title: S25 DSR - PL\,\nSecurity\, & Systems - years 1-3\n\nProgramming Languages\, Securit y\, & Systems doctoral student review for\nyears 1-3.\n\nSee email announc ements for additional details.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b2c80 DTSTART;TZID=America/New_York:20250508T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250508T120000 URL:https://csd.cmu.edu/academics/doctoral-resources/dsr-schedule LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Student Review (DSR) S25 - AI/Graphics/Theory CLASS:PUBLIC DESCRIPTION:Speaker: CSD Doctoral Student Review (DSR)Talk Title: S25 DSR - AI\,\nGraphics\, & Theory - years 1-3\n\nAI\, Graphics\, Theory student d octoral student review for years 1-3.\n\nSee email announcements for addit ional information.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b2f30 DTSTART;TZID=America/New_York:20250121T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250121T130000 URL:https://db.cs.cmu.edu/events/splitsql-practical-pushdown-cache-for-data lake-analytics-xiangpeng-hao LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Database Seminar - Xiangpeng Hao CLASS:PUBLIC DESCRIPTION:Speaker: XIANGPENG HAO\, Ph.D. Student\, Computer Sciences Depa rtment\,\nUniversity of Wisconsin-Madison\n\nTalk Title: SplitSQL: Practic al Pushdown Cache for DataLake Analytics\n\nModern data analytics embrace a disaggregated architecture which\ndecouples storage\, cache\, and comput e into network-connected\nindependent components. With disaggregated cache \, a key design\ndecision is whether to push down query predicates to the cache server.\nWithout predicate pushdown\, the cache must send all data t o compute\nnodes\, creating network bottlenecks. \n\nWith predicate pushd own\, the cache server evaluates predicates on\ncached data\, but its limi ted computational resources become the\nbottleneck. In this talk\, we intr oduce SplitSQL\, a pushdown cache\nsystem with efficient predicate evaluat ion. Our system is built upon a\nsurprising observation: pushdown cost is dominated by decoding data\,\nnot predicate evaluations. SplitSQL reduces decoding overhead by\ntranscoding storage formats (like Parquet) into a ca che-optimized\nformat that enables predicate evaluation on encoded data an d supports\nefficient\, fine-grained decoding. \n\nImplemented on Apache DataFusion\, SplitSQL achieves both low network\ntraffic and significantly reduced computational overhead compared to\nconventional pushdown systems . Experiments on ClickBench show that\nSplitSQL's cache-specific format de livers up to 3x end-to-end\nperformance improvement while maintaining comp ression ratio on par\nwith the original storage format. \n\n— \n\nXian gpeng Hao is a PhD student at the University of Wisconsin-Madison\nstudyin g computer science with a focus on database/storage systems.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b3382 DTSTART;TZID=America/New_York:20250509T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250509T170000 URL:https://csd.cmu.edu/academics/doctoral-resources/dsr-schedule LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Student Review (DSR) - S25 General Meeting CLASS:PUBLIC DESCRIPTION:Speaker: CSD Doctoral Student Review (DSR)Talk Title: Doctoral Student\nReview General Meeting - years 4-n\n\nGeneral Meeting years 4-n.  \n\nSee email announcements.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b362d DTSTART;TZID=America/New_York:20250120T000000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250120T235900 URL:https://www.cmu.edu/student-diversity/programs/MLK/index.html?utm_campa ign=2025-01-09+Piper&utm_source=events&utm_medium=all&utm_content=campus_0 _0&utm_term=0_0&utm_id=Dr.+Martin+Luther+King\,+Jr.+Community+Celebration SUMMARY:Martin Luther King\, Jr. Day Observance 2025 CLASS:PUBLIC DESCRIPTION:Speaker: University ClosedCMU will observe Martin Luther King\, Jr. Day\non Monday\, January 20\, but a variety of programs Celebrating t he Life\n& Legacy of Dr. Martin Luther King\, Jr. will be offered from Jan uary\n20 through February 5.\nThe celebration will include the annual  Dr . Martin Luther Kinger\nKeynote Lecture featuring Admiral Michelle J. Howa rd on Wednesday\,\nFebruary 5 at 5:00 pm.\nAll events are free and open to the public unless otherwise specified\nwithin the program description or registration link.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b38f2 DTSTART;TZID=America/New_York:20250117T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250117T180000 LOCATION:Collaborative Commons\, 6th Floor Gates Hillman SUMMARY:Meetups and Hot Chocolate Social: SCS First Year Undergraduates CLASS:PUBLIC DESCRIPTION:Speaker: Hosted by the SCS Undergrad Advisory Council\n\nHey SC S class of 2028! \n\nJoin us  for our MAHCS event (Meetups And Hot Choco late Social)\nbrought to you by the SCS Undergrad Advisory Council!  We'l l be\nhaving free hot chocolate and sweet treats for everyone to enjoy so\ nthat we can all meet up before classes get busy and welcome the new\nseme ster together. We hope to see you there!  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b3c10 DTSTART;TZID=America/New_York:20250117T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250117T133000 LOCATION:Gordon Bell Conference Room\, Gates Hillman 5117 SUMMARY:Doctoral Thesis Proposal - Long Pham CLASS:PUBLIC DESCRIPTION:Speaker: LONG PHAM\, Ph.D. Student\, Computer Science Departmen t\,\nCarnegie Mellon University\n\nTalk Title: Hybrid Resource-Bound Analy ses of Programs\n\nResource-bound analysis aims to infer symbolic bounds o f worst-case\nresource usage (e.g.\, running time\, memory\, and energy) o f programs as\nfunctions of program inputs. Resource analysis has numerous \napplications\, including job scheduling in cloud computing and\npreventi on of side-channel attacks. Various resource analysis\ntechniques have bee n developed\, and they have unique strengths and\nweaknesses that compleme nt each other. (Automatic) static resource\nanalysis\, which analyzes the source code of programs\, is sound: if it\nsuccessfully infers a cost boun d\, it is guaranteed to be a valid\nbound. However\, every static analysis technique is incomplete: there\nexists a program that the analysis techni que cannot handle. Meanwhile\,\ndata-driven analysis\, which statistically analyzes cost measurements\nobtained by running programs on many inputs\, can infer a candidate\ncost bound for any program. However\, it does not guarantee soundness\nof inference results. \n\nTo overcome limitations of individual analysis techniques\, I propose\nhybrid resource analysis\, wh ich integrates two complementary analysis\ntechniques to retain their stre ngths while mitigating their respective\nweaknesses. The user first specif ies which analysis techniques are\nused to analyze which code fragments an d quantities. Hybrid analysis\nthen performs its constituent analysis tech niques on their respective\ncode fragments and quantities. Finally\, their inference results are\ncombined into an overall cost bound. \n\nThe deve lopment of hybrid resource analysis has been driven by the\ndesire to go b eyond Automatic Amortized Resource Analysis (AARA)\, a\nstate-of-the-art t ype-based static resource analysis technique. I\nstart by proving polynomi al-time completeness of AARA. I next\nintroduce Bayesian data-driven analy sis\, which conducts Bayesian\ninference on cost measurements to infer a p osterior distribution of\nsymbolic cost bounds. I then present the first h ybrid resource\nanalysis\, Hybrid AARA\, followed by a discussion of its l imitations. To\novercome these limitations\, I introduce the second hybrid resource\nanalysis\, resource decomposition. I additionally describe Swif tlet\,\nwhich instantiates the resource-decomposition framework with AARA and\nBayesian resource analysis. \n\nFinally\, for proposed work\, my col laborators and I plan to develop\ndata-driven-analysis for statistically i nferring not only a worst-case\nsymbolic cost bound but also a worst-case input generator\, which is a\nprogram generating worst-case program inputs of various sizes. \n\nThesis Committee\n\nJan Hoffmann (Chair)\n\nFeras Saad\n\nMatt Fredrikson\n\nFrancois Pottier (Inria Paris Centre)\n\nAdditi onal Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b4117 DTSTART;TZID=America/New_York:20250117T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250117T113000 LOCATION:Gates Hillman 81092 SUMMARY:Theory Seminar - Allen Liu CLASS:PUBLIC DESCRIPTION:Speaker: ALLEN LIU\, Ph.D. Student\, Electrical Engineering and Computer\nScience Department\, Massachusetts Institute of Technology\n\nT alk Title: The Sudden Death of Entanglement in Quantum Gibbs States\n\nEnt anglement is the quintessential property that distinguishes quantum\nsyste ms from classical ones. Yet\, entanglement remains a mysterious\nphenomeno n\, especially in large many-body quantum systems.  In this\ntalk\, we ai m to understand at what temperatures can quantum systems\nexhibit entangle ment. We show that Gibbs states of local Hamiltonians\nare unentangled abo ve a critical constant temperature\, depending only\non the geometry of th e Hamiltonian\, independent of the system size.\nThis implies a surprising \, universal law\, that there is a sudden death\nof quantum entanglement\, and upends conventional wisdom about the\npresence of short-range quantum correlations in Gibbs states.  Our\nresult stems from an algorithmic fra mework that also gives us an\nefficient\, almost classical algorithm for p reparing the Gibbs state\nabove a critical temperature. \n\n— \n\nAlle n Liu is currently a graduate student in EECS at MIT where he is\nin his f ifth year\, working under the wonderful supervision of Ankur\nMoitra. He a lso completed his undergraduate degree (in mathematics) at\nMIT. He is gen erally interested in algorithms and learning theory\,\nfocusing on develop ing algorithms for machine learning with provable\nguarantees. He is also interested in applying tools from learning\ntheory to understand problems in quantum information. HIs work is\nsupported by an NSF Graduate Research Fellowship\, a Hertz Fellowship\,\nand a Citadel GQS Fellowship.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b45db DTSTART;TZID=America/New_York:20250114T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250114T163000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Thesis Proposal - Hojin Park CLASS:PUBLIC DESCRIPTION:Speaker: HOJIN PARK\, Ph.D. Student\, Computer Science Departme nt\,\nCarnegie Mellon University\n\nTalk Title: Cost-Efficient Storage and Caching in Public Clouds\n\nPublic clouds are widely adopted for their sc alability\, flexibility\,\nand reduced operational overhead\, but optimizi ng costs while\nmaintaining performance remains a significant challenge. T his thesis\naddresses the problems of reducing storage cluster costs in pu blic\nclouds and minimizing cross-cloud/region data access costs by\nexplo iting the elasticity and diversity of public cloud resources\ncombined wit h real-time workload monitoring. \n\nFirst\, this work addresses the chal lenge of reducing storage cluster\ncosts in public clouds through Mimir\, an automated system that\nleverages heterogeneous storage types to determi ne storage cluster\nconfigurations that reduce costs while meeting workloa d performance\nrequirements. Next\, this thesis introduces Macaron\, an\na uto-configuring caching system designed for cross-cloud/region data\nacces s. By dynamically adjusting cache capacity based on workload\ncharacterist ics\, Macaron minimizes remote data access costs and\nensures low latency. Utilizing the object storage for caching storage\ntype\, Macaron demonstr ates substantial cost savings compared to\nexisting cross-cloud/region dat a access approaches. \n\nBuilding on Macaron\, I propose a novel prefetch ing design for\ncross-region data access in public clouds that includes an object\ngrouping-based prefetching algorithm and dynamic cache space\nall ocation for demand-requested and prefetched data. This design aims\nto mit igate cache pollution\, reduce wasted prefetches\, and improve\ndata acces s latency while minimizing additional costs. \n\nBy addressing cost and p erformance optimization in public cloud\nstorage and cross-cloud/region da ta access systems\, this thesis\nenables reducing costs of utilizing publi c cloud resources for storage\nand caching. \n\nThesis Committee\n\nGeorg e Amvrosiadis (Co-Chair)\n\nGregory R. Ganger (Co-Chair)\n\nJignesh M. Pat el\n\nCarlo Curino (Microsoft Research)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b4a33 DTSTART;TZID=America/New_York:20250115T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250115T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20250115.html LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar - Mehrdad Ghadiri CLASS:PUBLIC DESCRIPTION:Speaker: MEHRDAD GHADIRI\, Postdoctoral Associate\, Laboratory for\nInformation & Decision Systems\, Institute for Data\, Systems\, and\n Society and Operations Research Center\, Massachusetts Institute of\nTechn ology\n\nTalk Title: Non-Asymptotic Guarantees for Treatment Effect Estima tion\nUsing Regression AdjustmentThe design and analysis of randomized\nex periments are fundamental to numerous fields\, ranging from the\nphysical and social sciences to industrial applications. Despite the\nlong history of randomized controlled trials in science\, estimating\nthe treatment eff ect on a finite population remains poorly understood\nwhen auxiliary varia bles are used to reduce error\, and the behavior of\nmany estimators has o nly been analyzed asymptotically. Recently\,\nHarshaw\, Sävje\, Spielman\ , and Zhang demonstrated that non-asymptotic\nvariance bounds can be estab lished for the average treatment effect\nestimation problem over a finite population by designing experiments\nusing discrepancy minimization techni ques.\nIn this talk\, we show that a simple Bernoulli design can achieve\n comparable bounds through regression adjustment techniques based on\nspect ral sparsification. Furthermore\, we discuss the advantages of\nusing a si mple design\, potential extensions of our approach\, and\nsupporting empir ical results.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b4da3 DTSTART;TZID=America/New_York:20250114T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250114T135000 URL:https://privacy.s3d.cmu.edu/masters/seminar/index.html LOCATION:Hamburg Hall 1002 and Zoom SUMMARY:Current Topics in Privacy - Arvind Narayanan CLASS:PUBLIC DESCRIPTION:Speaker: ARVIND NARAYANAN \, Professor of Computer Science\, Di rector\,\nCenter for Information Technology Policy\, Princeton University\ n\nTalk Title: How Predictive Decision Making Goes Wrong\n\nMany automated decision-making systems in criminal justice\,\nhealthcare\, finance\, and other domains follow the same logic: machine\nlearning is used to make pr edictions about people's future — who\nwill commit a crime\, pay back a loan\, fall sick\, etc. — and those\npredictions are used to make decisi ons about them. In this talk\, I\nwill explain why this type of automated decision making has inherent\,\nrecurring flaws\, resulting in widespread harm to people. The talk is\nbased on the paper Against Predictive Optimiz ation and the book AI\nSnake Oil. \n\n— \n\nArvind Narayanan is a prof essor of computer science at Princeton\nUniversity and the director of the Center for Information Technology\nPolicy. He is a co-author of the book AI Snake Oil and a newsletter of\nthe same name which is read by 50\,000 r esearchers\, policy makers\,\njournalists\, and AI enthusiasts. He previou sly co-authored two widely\nused computer science textbooks: Bitcoin and C ryptocurrency\nTechnologies and Fairness in Machine Learning. Narayanan le d the\nPrinceton Web Transparency and Accountability Project to uncover ho w\ncompanies collect and use our personal information. His work was among\ nthe first to show how machine learning reflects cultural stereotypes\,\na nd his doctoral research showed the fundamental limits of\nde-identificati on. Narayanan was one of TIME's inaugural list of 100\nmost influential pe ople in AI. He is a recipient of the Presidential\nEarly Career Award for Scientists and Engineers (PECASE). \n\nFaculty Hosts:  Hana Habib\, Norm an Sadeh\n\nCo-Sponsored by the Master's in Privacy Engineering Program an d the\nBlock Center for Technology and Society \n\nIn Person and Zoom Par ticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b51ff DTSTART;TZID=America/New_York:20250106T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250106T140000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 (and adjacent spaces) SUMMARY:SCS Teaching & Education Meeting CLASS:PUBLIC DESCRIPTION:Talk Title: SCS Teaching & Education Meeting\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b5492 DTSTART;TZID=America/New_York:20241224T000000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250102T000000 SUMMARY:Winter Break - F24 CLASS:PUBLIC DESCRIPTION:It will be quiet on campus: no classes or normal business hours for a\nshort while.\nWe will resume normal business hours January 2\, 202 5.\nClasses will resume Monday\, January 13\, 2025.\nRefresh. Reset. Reboo t for 2025!\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b5704 DTSTART;TZID=America/New_York:20241223T070000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241223T235900 SUMMARY:Community Appreciation Day (F24) - University Closed CLASS:PUBLIC DESCRIPTION:Talk Title: Community Appreciation Day - University Closed\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b593a DTSTART;TZID=America/New_York:20250123T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250123T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting - January 2025 CLASS:PUBLIC DESCRIPTION:Talk Title: CSD Faculty Meeting\n\nFaculty Meeting\n\nSee email announcements.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b5c0c DTSTART;TZID=America/New_York:20250113T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20250113T090000 SUMMARY:First Day of Classes S25 CLASS:PUBLIC DESCRIPTION:Talk Title: S25 First Day of Classes\n\nWelcome Back! \n\nFirs t day of classes for Spring 2025 semester\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b5e6b DTSTART;TZID=America/New_York:20241212T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241212T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - Alireza Shirzad CLASS:PUBLIC DESCRIPTION:Speaker: ALIREZA SHIRZAD\, PH.D. Student\, Department of Comput er and\nInformation Science\, University of Pennsylvania\n\nSNARKs are pow erful cryptographic primitives that allow a prover to\nproduce a succinct proof of a computation. Two key goals of SNARK\nresearch are to minimize t he size of the proof and to minimize the\ntime required to generate the pr oof. In this work\, we present new\nSNARK constructions that push the fron tier on both of these goals. \n\nOur first construction\, Pari\, is a SNA RK that achieves the smallest\nproof size amongst all known SNARKs. Specif ically\, Pari achieves a\nproof size of just two group elements and two fi eld elements\, which\,\nwhen instantiated with the BLS12-381 curve\, total s just 160 bytes\,\nsmaller than that of Groth16 [Groth\, EUROCRYPT '16] a nd Polymath\n[Lipmaa\, CRYPTO '24]. \n\nOur second construction\, Garuda\ , is a SNARK that reduces proof\ngeneration time by supporting\, for the f irst time\, arbitrary \"custom\"\ngates and free linear gates. To demonstr ate Garuda's performance\, we\nimplement and evaluate it\, and show that i t provides significant\nprover-time savings compared to both the state-of- the-art SNARKs\n(Groth16 and HyperPlonk [EUROCRYPT '22]). \n\nBoth constr uctions rely on a new cryptographic primitive:\n\"equifficient\" polynomia l commitment schemes that enforce that\ncommitted polynomials have the sam e representation in particular\nbases. We SNARKs are powerful cryptographi c primitives that allow a\nprover to produce a succinct proof of a computa tion. Two key goals of\nSNARK research are to minimize the size of the pro of and to minimize\nthe time required to generate the proof. In this work\ , we present new\nSNARK constructions that push the frontier on both of th ese goals. \n\nJoint work with Michel Dellepere & Pratyush Mishra.  Refe rence Paper\n\n—\n\nAlireza Shirzad is a first-year PhD student at Penn advised by Dr.\nPratyush Mishra. He is primarily interested in designing proof systems\nand SNARKs. Before coming to Penn\, He obtained his Master ’s degree\nin Secure communications and Cryptography in 2023 and a bache lor’s\ndegree in Electrical Engineering in 2021 from Sharif University o f\nTechnology. \n\nIn Person and Zoom Participation.  See announcement.\ n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b6353 DTSTART;TZID=America/New_York:20241212T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241212T130000 LOCATION:Gates Hillman 4405 SUMMARY:MSCS Thesis Defense - Nabe Efe Çekirge CLASS:PUBLIC DESCRIPTION:Speaker: NABE EFE ÇEKIRGE\, Master's Student\, Computer Scienc e\nDepartment\, Carnegie Mellon University\n\nTalk Title: Linear Sketches for Geometric LP-Type Problems\n\nLP-type problems such as the Minimum Enc losing Ball (MEB)\, Linear\nSupport Vector Machine (SVM)\, Linear Programm ing (LP)\, and\nSemidefinite Programming (SDP) are fundamental combinatori al\noptimization problems\, with many important applications in machine\nl earning applications such as classification\, bioinformatics\, and\nnoisy learning. We study LP-type problems in several streaming and\ndistributed big data models\, giving ϵ-approximation linear sketching\nalgorithms wit h a focus on the high accuracy regime with low\ndimensionality d\, that is \, when d < (1 / ε)0.999. Our main result is\nan O(ds) pass algorithm wit h O ( s ( √d / ε )3d/s) ⋅ poly (d\, log\n(1/ε)) space complexity\, f or any parameter s ∈ [1\, d log (√d /\nε)]\, to solve ϵ-approximate LP-type problems of O(d) combinatorial\nand VC dimension. Notably\, by tak ing s = d log (√d / ε)\, we achieve\nspace complexity polynomial in d a nd polylogarithmic in 1 / ϵ\,\npresenting exponential improvements in 1 / ϵ over current algorithms.\nWe complement our results by showing lower b ounds of (1 / ε)Ω(d) for\nany 1-pass algorithm solving the (1 + ε)-appr oximation MEB and linear\nSVM problems\, further motivating our multi-pass approach.  \n\nThesis Committee\n\nDavid Woodruff (Chair)\n\nRichard Pe ng\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b6703 DTSTART;TZID=America/New_York:20241211T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241211T133000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and 4300 / 4400 Corrido rs\,\nGates Hillman Center SUMMARY:Fall '24 SCS Thesis and Independent Study Poster Session and Senior \nThesis Research Presentations/Updates CLASS:PUBLIC DESCRIPTION:12:00 - 1:30 PM\n\n—  Please plan to attend our annual Fall SCS Thesis and Independent\nStudy Poster Session from on the Gates Hillman 4th floor near Rashid\nAuditorium.  Posters will be available for SCS se nior thesis students\nas well as independent study students. \n\n9:30 - 1 1:45 AM\n\n—  In addition\, you are welcome to join in at the SCS Senio r Thesis\nResearch Updates that will be held in GHC 4405. Senior thesis st udents\nwill present  5 minute presentations on their current progress an d\nexpected work for next semester. \n\nJoin in as your schedule permits. Faculty / Ph.D. students are\nencouraged to engage these students\, to as k questions\, and to provide\ninsights/advice on their specific research.  \n\nRefreshments provided.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b6a73 DTSTART;TZID=America/New_York:20241210T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241210T130000 LOCATION:Gates Hillman 9115 SUMMARY:5th Year MS Thesis Presentation - Tae Hoon Kim CLASS:PUBLIC DESCRIPTION:Speaker: TAE HOON KIM\, Master's Student\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: A Principled Framewor k for Pliable and Secure Speculation\nin Operating Systems\n\nTransient ex ecution attacks present an unprecedented threat to\ncomputing systems. Pro tecting the operating system (OS) is\nexceptionally challenging because a transient execution gadget in the\nOS can potentially leak the entire memo ry. \n\nIn this work\, we propose Perspective\, a principled framework fo r\nbuilding pliable and secure speculative execution defenses for the OS.\ nPerspective offers a pliable interface that allows the OS to\ncommunicate its security requirements to hardware defenses\, enabling\ntailored prote ction against transient execution attacks with little\nperformance overhea d. The design of Perspective is driven by a\ntaxonomy of transient executi on attacks in the OS kernel: (i) active\ntransient execution attacks\, whe re the attacker process exploits its\nown kernel thread to speculatively e xecute a transient execution\ngadget in the kernel\, and (ii) passive tran sient execution attacks\, \nwhere the attacker coerces the victim process ’s kernel thread to\nexecute a transient execution gadget. \n\nBased on the taxonomy\, Perspective introduces Data Speculation Views\n(DSVs) and Instruction Speculation Views (ISVs)\, to mitigate active\nand passive att acks\, respectively. DSVs define the ownership of kernel\ndata by a given execution context and block any speculative access to\ndata outside the DS V. ISVs define the set of kernel functions that can\nbe speculatively exec uted by a given execution context. Any\ntransmitter instructions—whose e xecution could leak secrets\, such as\nload instructions—that belong to kernel functions outside the ISVs\nare blocked from speculative execution. ISVs open up new opportunities\nof (i) swiftly patching gadgets in the OS \, (ii) reducing the surface\nof passive attacks\, and (iii) speeding up t he process of auditing\ntransient execution gadgets in the OS. \n\nWe bui ld Perspective’s software components in the Linux kernel and\nmodel the hardware components in gem5. We evaluate the security and\nperformance of Perspective on a set of microbenchmarks and datacenter\napplications. Pers pective has an execution overhead over an\nunprotected kernel of only 3.5% on microbenchmarks and only 1.2% on\ndatacenter applications.\n\nThesis C ommittee\n\nDimitrios Skarlatos (Chair)\n\nWenting Zheng\n\nAdditional Inf ormation\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b6f3c DTSTART;TZID=America/New_York:20241210T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241210T123000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-anup-agarwal LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Proposal - Anup Agarwal CLASS:PUBLIC DESCRIPTION:Speaker: ANUP AGARWAL\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Designing Network Contro l Algorithms with Performance\nGuarantees\n\nControl algorithms\, such as those used in congestion control (CC) and\nadaptive-bitrate (ABR) streamin g\, make critical decisions that\ninfluence performance\, user experience\ , and revenue in networked\napplications. Despite their importance\, these algorithms are often\ndeveloped through heuristics and intuition\, leadin g to implicit\nassumptions about network environments and a lack of formal guarantees\nabout their performance. The result is anywhere from silent t o\nembarrassing performance degradation (e.g.\, the recent Netflix\nlivest ream of Paul vs Tyson).\n\nIn this proposed thesis\, we design tools (both mathematical and\ncomputational) that make formal performance guarantees possible.\nNetwork control algorithms often operate in partially observabl e\nenvironments\, e.g.\, they do not explicitly know the network topology\ ,\nrouting\, link capacities\, or what other flows they share the network\ nwith. The mathematical tools allow us to reason about what information\na lgorithms may infer from their partial observations. The\ncomputational to ols build on these mathematical tools to semi-automate\nthe design of cont rol algorithms\, using techniques from program\nsynthesis\, and game theor y. By combining our tools with techniques in\nnetwork calculus and formal methods\, we precisely state performance\nobjectives and environment assum ptions\, enabling us to construct\nproofs about the performance of the des igned control algorithms.\n\nThrough this work\, we design new CC algorith ms that provide\nperformance bounds on networks where existing algorithms struggle to\nguarantee even 1% utilization. Our systematic methodology als o led us\nto discover and prove new fundamental tradeoffs in congestion co ntrol\,\nincluding a tradeoff between loss and convergence time on network s\nwith jitter and short buffers\, and a tradeoff between fairness and\nro bustness vs. latency and generality of CC algorithms.\n\nThe majority of t he completed work deals with single-agent congestion\ncontrol. In the prop osed work\, we extend our tools to reason about\nmulti-agent control envir onments and ABR algorithms. \n\nThesis Committee\n\nSrinivasan Seshan (Ch air)\n\nVyas Sekar\n\nJustine Sherry\n\nPhilip Brighten Godfrey (Universit y of Illinois Urbana-Champaign)\n\n \n\nAdditional Information\n\nIn Pers on and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b73f2 DTSTART;TZID=America/New_York:20241210T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241210T120000 LOCATION:Gates Hillman 7501 SUMMARY:5th Year MS Thesis Presentation - Justin Zhang CLASS:PUBLIC DESCRIPTION:Speaker: JUSTIN ZHANG\, Master's Student\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Secure Convertible Co des\n\nLarge-scale distributed storage systems (DSS) make use of erasure\n codes to enforce fault tolerance in the event of node failure.  Due\nto o bserved changing failure rates within these systems\, code\nredundancy tun ing\, or code conversion has been shown to reduce storage\ncost.  Prior w ork has developed bounds and constructions across many\nparameters for con vertible codes\, a class of erasure codes optimizing\neither the access or bandwidth costs of conversion. \n\nIn this thesis\, we investigate the i nformation-theoretic security of\nconvertible codes under the presence of an eavesdropper whom we\nenforce to learn nothing of the stored message. W hile current\nconvertible code constructions are inherently insecure since they are\nsystematic\, we present novel constructions that augment existi ng\ncost-optimal convertible codes with perfect eavesdropper security. \n \nFurthermore\, we prove that our constructions maximizes the amount of\ni nformation that can be stored on such a system\, and we give\nadditional c onstructions and bounds for secure codes when additional\ninformation is k nown about the distribution of eavesdropped nodes. \n\nThesis Committee\n \nRashmi Vinayak (Chair)\n\nAayush Jain\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b77d0 DTSTART;TZID=America/New_York:20241209T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241209T103000 LOCATION:Gates Hillman 8102 SUMMARY:MSCS Thesis Defense - Aditya Chanana CLASS:PUBLIC DESCRIPTION:Speaker: ADITYA CHANANA\, Master's Student\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Analyzing the OODA Loop of an Edge-enabled Autonomous\nDrone System\n\nThe \"Observe\, Orient \, Decide\, Act\" (OODA) loop encapsulates the\nagility of cyber-physical or cyber-human systems that depend on\ncontinuous iterations of these step s. Systems with faster OODA loops\nreact more quickly to changes in their environment. This work analyzes\nthe OODA loop of the SteelEagle edge-enab led autonomous drone system\,\nwhich transforms consumer aerial photograph y drones into fully\nautonomous UAVs by offloading computation to the edge using a drone\npayload with cellular connectivity. We identify bottleneck s and\nopportunities for optimization\, leading to a faster SteelEagle OOD A\nloop and thus improved performance in active vision tasks such as\nobst acle avoidance and object tracking. This enables the drone to fly\nsafely at higher speeds in crowded spaces\, increasing the practicality\nof Steel Eagle drones in applications such as search and rescue and\ninfrastructure inspection. Our findings show that the hardware\nencoding of the RTSP vid eo stream on the drone makes up about\ntwo-thirds of the drone-to-cloudlet latency. \n\nBecause of its dependence on offloading\, SteelEagle is cur rently\nlimited in its ability to operate in degraded network conditions. To\nmitigate these limitations\, we analyze the use of onboard computation \nwith SteelEagle by considering a new payload that can run\nfloat16-quant ized DNNs. We discuss how onboard computational abilities\ncan be combined with offloading to achieve an optimal system based on\ncomputation accura cy\, energy efficiency\, and latency. \n\nThesis Committee\n\nMahadev Sat yanarayanan (Chair)\n\nPadmanabhan Pillai\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b7bd8 DTSTART;TZID=America/New_York:20241206T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241206T170000 URL:https://www.cmu.edu/math/news-events/calendar.html#event=77803855\;inst ance=20241206160000?popup=1&lang=en-US LOCATION:Learning Hall\, Doherty Hall A302 SUMMARY:Frontiers of Mathematical Sciences Colloquium - Kevin Mark Buzzard CLASS:PUBLIC DESCRIPTION:Speaker: KEVIN MARK BUZZARD\, Professor of Pure Mathematics\, D epartment\nof Mathematics\, Faculty of Natural Sciences\, Imperial College \n\nTalk Title: Will computers prove theorems?\n\nDecades ago\, they digit ised music. At the time\, it just seemed like an\nexcuse to sell CDs at a higher price than vinyl records. But then came\nthe internet\, and now we have access to millions of songs on our\nphones. Digitisation of music ult imately turned the world of music on\nits head. \n\nDecades ago\, they al so digitised mathematics. For many years\,\nmathematicians didn't care. Bu t then came language models\, and now\nDeepMind can answer 4/6 questions o n the 2024 International Maths\nOlympiad and there are no signs of things slowing down. Is mathematics\ngoing to be turned on its head by these tool s? In this talk\, which is\nwritten for mathematicians and which will assu me no background in AI\nor computer science\, I will explain what it means to digitise\nmathematics using theorem provers\, I'll explain how these s ystems are\nnow engaging with modern mathematical ideas in several differe nt\nareas\, I'll talk about the benefits these systems have brought to\nma thematics\, and I'll discuss what is hype\, what might actually happen\nin the next few years\, and how we as mathematicians can help make it\nhappe n. \n\n— Kevin Mark Buzzard is a British mathematician and currently a\ nprofessor of pure mathematics at Imperial College London. He\nspecialises in arithmetic geometry and the Langlands program. \nAdditional informati on\n\nPresented by the Department of Mathematics Sciences\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b7f9e DTSTART;TZID=America/New_York:20241206T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241206T163000 URL:https://brain.andrew.cmu.edu/seminar LOCATION:Baker Hall 340A and Zoom SUMMARY:brAIn Seminar - David Lipshutz CLASS:PUBLIC DESCRIPTION:Speaker: DAVID LIPSHUTZ\, Assistant Professor (Incoming)\, Depa rtment of\nNeuroscience\, Baylor College of Medicine\n\nTalk Title: Compar ing image representations in terms of their local\ngeometry\n\nImage repre sentations (artificial or biological) are often compared in\nterms of thei r global geometry\; however\, representations with similar\nglobal structu re can have strikingly different local geometries. Here\,\nwe propose a fr amework for comparing a set of image representations in\nterms of their lo cal geometries. \n\nWe quantify the local geometry of a representation us ing the Fisher\ninformation matrix\, a standard statistical tool for chara cterizing the\nsensitivity to local stimulus distortions\, and use this as a substrate\nfor a metric on the local geometry in the vicinity of a base image.\nThis metric may then be used to optimally differentiate a set of\ nmodels\, by finding a pair of \"principal distortions\" that maximize the \nvariance of the models under this metric. We use this framework to\ncomp are a set of simple models of the early visual system\, identifying\na nov el set of image distortions that allow immediate comparison of\nthe models by visual inspection. \n\nIn a second example\, we apply our method to a set of deep neural\nnetwork models and reveal differences in the local ge ometry that arise\ndue to architecture and training types. These examples highlight how\nour framework can be used to probe for informative differen ces in\nlocal sensitivities between complex computational models\, and sug gest\nhow it could be used to compare model representations with human\npe rception. \n\n— \n\nDavid Lipshutz is an incoming assistant professor of Neuroscience at\nBaylor College of Medicine. He's currently an associa te research\nscientist at the Flatiron Institute where he works with Eero\ nSimoncelli and Dmitri Chklovskii. Prior to working in Neuroscience\, he\n received his Ph.D. in mathematics from UCSD and held postdoctoral\npositio ns at Brown University (Applied Math) and the Technion\n(Electrical Engine ering). Additional Information\n\nCommunal Viewing in Baker 340A and Zoom Participation.  See\nannouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b83e7 DTSTART;TZID=America/New_York:20241206T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241206T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Gates Hillman 8102 and Zoom SUMMARY:AI Institute for Societal Decision Making Seminar - Alex London CLASS:PUBLIC DESCRIPTION:Speaker: ALEX LONDON\, K&L Gates Professor of Ethics and Comput ational\nTechnologies\, Department of Philosophy\, Carnegie Mellon Univers ity\n\nTalk Title: When AI Can No Longer Help: Agency Transfer Agreements for\nAssistive AI Systems\n\nAs the populations of many high-income countr ies age there is growing\ninterest in artificial intelligence (AI) systems that would support\nolder adults and extend their ability to live indepen dently. This\nincludes AI systems that function as personal assistants hel ping to\nmanage finances\, pay bills\, organize health-related information \,\nassist with medications\, monitor health status\, manage schedules\, o r\ncall for emergency assistance. \n\nThis talk identifies strategies tha t AI systems might use to extend\nthe decisional capacity of older adults and with this their ability to\nmanage important cognitive tasks required for independent living. It\nthen introduces the idea of an “agency trans fer point\,” defined as\nthe point at which responsibility for a specifi c task must be\ntransferred to another person because the older adult’s cognitive\ndecline exceeds the ability of an AI system to extend their dec isional\ncapacity for that task. As the ability of AI systems to provide s uch\nsupport increases\, so does the probability that some individuals wil l\nreach an agency transfer point while living independently. \n\nThis ta lk provides an ethical analysis of strategies for addressing\nthe risks to older adults from being in a situation where they are no\nlonger capable of managing a task\, but no agent has been identified to\nassume responsib ility for that task. This includes a discussion of\nrequiring\, as a condi tion for using such systems\, that older adults\ncreate an “advance dire ctive” that identifies the person to whom\nthey prefer to transfer respo nsibility once an agency transfer point\nhas been reached. \n\n— \n\nA lex John London is the K&L Gates Professor of Ethics and\nComputational Te chnologies at Carnegie Mellon University. An elected\nFellow of the Hastin gs Center\, Professor London’s work focuses on\nethical and policy issue s surrounding the development and deployment\nof novel technologies in med icine\, biotechnology and artificial\nintelligence. His book\, For the Com mon Good: Philosophical Foundations\nof Research Ethics is available in ha rd copy from Oxford University\nPress and is available in PDF as an open a ccess title. He is a member\nof the World Health Organization (WHO) Expert Group on Ethics and\nGovernance of AI whose “Guidance on Large Multi-Mo dal Models” was\npublished in 2024 and whose report\, “Ethics and gove rnance of\nartificial intelligence for health” was published in 2021. Fr om\n2022–2023 he was a member of the U.S. National Academy of Medicine\n Committee on Creating a Framework for Emerging Science\, Technology\,\nand Innovation in Health and Medicine\, whose report \"Toward Equitable\nInno vation in Health and Medicine: A Framework” was published in\n2023. He i s currently a co-leader of the ethics core for the NSF AI\nInstitute for C ollaborative Assistance and Responsive Interaction for\nNetworked Groups ( AI-CARING).\n\n REGISTER→  Information on joining the meeting will be sent upon\nregistration.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b8989 DTSTART;TZID=America/New_York:20241205T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241205T170000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-asher-trockman LOCATION:Scott Hall 6002 and Zoom SUMMARY:Doctoral Thesis Proposal - Asher Trockman CLASS:PUBLIC DESCRIPTION:Speaker: ASHER TROCKMAN\, Ph.D. Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Mimetic Initialization for Deep Neural Networks\n\nWhile neural network weights are typically in itialized randomly from\nunivariate distributions\, pre-trained weights of ten have\nvisually-discernible multivariate structure. We propose a techni que\ncalled \"mimetic initialization\" that aims to replicate such structu res\nwhen initializing convolutional networks (CNNs)\, Transformers\, and\ nState Space Models (SSMs). For CNNs\, we handcraft a class of\nmultivaria te Gaussian distributions to initialize filters for\ndepthwise convolution al layers\; for Transformers\, we initialize the\nquery and key weights fo r self-attention layers such that their\nproduct approximates the identity \; and for SSMs\, we initialize layers\nto approximate simple linear atten tion. Mimetic initialization\nsubstantially reduces training time and incr eases final accuracy on\nvarious common small-scale benchmarks.\n\nOur tec hnique enables us to almost close the gap between untrained and\npre-train ed Vision Transformers on small datasets like CIFAR-10\,\nachieving up to a 6% gain in accuracy through initialization alone.\nFor convolutional net works like ConvMixer and ConvNeXt\, we observe\nimprovements in accuracy a nd reductions in training time\, even when\nconvolutional filters are froz en (untrained) after initialization. For\nSSMs\, mimetic initialization su bstantially improves generalization\nabilities on synthetic language tasks like copying and associative\nrecall. Overall\, our findings suggest that the benefits of\npre-training can be separated into two components: servi ng as a good\ninitialization and storing transferable knowledge\, with the former\nbeing simple enough to (at least partially) capture by hand in\nc losed-form.\n\nThesis Committee\n\nZico Kolter (Chair)\n\nAlbert Gu\n\nAdi ti Raghunathan\n\nSébastien Bubeck (OpenAI)\n\n \n\nAdditional Informati on\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b8e0a DTSTART;TZID=America/New_York:20241205T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241205T163000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Mehrabian Collaborative Innovation Center 1301 and Zoom (special\n time/location) SUMMARY:Crypto Seminar - Jiahui Liu CLASS:PUBLIC DESCRIPTION:Speaker: JIAHUI LIU\, Senior Researcher\, Quantum Computing\, F ujitsu\nResearch\n\nTalk Title: The Black-Box Simulation Barrier Persists in a Fully\nQuantum World\n\nZero-Knowledge (ZK) protocols have been a sub ject of intensive study\ndue to their fundamental importance and versatili ty in modern\ncryptography. However\, the inherently different nature of q uantum\ninformation significantly alters the landscape\, necessitating a\n re-examination of ZK designs.\n\nA crucial aspect of ZK protocols is their round complexity\,\nintricately linked to  simulation\, which forms the foundation of\ntheir formal definition and security proofs. In the *post-q uantum*\nsetting\, where honest parties and their communication channels a re all\nclassical but the adversaries could be quantum\, Chia\, Chung\, Li u\, and\nYamakawa [FOCS'21\, QIP'22] demonstrated the non-existence of\nco nstant-round  black-box-simulatable ZK arguments (BBZK) for NP\nunless NP is in BQP. However\, this problem remains widely open in the\nfull-fledge d quantum future that will eventually arrive\, where all\nparties (includi ng the honest ones) and their communication are\nnaturally quantum.  \n\n Indeed\, this problem is of interest to the broader theory of quantum\ncom puting. It has been an important theme to investigate how quantum\npower f undamentally alters traditional computational tasks\, such as\nthe uncondi tional security of Quantum Key Distribution and the\nincorporation of Obli vious Transfers in MiniQCrypt. Moreover\, quantum\ncommunication has led t o round compression for commitments and\ninteractive arguments. Along this line\, the above problem is of great\nsignificance in understanding wheth er quantum computing could also\nchange the nature of ZK protocols in some fundamental manner. \n\nWe resolved this problem by proving that only la nguages in BQP admit\nconstant-round *fully-quantum* BBZK. This result hol ds significant\nimplications. Firstly\, it illuminates the nature of quant um\nzero-knowledge and provides valuable insights for designing future\npr otocols in the quantum realm. Secondly\, it relates ZK round\ncomplexity w ith the intriguing problem of BQP vs QMA\, which is out of\nthe reach of p revious analogue impossibility results in the classical\nor post-quantum s etting. Lastly\, it justifies the need for the\n*non-black-box* simulation techniques or the relaxed security notions\nemployed in existing constant -round fully-quantum BBZK protocols.\n\nJoint work with Kai-Min Chung\, Na i-Hui Chia\, Xiao Liang.\n\nIn Person and Zoom Participation.  See announ cement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b92fa DTSTART;TZID=America/New_York:20241205T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241205T160000 URL:https://aco.math.cmu.edu/abs-24-25/dec5.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Jakob Hofstad CLASS:PUBLIC DESCRIPTION:Speaker: JAKOB HOFSTAD\, Ph.D. Student\, Department of Mathemat ical\nSciences\, Carnegie Mellon University\n\nTalk Title: A Note on Two-P oint Concentration of the Independence\nNumber of G_n\,mIt is well known t hat\, for any constant probability p\,\nthere exists a function k(n) such that the independence number α of\nthe binomial random graph Gn\,p is con centrated on two values\; i.e.\,\nα(Gn\,p) equals k(n) or k(n)+1 with hig h probability (that is\, with\nprobability converging to 1 as n approaches infinity). This result was\nproved independently by Erdös and Bollobás and by Matula in the\n1970's by using standard first and second moment tec hniques on the\nrandom variable counting the number of independent sets of a given\nsize. Although not emphasized in the literature\, this same stra tegy\ncan be used to prove the above statement if p=p(n) > n−γ for any\ nconstant γ < 1/3. In this talk I will discuss the two consecutive\nresul ts which\, in combination\, gives the extent of concentration of\nα(Gn\,p ) for p as small as n−3/4+ϵ\, in particular the latter paper\,\nwhich e stablishes two-point concentration of α(Gn\,m) for m > n5/4+ϵ\nand gives a threshold at which α(Gn\,p) is no longer two-point\nconcentrated.\nJoi nt work with Tom Bohman.4:00 pm → Jane Street-sponsored tea and\ncookies at 4:00 pm in the Math Lounge\, Wean 6220\n                → bring your own mug if you have one\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b96af DTSTART;TZID=America/New_York:20241205T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241205T153000 LOCATION:Gordon Bell Conference\, Gates Hillman 5117 SUMMARY:Doctoral Thesis Proposal - Emre Yolcu CLASS:PUBLIC DESCRIPTION:Speaker: EMRE YOLCU\, Ph.D. Student\, Computer Science Departme nt\,\nCarnegie Mellon University\n\nTalk Title: Weak Extended Resolution a nd Nonautomatability in Proof\nComplexity\n\nThe satisfiability problem fo r propositional logic (SAT) has been a\ncentral topic in computer science for many decades. It is arguably the\ncanonical NP-complete problem: reduc tions from many other problems in\nNP to SAT are often straightforward. As a consequence\, a reasonable\nstrategy when trying to solve a problem in NP is to reduce it to SAT\nand to try to solve the resulting SAT problem i nstead. This strategy\nturns out to be surprisingly effective thanks to th e effectiveness of\nimplementations of heuristic algorithms for SAT\, comm only known as SAT\nsolvers. Those solvers are expected to output proofs to certify their\nanswers\, and in this sense they are proof search algorith ms. Proof\ncomplexity\, the branch of computational complexity that studie s the\nlengths of proofs in propositional proof systems\, offers a way to\ nanalyze the performance of SAT solvers. \n\nIn this thesis proposal\, we focus on proof complexity theoretic\nproblems from two \"contrasting\" en ds of a spectrum: those motivated by\nSAT solving\, aiming to improve heur istic proof search\, and those\narising from computational complexity\, ai ming to show that proof\nsearch is hard in the worst case. In particular\, we focus on the\nfollowing two lines of research: proof complexity of a f amily of\nsystems that we refer to as weak extended resolution systems and the\nautomatability problem from computational complexity. The first\ninv olves proof systems that formalize restricted versions of the\nability to make assumptions that hold without loss of generality\,\ncommonly used inf ormally to shorten proofs. The second is the question\nof whether efficien t proof search is possible in the worst case. We\ngive a complete descript ion of the relative strengths of core weak\nextended resolution systems by introducing a high-level recipe for\nproving separations. We introduce a technique to reduce the hardness\nof automatability between proof systems in a black-box manner\; we give\nan example application and propose an app roach to lift hardness\nresults from weaker to stronger proof systems. \n \nThesis Committee \n\nMarijn J. H. Heule (Chair)\n\nJeremy Avigad\n\nRya n O'Donnell\n\nSamuel R. Buss (University of California San Diego)\n\nAddi tional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b9b44 DTSTART;TZID=America/New_York:20241205T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241205T130000 URL:https://www.pdl.cmu.edu/SDI/2024/120524.html LOCATION:Panther Hollow Room 4105\, Mehrabian Collaborative Innovation Cent er SUMMARY:SDI Seminar - Josh Fried CLASS:PUBLIC DESCRIPTION:Speaker: JOSH FRIED\, Ph.D. Student\, Computer Science & Artifi cial\nIntelligence Laboratory\, Massachusetts Institute of Technology\n\nT alk Title: Making Kernel Bypass Practical for the Cloud with Junction\n\nK ernel bypass systems deliver significant improvements in throughput\nand t ail latency for network-intensive applications\, often\noutperforming trad itional operating systems (OSes) by an order of\nmagnitude. However\, thes e performance gains come at a steep cost: they\nrely on dedicated resource s (e.g.\, spinning cores\, pinned memory) and\ndemand extensive applicatio n rewriting. These limitations make kernel\nbypass systems impractical for cloud operators\, who prioritize dense\napplication packing and seek to a void costly code modifications. \n\nIn this talk\, I will introduce Junct ion\, a kernel bypass system that\novercomes these challenges\, supporting thousands of instances on a\nsingle machine while remaining compatible wi th unmodified Linux\napplications. Junction leverages modern NIC features to minimize\nmemory and monitoring overheads\, enabling dense packing of i nstances.\nIt incorporates a library operating system that implements the Linux\nsystem call interface\, allowing compatibility with existing\nappli cations while relying on kernel bypass for efficient OS\nfunctionality. Ju nction matches or exceeds the performance of\nstate-of-the-art kernel bypa ss systems without requiring application\nporting. It improves throughput\ , latency\, and CPU efficiency for\nunmodified network-intensive datacente r applications and seamlessly\nsupports popular frameworks like Go\, Pytho n\, and Java. \n\n— \n\nJosh Fried is a senior PhD student at MIT CSAI L advised by Adam Belay.\nHe is broadly interested in operating systems\, networks\, and\ndistributed systems\, with a focus on improving performanc e and\nefficiency for datacenters through better operating system design.  \n\nFaculty Host:  Justine Sherry\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91b9f72 DTSTART;TZID=America/New_York:20241205T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241205T130000 LOCATION:ASA Conference Room\, Gates Hillmn 6115 SUMMARY:EconCS Seminar - Jaisun Li CLASS:PUBLIC DESCRIPTION:Speaker: JIASUN LI\, Visiting Faculty\, Computer Science Depart ment\,\nCarnegie Mellon University\, and\, Associate Professor of Finance (on\nsabbatical) \, George Mason University\n\nTalk Title: Profit Division Among Privately-Informed Agents and Beyond\n\nWhen n privately-informed investors (or AI agents on behalf of them)\njointly invest in a risky pro ject\, how should they divide profit? Is\nthe familiar practice of dividi ng in proportion to initial\ninvestments really optimal? In a simple examp le\, I demonstrate that\nthe Pareto dominant outcome may only be implement ed by a Bayesian Nash\nequilibrium when agents equally share profits\, reg ardless of their\ninitial contributions or private information precisions.  \n\nIn the first half of the talk\, I will prove this statement\nanalyti cally under fairly standard utility/signal assumptions\, and\nreconcile it with the familiar practice of dividing according to\ninitial contribution s. Then in the second half of the talk\, I will\nopen up to several potent ial directions for further generalization and\ndiscuss ongoing work along them\, envisioning how they may inspire (1)\nthe co-design of statistical learning algorithms and contracts and (2)\novercoming numerous impossibili ty results in mechanism design with a\nrelaxed solution concept of Bayesia n Nash equilibrium of contingent\nschedules. I will also discuss potential applications\, including the\ndesign of decentralized autonomous organiza tions (DAO)\, the\naugmentation of secure multi-party computation (MPC)\, the development\nof intent-based markets\, and the design of crowdfunding for\nharnessing the \"wisdom of the crowd\"\, etc. \n\n— About the Sem inar:  This seminar is designed for those working on\nor otherwise intere sted in topics at the intersection of computer\nscience and economics (alg orithmic game theory\, computational social\nchoice\, mechanism design\, e tc).  \n\nJoin in to learn more about this new series.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ba375 DTSTART;TZID=America/New_York:20241204T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241204T153000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Speaking Skills Talk - Meng-Chieh (Jeremy) Lee CLASS:PUBLIC DESCRIPTION:Speaker: MENG-CHIEH (JEREMY) LEE\, Ph.D. Student\, Computer Sci ence\nDepartment\, Carnegie Mellon University\n\nTalk Title: Accurate\, Ro bust\, and Interpretable Graph Mining\n\nHow can we solve semi-supervised node classification in various graphs\npossibly with noisy features and st ructures? Graph neural networks\n(GNNs) have succeeded in many graph minin g tasks\, but their\ngeneralizability to various graph scenarios is limite d due to the\ndifficulty of training\, hyperparameter tuning\, and the sel ection of a\nmodel itself. \n\nIn this talk\, I will present a carefully- designed simple model SlimG\nfor solving semi-supervised node classificati on. It exhibits the\nfollowing desirable properties: accurate\, robust\, f ast\, scalable\, and\ninterpretable. \n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ba729 DTSTART;TZID=America/New_York:20241204T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241204T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20241204.html LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Lunch Seminar - Renfei Zhou CLASS:PUBLIC DESCRIPTION:Speaker: RENFEI ZHOU\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Dynamic \"Succincter\"The seminal work \"Succincter\" by\nPătraşcu presents a way to store an aug mented B-tree with only two\nbits of redundancy while supporting queries e fficiently. It is a\ngeneric and powerful tool for designing static succin ct data\nstructures. We extend \"Succincter\" to support dynamic operation s\,\nachieving the same space bounds as the static \"Succincter\" and the\ noptimal time bounds\, enabling numerous applications. Our technique\naddr esses a key challenge in dynamic succinct data structures:\nmanaging varia ble-size components within a contiguous piece of memory.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91baa75 DTSTART;TZID=America/New_York:20241204T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241204T110000 LOCATION:Gates Hillman 7101 and Zoom SUMMARY:5th Year MS Thesis Presentation - Xin Yue Amanda Li CLASS:PUBLIC DESCRIPTION:Speaker: XIN YUE AMANDA LI\, Master's Student\, Computer Scienc e\nDepartment\, Carnegie Mellon University\n\nTalk Title: WaybackUI: A Dat aset to Support the Longitudinal Analysis\nof User Interfaces\n\nHow do us er interfaces (UIs) change over time? Understanding the\nevolution of UIs is essential for assessing the impact on users and\nautomated systems that interact with them. To this end\, we collected\nWaybackUI\, a dataset of nearly 100\,000 UIs from 2014 -- 2024 mined\nfrom publicly available inter net archives\, paired with rendered\nscreenshots and programmatically-extr acted semantics. The resulting\ndata allowed us to analyze how a decade of UI changes has impacted i)\nvisual design\, ii) accessibility\, and iii) automated systems that\ninteract with UIs. Examples of our findings includ e: i) a growing\npreference for muted color palettes in visual design\, ii ) an increase\nin the number of highly inaccessible web pages in recent ye ars\, and\niii) quantifying the impact of data drift on the performance of UI\nunderstanding models. We conclude with a discussion of how WaybackUI\ ncan enable future data-driven discovery and understanding of UI\ntrends.  \n\nThesis Committee Jeffrey Bigham (Chair)\n\nNikolas Martelaro\n\nAddi tional Information\n\nIn Person and Zoom Participation.   See announceme nt.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bae34 DTSTART;TZID=America/New_York:20241203T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241203T130000 LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Doctoral Speaking Skills Talk - William Zhang CLASS:PUBLIC DESCRIPTION:Speaker: WILLIAM ZHANG\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: The Holon Approach to H olistic Database Optimization\n\nThe optimal configuration of a database m anagement system (DBMS)\ndepends on its workload\, database contents\, and hardware. However\,\nthese constantly change over time. It is increasingl y difficult for\nhumans to reason about those changes with the growing com plexity and\nvariety of tunable DBMS system components. As such\, this nec essitates\nusing automated machine learning agents to tune the DBMS. \n\n The challenge is that a tuning agent must explore a large action space\nto construct and try out promising configurations while learning how\nto opt imize for a given workload and objective (e.g.\, minimize\nlatency) functi on. As the tuner reasons across DBMS tunable aspects\n(e.g.\, knobs and in dexes)\, the number of available combined actions\ngrows combinatorially. In order to manage this explosion\, existing\ntechniques optimize each tun able aspect individually and in isolation\nfrom one another. They then att empt to compose the local optima\ndiscovered by each tuner into a holistic configuration. However\, this\nprocess does not guarantee finding global optima. \n\nRather than composing bespoke tuners\, we should use a holist ic model\nto simultaneously reason across multiple configuration spaces (i .e.\,\nconsider multiple tuning decisions simultaneously). In order to man age\nthe space's complexity\, we make a critical insight. Although the\nnu mber of unique actions in the space is large\, many share similar\npropert ies. This similarity\, derived from performance estimates or\ndomain knowl edge\, enables an agent to reduce the effective space by\ntransferring obs ervations from one action to similar actions. With\nthis holistic techniqu e and considering orders of magnitude more\ncomplex and varied tunable opt ions\, we achieve up to 53% workload\nreduction over state of the art tune rs for tuning PostgreSQL on\nanalytical workloads. We will conclude the ta lk with a brief\ndiscussion on future directions and next steps. \n\nPres ented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bb260 DTSTART;TZID=America/New_York:20241203T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241203T130000 LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Artificial Intelligence Seminar - Robin Jia CLASS:PUBLIC DESCRIPTION:Speaker: ROBIN JIA\, Assistant Professor of Computer Science\, Thomas\nLord Department of Computer Science\, University of Southern Calif ornia\n\nTalk Title: Auditing\, understanding\, and leveraging large langu age\nmodels\n\nThe rise of large language models offers opportunities to b oth\nscientifically study these complex systems and apply them in novel\nw ays. In this talk\, I will describe my group’s recent work along\nthese lines. First\, I will discuss data watermarks\, a statistically\nrigorous technique for auditing a language model’s training data\nbased only on b lack-box model queries. Then\, we will investigate how\nlanguage models me morize training data: based on results from two\ncomplementary benchmarks\ , I will demonstrate the viability of\nlocalizing memorized data to a spar se subset of neurons. Next\, I will\nprovide a mechanistic account of how pre-trained language models use\nFourier features to solve arithmetic prob lems\, and how pre-training\nplays a critical role in these mechanisms. Fi nally\, I will show how to\nleverage the complementary strengths of large language models and\nsymbolic solvers to handle complex planning tasks. \ n\n— \n\nRobin Jia is an Assistant Professor of Computer Science at the \nUniversity of Southern California. He received his Ph.D. in Computer\nSc ience from Stanford University\, where he was advised by Percy Liang.\nHe has also spent time as a visiting researcher at Facebook AI\nResearch\, wo rking with Luke Zettlemoyer and Douwe Kiela. He is\ninterested broadly in natural language processing and machine\nlearning\, with a focus on scient ifically understanding NLP models in\norder to improve their reliability. Robin’s work has received best\npaper awards at ACL and EMNLP. \n\nIn P erson and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bb685 DTSTART;TZID=America/New_York:20241202T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241202T173000 URL:https://db.cs.cmu.edu/events/building-blocks-apache-opendal-databend-xu anwo LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Xuanwo CLASS:PUBLIC DESCRIPTION:Speaker: XUANWO\, Apache OpenDAL PMC Chair\, and Database Engin eer\,\nDatabend Labs\n\nTalk Title: Apache OpenDAL: One Layer\, All Storag e\n\nApache OpenDAL is an Open Data Access Layer that enables seamless\nin teraction with diverse storage services\, guided by its mission of\n\"One Layer\, All Storage\" and core tenets of being open\, solid\, fast\,\nand extensible to serve various users from infrastructure builders to\napplica tion developers. In this talk\, we will explain OpenDAL in more\ndetail an d describe the abstractions it builds. We will discuss how\nOpenDAL helps developers build database systems. \n\n— \n\nXuanwo is an ASF Member\, Apache OpenDAL PMC Chair\, and Database\nEngineer at Databend Labs. \n\n This talk is part of the Database Building Blocks Seminar\n\nZoom Particip ation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bba1d DTSTART;TZID=America/New_York:20241202T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241202T163000 URL:https://www.ri.cmu.edu/event/practical-challenges-and-recent-advances-i n-data-attribution/ LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Jiaqi Ma CLASS:PUBLIC DESCRIPTION:Speaker: JIAQI MA\, Assistant Professor\, School of Information \nSciences\, University of Illinois Urbana-Champaign\n\nTalk Title: Practi cal Challenges and Recent Advances in Data\nAttribution\n\nData plays an i ncreasingly crucial role in both the performance and\nthe safety of AI mod els. Data attribution is an emerging family of\ntechniques aimed at quanti fying the impact of individual training data\npoints on a model trained on them\, which has found data-centric\napplications such as training data c uration\, instance-based\nexplanation\, and copyright compensation. In thi s talk\, I will explore\npractical challenges of deploying data attributio n in real-world\napplications. \n\nIn the first part\, I will examine the adversarial robustness of data\nattribution methods\, particularly in the context of fairly\ncompensating training data providers. Our study reveal s a critical\nvulnerability\, demonstrating how malicious data providers c an\nmanipulate these data to unfairly inflate their compensation. \n\nIn the second part\, I will address the limitations in the flexibility\nof ex isting influence function approaches and introduce a novel method\nthat ex tends data attribution to broader machine learning paradigms\,\nincluding survival analysis and contrastive learning. If time permits\,\nI will also briefly introduce our efforts to tackle challenges related\nto computatio nal efficiency and group effects in data attribution\, and\ndiscuss the cu rrent advancements and open problems in this field. \n\n— \n\nJiaqi Ma is an Assistant Professor at the University of Illinois\nUrbana-Champaign (UIUC). His research interests lie in the broad area\nof trustworthy AI\, with recent focuses including data attribution\,\nmachine unlearning\, ex plainable machine learning\, and training data\ncuration. Jiaqi's work has been recognized with the Gary M. Olson\nOutstanding Student Award from Un iversity of Michigan and a Best Paper\nAward from the DPFM Workshop at ICL R 2024. Prior to joining UIUC\,\nJiaqi earned his PhD from the University of Michigan and worked as a\npostdoctoral researcher at Harvard University .\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bbe6d DTSTART;TZID=America/New_York:20241202T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241202T141500 LOCATION:Gates Hillman 9115 SUMMARY:5th Year Master Thesis Presentation - Prakruthi Pradeep CLASS:PUBLIC DESCRIPTION:Speaker: PRAKRUTHI PRADEEP\, Master's Student\, Computer Scienc e\nDepartment\, Carnegie Mellon University\n\nTalk Title: An Image-Based A utomated Assessment of Cutaneous\nDermatomyositis Severity\n\nDermatomyosi tis is a rare autoimmune disease characterized by chronic\nmuscle inflamma tion\, weakness\, and skin rashes. Cutaneous\nDermatomyositis (CDM)\, the skin manifestation of the disease\,\ntypically presents as purple or red r ashes on the eyelids\, joints\,\nknuckles\, and other areas\; while there is no cure\, treatment can\nalleviate symptoms\, and monitoring disease pr ogression is crucial.\n\nThis study introduces a novel image-based approac h for assessing CDM\nseverity\, aiming to create an objective\, predictive model based on\ndermatological images\, with expert assessments of the Cu taneous\nDermatomyositis Activity Score as the gold standard. Through our\ ncollaboration with researchers at the University of Pittsburgh Medical\nC enter\, we analyze a dataset of high-resolution in-clinic hand images\nfro m 27 Dermatomyositis (DM) patients. Key clinical features\,\nincluding the extent\, intensity and texture of the rash\, are analyzed\nalongside CNN- based image features\, enabling a comprehensive\nassessment of disease sev erity. We evaluate multiple state-of-the-art\nimage classification models\ , fine-tuning them on our dataset to\noptimize performance. Our approach i ncludes utilizing semantic image\nsegmentation to accurately delineate ras h boundaries\, with significant\nimprovements achieved through this integr ation.\n\nOur study lays the groundwork for the use of patient-taken image s for\nremote monitoring\, demonstrating the potential for patients to tra ck\ntheir condition at home. Finally\, we evaluate the potential for\nself -learning algorithms\, incorporating new images and expert feedback\nto im prove our model’s predictive power. By combining clinical\ninsights with advanced image analysis\, this work contributes to\nimproved automated as sessment of CDM and better monitoring of disease\nprogression. \n\nThesis Committee\n\nArtur W. Dubrawski (Chair)\n\nBhiksha Raj \n\nAdditional In formation\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bc29f DTSTART;TZID=America/New_York:20241202T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241202T120000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Justin Alexander Whitehouse CLASS:PUBLIC DESCRIPTION:Speaker: JUSTIN ALEXANDER WHITEHOUSE\, Ph.D. Candidate \, Compu ter\nScience Department\, Carnegie Mellon University\n\nTalk Title: Modern Martingale Methods: Theory and Applications\n\nMartingale concentration i s at the heart of sequential statistical\ninference. Due to their time-uni form concentration of measure\nproperties\, martingales allow researchers to perform inference on\nhighly correlated data as it is adaptively collec ted over time. Many\nstate-of-the-art results in areas such as differentia l privacy\,\nmulti-armed bandit optimization\, causal inference\, and onli ne learning\nboil down to (a) finding an appropriate\, problem-dependent m artingale\nand (b) carefully bounding its growth. Despite the important ro les\nmartingales and time-uniform concentration of measure play in modern\ nstatistical tasks\, applications of martingale concentration are\ntypical ly ad-hoc. Often\, poorly chosen martingale concentration\ninequalities ar e applied\, which results in suboptimal\, even vacuous\nrates in sequentia l estimation problems. \n\nThe focus of this thesis is twofold. In the fi rst part of this thesis\,\nwe provide simple yet powerful frameworks for c onstructing\ntime-uniform martingale concentration inequalities in univari ate\,\nmultivariate\, and even sometimes infinite-dimensional settings. Th e\ninequalities contained herein can be applied to processes with both\nli ght-tailed and heavy-tailed increments\, and follow from simple\ngeometric arguments. The second part of this thesis is focused on\napplying marting ale methods and time-uniform martingale concentration\nto practically rele vant data science tasks. In particular\, we show\nthat\, by appropriately applying martingale concentration\, one can\nobtain salient improvements o ver the state-of-the-art in both\ndifferentially private machine learning and kernel bandit optimization\ntasks. In sum\, the hope is to give a read er a start to finish view of\nhow to derive and apply time-uniform marting ale concentration in\nmodern statistical research. \n\nThesis Committee\n \nZhiwei Steven Wu (Co-Chair)\n\nAaditya Ramdas (Co-Chair)\n\nAarti Singh\ n\nCsaba Szepesvari (University of Alberta)\n\nEmilie Kaufmann (Universit é de Lille)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bc757 DTSTART;TZID=America/New_York:20241127T000000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241130T000000 SUMMARY:Thanksgiving Recess - 2024 CLASS:PUBLIC DESCRIPTION:Speaker: No Classes during Thanksgiving Break\, Classes/Normal Business\nHours Resume Monday\, December 2\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bc9a5 DTSTART;TZID=America/New_York:20241126T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241126T133000 URL:https://www.cs.cmu.edu/~theorylunch LOCATION:Gates Hillman 7101 (special time / location) SUMMARY:Theory Lunch Seminar - Rajesh Jayaram CLASS:PUBLIC DESCRIPTION:Speaker: RAJESH JAYARAM\, Research Scientist\, Google NYC\n\nTa lk Title: Nearest Neighbor Search for the Earth Mover’s Distance\n\n In this talk\, we discuss the first improvement in approximation for\nneares t neighbor search under the Earth Mover’s Distance (EMD) in\nover a deca de. Given two sets of s vectors A\,B in high dimensional\nspace (Rd)\, the EMD between A and B is the minimum cost of a perfect\nmatching between th e vectors in A and B where the edge weights are\ngiven by the distances in Euclidean space. EMD is a classic metric in\ncomputer science (dating bac k over 100 years to the Hungarian\nalgorithm of Jacobi)\, and a standard d istance between two sets of\npoints. In nearest neighbor search\, one has a collection of n such\nsets A1\,…\,An\, which one pre-processes so that given a query set Q\,\none can quickly return a set Ai whose distance (un der EMD) is within a\nC-factor of the nearest neighbor to Q. To date\, the only algorithm\nwith sublinear O(neps) query time was given by Andoni\, I ndyk\, and\nKrauthgamer ([AIK\, SODA 08])\, who designed a (data-independe nt)\nlocality sensitive hash function (LSH) for EMD with approximation\nO( log2 s). In this work\, we improve this approximation to Õ(log s) in\nthe same runtime by designing the first data-dependent LSH functions\nfor EMD . The talk will discuss the main techniques behind sublinear\nalgorithms f or EMD\, including the tree embeddings techniques of\n[AIK’08] and [Chen \, Jayaram\, Levi\, Waingarten STOC ‘22]\, as well\nas the key insights needed to glue them together into an improved LSH\nfor EMD. \n\nJoint wor k with Erik Waingarten and Tian Zhang (STOC ‘24)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bcd47 DTSTART;TZID=America/New_York:20241125T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241125T173000 URL:https://db.cs.cmu.edu/events/building-blocks-greptimedb-ruihang-xia LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Ruihang Xia CLASS:PUBLIC DESCRIPTION:Speaker: RUIHANG XIA\, Apache DataFusion PMC Member\,\, and Dev eloper\,\nGreptime\n\nTalk Title: Implement\, Integrate and Extend a Query Engine\n\nGreptimeDB uses Apache DataFusion and many other common buildin g\nblocks in its implementation. This talk will focus on managing the\nque ry aspect of a (time-series) database across various parts. We have\nexten ded DataFusion to implemenet PromQL\, add grammar candies to SQL\,\ncooper ate with external secondary indexes and write domain-specific\noptimizer r ules etc. Each of above is extended in a different stage of\nquery executi on. In addition to new features\, we'll also discuss using\nDataFusion and Arrow as frameworks for implementing and optimizing\ncustom plans. \n\n — \n\nRuihang Xia is an Apache DataFusion PMC Member\, developer at\nGr eptime. \n\nThis talk is part of the Database Building Blocks Seminar Ser ies\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bd096 DTSTART;TZID=America/New_York:20241125T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241125T163000 URL:https://www.ri.cmu.edu/event/high-resolution-cloth-simulation-in-millis econds-efficient-gpu-cloth-simulation-with-non-distance-barriers-and-subsp ace-reuse-interactions/ LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Yin Yang CLASS:PUBLIC DESCRIPTION:Speaker: YIN YANG\, Associate Professor\, Kahlert School of Com puting\,\nUniversity of Utah\n\nTalk Title: High-resolution cloth simulati on in milliseconds:\nEfficient GPU Cloth Simulation with Non-distance Barr iers and Subspace\nReuse Interactions\n\nWe show how to push the performan ce of high-resolution cloth\nsimulation\, making the simulation interactiv e (in milliseconds) for\nmodels with one million degrees of freedom (DOFs) while keeping every\ntriangle untangled. The guarantee of being penetrati on-free is\ninspired by the interior-point method\, which converts the ine quality\nconstraints to barrier potentials. Nevertheless\, we propose a ma jor\noverhaul of this modality by defining a novel and simple barrier\nfor mulation which does not depend on the distance between mesh\nprimitives. S uch a non-distance barrier model allows a new way to\nintegrate collision detection into the simulation pipeline. Another\ncontributor to the perfor mance boost comes from the so-called subspace\nreuse strategy. This is bas ed on the observation that low-frequency\nstrain vibrations are near ortho gonal to the deformation induced by\ncollisions or self-collisions\, often of high frequency. Subspace reuse\nthen takes care of low-frequency resid uals\, while high-frequency\nresiduals can also be effectively smoothed by GPU-based iterative\nsolvers. We show that our method outperforms existin g fast cloth\nsimulators by nearly one order while keeping the entire simu lation\npenetration-free and producing high-equality animations of\nhigh-r esolution models. \n\n— \n\nDr. Yin Yang is currently an Associate Pro fessor with the Kahlert\nSchool of Computing at the University of Utah. Be fore joining the U\,\nhe was a faculty member at Clemson University and Un iversity of New\nMexico. He received Ph.D. degree of Computer Science from The\nUniversity of Texas\, Dallas in 2013 (the awardee of David Daniel\nF ellowship Prize). He was a Research/Teaching Assistant at UT Dallas\nas we ll as UT Southwestern Medical Center. His research mainly focuses\non real -time physics-based computer graphics\, animation and simulation\nwith a s trong emphasis on interdisciplinarity. He was a Research\nIntern in Micros oft Research Asia in 2012. He received NSF CRII (2015)\nand CAREER (2019) awards.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bd4bd DTSTART;TZID=America/New_York:20241122T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241122T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Gates Hillman 8102 and Zoom SUMMARY:AI Institute for Societal Decision Making Seminar - Nikhil Garg CLASS:PUBLIC DESCRIPTION:Speaker: NIKHIL GARG\, Assistant Professor\, Operations Researc h and\nInformation Engineering (ORIE)\, Cornell Tech\n\nTalk Title: Engine ering Societal Systems for Efficiency and Equity\n\nModern societal and go vernmental decisions — what gets built and\nmaintained\, who gets access to what school\, where people interview and\nwork — are driven by compu tational systems. While much of the focus\nis on learning from data\, many of the real-world challenges appear\nbefore and after the learning algori thm: can we collect accurate\,\nunbiased\, up-to-date data\; and then how should we use individually\naccurate/noisy predictions to make efficient\, equitable societal\ndecisions\, given resource constraints and population (market and\nstrategic) effects? \n\nIn this talk\, I'll discuss our the oretical\, empirical\, and deployment\nwork in tackling these challenges\, in the context of crowdsourcing\nreporting and recommendation-matching sy stems\, in collaboration with\nNYC government agencies\, a university admi ssions team\, and a platform\nto help discharge patients to long-term care facilities. \n\n— \n\nNikhil Garg is an Assistant Professor of Operat ions Research and\nInformation Engineering at Cornell Tech as part of the Jacobs\nInstitute. He uses algorithms\, data science\, and mechanism desig n\napproaches to study democracy\, markets\, and societal systems at large .\nNikhil has received the NSF CAREER\, INFORMS George Dantzig\nDissertati on Award\, an honorable mention for the ACM SIGecom\ndissertation award\, several other best paper awards\, and Forbes 30\nunder 30 for Science. He received his PhD from Stanford University and\nhas spent considerable coll aborating with government agencies and\nnon-profits. \n\nREGISTER → to join the Zoom Meeting \n\nIn Person and Zoom Participation.  See announc ement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bd8ab DTSTART;TZID=America/New_York:20241122T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241122T113000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Seminar - Xinyu (Norah) Tan CLASS:PUBLIC DESCRIPTION:Speaker: XINYU (NORAH) TAN\, Ph.D. Student\, Department of Math ematics\,\nMassachusetts Institute of Technology\n\nTalk Title: Incompress ibility and spectral gaps of random circuits\n\nPick your favorite n-qubit gate set that is universal (e.g.\, the set\nof all 2-qubit unitaries)\, a nd then sample L gates independently and\nuniformly at random from this ge nerating set. Can you approximate the\nproduct of these L gates using fewe r than L 2-qubit gates? In other\nwords\, can you compress this random cir cuit? This is a fundamental\nquestion related to the growth of (robust) qu antum circuit complexity\nin random circuits. We show for up to exponentia lly large L that\nrandom quantum circuits cannot be implemented with fewer than\nL/poly(n) gates.   \n\nWe prove this by showing that the 2-local random circuits form\nmultiplicative-error unitary t-designs with O(t*poly (n)) gates. More\nconcretely\, we prove a t-independent spectral gap for t he t-th moment\noperator of the random walk by reducing random quantum cir cuits to a\nmore structured walk: a modification of the “PFC ensemble” from\n[MPSY24] together with an expander on the alternating group due to\ nKassabov [Kas07a]\, for which we give an efficient implementation using\n reversible circuits. \n\nBased on joint work with Chi-Fang Chen\, Jeongwa n Haah\, Jonas\nHaferkamp\, Yunchao Liu\, and Tony Metger\n\nReference Pap er\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bdc4f DTSTART;TZID=America/New_York:20241121T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241121T180000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Katayanagi Distinguished Lecture - Noam Brown CLASS:PUBLIC DESCRIPTION:Speaker: NOAM BROWN\, Research Scientist\, OpenAI\n\nTalk Title : Learning to Reason with LLMs\n\nLarge language models (LLMs) have demons trated remarkable capabilities\nin generating coherent text and completing various natural language\ntasks. Nevertheless\, their ability to perform complex\, general\nreasoning has remained limited. \n\nIn this talk\, I w ill describe OpenAI's new o1 model\, an LLM trained\nvia reinforcement lea rning to generate a hidden chain of thought\nbefore its response. We have found that the performance of o1\nconsistently improves with more reinforc ement learning compute and\nwith more inference compute. o1 surpasses prev ious state-of-the-art\nmodels in a variety of benchmarks that require reas oning\, including\nmathematics competitions\, programming contests\, and a dvanced science\nquestion sets. I will discuss the implications of scaling this\nparadigm even further.\n\n—\n\nNoam Brown is a research scientist at OpenAI investigating reasoning\nand multi-agent AI. He co-created Libr atus and Pluribus\, the first AIs\nto defeat top humans in two-player no-l imit poker and multiplayer\nno-limit poker\, respectively\, and Cicero\, t he first AI to achieve\nhuman-level performance in the natural language st rategy game\nDiplomacy. He has received the Marvin Minsky Medal for Outsta nding\nAchievements in AI\, was named one of MIT Tech Review's 35 Innovato rs\nUnder 35\, and his work on Pluribus was named by Science as one of the \ntop 10 scientific breakthroughs of 2019. Noam received his PhD from\nCar negie Mellon University. \n\nAbout the Lecture:  The Katayanagi Lectures recognize the best and\nthe brightest in the field of computer science an d are presented by\nthe School of Computer Science at Carnegie Mellon Univ ersity in close\ncooperation with the Tokyo University of Technology (TUT) . The\nlectures recognize both senior and junior talent.  The series were \nestablished through a gift from Japanese entrepreneur and education\nadv ocate\, Mr. Koh Katayanagi\, who founded TUT and other technical\ninstitut ions in Japan over many multiple decades.  We are delighted\nto have TUT as partners.  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91be0ab DTSTART;TZID=America/New_York:20241121T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241121T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - John Kolesar CLASS:PUBLIC DESCRIPTION:Speaker: JOHN KOLESAR\, Ph.D. Student\, Department of Computer Science\,\nYale University\n\nTalk Title: Zero-Knowledge Proofs for SMT Th eorems and Regular\nExpression Equivalence\n\nZero-knowledge (ZK) protocol s allow software developers to provide\nproofs of their programs' correctn ess to other parties without\nrevealing the programs themselves.  Most pr ior work in the domain of\nZK verification has focused on the encoding of witnesses for\nsatisfiability proofs\, but the task of encoding proof tree s for\ndemonstrating that a formula is unsatisfiable has received\ncompara tively little attention.  We introduce ZKSMT\, the first\nzero-knowledge protocol designed for encoding unsatisfiability proofs\ngenerated by SMT s olvers.  We design a virtual machine (VM) structure\nfor proofs that enab les the prover to hide the structure of a proof\nwhile still manipulating complex ASTs for terms and formulas.  We use\nZKSMT to validate SMT proof s involving equality with uninterpreted\nfunctions (EUF) and linear intege r arithmetic (LIA).  In terms of\nperformance\, ZKSMT achieves a three-or der-of-magnitude improvement\nover general-purpose ZK protocols. \n\nAlon g with SMT proof validation\, we tackle another novel problem for\nZK proo fs:  the problem of checking whether two regular expressions\nare equival ent.  We introduce Crepe\, the first ZK protocol for\nencoding regular ex pression equivalence proofs.  For the proofs that\nCrepe validates\, we i ntroduce a new custom calculus of proof rules\nbased on regular expression derivatives and coinduction.  We prove\nthe calculus sound and complete\ , and we provide a decision procedure\nfor generating proofs in our format .  We test our proof generator on\na suite of thousands of equivalent reg ular expression pairs\, and we\ntest Crepe on the proofs that we generate.   The proof generator\nscales effectively\, and Crepe itself can validate large equivalence\nproofs in only a few seconds each. \n\n— \n\nJohn Kolesar is a fifth-year Ph.D. student in computer science at Yale\nUnivers ity advised by Ruzica Piskac.  His dissertation research\ncombines automa ted program verification and zero-knowledge proofs. \nHis other research interests include functional programming\, symbolic\nexecution\, automatic program repair\, and software-defined\nnetworking.  He has held research intern positions at both Microsoft\nand Amazon Web Services.  Before com ing to Yale\, he earned a\nbachelor's degree at Cornell\, where he did res earch work with Nate\nFoster. \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91be544 DTSTART;TZID=America/New_York:20241121T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241121T163000 URL:https://brain.andrew.cmu.edu/seminar LOCATION:Baker Hall 336B and Zoom SUMMARY:brAIn Seminar - Tyler Bonnen CLASS:PUBLIC DESCRIPTION:Speaker: TYLER BONNEN\, Presidential Postdoctoral Fellow\, Berk eley AI\nResearch Lab\, Electrical Engineering and Computer Sciences Depar tment\,\nUniversity of California\, Berkeley\n\nTalk Title: A science of f ailure: understanding the gap between humans\nand vision models\n\nDeep le arning frameworks have been widely adopted to study primate\nvision. This includes methods that successfully predict neural\nresponses throughout ve ntral temporal cortex (VTC). Nonetheless\, there\nis a persistent gap betw een these 'VTC-like' models and human\nperformance in many visual tasks (e .g. 3D shape inferences). This is\ngenerally interpreted as a shortcoming of these 'VTC-like' models. In\nthis talk\, I'll suggest that this interpr etation is misguided\; this\nhuman-model gap challenges common assumptions about VTC function\,\nwhile providing tools that can be used to better un derstand the\ndownstream neural structures and algorithms that enable visu al\nperception. I'll support this claim through a series of modeling\nexpe riments that incorporate behavioral\, electrophysiological\, and\nlesion d ata. These results provide clear architectural and algorithmic\nconstraint s for models of primate vision—a departure from prevailing\napproaches i n both the neurosciences and computer vision.\n\n—\n\nTyler Bonnen compl eted his PhD at Stanford University co-advised by\nAnthony Wagner and Dan Yamins. He is currently a postdoc within\nBerkeley's AI Research (BAIR) la bs\, co-advised by Alexei Efros\, Angjoo\nKanazawa\, and Jitendra Malik.  \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91be8fe DTSTART;TZID=America/New_York:20241121T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241121T160000 URL:https://aco.math.cmu.edu/abs-24-25/nov21.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Anton Dochtermann CLASS:PUBLIC DESCRIPTION:Speaker: ANTON DOCHTERMANN\, Associate Professor\, Math Departm ent\,\nCollege of Science and Engineering\, Texas State University\n\nTalk Title: Parking functions and h-vectors of matroids\n\nThe h-vector of a m atroid M is an important invariant related to the\nindependence complex of M\, which can also be covered as an evaluation\nof its Tutte polynomial. A well-known conjecture of Stanley posits\nthat the h-vector of a matroid is a \"pure O-sequence\"\, meaning that it\ncan be recovered by counting f aces of a pure multicomplex. Merino has\nestablished Stanley's conjecture for the case of cographic matroids\nvia a connection to chip-firing on gra phs and the concept of a\nG-parking function. Inspired by these constructi ons\, we introduce the\nnotion of a cycle system for a matroid M . This le ads to a collection\nof integer sequences that we call \"parking functions \" for M\, which we\nshow are in bijection with the set of bases of M. We study maximal\ncoparking functions\, and also how cycle systems behave und er deletion\nand contraction. This leads to a proof of Stanley’s conject ure for\nthe case of matroids that admit cycle systems (which for instance \nincludes graphic matroids of K_{3\,3}-minor free graphs). \n\nThis is j oint work with Scott Cory\, Solis McClain\, and David\nPerkinson.\n\n4:00 pm →  Jane Street sponsored Tea & COokies followin Wean 7220\n(bring yo ur own mug).  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bec9e DTSTART;TZID=America/New_York:20241121T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241121T130000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate - Babu Pillai CLASS:PUBLIC DESCRIPTION:Speaker: BABU PILLAI\, Research Scientist\, Intel Labs\n\nTalk Title: Out-of-the-Box Thinking in Systems Design and\nImplementation\n\nAs systems scientists\, we are generally taught to think about and\ndesign s ystems to be broadly applicable and maximally useful for a\nwide range of applications.  While this is usually the right mindset\nto have\, sometim es it is necessary to break with conventional\napproaches and any hope of broad applicability in order to find\npractical solutions to hard real-wor ld problems.  In this talk\, we\nwill look at two case studies of systems that buck convention to\nachieve target performance goals.  We will firs t take a deep dive\ninto the Google File System (GFS)\, a seminal work tha t defied many of\nthe ideas on how scalable and reliable file systems shou ld be built. \nGFS is a topic we occasionally cover in lectures in 15-440 /640\n(Distributed Systems).  We will then discuss a research prototype f or\nrapid deployment of custom virtual machines in edge infrastructure. \ nBoth of these systems demonstrate that with a clear focus on the\ncritica l use case\, relaxation of generality\, and some clever\nengineering\, we can build systems to meet seemingly impossible\ngoals.  \n\n— \n\nDr. Padmanabhan (Babu) Pillai has been a research scientist at Intel\nLabs fo r the past 21 years.  Babu received his M.S.  and Ph.D. in\ncomputer sci ence from University of Michigan (in 1999 and 2004)\, and\nholds a B.S. in electrical and computer engineering from Carnegie\nMellon University (rec eived in 1996).  His core research interests\nare in distributed systems\ , edge computing\, and low-power systems. \nHis recent work includes the use of edge computing to perform\nreal-time understanding of the world and the development of\ndistributed\, performant\, parallel array frameworks for Python.  Babu\nhas maintained close ties to academic research\, and h as approximately\n100 publications\, including three papers awarded Best P aper / Runner\nup\, and one paper that received the SIGMOBILE Test-of-Time award in\n2023.  \n\nFor the past 10 years\, he has been co-instructing the Spring semester\n15-440/640 (Distributed Systems) courses with Profes sor Mahadev\nSatyanarayanan.  He hopes to apply his broad research and sy stems\nexperience to teach core systems principles and systems building\ns kills at the undergraduate and graduate levels.  Outside of work\nBabu is a husband and father of two\, and aspires to be an avid\ngardener and cyc list. Faculty Host: Mark Stehlik \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bf121 DTSTART;TZID=America/New_York:20241121T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241121T130000 URL:https://www.pdl.cmu.edu/SDI/2024/112124.html LOCATION:Newell-Simon 3305 and Zoom SUMMARY:SDI Seminar - Tianyin Xu CLASS:PUBLIC DESCRIPTION:Speaker: TIANYIN XU\, Assistant Professor\, Department of Compu ter\nScience\, University of Illinois at Urbana-Champaign\n\nTalk Title: S oftware Reliability in Emerging Cloud Computing Paradigms\n\nCloud system reliability has been a grand challenge in the past decade\ndue to prevalen t\, inevitable hardware faults\, software bugs\, and\nmisconfigurations. E merging computing paradigms such as microservices\nand serverless computin g further expand reliability challenges by\nsignificantly increasing the c omplexity of system interactions and\ndependencies\, with new failure doma ins. In this talk\, I will present\nour work on improving the reliability of cloud infrastructure systems\nsuch as Kubernetes and cloud-native appli cations running atop them via\nsoftware testing\, model checking\, and for mal verification.\nSpecifically\, I will present push-button testing techn iques for\nsystematically checking safety and liveness properties of exist ing\nsystem components and discuss the promises of building new components \nwith formally verified correctness properties towards a vision of\ntruly reliable cloud infrastructures and systems.\n\n—\n\nTianyin Xu is an As sistant Professor of Computer Science at the\nUniversity of Illinois at Ur bana-Champaign (UIUC). His research\nfocuses on building reliable computer systems that empower\nnext-generation cloud and datacenter computing. He has been on the\nUIUC List of Teachers Ranked as Excellent for eight times . His work\nreceives Jay Lepreau Best Paper Awards at OSDI 2024 and 2016\, a Best\nPaper Award at ASPLOS 2020\, two SIGSOFT Distinguished Paper Awar ds at\nISSTA 2021 and FSE 2021\, a Gilles Muller Best Artifact Award at\nE uroSys 2023\, and a CACM Research Highlight. He is also a recipient of\nth e C.W. Gear Outstanding Junior Faculty Award\, a Dean's Award for\nExcelle nce in Research\, NSF CAREER Award\, an Intel Rising Star Faculty\nAward\, and a Facebook Distributed Systems Research Award. \n\nMore information.  \n\nFaculty Host:  Dimitrios Skarlatos \n\nIn Person and Zoom Particip ation. See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bf582 DTSTART;TZID=America/New_York:20241120T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241120T150000 URL:https://db.cs.cmu.edu/events/the-rise-of-data-streaming-platforms LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Jun Rao CLASS:PUBLIC DESCRIPTION:Speaker: JUN RAO\, Co-founder\, COnfluent\n\nTalk Title: The Ri se of Data Streaming Platforms\n\nApache Kafka and Apache Flink are poweri ng a new category of data\ninfrastructure called data streaming platform ( DSP). This provides an\nopportunity for each enterprise to take actions on what’s happening\nin its business in real time. I will first provide an overview of DSP.\nDSP has both similarities and differences to database s ystems. I will\nshow how existing database technologies can be used in thi s new\nplatform and some of the unique problems that DSP needs to solve. I \nwill conclude with some of the new opportunities on DSP. \n\n— \n\nJ un Rao is a co-founder of Confluent\, a company that provides a data\nstre aming platform based on Apache Kafka. Before Confluent\, Jun Rao\nwas a se nior staff engineer at LinkedIn where he led the development\nof Kafka. Be fore LinkedIn\, Jun Rao was a researcher at IBM’s Almaden\nresearch cent er\, where he conducted research on database and\ndistributed systems. Jun Rao is a committer of Apache Kafka and Apache\nCassandra. \n\nZoom Parti cipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bf91a DTSTART;TZID=America/New_York:20241120T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241120T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Lunch Seminar - Jeff Xu CLASS:PUBLIC DESCRIPTION:Speaker: JEFF XU\, Ph.D. Student\, Computer Science Department\ , Carnegie\nMellon University\n\nTalk Title: SoS Lower Bounds for Coloring Random Graphs\n\nColoring for random graph from G(n\,1/2) is a classic ex ample\nexhibiting an Information v. Computation gap: it has chromatic numb er\nof θ(n/log n) w.p. 1-o(1) while the best efficiently certifiable\nlow er bound is θ(√{n}) achieved by basic SDP. Unlike the related\nquestion of independence number where we have essentially matching\nhardness and a lgorithms in various restricted models\, our hardness\nevidence for colori ng was surprisingly limited to the basic SDP. I\nwill talk about lower bou nds against Sum-of-Squares algorithms showing\nthat the order root-n bound is essentially the best we can hope for in\npoly time. \n\nIn the past f ew years\, low-degree polynomial method has become a\npopular heuristic in predicting hardness for average-case problems\,\nbut it also may be fairl y limited in various problem-specific\napplications: for example\, colorin g on dense random graph is a problem\nwhose natural distinguishing variant is “easy”. Against the\nconventional wisdom that some form of low-deg ree hardness is a\nnecessary precursor for SoS lower bounds\, we circumven t the lack of\nlow-degree hardness\, and design our SoS lower bounds via a systematic\ndeparture from the “pseudo-calibration” recipe. \n\nThis is based on a joint work with Aaron Potechin (UChicago).\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91bfcce DTSTART;TZID=America/New_York:20241119T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241119T163000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Machine Learning / Duolingo Seminar - Johan Ugander CLASS:PUBLIC DESCRIPTION:Speaker: JOHAN UGANDER\, Associate Professor\, Department of Ma nagement\nScience & Engineering\, Institute for Computational & Mathematic al\nEngineering\, School of Engineering\, Stanford University\n\nTalk Titl e: Bridging-based fact-checking moderates the diffusion of\nfalse informat ion on social media\n\nSocial networks scaffold the diffusion of informati on on social media.\nMuch attention has been given to the spread of true v s. false content\non social media\, including the structural differences b etween their\ndiffusion patterns. However\, much less is known about how p latform\ninterventions on false content alter the diffusion of such conten t. \n\nIn this work\, we estimate the causal effects of a novel fact-chec king\nfeature\, Community Notes\, adopted by Twitter (now X) to solicit an d\nvet crowd-sourced fact-checking notes for false content. An important\n aspect of this feature is its use of a bridging-based decision\nalgorithm whereby fact-checking notes are shown only if they are seen\nas broadly in formative and helpful by users from across the political\nspectrum. To est imate the causal effect of bridging-based\nfact-checking\, we gather detai led time series data for 40\,000 posts\nfor which notes have been proposed and use synthetic control methods\nto produce counterfactual estimates of a range of diffusion-based\noutcomes. We find that attaching fact-checkin g notes significantly\nreduced the reach of and engagement with false cont ent. In reducing\nreach\, we observe that diffusion trees for fact-checked content are\nless deep\, but not less broad\, than synthetic control esti mates for\nnon-fact-checked content with similar reach. \n\nThis finding contrasts notably with differences between false vs. true\ncontent\, where false information diffuses farther\, but with structural\npatterns that a re otherwise indistinguishable from those of true\ninformation\, condition al on reach. \n\n— \n\nJohan Ugander is an Associate Professor at Stan ford University in the\nDepartment of Management Science & Engineering\, w ithin the School of\nEngineering. His research develops algorithmic and st atistical\nframeworks for analyzing social networks\, social systems\, and other\nlarge-scale social and behavioral data. \n\nPrior to joining the Stanford faculty he was a postdoctoral researcher\nat Microsoft Research R edmond 2014-2015 and held an affiliation with\nthe Facebook Data Science t eam 2010-2014. He obtained his Ph.D. in\nApplied Mathematics from Cornell University in 2014. His awards\ninclude a NSF CAREER Award\, a Young Inves tigator Award from the Army\nResearch Office (ARO)\, several Best Paper Aw ards\, and the 2016 Eugene\nL. Grant Undergraduate Teaching Award from the Department of\nManagement Science & Engineering. \n\nFaculty Host:  Nih ar Shah \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c01bd DTSTART;TZID=America/New_York:20241119T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241119T140000 URL:https://csd.cmu.edu/calendar/master-of-science-in-computer-science-thes is-defense-hongwei-tu LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:MSCS Thesis Defense - Hongwei Tu CLASS:PUBLIC DESCRIPTION:Speaker: HONGWEI TU\, Master's Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Learning Genome-Wide In teractions of Intrinsically\nDisordered Proteins with DNA Using U-DisCo\n\ nProteins are essential regulators of cellular processes. Intrinsically\nd isordered proteins (IDPs)\, despite lacking stable tertiary structures\nun der physiological conditions\, play crucial yet often underexplored\nroles in biological processes. With recent experimental advances like\nDisP-seq for probing IDP-DNA binding\, there is a pressing need for\nefficient\, i nterpretable computational methods to identify sequence\ndeterminants of I DP-DNA interactions and analyze their cooperative\neffects on gene regulat ion. \n\nTo address this\, we develop U-DisCo\, a novel deep learning mod el that\npredicts base-resolution IDP-DNA binding profiles directly from D NA\nsequences. Leveraging a U-Net architecture\, U-DisCo captures both\nlo cal base-level interactions and long-range dependencies up to 20\nkilobase s with high accuracy and computational efficiency\,\noutperforming the bas eline BPNet. By incorporating ATAC-seq data\,\nU-DisCo enables robust cros s-cell type predictions as a multimodal\nframework. U-DisCo identified key IDP-binding motifs\, revealing\ndistinct interaction patterns and coopera tive behaviors across\ndifferent IDPs. Interestingly\, we observed short-r ange interactions\nfor motifs like AP-2 and EWS-FLI1 (single GGAA motif)\, while others\nexhibited independent\, enhancer-like functions. Further an alysis\nrevealed that some IDPs favored certain strand orientations\,\nsug gesting their involvement in specific regulatory mechanisms.\nOverall\, U- DisCo is the first computational approach to explore\nmultiple IDPs within a single cell type\, offering a versatile\nframework for studying IDP-med iated gene regulation and genome-wide\nregulatory elements. \n\nThesis Co mmittee\n\nJian Ma (Chair)\n\nLei Li\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c0626 DTSTART;TZID=America/New_York:20241119T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241119T113000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-nuno-sabino LOCATION:Gates Hillman 7101 SUMMARY:Doctoral Thesis Proposal - Nuno Sabino CLASS:PUBLIC DESCRIPTION:Speaker: NUNO SABINO\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Improving Code-Injection Vulnerability Detection &\nConfirmation in JS Programs via Program Analysi s\n\nApplications written in JavaScript are often vulnerable to a range of \nsecurity threats. On the frontend\, DOM-based Cross-Site Scripting\n(DOM -XSS) allows attackers to inject malicious JavaScript code into a\nwebpage . On the backend\, arbitrary code execution (ACE) and arbitrary\ncommand i njection (ACI) enable attackers to execute arbitrary code or\ncommands on the server. Exploiting such vulnerabilities can lead to\nsevere consequenc es\, including unauthorized access to sensitive data\nand even full system compromise.\n\nEach potential flow identified by these tools traces a pro gram path\nwhere attacker-controlled input\, such as a URL\, reaches a sen sitive\nfunction that may lead to arbitrary code execution. DTA requires\n finding a concrete input that demonstrates a potential flow in the\ntarget application\, but prior work fails to thoroughly explore program\npaths. In the backend\, these tools miss ACI and ACE that require\ninputs with co mplex structure. We develop a novel type- and\nstructure-aware fuzzing tec hnique to explore Node.js packages\, and an\nenumerator to synthesize synt actically valid payloads for ACE\nvulnerabilities. Incorporating these com ponents on prior work\nNodeMedic led to finding 2257 potential flows and c onfirm\nvulnerabilities in 766 Node.js packages.\n\nA unique challenge in exploring frontend code is that program behavior\nmay depend on user actio ns on the webpage. To address this\, we develop\na fuzzer to interact with the target webpage and evaluated it against\n43\,436 popular pages. Furth ermore\, we found that including optional\nGET parameters in the target UR L uncovers significantly more DOM-XSS\nvulnerabilities. This led us to use dynamic symbolic execution to\nautomatically synthesize GET parameters sa tisfying program\nconstraints. Compared to our replication of prior work D OMsday\, the\nfuzzer increases potential DOM-XSS flows found by 37% and co nfirms 57%\nmore vulnerabilities.\n\nFinally\, we find that non-exploitabl e potential flows may still hint\ntowards real vulnerabilities that requir e additional steps to confirm\,\nsuch as bypassing sanitization measures a nd extending the attacker’s\ncapabilities by executing other program par ts. Thus\, we propose the\ndesign and implementation of exploration strate gies that efficiently\nexplores the program to discover an exploitable pat h\, using\ninformation from a given potential flow that we assume to have found\nalready. \n\nThesis Committee\n\nLimin Jia (Chair)\n\nPedro Adão (Co-chair\, Instituto Superior Técnico)\n\nRui Maranhão (Co-chair\, Univ ersidade do Porto)\n\nLujo Bauer\n\nRuben Martins\n\nJosé Fragoso (Instit uto Superior Técnico)\n\nCristian-Alexandru Staicu (CISPA Helmholtz Cente r for Information\nSecurity)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c0b58 DTSTART;TZID=America/New_York:20241118T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241118T173000 URL:https://db.cs.cmu.edu/events/building-blocks-glaredb-sean-smith LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Sean Smith CLASS:PUBLIC DESCRIPTION:Speaker: SEAN SMITH\, Founder\, GlareDB\n\nTalk Title: Biting t he Bullet: Rebuilding GlareDB from the Ground Up\n\nGlareDB is a database system enabling querying across a variety of\ndata sources\, including Sno wflake\, Postgres\, and more. Building on top\nof DataFusion let us get to an early product very quickly. But not\neverything is sunshine and roses. In this talk\, we'll explore some of\nthe limitations we hit with DataFus ion\, and how we plan to address\nthose in our upcoming engine Bullet. \n \n— \n\nSean Smith is Founder at GlareDB\n\nFaculty Hosts:  Jignesh Pa tel\, Andy Pavlo \n\nThis talk is part of the Database Building Blocks Se minar Series\n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c0e9e DTSTART;TZID=America/New_York:20241118T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241118T163000 URL:https://csd.cmu.edu/calendar/master-of-science-in-computer-science-thes is-defense-ankit-gupta LOCATION:Newell-Simon 4305 SUMMARY:MSCS Thesis Defense - Ankit Gupta CLASS:PUBLIC DESCRIPTION:Speaker: ANKIT GUPTA\, Master's Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Analyzing Multimodal M achine Learning Model Performance\nand Evaluation Metrics for Medical Repo rt Generation\n\nAs a result of recent advancements in foundation models\, including\nlarge vision-language models\, several researchers have explor ed\nmethods of combining multiple modalities of data as inputs for visual\ nquestion answering. One key application of visual question answering\nin the context of the healthcare domain is automated medical report\ngenerati on\, where x-ray images and text-based symptom data for a\npatient might b e provided as inputs\, with the intention of generating\na relevant medica l report as an output. However\, very few studies\nanalyze the performance of these models alongside uni-modal\nencoder-decoder models\, and even fe wer compare the performance of\nthese multimodal models depending on wheth er they are provided symptom\ninformation as an input. Furthermore\, past studies often use simple\nevaluation metrics that look at n-gram overlaps\ , such as BLEU and\nROUGE scores\, which are not effective for generative foundation models\nthat can generate different sentences with the same sem antic meaning.\n\nIn this paper\, we present two main contributions. First \, we compare\nthe performance of a variety of approaches for generating m edical\nreports on a dataset of Chest X-Ray medical reports\, including an \nencoder-decoder model\, a multimodal model without symptom data\, and a\ nmultimodal model with symptom data. Second\, we design a new metric for\n evaluating the similarity between generated and reference medical\nreports using medical term transformation\, sentence embeddings\, and\ncosine vec tor similarity. Our results show that multimodal approaches\nto medical re port generation far outperform encoder-decoder\napproaches\, and providing symptom data slightly improves accuracy for\ngenerated medical reports. W e also find that our evaluation metric\nmore closely measures similarity b etween generated and reference\nmedical reports than standard techniques\, as evidenced by both\nquantitative and qualitative case-study comparisons .\n\nThis research pushes the frontier of medical report generation by\nfu rther reinforcing the accuracy benefits of using multimodal models\nwith s ymptom inputs and introducing a more comprehensive\, customized\nscoring m etric for evaluating generated medical reports. \n\nThesis Committee\n\nM in Xu (Chair)\n\nMartin Zhang\n\nBryan Wilder\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c1356 DTSTART;TZID=America/New_York:20241118T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241118T163000 URL:https://www.ri.cmu.edu/event/generative-modelling-for-3d-multimodal-und erstanding-of-human-physical-interactions/ LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Srinath Sridhar CLASS:PUBLIC DESCRIPTION:Speaker: SRINATH SRIDHAR\, Assistant Professor\, Department of Computer\nScience\, Brown University\n\nTalk Title: Generative Modelling f or 3D Multimodal Understanding of\nHuman Physical Interactions\n\nGenerati ve modelling has been extremely successful in synthesizing\ntext\, images\ , and videos. Can the same machinery also help us better\nunderstand how t o physically interact with the multimodal 3D world? In\nthis talk\, I will introduce some of my group's work in answering this\nquestion. I will fir st discuss how we can enable 2D image generation\nmodels to edit images in a 3D-aware manner\, and how to generate audio\nfor muted egocentric video s. I will then zoom in specifically on hand\ninteractions by introducing ( 1) FoundHand\, a large-scale generative\nmodel for synthesizing realistic 2D hand images\, and (2) GigaHands\, a\nnew large-scale 3D hand activities dataset designed to push the\nboundary of hand interaction modeling. Fina lly\, I will conclude with\nan outlook of the future of generative modelin g for understanding 3D\nhuman interactions.  \n\n— \n\nSrinath Sridhar is an Assistant Professor of Computer Science at Brown\nUniversity\, wher e he leads the Interactive 3D Vision & Learning Lab.\nHe received his PhD at the Max Planck Institute for Informatics and\nwas subsequently a postdo ctoral researcher at Stanford. His research\ninterests are in 3D computer vision and machine learning.\nSpecifically\, his group focuses on visual u nderstanding of 3D human\nphysical interactions with applications ranging from robotics to mixed\nreality. He is a recipient of the NSF CAREER award \, a Google Research\nScholar award\, and his work received the Eurographi cs Best Paper\nHonorable Mention. He spends part of his time as a visiting academic\nat Amazon Robotics\, and has previously spent time at Microsoft \nResearch Redmond and Honda Research Institute.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c175c DTSTART;TZID=America/New_York:20241118T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241118T160000 LOCATION:Hall of Arts 160 and Zoom SUMMARY:Teaching Track Faculty Candidate - Ranysha Ware CLASS:PUBLIC DESCRIPTION:Speaker: RANYSHA WARE\, Postdoctoral Teaching Fellow\, Computer Science\nDepartment\, Carnegie Mellon University\n\nTalk Title: Teaching Demo: Data Analysis – Analyzing and Visualizing\n(15-110 Lecture)\n\n15- 110 is a broad computer science overview for non-majors with little\nprogr amming experience\, taught in Python. In this interactive\n50-minute lectu re\, we cover the practicalities of doing basic data\nanalysis in Python. After this lecture students should be able to: 1)\nPerform basic analyses of data\, including calculating statistics and\nprobabilities\, to answer simple questions\, 2) Choose an appropriate\nvisualization to create based on the number of dimensions and data\ntypes\, and 3) Create simple matplo tlib visualizations. You can see the\nlectures leading up to this one on t he course schedule here.\n\n— \n\nDr. Ranysha Ware is a Teaching Postdo ctoral Fellow at CMU passionate\nabout effective and inclusive computer sc ience education. She was\npreviously a PhD student at CMU\, advised by Jus tine Sherry and\nSrinivasan Seshan. Her research focused on building bette r tools to\nunderstand modern congestion control algorithms. Her work has received\nnumerous awards including the IRTF Applied Networking Research P rize\,\nApplied Networking Research Prize\, a Facebook Emerging Scholars A ward\,\nand a National GEM Consortium PhD Fellowship. Outside of research\ ,\nduring her time at CMU\, she was a TA for 15-441/641 (Networking and\nt he Internet)\, 15-112 (Fundamentals of Programming and CS)\, and 07-300\n( Research and Innovation in CS).  In addition\, she was an instructor\nof record for 15-112 in Summer 2023 and currently is a co-instructor\nfor 15- 110 (Principles of Computing) with Professor Michael Taylor.\nDr. Ware see ks to broaden participation in CS by combining pedagogical\nresearch with care and compassion. Outside of work\, she is an avid\nboard gamer. \n\nI n Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c1b74 DTSTART;TZID=America/New_York:20241118T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241118T150000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-long-pham LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:Doctoral Thesis Proposal - to be rescheduled CLASS:PUBLIC DESCRIPTION:Speaker: LONG PHAM \, Ph.D. Student\, Computer Science Departme nt\,\nCarnegie Mellon University\n\nTalk Title: Hybrid Resource-Bound Anal yses of Programs\n\nResource-bound analysis aims to infer symbolic bounds of worst-case\nresource usage (e.g.\, running time\, memory\, and energy) of programs as\nfunctions of program inputs. Resource analysis has numerou s\napplications\, including job scheduling in cloud computing and\nprevent ion of side-channel attacks. Various resource analysis\ntechnique have bee n developed\, and they have unique strengths and\nweaknesses that compleme nt each other. (Automatic) static resource\nanalysis\, which analyzes the source code of programs\, is sound: if it\nsuccessfully infers a cost boun d\, it is guaranteed to be a valid\nbound. However\, every static analysis technique is incomplete: there\nexists a program that the analysis techni que cannot handle. Meanwhile\,\ndata-driven analysis\, which statistically analyzes cost measurements\nobtained by running programs on many inputs\, can infer a candidate\ncost bound for any program. However\, it does not guarantee soundness\nof inference results. \n\nTo overcome limitations of individual analysis techniques\, I propose\nhybrid resource analysis\, wh ich integrates two complementary analysis\ntechniques to retain their stre ngths while mitigating their respective\nweaknesses. The user first specif ies which analysis techniques are\nused to analyze which code fragments an d quantities. Hybrid analysis\nthen performs its constituent analysis tech niques on their respective\ncode fragments and quantities. Finally\, their inference results are\ncombined into an overall cost bound. \n\nThe deve lopment of hybrid resource analysis has been driven by the\ndesire to go b eyond Automatic Amortized Resource Analysis (AARA)\, a\nstate-of-the-art t ype-based static resource analysis technique. I\nstart by proving polynomi al-time completeness of AARA. I next\nintroduce Bayesian data-driven analy sis\, which conducts Bayesian\ninference on cost measurements to infer a p osterior distribution of\nsymbolic cost bounds. I then present the first h ybrid resource\nanalysis\, Hybrid AARA\, followed by a discussion of its l imitations. To\novercome these limitations\, I introduce the second hybrid resource\nanalysis\, resource decomposition. I additionally describe Swif tlet\,\nwhich instantiates the resource-decomposition framework with AARA and\nBayesian resource analysis. Finally\, for proposed work\, my\ncollabo rators and I plan to develop data-driven-analysis for\nstatistically infer ring not only a worst-case symbolic cost bound but\nalso a worst-case inpu t generator\, which is a program generating\nworst-case program inputs of various sizes. \n\nThesis Committee\n\nJan Hoffmann (Chair)\n\nFeras Saad \n\nMatt Fredrikson\n\nNadia Polikarpova (University of California\, San D iego)\n\n \n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c207c DTSTART;TZID=America/New_York:20241118T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241118T130000 URL:https://www.cylab.cmu.edu/events/2024/11/18-seminar-olubenga.html LOCATION:Panther Hollow Room 4105\, Mehrabian Collaborative Innovation Cent er SUMMARY:CyLab Seminar - Dr. Olugbenga Moses Anubi CLASS:PUBLIC DESCRIPTION:Speaker: OLUBENGA MOSES ANUBI\, Associate Professor\, and Direc tor\,\nResilient and Autonomous Systems Lab\, Department of Electrical and \nComputer Engineering\, Florida State University\n\nTalk Title: Generativ e Vulnerability Assessment of cyber-physical\nSystems\n\nReliability\, sec urity\, and resiliency of the cyber-physical systems\nhas been an active a rea of research\, in one form or the other. With\nevery advancement in ena bling technology comes new challenges in terms\nof reliability and resilie ncy. The past few decades have seen a sharp\nrise in communication and com putational technologies applied to\nlarge-scale systems. The cyber-physica l layer grows exponentially as\nthese systems undergo a transformation to an extensive dependence on\ndistributed resources with associated digital control and\ncommunication interfaces\, many of these located beyond the g rid\nedge. \n\nConsequently\, the problem of securing cyber-physical syst ems has\nbecome technically more challenging on multiple fronts: (1) Ubiqu ity\nof internet of things (IoT) and industrial IoT (IIoT) devices\nincrea ses the vulnerability and threat landscape exponentially\; (2)\nThe increa sed variability due to the introduction of distributed\nresources makes co ntingency analysis challenging\; (3) Despite\nadvancements in computation technology\, the sheer size of the cyber\nlandscape often renders computat ional processes infeasible\; and (4)\nThe tight coupling of the cyber and physical components makes pure\ninformation technology (IT) or operational technology (OT)-based\nsecurity technologies less effective. \n\nTo addr ess these challenges\, current state-of-the-art approaches often\nincorpor ate AI/ML technologies to detect anomaly in the system\noperational data. This has shown significant promise with high true\npositive rates when the available data adequately captures the\noperating conditions of the syste m. However\, performance degradation\nsets in when new operation regimes\, not represented in the available\ndataset\, are encountered. Unfortunatel y\, this is well known and\nexploited by malicious attackers to mimic extr eme event situations\,\nforcing the system into a panic mode. \n\nIn this talk\, I will discuss some recent results leveraging the\ncyber-physical nature to develop algorithms for holistic vulnerability\nassessment of the se systems. This approach seamlessly merges\ndata-driven machine learning models\, for the cyber layer\, with domain\nknowledge physics-based models \, for the physical layer\, to\nsimultaneously achieve high accuracy and h igh generalizability for\nunderstanding the effects of both known and unkn own extreme events.\nThis knowledge promises to enable the development of resilient control\nalgorithms to enable the systems to survive malicious o r natural\nextreme events while sustaining critical functions. \n\n— \ n\nDr. Olugbenga Moses Anubi is an Associate Professor of Electrical and\n Computer Engineering at the Florida State University\, with\naffiliations with the Center for Advanced Power Systems (CAPS) and the\nCenter for Inte lligent Systems\, Controls and Robotics (CISCOR). He is\nthe director of t he Resilient and Autonomous Systems Lab (RASLab). He\nwas a Lead Control S ystems Engineer at the GE Global Research Center\,\nNY. \n\nHis work has resulted in more than 15 patents and several recognitions\nincluding the 2 023 FAMU-FSU College of Engineering Faculty Rising Star\nAward\, the GE Te chnology Award (Physical+Digital)\, the Connected\nControls Technical Achi evement Award\, the Whitney Award and the\nDushman Technology Award. His w ork is supported by DOE\, ONR\, DARPA\,\nDOT and several industry partners . \n\nHis research interests include control of autonomous systems and\nr esilient cyber-physical systems with applications to energy\,\ntransportat ion and other critical infrastructures. He is an inducted\nsenior member o f the National Academy of Inventors (NAI) and a senior\nmember of IEEE. \ n\nHost:  Michael Lisanti \n\nIn Person and Zoom Participation.  See an nouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c2623 DTSTART;TZID=America/New_York:20241117T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241117T140000 URL:https://www.carnegielibrary.org/event/cmu-presents-exploring-ai-and-g-a i-in-our-communities/ LOCATION:South Wing Reading Room\, 2nd Floor\, Carnegie Library\, 4400 Forb es\nAvenue SUMMARY:AI-SDM: Public Speaker Series - AI and Generative AI CLASS:PUBLIC DESCRIPTION:Talk Title: Exploring Artificial Intelligence (AI) and Generati ve AI\nin our Communities\n\nJoin us for this illuminating session with CM U’s NSF-funded AI\nInstitute for Societal Decision Making (AI-SDM) as we explore\nArtificial Intelligence (AI) and Generative AI in our surroundin g\ncommunities.\n\nREGISTER →  registration suggested for in-person\nat tendanceREGISTER  - Virtual Attendees  →  must register in\nadvance t o receive the meeting link to participate.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c290c DTSTART;TZID=America/New_York:20241115T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241115T140000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-benjamin-stoler LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Proposal - Benjamin Stoler CLASS:PUBLIC DESCRIPTION:Speaker: BENJAMIN STOLER\, Ph.D. Student\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Towards Robust Autono mous Driving and Social Robot\nNavigation via Enhanced Data Utilization\n\ nAutonomous robots—including self-driving vehicles\, sidewalk delivery\n robots\, and more—must navigate among humans in a safe and\nsocially-com pliant manner. Current approaches for building and\nevaluating such autono mous systems rely on data-driven techniques\;\nhowever\, a generalization gap emerges\, as methods trained in these\ntraditional paradigms are unabl e to cope with unexpected real-world\nscenarios. Therefore\, this thesis a ims to develop improved evaluation\nsettings and methodologies to increase and assess robustness in\nautonomous robot navigation against these chall enges. This thesis\nproposal describes several completed works that assess and improve\ndifferent facets of robustness in autonomy:\n\nFor robustnes s against perception errors affecting downstream motion\nprediction\, we c onstruct a framework for converting top-down\npedestrian trajectory datase ts into a more challenging first-person\nview perspective. We then develop a correction module to account for\nthe resulting errors\, trained end-to -end with trajectory prediction\napproaches.For robustness against out-of- distribution\, safety-relevant\nscenarios\, we create a hierarchical chara cterization method which\nleverages counterfactual probes to find hidden s afety-relevant\nscenarios in large datasets. We then address the induced\n generalization gap by incorporating the characterizations into\ndownstream trajectory prediction models' inductive biases.For\nrobustness against ad versarial\, safety-critical scenarios\, we develop\na reactive\, skill-bas ed adversary policy which leverages a learned\,\nmulti-faceted criticality objective to perturb existing scenarios. We\nthen train ego policies in a closed-loop manner against these\ngenerated scenarios\, demonstrating imp roved downstream ego\nperformance.\n\nThis proposal concludes by outlining and discussing proposed works to\nfurther advance robustness in autonomou s navigation. These works\ninclude enhancements in scenario characterizati on and\nout-of-distribution generalization\, as well as novel formulations of\nrealism as an objective in safety-critical generation. \n\nThesis Co mmittee\n\nJean Oh (Chair)\n\nSebastian Scherer\n\nReid Simmons\n\nJonatha n Francis (Bosch Center for Artificial Intelligence)\n\n \n\nAdditional I nformation\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c2e7e DTSTART;TZID=America/New_York:20241114T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241114T180000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SCS Distinguished Alumni / Bruce Nelson Distinguished Lecture -\nCa mille Fournier CLASS:PUBLIC DESCRIPTION:Speaker: CAMILLE FOURNIER\, Distinguished Technology Leader\, a nd\nBestselling Author\n\nTalk Title: Engineering Your Own Path: From Univ ersity to Universal\nImpact\n\nCamille Fournier is a distinguished technol ogy leader who most\nrecently served as Global Head of Engineering and Arc hitecture for the\nCommercial and Investment Bank at JPMorgan Chase\, wher e she led an\n800+ person engineering organization. \n\nAn accomplished a uthor of the bestselling book \"The Manager's Path\"\nand \"Platform Engin eering: A Guide for Technical\, Product\, and People\nLeaders\"\, she prev iously led Platform Engineering at Two Sigma and was\nthe CTO of Rent the Runway. Camille is known for her expertise in\ndistributed systems\, engin eering management\, and scaling both\ntechnology and teams. She has contri buted to open source including the\nApache ZooKeeper project\, served on t he founding technical oversight\ncommittee for the Cloud Native Compute Fo undation\, and is currently on\nthe board of the ACM Queue. She holds a BS in Computer Science from\nCarnegie Mellon University\, and an MS in Compu ter Science from the\nUniversity of Wisconsin-Madison.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c320f DTSTART;TZID=America/New_York:20241114T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241114T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - Eli Goldin CLASS:PUBLIC DESCRIPTION:Speaker: ELI GOLDIN\, Ph.D. Student\, Department of Computer Sc ience\,\nNew York University\n\nTalk Title: CountCrypt: Quantum Cryptograp hy between QCMA and PPWe\nconstruct a quantum oracle relative to which BQP = QCMA but\nquantum-computation-classical-communication (QCCC) key exchan ge\, QCCC\ncommitments\, and two-round quantum key distribution exist. We also\nconstruct an oracle relative to which BQP = QMA\, but quantum lightn ing\n(a stronger variant of quantum money) exists. This extends previous\n work by Kretschmer [Kretschmer\, TQC22]\, which showed that there is a\nqu antum oracle relative to which BQP = QMA but pseudorandom state\ngenerator s (a quantum variant of pseudorandom generators) exist.We\nalso show that QCCC key exchange\, QCCC commitments\, and two-round\nquantum key distribu tion can all be used to build one-way puzzles.\nOne-way puzzles are a vers ion of \"quantum samplable\" one-wayness and\nare an intermediate primitiv e between pseudorandom state generators\nand EFI pairs\, the minimal quant um primitive. In particular\, one-way\npuzzles cannot exist if BQP = PP.Ou r results together imply that aside\nfrom pseudorandom state generators\, there is a large class of quantum\ncryptographic primitives which can exis t even if BQP = QCMA\, but are\nbroken if BQP = PP. Furthermore\, one-way puzzles are a minimal\nprimitive for this class. We denote this class \"Co untCrypt\".In Person\nand Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c35b8 DTSTART;TZID=America/New_York:20241114T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241114T163000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:Special Talk - Christos Faloutsos CLASS:PUBLIC DESCRIPTION:Speaker: CHRISTOS FALOUTSOS\, Fredkin University Professor of C omputer\nScience \, Computer Science Department\, Carnegie Mellon Universi ty\n\nTalk Title: Lessons Learned from 1\,000 Research Paper Rejections\n\ nWhat do we need to do to polish the write-up of our research papers? \n\ nPapers that are otherwise worth publishing often get rejected due to\npre sentation pitfalls. After about 1\,000 rejections over the past 30+\nyears \, some patterns started to emerge. In this talk\, we discuss the\nthree m ost important of these patterns\, their potential remedies\, as\nwell as a list of an additional dozen of patterns and remedies. \n\n— \n\nChris tos Faloutsos is a Professor in the Computer Science and Machine\nLearning Departments at Carnegie Mellon University and an Amazon\nScholar. He is t he recipient of the Fredkin Professorship in\nArtificial Intelligence (202 0)\; he has received the Presidential Young\nInvestigator Award by the Nat ional Science Foundation (1989)\, the\nResearch Contributions Award in ICD M 2006\, the SIGKDD Innovations\nAward (2010)\, the PAKDD Distinguished Co ntributions Award (2018)\, 31\n“best paper” awards (including 8 “tes t of time” awards)\, and\nfour teaching awards. \n\nEight of his advise es or co-advisees have received KDD or SCS\ndissertation awards. He is an ACM Fellow\, he has served as a member of\nthe executive committee of SIGK DD\; he has published over 500 refereed\narticles\, 17 book chapters and t hree monographs. He holds 12 patents\n(and several more are pending)\, and he has given over 50 tutorials and\nover 25 invited distinguished lecture s. His research interests include\nlarge-scale data mining with emphasis o n graphs and time sequences\;\nanomaly detection\, tensors\, and fractals.  \n\nREGISTER → open to all SCS PhD Students and Post-docs\n\nAdditiona l Information\n\nHost: The Computer Science Department\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c39be DTSTART;TZID=America/New_York:20241114T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241114T160000 URL:https://csd.cmu.edu/calendar/doctoral-speaking-skills-talk-cayden-codel LOCATION:Newell-Simon 3305 SUMMARY:Doctoral Speaking Skills Talk - Cayden Codel CLASS:PUBLIC DESCRIPTION:Speaker: CAYDEN CODEL\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Verified Substitution Re dundancy Checking for SAT Solving\n\nOne reason for the widespread adoptio n of SAT solvers is that they are\ntrustworthy: their answers can be check ed with verified software. In\nparticular\, many SAT solvers can emit proo f certificates of\nunsatisfiability that are efficient to check. However\, the standard\nproof systems in use today struggle to succinctly express p roofs for\nproblem instances with a high degree of symmetry. \n\nIn this talk\, we discuss our recent work on proof checking tools for\nthe substit ution redundancy (SR) proof system. We discuss a few\nproblems that admit short SR proofs\, as well as how we can express and\ncheck those proofs. O ur verified proof checker was developed in the\nLean theorem prover. \n\n Presented in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c3d12 DTSTART;TZID=America/New_York:20241114T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241114T160000 URL:https://aco.math.cmu.edu/abs-24-25/nov14.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - J. Andrew Newman CLASS:PUBLIC DESCRIPTION:Speaker: J. ANDREW NEWMAN\, Postdoctoral Research and Instructo r\,\nDepartment of Mathematical Sciences\, Carnegie Mellon University\n\nT alk Title: The Turán problem for spheres\n\nThe homeomorph Turán problem is the extremal hypergraph problem to\ndetermine the maximum number of fa cets in a pure d-dimensional\nsimplicial complex on n vertices that does n ot contain a subcomplex\nhomeomorphic to some fixed d-dimensional topologi cal space. The d = 1\ncase of this problem (i.e. subdivisions of a fixed g raph) was settled\ndecades ago by Mader\, and in the last few years there has been\nsubstantial progress for the d = 2 case by many different resear chers.\nIn this talk I will outline some of this recent progress and then turn\nattention to joint work with Marta Pavelka in which we study the\nho meomorph Turán problem for the d-dimensional sphere and tie the\nproblem to an important enumeration question of Gromov. \n\n4:00 pm → Tea and C ookies sponsored by Jane Street\, Math Lounge Wean\n6220 (bring your own m ug if you have one)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c4059 DTSTART;TZID=America/New_York:20241113T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241113T150000 URL:https://db.cs.cmu.edu/events/evolution-of-the-storage-engine-for-spanne r-an-exabyte-scale-database-system/ LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - David F. Bacon CLASS:PUBLIC DESCRIPTION:Speaker: DAVID F. BACON\, Principal Engineer\, Leads\, Spanner Storage\nEngine Team\, Google\n\nTalk Title: Evolution of the Storage Engi ne for Spanner\, an\nExabyte-scale Database System\n\nI’ll describe the design of Spanner’s new storage engine\, Ressi\,\nwhich replaced untyped sorted string tables (inherited from Bigtable)\nwith a strongly typed SQL -native representation. Live migration of 6\nexabytes of data and multiple billion-user products to the new engine\nposed unique challenges. Sound m ethodology from experimental computer\nscience was the key to its success. The simplicity and power of\ndeclarative queries combined with strongly c onsistent transactional\nsemantics has scaled to many thousands of machine s running an\naggregate of over 2 billion queries per second for some of t he largest\napplications in the world. While challenges emerge as we conti nue to\nscale\, I argue that the dominant obstacle to achieving zettabyte scale\ndatabases is in experimental methodology rather than in the underly ing\ntechnical problems themselves. \n\n— \n\nDavid F. Bacon leads Goo gle’s Spanner storage engine team\,\nresponsible for over 70% of the tot al fleet-wide cost of Spanner. His\ncurrent work includes compression\, RA M efficiency\, ASIC support for\ndatabases\, protection against “mercuri al cores”\, and tools for\npredicting fine-grained impact of software an d hardware changes. Prior\nto Google\, he worked at IBM Research on progra mming language design\,\noptimization\, and hardware synthesis. He was nam ed an ACM Fellow for\npioneering work on real-time garbage collection. He holds a Ph.D. from\nUC Berkeley\, and his thesis work on optimizing virtua l functions is\nused in most modern C++ and Java compilers. He has publish ed over 80\npapers and holds 29 patents. \n\nZoom Participation.  See an nouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c4442 DTSTART;TZID=America/New_York:20241113T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241113T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Lunch Seminar — Richard Peng CLASS:PUBLIC DESCRIPTION:Speaker: RICHARD PENG\, Associate Professor\, Computer Science\ nDepartment\, Carnegie Mellon University\n\nTalk Title: Krylov Space Metho dsIn this talk I will survey (block)\nKrylov methods for solving systems o f linear equations. Such methods\nhave close connections with conjugate gr adient\, Lanczos method\,\nWiedemann's algorithm\, and can be viewed as ge neralizations of the\npower method. What I know about these mostly occur o ver reals\, so I'll\nmainly focus on the continuous setting. However\, I w ill also attempt\nto discuss how such algorithms work over finite fields\, and\nconnections with matrix rank\, eigenvalues\, and minimum polynomials .\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c473f DTSTART;TZID=America/New_York:20241112T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241112T135000 URL:https://privacy.cs.cmu.edu/masters/seminar/index.html LOCATION:Hamburg Hall 1002 and Zoom SUMMARY:Privacy Seminar - Frank Lamar CLASS:PUBLIC DESCRIPTION:Speaker: FRANK LAMAR\, Technical Manager\, Cloud Privacy Operat ions and\nEngineering Tooling\, Google\,\n\nTalk Title: Patenting Your PET s\n\nThe patent process can be confusing at best and overwhelming at worst .\nThis presentation aims to cut through the confusion and provide a\nclea r\, actionable understanding of how to protect your innovations. In\nthis talk Frank will walk users through one of his patents from\nideation to is suance. Key Takeaways:\n\nUnderstanding the types of patentsSteps to go fr om Idea to patent\nsubmissionPitfalls to avoid during the Patent submissio n process\n\n— Frank Lamar manages the Cloud Privacy Operations and Engi neering\ntooling team at Google. The team is responsible for architecting and\nimplementing privacy-enhancing features for Google Cloud Platform.\nU nder Frank's leadership the team has been issued 5 patents and\nproduced o ne defensible publication. Frank has previously led\nengineering teams at Delta Air Lines and The Home Depot. Frank is a\nmember of the Internationa l Association of Privacy Professionals and a\ncurrent CMU Privacy Engineer ing MSIT student. \n\nFall 2024 Privacy Seminars are presented by the Mas ters in Privacy\nEngineering Program and Co-Sponsored by the Carnegie Bosc h Institute\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c4ada DTSTART;TZID=America/New_York:20241111T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241111T173000 URL:https://db.cs.cmu.edu/events/building-blocks-building-influxdb-3-0-with -the-fdap-stack-apache-flight-datafusion-arrow-and-parquet-paul-dix/ LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Paul Dix CLASS:PUBLIC DESCRIPTION:Speaker: PAUL DIX\, Chief Technology Officer\, InfluxData\n\nTa lk Title: Building InfluxDB 3.0 with the FDAP Stack: Apache Flight\,\nData Fusion\, Arrow and Parquet\n\nPaul Dix is CTO of InfluxData and the creato r of InfluxDB. He lives in\nNYC because of personal reasons. \n\nThis tal k is part of the Database Building Blocks Seminar\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c4dca DTSTART;TZID=America/New_York:20241111T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241111T150000 URL:https://db.cs.cmu.edu/events/ai-vector-search-in-the-oracle-database LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Shasank Chavan CLASS:PUBLIC DESCRIPTION:Speaker: SHASANK CHAVAN\, Vice President\, In-Memory Data and A I\nTechnologies group\, Oracle\n\nTalk Title: Vector Search in Oracle\n\nA I Vector Search in Oracle Database is a new\, transformative way to\nintel ligently\, efficiently\, and accurately search business data by\nusing AI techniques to search data by semantics\, or meaning. With the\ninclusion o f a new VECTOR data type\, new approximate search indexes\,\nand new SQL o perators and extensions\, enterprise companies can quickly\nand easily lev erage AI Vector Search to build modern\, generative AI\napplications in ju st a few lines of SQL. And with this simplicity\ncomes power\, as AI Vecto r Search is fully integrated with Oracle’s\nenterprise-grade functionali ty such as transactions and scale-out\ncapabilities. This talk will focus on the technology behind Oracle AI\nVector Search\, with a deep dive into its underlying mechanics and\noverall benefits. \n\n— \n\nShasank Chav an is the Vice President of the In-Memory Data and AI\nTechnologies group at Oracle. He leads an organization of brilliant\nengineers working on the nexus between AI systems and modern\ndatabases. His team is currently hyp er-focused on developing the\nnext-generation\, AI-centric data storage en gine\, designed for\nin-memory OLTP\, Analytics and Vector Search capabili ties to power the\nAI and Generative AI revolution to come. Shasank earned his BS/MS in\nComputer Science at the University of California\, San Dieg o. He has\naccumulated 50+ patents over a span of 26 years working on syst ems\nsoftware technology. \n\nFaculty Hosts:  Jignesh Patel\, Andy Pavlo This talk is part of the\nDatabase Building Blocks Seminar\n\nZoom Partic ipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c519e DTSTART;TZID=America/New_York:20241111T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241111T140000 LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:SCS Teaching Track Candidate - Maryam Hedayati CLASS:PUBLIC DESCRIPTION:Speaker: MARYAM HEDAYATI\, Ph.D. Candidate\, Center for Compute r Science\nand Learning Sciences\, Northwestern University\n\nTalk Title: Teaching Demonstration: Binary search trees\n\nBinary search trees (BSTs) provide an efficient way to manage and\nsearch through data. In this teach ing demonstration\, I will explore\nthe motivation for using BSTs and intr oduce students to BST\noperations. I will also talk about how BSTs can be used to implement\nsets. This teaching demonstration will target students who are\nfamiliar with some Python programming and are just beginning to l earn\nabout data structures. \n\n— \n\nMaryam Hedayati is a PhD candid ate in Computer Science and Learning\nSciences at Northwestern University\ , working with Matthew Kay in the\nMU Collective lab. My current research primarily focuses on trying to\nbetter understand data visualization liter acy. I have taught two\ncourses at Northwestern as instructor of record: a n introductory\ncomputer science course taught in Racket\, and an introduc tion to\ncomputer science research for undergraduates. I have also TAed si x\nclasses at Northwestern: CS1 (x2)\, CS1.5\, CS2\, and data visualizatio n\n(x2). Before my PhD\, I received a B.A. from Carleton College in\ncompu ter science and psychology. I also spent some time working as a\ntech cons ultant at Pariveda Solutions in New York\, and was a staff\nmember on two iterations of a computer science study abroad program in\nthe UK. Before c oming to Northwestern\, I received a B.A. from Carleton\nCollege in comput er science and psychology\, where I was a member of\nthe Carleton Percepti on Lab. I also spent time working as a tech\nconsultant at Pariveda Soluti ons in New York\, and was a staff member\non two iterations of a computer science study abroad program in the\nUK. Outside of work\, I enjoy watchin g musicals\, sending postcards\,\ndrinking tea\, sewing\, and hanging out with my cat. \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c5649 DTSTART;TZID=America/New_York:20241111T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241111T130000 URL:https://www.cylab.cmu.edu/events/2024/11/11-seminar-scheffler.html LOCATION:Pather Hollow 4105\, Mehrabian Collaborative Innovation Center and Zoom SUMMARY:CyLab Seminar - Sarah Scheffler CLASS:PUBLIC DESCRIPTION:Speaker: SARAH SCHEFFLER\, Assistant Professor\, Software and S ocietal\nSystems Department and\, Department of Engineering and Public Pol icy\,\nCarnegie Mellon University\n\nTalk Title: Verifiable and Private Co ntent Moderation for End-to-End\nEncryption\n\nEnd-to-end encryption (E2EE ) prevents service providers from directly\naccessing plaintext user conte nt like messages or files\, providing\nvaluable privacy and security benef its to users of these services.\nHowever\, those privacy protections also make it difficult for the\nservice to perform content moderation to identi fy\, label\, monitor\, or\nremove content that is offensive\, harmful\, or illegal. Some proposals\nhave been made for doing content moderation unde r encryption in a\nprivacy-preserving way\, but these often require subtle adjustments to\nthe security and privacy guarantees of E2EE. \n\nThis ta lk will begin with a systematization of content moderation\nunder E2EE\, a nd will proceed to discuss two important and\nunderexamined possibilities for content moderation in this setting\nthat mitigate many of the privacy concerns inherent in E2EE\nmoderation: verifiability of moderation\, and m easuring aggregate\ntrends. For the verification component\, we will discu ss cryptographic\nprotocols for (1) verifying approval of matchlists by ex ternal\norganizations\, (2) guaranteed notification of detection\, and (3) proof\nof non-membership of benign content in matchlists. For aggregate\n measurement\, we will describe ongoing efforts with journalists to\nbuild ethical data donation models for E2EE messages that enable\njournalists to identify aggregate \"trends\" in E2EE messages without\never revealing in dividual message content\, using a combination of\ndifferential privacy an d secure multi-party computation over donated\nmessage embeddings. \n\nTh is talk is based upon three works which were completed or are\nongoing\; t his is joint work with Micha Gorelick\, Palak Jain\, Anunay\nKulshrestha\, Jonathan Mayer\, and Madelyne Xiao. \n\n— \n\nSarah Scheffler is an A ssistant Professor at CMU\, joint between the\nDepartment of Engineering a nd Public Policy\, and the Software and\nSocietal Systems Department\, and is a core faculty member of CyLab.\nHer research centers on applied crypt ography and its intersection with\npolicy\, especially content moderation with verification and privacy\,\nend-to-end encryption\, compelled decrypt ion\, privacy-respecting and\nverifiable data journalism\, and private and secure autonomous systems.\nShe also does research in “pure” applied cryptography\, including\nworks on zero-knowledge proofs and multi-party c omputation. She holds\na Ph.D. in computer science from Boston University. Prior to joining\nCMU in 2024 she was a postdoc at MIT’s Internet Polic y Research\nInitiative\, and in Princeton University’s Center for Inform ation\nTechnology Policy. \n\nIn Person and Zoom Participation.  See ann ouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c5b6a DTSTART;TZID=America/New_York:20241108T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241108T160000 URL:https://www.industry-academia.org/cmu-2024.html LOCATION:ASA Conference Room\, Gates HIllman 6115 SUMMARY:IAP CMU Workshop on the Future of AI and Security in the Cloud CLASS:PUBLIC DESCRIPTION:Speaker: Held in Conjunction with the Industry-Academia Partner ship\n(IAP) and CyLab\n\nWhat is a “Cloud Workshop”?  It’s a full d ay of talks by\nleading experts in academia and industry working in AI and machine\nlearning\, hardware acceleration\, operating systems\, networkin g\, Big\nData\, security\, storage and data management — all in the cont ext of\ndistributed systems — from mammoth data centers to tiny IoT devi ces\non the edge of the cloud.  Expect a full day of talks by leading\ne xperts on the CMU faculty and industry on the status and future of AI\nand Security in the cloud and edge - ranging from infrastructure to\napplic ations - AI and machine learning\, hardware acceleration\,\nnetworking\, security and storage.  The audience is a mix of CMU\nfaculty\, graduate s tudents in CS and ECE\, and industry. \n\nBetween the morning and aftern oon sessions\, we will host a research\nposter session for PhD students T he 2024 AI Workshop is co-organized\nby:  IAP and Professor Riccardo Pacc agnella\, Carnegie Mellon\nUniversity\, in collaboration with CyLab. \n\n Additional Information  \n\n→   Please check back regularly for Regi stration Information and\nAgenda The Industry-Academia Partnership (IAP) i s an association\nfounded in 2013 that brings together industry and univer sity experts\nto pursue research in Web 2.0 and 3.0 applications and infra structure\,\nincluding AI and machine learning\, hardware acceleration\, n etworking\,\nsecurity\, and storage.  The IAP hosts workshops and webinar s on\nuniversity campuses with experts from academia and industry to discu ss\ntheir work and spark new ideas and activities.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c5f49 DTSTART;TZID=America/New_York:20241107T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241107T181500 LOCATION:McConomy Auditorium\, Cohon University Center and Livestream SUMMARY:Raj Reddy Artificial Intelligence Lecture - Raquel Urtasun CLASS:PUBLIC DESCRIPTION:Speaker: RAQUEL URTASUN\, Founder and CEO\, Waabi\, and Full Pr ofessor\,\nUniversity of Toronto\n\nTalk Title: Bringing Generative AI to the Physical World\n\nMartial Hebert\,  Dean of the School of Computer Sc ience and\nUniversity Professor of Robotics invites you to the Raj Reddy\n Artificial Intelligence Lecture  honoring Raj Reddy (HON 2022)\, Moza\nBi nt Nasser University Professor of Computer Science and Robotics.\n\nWith d istinguished guest \n\n  Raquel Urtasun\n\n   Founder and CEO\, Waabi and Full Professor\, University of Toronto\n\n— Bringing Generative AI to the Physical World\n\nAdvances in AI over recent years have been nothin g short of\nremarkable\, but we've only scratched the surface. The next fr ontier\nfor AI will see the technology move from the virtual world and int o\nthe physical world. Join Raquel as she explores how generative AI is\nu nlocking the future of robotics\, starting with self-driving.\n\nREGISTER → by Monday\, 4 November 2024 \n\nAbout the Speaker:   Raquel Urtasun is the Founder and CEO\nof Waabi. A world-renowned expert within the fie ld of AI\, Raquel is\npioneering the application of generative AI technolo gy to build\ninnovative self-driving solutions. Raquel is a Full Professor in the\nDepartment of Computer Science at the University of Toronto and \nco-founded the Vector Institute for AI together with Geoff Hinton.\nOv er her 25 year career within AI\, she has been the recipient of\nseveral h igh profile and prestigious awards including a\nLonguet-Higgins Prize (tes t of time award)\, an Everingham Prize\, an\nNSERC EWR Steacie Award (best scientist in Canada)\, two NVIDIA\nPioneers of AI Awards\, three Google F aculty Research Awards\, an Amazon\nFaculty Research Award\, two Best Pape r Awards at CVPR\, the leading\nconference in computer vision\, and more. She has published more than\n200 papers in AI\, which have been cited more than 70\,000 times\, giving\nher an h-index of 117\, and she has filed mo re than 100 patents in AI\nfor self-driving. \n\nPrior to founding Waabi\ , Raquel was the Chief Scientist and Head of\nR&D at Uber ATG and a Canada Research Chair in Machine Learning and\nComputer Vision. Raquel received her Ph.D. degree from the Computer\nScience department at Ecole Polytechni que Fédérale de Lausanne\n(EPFL) in 2006 and did her postdoc at MIT and UC Berkeley. In 2024\,\nRaquel was included on the CNBC Changemakers: Wome n Transforming\nBusiness list. In 2023\, she was named one of the TIME100 Most\nInfluential People in AI\, made Business Insider’s AI 100 list of Top\nPeople in AI\, and was awarded the Ontario Chamber of Commerce’s\nE merging Tech CEO Award and the Order of Ontario\, the highest civilian\nho nor in the province. In 2021\, she was named one of the 100 Women of\nImpa ct by Entrepreneur Magazine\, 100 People Transforming Business by\nBusines s Insider\, and Toronto’s Most Influential Torontonians\naccording to To ronto Life Magazine. She was also named Chatelaine 2018\nWoman of the Year and 2018 Toronto's top influencers by Adweek\nmagazine. \n\nAbout the Ra j Reddy Artificial Intelligence Lecture\n\nThe Raj Reddy Artificial Intell igence Lecture Series was established\nin honor of Raj Reddy and his criti cal work in AI during his now 50\nyear career at CMU. Reddy was the foundi ng Director of the Robotics\nInstitute from 1979 to 1991 and the Dean of S chool of Computer Science\nfrom 1991 to 1999. He has been active in AI res earch for over five\ndecades in the areas of AI\, Speech Understanding\, I mage Understanding\,\nRobotics\, Multi-sensor Fusion\, and Intelligent Age nts.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c64c3 DTSTART;TZID=America/New_York:20241107T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241107T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - João Ribeiro CLASS:PUBLIC DESCRIPTION:Speaker: JOÃO RIBEIRO\, Assistant Professor\, Departamento de\ nMatemática\, and \, Instituto Superior Técnico\, Universidade de Lisboa \n\nTalk Title: \"Noisy\" versus \"Bounded\" Leakage\n\nPhysical implement ations of cryptographic schemes are the target of\n“side-channel attacks ”\, which aim to extract some information\nabout secret components (e.g. \, a secret key) by exploiting hardware\nquirks. This has given rise to th e study of leakage-resilient\ncryptography\, whose goal is to design crypt ographic schemes that\nremain secure even when partial information about s ecret components is\nleaked to the adversary.  \n\nThere is\, however\, a mismatch between the theory and practice of\nleakage-resilient cryptogra phy. Theoretical work on leakage-resilience\nusually focuses on the bounde d leakage model\, where the adversary is\nallowed to learn an arbitrary t- bit output function of the secret key\,\nwith t being a predefined thresho ld. On the other hand\, real world\nside-channel attacks output very long transcripts that contain noisy\ninformation about the key. Ideally\, we wo uld like to say that every\ncryptographic scheme that is resilient to boun ded leakage is also\nresilient to “noisy” leakage\, for a useful defin ition of\n\"noisy\". \n\nIn this talk\, I will discuss recent work in thi s direction. \n\nThis is based on joint work with Gianluca Brian\, Antoni o Faonio\,\nMaciej Obremski\, Lawrence Roy\, Mark Simkin\, Maciej Skórski \,\nFrançois-Xavier Standaert\, and Daniele Venturi at Eurocrypt 2021 and \nCRYPTO 2024.\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c68de DTSTART;TZID=America/New_York:20241107T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241107T130000 URL:https://www.pdl.cmu.edu/SDI/index.shtml LOCATION:Panther Hollow Room 4105\, Mehrabian Collaborative Innovation Cent er SUMMARY:SDI Seminar - Aurojit Panda CLASS:PUBLIC DESCRIPTION:Speaker: AUROJIT PANDA\, Assistant Professor of Computer Scienc e\,\nCourant Institute\, New York University\n\nTalk Title: Detecting Prot ocol Implementation Bugs at Runtime\n\nDespite significant progress in ver ifying protocols\, services that\nimplement distributed protocols \, e.g.\ , Chubby or Etcd\, can exhibit\nsafety bugs in production deployments.  T hese bugs are often\nintroduced by programmers when converting protocol de scriptions into\ncode. \n\nIn this talk I will describe a new technique w e have been developing\nto identify these bugs at runtime: Runtime Protoco l Refinement\nChecking} (RPRC). RPRC systems observe a deployed service's runtime\nbehavior and notify operators when this behavior evidences a prot ocol\nimplementation bug\, allowing operators to mitigate the bugs impact and\ndevelopers to fix the bug.  \n\nWe have developed an algorithm for RPRC and implemented it in a system\ncalled Ellsberg that targets services that assume the asynchronous or\npartially synchronous model\, and fail-s top failures. We designed\nEllsberg so it makes no assumptions about how s ervices are\nimplemented\, and requires no additional coordination or comm unication.\nWe have used Ellsberg with three open source services: Etcd\, Zookeeper\nand Redis Raft. \n\n— \n\nAurojit Panda is an assistant pro fessor in the Computer Science\ndepartment at New York University working on systems and networking.\nHe received his PhD in 2017 from UC Berkeley\, where he was advised by\nScott Shenker.  He has received several awards\ , including  a VMware\nEarly Career Faculty Award\, a Google Research Sch olar Award\, an NSF\nCareer award\, best paper awards at EuroSys\, SIGCOMM and OSDI\, and a\nEuroSys test of time award. \n\nFaculty Host:  Justin e Sherry\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c6d73 DTSTART;TZID=America/New_York:20241106T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241106T150000 URL:https://db.cs.cmu.edu/events/snowflake-and-why-the-cloud-reshaped-the-a nalytics-industry LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Dan Sotolongo CLASS:PUBLIC DESCRIPTION:Speaker: DAN SOTOLONGO\, Principal Software Engineer\, Snowflak e\n\nTalk Title: Snowflake\, and why the Cloud reshaped the analytics\nind ustry\n\nSnowflake was the first data warehouse designed from scratch to t ake\nadvantage of Cloud economics. We'll talk about what that means\, why it\nwas such a big deal\, and how its design differs from the approaches\n taken by similar systems. Stay until the end for some bonus content on\nho w Snowflake is bringing stream processing into the DBMS. \n\n— \n\nDan Sotolongo is a Principal Software Engineer at Snowflake\, working\nto mak e Snowflake the best place to develop and operate data\npipelines. He enjo ys thinking about databases\, programming languages\,\nand distributed sys tems\, and cares about making people's lives better.\n\n Faculty Hosts:  Andy Pavlo and Jignesh Patel \n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c710f DTSTART;TZID=America/New_York:20241106T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241106T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Lunch Seminar - Keegan Harris CLASS:PUBLIC DESCRIPTION:Speaker: KEEGAN HARRIS\, Ph.D. Student\, Machine Learning Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Regret Minimization in Stackelberg Games with Side\nInformation\n\nAlgorithms for playing in Stac kelberg games have been deployed in\nreal-world domains including airport security\, anti-poaching efforts\,\nand cyber-crime prevention. However\, these algorithms often fail to\ntake into consideration the additional inf ormation available to each\nplayer (e.g. traffic patterns\, weather condit ions\, network\ncongestion)\, which may significantly affect both players ’ optimal\nstrategies. We formalize such settings as Stackelberg games w ith side\ninformation\, in which both players observe an external context before\nplaying. The leader commits to a (context-dependent) strategy\, an d the\nfollower best-responds to both the leader’s strategy and the\ncon text. We focus on the online setting in which a sequence of\nfollowers arr ive over time\, and the context may change from\nround-to-round. \n\nIn s harp contrast to the non-contextual version\, we show that it is\nimpossib le for the leader to achieve no-regret in the full adversarial\nsetting. M otivated by this result\, we show that no-regret learning is\npossible in two natural relaxations: the setting in which the sequence\nof followers i s chosen stochastically and the sequence of contexts is\nadversarial\, and the setting in which contexts are stochastic and\nfollower types are adve rsarial.   \n\nThis talk is based on the paper of the same name\, to ap pear at NeurIPS\n2024.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c77f2 DTSTART;TZID=America/New_York:20241105T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241105T170000 URL:https://www.cmu.edu/leadership/the-provost/provost-initiatives/democrac y-day/index.html SUMMARY:Democracy Day 2024 CLASS:PUBLIC DESCRIPTION:Speaker: No Classes (except classes after 5:00 pm will meet)\n\ nDemocracy Day is an opportunity to underscore the CMU community’s\nfocu s on institutional commitment to civic service and democracy on\nElection Day. \n\nOne of the greatest ways to observe Democracy Day is by exercisi ng\nyour right to vote\, so please make sure you take time on this day to\ ndo so if you are able. \n\nThere are no classes on Democracy Day before 5:00 pm and the entire\nCMU community —faculty\, staff and students — is encouraged to\nparticipate as their schedules allow. \n\nWatch for de tails on programming throughout the day.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c7aeb DTSTART;TZID=America/New_York:20241104T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241104T173000 URL:https://db.cs.cmu.edu/events/building-blocks-synnada-mehmet-ozan-kabak LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Mehmet Ozan Kabak CLASS:PUBLIC DESCRIPTION:Speaker: MEHMET OZAN KABAK\, Apache Datafusion PMC Member\, Co- founder\nand CEO\, Synnada\n\nTalk Title: Towards \"Unified\" Compute Engi nes: Opportunities and\nChallenges\n\nThe architecture diagram of a typica l data and AI infrastructure setup\noften features a primary compute engin e (e.g.\, Apache Spark) alongside\nan array of supplementary tools for obs ervability\, AI integration\,\nstreaming support\, memory management\, int eractivity\, and more. While\nthis modular architecture can be effective\, it also introduces\nchallenges around performance bottlenecks\, maintenan ce costs\, and\nintegration complexity. \n\nIn this talk\, we will explor e whether it is possible to simplify such\ncomplex architectures by addres sing some of the core engine-level\nlimitations that ultimately necessitat e this complex picture. \nDrawing from Synnada’s experience with Apache DataFusion\, we will\nexplore the concept of “unified” compute engine s\, highlighting\nboth the opportunities this approach presents and the ob stacles that\nremain. \n\n— \n\nMehmet is a Apache Datafusion PMC Memb er\, co-founder and CEO of\nSynnada. Faculty Hosts: Andy Pavlo\, Jignesh P atel \n\nThis talk is part of the Database Building Blocks Seminar Series .\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c7ee7 DTSTART;TZID=America/New_York:20241104T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241104T130000 URL:https://www.cylab.cmu.edu/events/2024/11/04-seminar-zhandry.html LOCATION:Panther Hollow Room\, Mehrabian Collaborative Innovation Center 41 05 SUMMARY:CyLab Seminar - Mark Zhandry CLASS:PUBLIC DESCRIPTION:Speaker: MARK ZHANDRY\, Senior Scientist\, Cryptography and Inf ormation\nSecurity Lab\, NTT Research\n\nTalk Title: Optimal Traitor Traci ng from Pairings\n\nWe use pairings over elliptic curves to give a collusi on-resistant\ntraitor tracing scheme where the sizes of public keys\, secr et keys\,\nand ciphertexts are independent of the number of users. Prior\n constructions from pairings had size at least N1/3. \n\n— \n\nMark Zha ndry is currently a Senior Scientist at NTT Research in the\nCryptography & Information Security Lab. He will be joining the\nfaculty at Stanford Un iversity starting Fall 2025\, and was previously\nan Assistant Professor o f Computer Science at Princeton University.\nMark's research focus is on c ryptography and quantum computing\,\nalthough he am broadly interested in all aspects of computer science\ntheory. \n\nFaculty Host:  Elaine Shi  \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c8266 DTSTART;TZID=America/New_York:20241101T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241101T120000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI-SDM - Student Brainstorming Session - Valerie Chen CLASS:PUBLIC DESCRIPTION:The AI Institute for Societal Decision Making (AI-SDM) —  wh ich\nbrings together AI and social sciences researchers to develop\nhuman- centric AI for societal good — hosts student-led brainstorming\ndiscussi on sessions. Our goal is to foster interdisciplinary\ncollaboration and ge nerate ideas on how AI can help solve societal\nproblems\, particularly fr om an angle of decision-making.\n\nWe welcome Valerie Chen\, Ph.D. Cand idate\, Machine Learning Department\nat Carnegie Mellon University who wil l discuss Human-AI Collaboration.\n—\n\nValerie Chen is a 5th year PhD s tudent advised by Ameet Talwalkar. She\nis also a visiting researcher at N YU Center for Data Science\, where\nshe works with He He. Her research foc uses on the principled design of\nhuman-AI teams\, aiming to build interac tive AI systems that enhance\nhuman capabilities and develop scalable eval uation methods for team\nsetups. Valerie's work has been recognized with p restigious awards\,\nincluding an NSF Graduate Research Fellowship and a C MU Presidential\nFellowship. She has previously interned at Microsoft Rese arch and\nworked with renowned researchers at Yale University\, IBM Resear ch\, and\nthe Naval Research Laboratory.  \n\nREGISTER \n\nIn Person an d Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c8605 DTSTART;TZID=America/New_York:20241101T000000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241103T000000 URL:https://web.cvent.com/event/348b7374-64e7-43a4-89c3-48e8917defc6/websit ePage:645d57e4-75eb-4769-b2c0-f201a0bfc6ce?sfmc_id=150409081&utm_campaign= Homecoming%202024%20Invitation%20-%202024-09-12%20-%20All%20Others&utm_id= 567502&utm_medium=email&utm_source=sfmc&utm_term=https%3A%2F%2Fcvent.me%2F xnX7rZ LOCATION:Carnegie Mellon Campus SUMMARY:Carnegie Mellon University Homecoming 2024 CLASS:PUBLIC DESCRIPTION:Speaker: Homecoming 2024\n\nCalling all Tartans! \n\nThe sched ule for Homecoming 2024 is live and registration is now open.\nHomecoming has something for everyone to enjoy\, including\nfamily-friendly activitie s\, the annual Alumni Awards\, the Homecoming\nfootball game and Spirit Da y. With a full schedule of activities\, you\ncan plan the perfect weekend for you. We look forward to celebrating\nwith you Nov. 1-2!\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c8911 DTSTART;TZID=America/New_York:20241031T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241031T160000 URL:https://aco.math.cmu.edu/abs-24-25/oct31.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Yuval Widgerson CLASS:PUBLIC DESCRIPTION:Speaker: YUVAL WIDGERSON\, Junior Fellow\, Department of Mathem atics\,\nETH Zürich\n\nTalk Title: Graph decompositions\, Ramsey theory\, and random graphs\n\nA basic result of probabilistic combinatorics\, orig inally due to\nErdös and Rényi\, is the determination of the threshold a t which the\nrandom graph Gn\,p contains a triangle with high probability. But one\ncan also ask more refined versions of this question\, where we a sk not\njust for one triangle but for many triangles which interact in\nco mplicated ways. For example\, what is the threshold at which we can\nno lo nger partition Gn\,p into two triangle-free subgraphs? \n\nIn this talk\, I will discuss the proof of the Kohayakawa–Kreuter\nconjecture\, which gives a general answer to all such questions. Rather\nsurprisingly\, a key step of the proof is a purely deterministic graph\ndecomposition statemen t\, closely related to classical results such as\nNash-Williams' tree deco mposition theorem\, whose proof uses techniques\nfrom combinatorial optimi zation and structural graph theory. \n\nBased on joint works with Micha C hristoph\, Eden Kuperwasser\, Anders\nMartinsson\, Wojciech Samotij\, and Raphael Steiner.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c8ca2 DTSTART;TZID=America/New_York:20241031T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241031T150000 URL:https://csd.cmu.edu/calendar/doctoral-speaking-skills-talk-zhengyao-lin LOCATION:Mehrabian Collaborative Innovation Center 2201 SUMMARY:Doctoral Speaking Skills Talk - Zhengyao Lin CLASS:PUBLIC DESCRIPTION:Speaker: ZHENGYAO LIN\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: FlowCert: Translation Va lidation for Asynchronous Dataflow\nvia Dynamic Fractional Permissions\n\n Coarse-grained reconfigurable arrays (CGRAs) have gained attention in\nrec ent years due to their promising power efficiency compared to\ntraditional von Neumann architectures. \n\nTo program these architectures using ordi nary languages such as C\, a\ndataflow compiler must transform the origina l sequential\, imperative\nprogram into an equivalent dataflow graph\, com posed of dataflow\noperators running in parallel. This transformation is challenging\nsince the asynchronous nature of dataflow graphs allows out-o f-order\nexecution of operators\, leading to behaviors not present in the\ noriginal imperative programs. \n\nIn this talk\, we address this challen ge by developing a translation\nvalidation technique for dataflow compiler s to ensure that the\ndataflow program has the same behavior as the origin al imperative\nprogram on all possible inputs and schedules of execution. We apply\nthis method to a state-of-the-art dataflow compiler targeting th e\nRipTide CGRA architecture. \n\nOur tool uncovers 8 compiler bugs where the compiler outputs incorrect\ndataflow graphs\, including a data race t hat is otherwise hard to\ndiscover via testing. After repairing these bugs \, our tool verifies\nthe correct compilation of all programs in the RipTi de benchmark\nsuite. \n\nPresented in Partial Fulfillment of the CSD Spea king Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c9062 DTSTART;TZID=America/New_York:20241030T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241030T130000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Lunch Seminar - Jason Li CLASS:PUBLIC DESCRIPTION:Speaker: JASON LI\, Assistant Professor\, Carnegie Mellon Unive rsity\n\nTalk Title: Minimum Isolating Cuts: A new tool for solving minimu m cut\nproblems\n\nMinimum cut problems are among the most well-studied qu estions in\ncombinatorial optimization. In this talk\, I will introduce a simple\nbut powerful new tool for solving minimum cut problems called the\ nminimum isolating cuts. I will show how this tool can be employed to\nobt ain faster algorithms for several fundamental min-cut problems\,\nnamely g lobal min-cut\, Steiner min-cut\, and all-pairs min-cut. For\nthese proble ms\, the new results represent the first improvement in\ntheir runtimes in several decades. \n\nThese results are in collaboration with Amir Abboud \, Robert\nKrauthgamer\, Danupon Nanongkai\, Thatchaphol Saranurak\, and O had\nTrabelsi.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c938d DTSTART;TZID=America/New_York:20241029T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241029T191500 LOCATION:Swartz Center for Entrepreneurship\, 3rd Floor\, Tepper Building SUMMARY:Pushing the Frontiers of AI and ML: From Theory to Real-World\nAppl ications CLASS:PUBLIC DESCRIPTION:Speaker: Hosted by: The Swartz Center for Entrepreneurship\, Pr oject\nOlympus\, Samsung Next\, Radical Ventures\n\nJoin the Swartz Center for Entrepreneurship\, Project Olympus\,  and\nSamsung Next and Radical Ventures for a private event that will gather\na curated audience of inves tors\, founders\, students and academics in\nconversation about the latest advancements in AI/ML\, and bridging the\ngap between cutting-edge resear ch and practical applications. \n\nThe discussion will cover the latest t hemes in AI\, including on-device\nAI\, agent/action models\, and the auto mation of training and\nfine-tuning processes for enterprises. Experts fro m academia and\nindustry will share insights on how theoretical AI/ML rese arch informs\npractical solutions\, from enhancing decision-making systems to\nenabling real-time data processing on edge devices. The panel will\na lso address the use of synthetic data for training models\, the\nchallenge s of scaling AI innovations from lab to market\, and the\nethical consider ations in deploying AI at scale. Join the panelists \nwho will discuss th e practical applications of their research in\nreal-world settings. Paneli sts\n\nTianqi Chen  —  Assistant Professor\, Machine Learning and Comp uter\nScience Departments\, Carnegie Mellon University\;  Chief Technolog ist\,\nOctoML​Matt Fredrikson  —  Associate Professor\, Computer Sci ence\nDepartment and Cylab\, Carnegie Mellon University\;  CEO\, Gray\nSw an​Vyas Sekar  —  Tan Family Professor of Electrical and\nComputer E ngineering\, Department of Electrical and Computer\nEngineering\, Carnegie Mellon University\;   Co-founder and Chief\nTechnologist\, Rockfish Data \; Chief Scientist\, Conviva​Moderator: \nDiane Choi  —  Investor a t Samsung Next\n\nREGISTER\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c9771 DTSTART;TZID=America/New_York:20241029T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241029T130000 URL:https://www.cs.cmu.edu/~aiseminar/ LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Artificial Intelligence Seminar - Virginia Smith CLASS:PUBLIC DESCRIPTION:Speaker: VIRGINIA SMITH\, Leonardo Associate Professor\, Machin e\nLearning Department\, Carnegie Mellon University\n\nTalk Title: A Reali ty Check for Vibes-Based ML Safety\n\nMachine learning applications are in creasingly reliant on black-box\npretrained models. To ensure safe use of these models\, techniques such\nas unlearning\, guardrails\, and watermark ing have been proposed to curb\nmodel behavior and audit usage. Unfortunat ely\, while these post-hoc\napproaches give positive safety ‘vibes’ wh en evaluated in\nisolation\, our work shows that existing techniques are q uite brittle\nwhen deployed as part of larger systems. In a series of rece nt works\,\nwe show that: (a) small amounts of auxiliary data can be used to 'jog'\nthe memory of unlearned models\; (b) current unlearning benchmar ks\nobscure deficiencies in both finetuning and guardrail-based\napproache s\; and (c) simple\, scalable attacks erode existing LLM\nwatermarking sys tems and reveal fundamental trade-offs in watermark\ndesign. Taken togethe r\, these results highlight major deficiencies in\nthe practical use of po st-hoc ML safety methods. We end by discussing\npromising alternatives to ML safety\, which instead aim to ensure\nsafety by design during the devel opment of ML systems. \n\n— \n\nVirginia Smith is the Leonardo Associa te Professor of Machine Learning\nat Carnegie Mellon University. Her curre nt work addresses challenges\nrelated to safety and efficiency in large-sc ale machine learning\nsystems. \n\nIn Person and Zoom Participation.  Se e announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c9b50 DTSTART;TZID=America/New_York:20241029T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241029T123000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-joseph-e-reeves LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Proposal - Joseph E. Reeves CLASS:PUBLIC DESCRIPTION:Speaker: JOSEPH E. REEVES\, Ph.D. Student\, Computer Science De partment\,\nCarnegie Mellon University\n\nTalk Title: Cardinality Constrai nts in Satisfiability Solving\n\nAutomated reasoning (AR) engines solve pr oblems represented in\nmathematics and logic stemming from a wide range of domains including\nhardware and software verification\, cryptography\, an d cloud security.\nBoolean satisfiability (SAT) solvers drive much of the reasoning\nbehind many AR engines\, but their input format\, a formula in\ npropositional logic\, can be limiting. High-level constraints must\nbe e ncoded into sets of simpler constraints\, clauses\, and finding\na good  encoding often requires expert knowledge. We propose\nextending the inpu t of SAT solvers to include cardinality constraints\,\nmoving encoding que stions from the user-side to the solver-side.\nCardinality constraints are a frequently occurring type high-level\nconstraint that represent countin g\, e.g.\, “at least k packages must\nbe delivered” or “you can wor k from home at most one day of the\nweek”.\n\nIn this proposal we disc uss four research problems arising from a\ncardinality-based input. First\ , we will develop a cardinality\nconstraint extraction tool that will conv ert previously encoded\nproblems into a cardinality-based normal form\, pr oviding backwards\ncompatibility for our solving techniques. Second\, we w ill engineer\ndynamic cardinality constraint encoding into the top-tier SA T solver\nCaDiCaL to improve performance on problems with many cardinality \nconstraints. Third\, we will explore the ways in which parallel solving\ ntechniques can leverage information within cardinality constraints to\nac hieve better problem partitioning. Fourth\, we will equip the\nextraction and solving with end-to-end proof checking through\nmodifications to exist ing proof systems and proof checkers. Our goal\nin investigating these fou r research problems is to support the claim\nthat a cardinality-based form at should be the standard input format\nfor modern SAT solvers.\n\nThesis Committee\n\nMarijn Heule (Chair)\n\nRandal Bryant\n\nRuben Martins\n\nArm in Biere (University of Freiburg) \n\nAdditional Information\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91c9fa9 DTSTART;TZID=America/New_York:20241028T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241028T173000 URL:https://db.cs.cmu.edu/events/building-blocks-exon-trent-hauck LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Trent Huack CLASS:PUBLIC DESCRIPTION:Speaker: TRENT HAUCKTalk Title: Exon: A Built for Purpose\nBioi nformatics Database\n\nWithout having to implement every component of a da tabase engine\,\nit’s now feasible to build databases that can lean into the\nidiosyncrasies of specific domains to deliver a better user\nexperie nce. Exon is one such databases. Thanks to DataFusion\, Exon can\ndeliver a complete database\, but also have capabilities bridge the gap\nbetween b ioinformatics and database systems. In this talk I’ll\ndiscuss some of t he features that make Exon specially adapted to\nbiodata and how those fea tures come about due to DataFusion’s\nability to be extended. \n\n—  \n\nTrent Hauck is the owner and operator of WHERE TRUE Technologies.\nBef ore starting WHERE TRUE\, he built large-scale data platforms for\nbioinfo rmatics companies. This talk is part of the Database Building\nBlocks Semi nar\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ca32e DTSTART;TZID=America/New_York:20241028T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241028T163000 LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Three AR/VR Talks CLASS:PUBLIC DESCRIPTION:Speaker: QITAO ZHAO\, VIMAL MOLLYN\, HYUNSUNG CHOTalk Title: Th ree AR/VR\nTalks\n\n► Qitao Zhao\, Master's Student\, MS in Computer Vis ion\, Robotics\nInstitute \n\n      —Sparse-view Pose Estimation a nd Reconstruction via\nAnalysis by Generative Synthesis\n\nThis talk will present our approach for reconstructing objects from\nsparse-view images c aptured in unconstrained environments. In the\nabsence of ground-truth cam era poses\, we will demonstrate how to\nutilize estimates from off-the-she lf systems and address two key\nchallenges: refining noisy camera poses in sparse views and\neffectively handling outlier poses.\n\nQitao Zhao is a second-year Master's student in Computer Vision at\nCMU\, RI\, advised by Prof. Shubham Tulsiani. His research focuses on\ncamera pose estimation an d 3D reconstruction in the wild. He holds a\nBachelor's degree from Shando ng University in China and was a visiting\nstudent at the University of Ce ntral Florida\, where he worked with\nProf. Chen Chen. Additional Informat ion\n\n► Vimal Mollyn\, Ph.D. Student\, Human Computer Interaction Insti tute\n\n      — EgoTouch: On-Body Touch Input Using AR/VR Headset\n Cameras\n\nIn augmented and virtual reality (AR/VR) experiences\, a user ’s arms\nand hands can provide a convenient and tactile surface for touc h\ninput. Prior work has shown on-body input to have significant speed\,\n accuracy\, and ergonomic benefits over in-air interfaces\, which are\ncomm on today. In this work\, we demonstrate high accuracy\, bare hands\n(i.e.\ , no special instrumentation of the user) skin input using just\nan RGB ca mera\, like those already integrated into all modern XR\nheadsets. Our res ults show this approach can be accurate\, and robust\nacross diverse light ing conditions\, skin tones\, and body motion (e.g.\,\ninput while walking ). Finally\, our pipeline also provides rich input\nmetadata including tou ch force\, finger identification\, angle of\nattack\, and rotation. We bel ieve these are the requisite technical\ningredients to more fully unlock o n-skin interfaces that have been\nwell motivated in the HCI literature but have lacked robust and\npractical methods.\n\nVimal Mollyn is a PhD stud ent in the Future Interfaces Group at\nCarnegie Mellon University where I ’m advised by Chris Harrison.\nI’m interested in creating new ways for people to interact with the\nworld using my background in sensing and mac hine learning. Previously\nI graduated with a Bachelors and Masters from I IT Madras\, where I\nmajored in Engineering Design and Data Science. Addit ional Information\n\n► Hyunsung Cho\, Ph.D. Student\, Human-Computer Int eraction Institute\n\n      —  Auptimize: Optimal Placement of Spa tial Audio Cues\nfor Extended Reality \n\nSpatial audio in Extended Reali ty (XR) provides users with better\nawareness of where virtual elements ar e placed\, and efficiently guides\nthem to events such as notifications\, system alerts from different\nwindows\, or approaching avatars. Humans\, h owever\, are inaccurate in\nlocalizing sound cues\, especially with multip le sources due to\nlimitations in human auditory perception such as angula r\ndiscrimination error and front-back confusion. This decreases the\neffi ciency of XR interfaces because users misidentify from which XR\nelement a sound is coming. To address this\, we propose Auptimize\, a\nnovel comput ational approach for placing XR sound sources\, which\nmitigates such loca lization errors by utilizing the ventriloquist\neffect. Auptimize disentan gles the sound source locations from the\nvisual elements and relocates th e sound sources to optimal positions\nfor unambiguous identification of so und cues\, avoiding errors due to\ninter-source proximity and front-back c onfusion. Our evaluation shows\nthat Auptimize decreases spatial audio-bas ed source identification\nerrors compared to playing sound cues at the pai red visual-sound\nlocations. We demonstrate the applicability of Auptimize for diverse\nspatial audio-based interactive XR scenarios. \n\nHyunsung Cho is a fourth-year Ph.D. student in the Human-Computer\nInteraction Inst itute (HCII) at Carnegie Mellon University\, advised by\nProf. David Lindl bauer. Her research focuses on designing\,\nimplementing\, and evaluating context-aware Extended Reality (XR)\ninterfaces and multimodal interaction techniques in XR to enable\nseamless\, unobtrusive human-computer interac tions. Her work combines\ncomputational modeling of human perception and b ehavior\, user-centered\ndesign\, and intelligent systems to create adapti ve interfaces for\ndiverse user contexts. Her research has received the Be st Paper Awards\nand Methods Recognition at ACM CSCW and ACM ISS. She hold s a M.S. and\nB.S. in Computer Science from KAIST. She has previously work ed as a\nResearch Scientist Intern at Meta's Reality Labs Research and Nok ia\nBell Labs' Pervasive Systems research group. Additional Information\n\ nThe VASC Seminar is sponsored in part by Meta Reality Labs Pittsburgh\n\n  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ca993 DTSTART;TZID=America/New_York:20241028T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241028T130000 URL:https://www.cylab.cmu.edu/events/2024/10/28-seminar-pellegrino.html LOCATION:Panther Hollow Room 4105\, Mehrabian Collaborative Innovation Cent er SUMMARY:CyLab Seminar - Giancarlo Pellegrino CLASS:PUBLIC DESCRIPTION:Speaker: GIANCARLO PELLEGRINO\, Tenured Faculty\, CISPA \, Helm holtz\nCenter for Information Security\n\nTalk Title: Program Analysis for Web Applications at Scale\n\nWeb applications are integral to our modern society\, serving a wide\nrange of purposes from social sharing to financi al transactions and\ncontrol of critical infrastructure. However\, the rap id development of\nnew functionalities has led to increased complexity\, w hich in turn has\ncreated significant security vulnerabilities exploited b y attackers\nfor unauthorized activities such as data theft\, fraud\, and illegal\ntransactions. In response\, it is crucial to discover\, analyze\, and\nlearn from software vulnerabilities to enhance the security of web\n applications. This involves both immediate detection and long-term\nknowle dge building. \n\nIn this talk\, I will highlight the work of my research group in\ndeveloping methods to detect\, study\, and learn from vulnerabi lities.\nWe will then focus on JAW\, an open-source framework for large-sc ale\nprogram analysis of client-side JavaScript. I will show how JAW can\n study vulnerabilities like client-side request hijacks and DOM\nclobbering \, showcasing JAW's capacity to assess actual risks at scale\,\nevaluate t he efficacy of security measures\, and contribute to secure\ndevelopment b y aggregating code patterns and real-world examples. \n\n— \n\nGiancar lo Pellegrino is a tenured faculty at CISPA Helmholtz Center\nfor Informat ion Security\, where he leads the application security\nresearch group. Pr eviously\, he was a visiting asst professor at\nStanford University as the first member of the CISPA-Stanford Center\nfor Cybersecurity. Giancarlo e arned his Ph.D. in Computer Science from\nEurecom. His research primarily focuses on identifying\, analyzing\, and\naddressing vulnerabilities in we b applications\, both at the\napplication and platform levels. Giancarlo s erved as a PC member for\nthe major security venues (e.g.\, IEEE SP\, CCS\ , and USENIX Security)\,\nas an area chair (USENIX Security 22-23)\, and i s serving as a PC\nco-chair for USENIX Security 2025.\n\n Faculty Host:  Lujo Bauer \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cadf7 DTSTART;TZID=America/New_York:20241028T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241028T130000 URL:https://pdl.cmu.edu/SDI/2024/102824.html LOCATION:Newell-Simon 3305 and Zoom SUMMARY:Special SDI Seminar - Sanidhya Kashyap CLASS:PUBLIC DESCRIPTION:Speaker: SANIDHYA KASHYAP\, Systems Researcher\, and Assistant Professor\n\, School of Computer and Communication Sciences\, EPFLÉcole\n Polytechnique Fédérale de Lausanne (EPFL)\n\nTalk Title: Towards Adaptiv e and Evolving Systems Software\n\nSpecializing the OS for application req uirements is becoming the need\nof the hour. In this talk\, I will present our ongoing effort to meet\nthese requirements dynamically. First\, I wil l present a new approach\nto designing a storage stack that allows file sy stem developers to\ndesign userspace file systems without compromising fil e system\nsecurity guarantees while at the same time ensuring direct acces s to\nnon-volatile memory (NVM) hardware. \n\nI will present a new file s ystem architecture that decouples file\nsystem design\, access control\, a nd metadata integrity enforcement\,\nproviding a clean structure to design file systems. Then\, I will\npresent KFlex\, a new approach to kernel ext ensibility that balances\nexpressivity and performance. KFlex separates th e safety of\nkernel-owned resources (e.g.\, kernel memory) from the safety of\nextension-specific resources (e.g.\, extension memory). This separati on\nenables KFlex to use distinct\, bespoke mechanisms to enforce each\nsa fety property\, which enables the offload of diverse functionality\nwhile incurring low runtime overheads. \n\n— \n\nSanidhya Kashyap is a syste ms researcher and an Assistant Professor at\nthe School of Computer and Co mmunication Sciences at EPFL. His\nresearch focuses on designing robust an d scalable systems software. He\nreceived the Vmware Early Career Faculty Award in 2021 and his Ph.D.\nfrom Georgia Tech in 2020. \n\nFaculty Host: Dimitrios Skarlatos \n\nIn Person and Zoom Participation.  See announce ment.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cb210 DTSTART;TZID=America/New_York:20241025T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241025T143000 URL:https://seminartracker.tepper.cmu.edu/SeminarDetail?SeminarId=1103 LOCATION:Tepper 5219 SUMMARY:Operations Research Seminar - Karthik Chandrasekaran CLASS:PUBLIC DESCRIPTION:Speaker: KARTHIK CHANDRASEKARAN\, Associate Professor\, Departm ent of\nIndustrial and Enterprise Systems Engineering\, University of Illi nois\,\nUrbana-Champaign\n\nTalk Title: Splitting-off in Hypergraphs\n\nTh e splitting-off operation in undirected graphs is a fundamental\nreduction operation that detaches all edges incident to a given vertex\nand adds ne w edges between the neighbors of that vertex while\npreserving their degre es. Lovász (1974) and Mader (1978) showed the\nexistence of this operatio n while preserving global and local\nconnectivities respectively in graphs under mild conditions. These\nresults have been influential in structural graph theory as an\ninduction tool and in graph algorithms as a recursion tool.   \n\nIn this talk\, I will introduce a splitting-off operation i n\nhypergraphs. The main result is that there exists a local connectivity\ npreserving complete splitting-off in hypergraphs and a strongly\npolynomi al-time algorithm to compute it in weighted hypergraphs. I\nwill outline t wo applications of our local connectivity preserving\nsplitting-off result in hypergraphs: (1) constructive characterization\nof k-hyperedge-connect ed hypergraphs and (2) alternate proof of an\napproximate min-max relation for max Steiner rooted-connected\norientation of graphs and hypergraphs ( due to Király and Lau\, 2008).\nOur proof of the approximate min-max rela tion for graphs circumvents\nthe Nash-Williams' strong orientation theorem and uses tools developed\nfor hypergraphs. As a special case of this appl ication\, I will present\na unified proof of Menger’s theorem for graphs and hypergraphs (edge\nversion).    \n\nBased on joint work with Krist of Berczi\, Tamas Kiraly\, and Shubhang\nKulkarni. \n\n — \n\nKarthik Chandrasekaran is an associate professor in the Department of\nIndustrial and Enterprise Systems Engineering and an affiliate in the\nDepartment of Computer Science at University of Illinois\,\nUrbana-Champaign. He receiv ed his bachelor’s in Computer Science\nfrom IIT Madras and Ph.D. in in A lgorithms\, Combinatorics\, and\nOptimization (ACO) from Georgia Tech. His Ph.D. thesis was awarded the\nDissertation Prize by the College of Comput ing at Georgia Tech and the\nBest Ph.D. Thesis Award by the Sigma Xi chapt er of Georgia Tech. Prior\nto joining UIUC\, he was a Simons postdoctoral fellow at Harvard\nUniversity. His research focuses on fundamental problem s in\nCombinatorial Optimization and Algorithms and is supported by the\nN ational Science Foundation.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cb6c1 DTSTART;TZID=America/New_York:20241025T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241025T113000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Seminar - To be rescheduled CLASS:PUBLIC DESCRIPTION:Speaker: TONY METGER - To Be Rescheduled\, PhD. Student\, Insti tute for\nTheoretical Physics \, ETH Zurich\n\nTalk Title: Pseudorandom un itaries\, t-designs\, and the\nincompressibility of random circuits\n\nUni formly random unitaries\, i.e. unitaries drawn from the Haar\nmeasure\, ha ve many useful properties\, but cannot be implemented\nefficiently. This h as motivated a long line of research into random\nunitaries that \"look\" sufficiently Haar random while also being\nefficient to implement. Two dif ferent notions of derandomisation have\nemerged: t-designs are random unit aries that information-theoretically\nreproduce the first t moments of the Haar measure\, and pseudorandom\nunitaries (PRUs) are random unitaries th at are computationally\nindistinguishable from Haar random. \n\nI will ex plain a simple unified construction of both t-designs and\nPRUs from the “PFC ensemble”\, the concatenation of a random\nClifford unitary\, a r andom binary phase\, and a random computational\nbasis state permutation. Then\, I will show how the PFC ensemble helps\nus to resolve a long-standi ng open question about the spectral gap of\nrandom quantum circuits\, impl ying that a random quantum circuit is\nessentially incompressible. This pr oves the Brown-Susskind conjecture\nfrom black hole physics.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cba67 DTSTART;TZID=America/New_York:20241024T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241024T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room and Zoom SUMMARY:Crypto Seminar - Surya Mathialagan CLASS:PUBLIC DESCRIPTION:Speaker: SURYA MATHIALAGAN\, Ph.D. Student\, Theory Group\, Ele ctrical\nEngineering & Computer Science Department\, Massachusetts Institu te of\nTechnology\n\nTalk Title: Universal SNARGs for NP from Proofs of Co mpleteness\n\nIn this work\, we give new constructions of succinct non-int eractive\narguments SNARGs for NP in the settings of both non-adaptive and \nadaptive soundness. First\, we construct a succinct non-interactive\narg ument system (SNARG) for any NP language L\, and prove the\nnon-adaptive s oundness assuming the security of an FHE scheme\, a batch\nargument (BARG) scheme\, as well as the existence of any two-message\nargument system for L where the prover’s message is succinct\, and\nthe completeness proper ty has a polynomial-size Extended Frege proof.\nOur SNARG is universal in the sense that the construction does not\ndepend on the two-message argume nt system. The set-up of our SNARG\nscheme is transparent (i.e. no private randomness). Beyond\nuniversality\, we note that weaker primitives such a s designated\nverifier SNARGs\, and witness encryption both imply such 2-m essage\narguments. Such an amplification of these primitives was not known \nbefore. \n\nIn the adaptive setting\, we also show how to convert any a daptively\nsound designated verifier SNARG into publicly verifiable SNARGs with\nadaptive soundness\, assuming the underlying designated verifier SN ARG\nhas a polynomial-size Extended Frege proof of completeness. \n\nOur framework yields several corollaries\, including:\n\na SNARG for NP with a transparent CRS and non-adaptive soundness\,\nassuming LWE and the (non-e xplicit) existence of any witness\nencryption for NP that has a polynomial -size 'Extended Frege proof of\ncorrectness'. As a corollary\, we obtain S NARGs for NP under the\nevasive LWE and subexponential LWE assumptions\, w ith a (long)\ntransparent CRS and non-adaptive soundness.a SNARG for UP (N P language\nwith unique witnesses) with a long (and even transparent!) CRS and\nadaptive soundness under the evasive LWE and subexponential LWE\nass umptions.a SNARG for NP with a short CRS and non-adaptive soundness\nassum ing LWE\, FHE\, and the (non-explicit) existence of any hash\nfunction tha t makes Micali's SNARG construction sound.\n\nWe prove our results by exte nding the encrypt-hash-and-BARG paradigm\nof [Jin-Kalai-Lombardi-Vaikuntan athan\, STOC '24]\; in this work\, we use\nExtended Frege proofs as a secu rity reduction from one argument system\nto another\, rather than as an ou tright security proof. Our universal\nconstruction suggests that the encry pt-hash-and-BARG construction can\nbe viewed as a “best possible SNARG'' . \n\nBased on the joint work with Zhengzhong Jin\, Yael Tauman Kalai\, a nd\nAlex Lombardi.\n\nIn Person and Zoom Participation.  See announcement .\n\nBased on the joint work with Zhengzhong Jin\, Yael Tauman Kalai\, and \nAlex Lombardi.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cbf33 DTSTART;TZID=America/New_York:20241024T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241024T160000 URL:https://aco.math.cmu.edu/abs-24-25/oct24.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Ramon Van Handel CLASS:PUBLIC DESCRIPTION:Speaker: RAMON VAN HANDEL\, Associate Professor\, Program in Ap plied &\nComputational Mathematics\, and Operations Research & Financial\n Engineering\, Princeton University\n\nTalk Title: A new approach to strong convergence\n\nIt was conjectured by Alon in the 1980s that random d-regu lar graphs\nhave the largest possible spectral gap (up to negligible error ) among\nall d-regular graphs. This conjecture was proved by Friedman in 2 004\nin major tour de force. In recent years\, deep generalizations of\nFr iedman's theorem\, such as strong convergence of random permutation\nmatri ces due to Bordenave and Collins\, have played a central role in a\nseries of breakthrough results on random graphs\, geometry\, and\noperator algeb ras. \n\nIn this talk\, I will discuss a surprisingly simple new approach to\nsuch results that is almost entirely based on soft arguments. This\na pproach makes it possible to address previously inaccessible\nquestions: f or example\, it enables a sharp understanding of the large\ndeviation prob abilities in Friedman's theorem\, and establishes strong\nconvergence of v ery high-dimensional representations of the symmetric\nand classical group s. I will aim to explain some of these results and\nthe basic ideas on wh ich they are based. \n\nJoint work with Chi-Fang Chen\, Jorge Garza-Varga s\, Joel Tropp. \n\n→ 4:00 pm Tea & Cookies in the Math Lounge - sponso red by Jane\nStreet \n\n→ Bring your own mug if you have one\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cc2f1 DTSTART;TZID=America/New_York:20241024T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241024T152000 LOCATION:Newell-Simon 3002 SUMMARY:15-499/899 Special Topics: Social Agents Guest Lecture CLASS:PUBLIC DESCRIPTION:Speaker: MARTEN SAPP\, Assistant Professor\, Language Technolog ies\nInstitute\, Carnegie Mellon University\n\nTalk Title: Promise and Pit falls of Safe and Social LLM Agents\n\nLLMs are increasingly playing agent ic and social roles by interacting\nwith users and tools. In this talk\, I will discuss the shortcomings\nand pitfalls of LLM-based agents. First\, I will introduce Sotopia\, a\nnew framework for social simulation with LLM agents\, and Sotopia-Eval\,\na multi-dimensional framework for quantitati vely evaluating social\nskills of LLM in interactions. Showing in Sotopia that LLM agents are\nstill far from humans' interaction skills\, I will th en discuss the\ncrucial role of information asymmetry\, showing that LLMs cannot\nproperly deal with the information asymmetry that is present in re al\ninteractions. Finally\, I will introduce HAICosystem\, a new framework \nfor simulating and evaluating the safety of LLM agents in user-AI-tool\n interactions\, showcasing the many safety issues that could arise from\nin creasingly autonomous and tool-using agents. I will conclude with\nsome fu ture directions towards safe and social LLM agents. \n\nCourse / Faculty Host: Hirokazu Shirado\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cc64f DTSTART;TZID=America/New_York:20241024T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241024T120000 URL:https://pdl.cmu.edu/SDI/2024/102424.html LOCATION:Newell-Simon 3305 SUMMARY:Special SDI seminar - Rishabh Iyer CLASS:PUBLIC DESCRIPTION:Speaker: RISHABH IYER\, Postdoctoral Researcher\, Berkeley NetS ys Lab \,\nUniversity of California\, Berkeley\n\nTalk Title: Performance Clarity for Systems Software and Hardware\n\nPerformance is increasingly b ecoming a first-class requirement in\nsystem design. However\, system deve lopers today lack the techniques\nand tools to reason precisely about the expected performance behavior\nof both the code they write and the hardwar e they run their code on.\nWidely used tools such as profilers and simulat ors provide an\nincomplete understanding of performance\, leading to hiccu ps and\nmeltdowns in production when the workload or runtime environment\n changes in unpredicted ways. \n\nIn this talk\, I will introduce a new ap proach to reasoning about the\nperformance of systems software and hardwar e and make the case for\ndesigning new abstractions that precisely capture a system's expected\nperformance behavior. I will present two tools that leverage these\nabstractions to help developers answer frequently-asked \" what-if\"\nperformance questions. \n\nFirst\, I will discuss CFAR\, a too l that enables developers to reason\nprecisely about how their code\, as w ell as third-party code\, uses the\nCPU cache. Then\, I will introduce LTC \, a tool that enables developers\nto reason about the potential performan ce benefits provided by\nhardware accelerators without requiring them to p urchase the\naccelerator or port their code to it. The improved performanc e\nvisibility provided by both CFAR and LTC has tangible benefits: for\nin stance\, we used CFAR to identify several cache-inefficient access\npatter ns and performance bugs (including in the Linux kernel's TCP\nstack) and L TC to speed up compilation for ML accelerators by 5-41x. \n\n— \n\nRis habh Iyer is a postdoctoral researcher at UC Berkeley\, working with\nSylv ia Ratnasamy and Scott Shenker. He received his bachelor's degree\nfrom II T Bombay\, and his PhD from EPFL\, under the supervision of\nGeorge Candea and Katerina Argyraki. Rishabh's research is centered\naround developing techniques that enable developers to reason\nprecisely about the performan ce behavior of their systems before they\nare deployed. His dissertation w ork introduced the notion of latency\ninterfaces---simple\, succinct progr ams that summarize a system's\nlatency behavior just like semantic interfa ces such as code\ndocumentation and specifications summarize functionality ---and was\nawarded the ACM SIGOPS Dennis M. Ritchie Award\, the Eurosys R oger\nNeedham PhD Award\, and the Dimitris N. Chorafas Award. \n\nFaculty Host:  Jan Hoffman\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ccb61 DTSTART;TZID=America/New_York:20241023T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241023T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:Theory Lunch Seminar - David Wu CLASS:PUBLIC DESCRIPTION:Speaker: DAVID WU\, Ph.D. Student\, Electrical Engineering and Computer\nSciences\, University of California\, Berkeley\n\nTalk Title: Lo cally stationary distributions: inference and\noptimization beyond rapid m ixing\n\nMarkov chain sampling algorithms such as MCMC are an important\na lgorithmic primitive for inference\, optimization\, and statistical\nestim ation in high-dimensional settings. To guarantee downstream\nperformance\, it suffices for MCMC to converge\, or mix to a target\ndistribution. Unfo rtunately\, in many relevant scenarios\, the mixing\ntime of canonical MCM C algorithms is exponential in the dimension\,\nrendering this theoretical guarantee impractical. \n\nIn this talk\, we study Markov chains with po tentially slow mixing\ntimes\, based on an analogy to stationary points in optimization. In\nparticular\, under mild assumptions\, any reversible Ma rkov chain\nconverges to a “locally stationary” distribution in polyno mially\nmany steps. Such locally stationary distributions enjoy various\np roperties that make them amenable to analysis in interesting\ninference an d optimization settings. Using this framework\, we prove\nthat canonical M CMC algorithms with exponential mixing time can\nnevertheless recover (1) large independent sets in triangle free\ngraphs and (2) good estimates of the hidden communities in a\n2-community stochastic block model. Our analy sis reveals a connection\nbetween a certain MCMC algorithm and the ubiquit ous power iteration\nfor spiked matrix models.   \n\nBased on joint work with Kuikui Liu (MIT)\, Sidhanth Mohanty (MIT)\,\nAmit Rajaraman (MIT)\, and Prasad Raghavendra (UC Berkeley)\n(arXiv:2405.20849)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cd063 DTSTART;TZID=America/New_York:20241022T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241022T130000 URL:https://csd.cmu.edu/calendar/joint-speaking-skills-talk-database-lunch- group-talk-wan-shen-lim LOCATION:Blelloch-Skees Conferences Room\, Gates Hillman 8115 SUMMARY:Joint Speaking Skills Talk / Database Lunch Group Talk - Wan Shen L im CLASS:PUBLIC DESCRIPTION:Speaker: WAN SHEN LIM\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Accelerating Machine Lea rning for Database Systems\n\nDatabase tuning is the process of finding be tter configurations to\noptimize the performance of a database management system (DBMS). The\nsize and complexity of the tuning search space makes i t difficult for\na human to manually discover good configurations. This ne cessitates\nthe use of automated methods that rely on machine learning (ML ) models\nto predict the DBMS's run-time behavior. These ML models enable the\nevaluation of candidate configurations without the expensive executio n\nof queries. \n\nHowever\, the high cost of obtaining the training data for these ML\nmodels make them impractical for real-world deployments. Fi rst\,\ngenerating the training data itself requires expensive query\nexecu tion. Second\, the training data is \"specialized\" as it depends on\ninst ance-specific factors such as the workload\, schema\, database\nversion\, and more. Unlike other ML tasks\, a pre-trained model cannot\nbe downloade d off the internet\; each database deployment must collect\nits own expens ive training data from scratch. This problem is\nexacerbated by the freque ncy at which model invalidation due to\nchanges in the DBMS's environment.  \n\nConsequently\, training data generation has become a major bottlenec k\nin machine learning for database research\, taking weeks or even months \nof time. To mitigate this problem\, we make the critical observation\nth at training data does not require accurate query results (unlike\nordinary query execution). This allows us to modify query execution\nsemantics wit h a \"training data mode\" to approximate and eliminate\nrepetition from t he training data generation process\, achieving up to\n268x speedup with m odest degradation in model accuracy. \n\nHaving addressed this bottleneck \, we will also briefly discuss what\nthe next challenge is and our next s teps. \n\nPresented as part of the Database Group Lunch Talks\n\nPresente d in Partial Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cd545 DTSTART;TZID=America/New_York:20241021T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241021T173000 URL:https://db.cs.cmu.edu/events/building-blocks-accelerating-data-and-ai-w ith-spice-ai-open-source-software-luke-kim LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Luke Kim CLASS:PUBLIC DESCRIPTION:Speaker: LUKE KIM\, Co-founder and Chief Executive Officer\, Sp ice AI\n\nTalk Title: Accelerating Data and AI with Spice.ai Open-Source\n Software\n\nSpice.ai OSS is an open-source\, portable runtime designed to simplify\nbuilding data and AI applications. It’s built on industry lead ing\ntechnologies like Apache DataFusion\, Apache Arrow\, DuckDB and\nSQLi te. \n\nIn this talk\, we tell the story of building neurofeedback system s\, to\noperating DuckDB at cloud-scale\, to building Spice.ai OSS for the \nintersection of high-performance data query and ML-inference. We\nintrod uce Spice.ai OSS\, demo some of its capabilities and use-cases\,\nexplore the design principles and architecture of the platform\, and go\ndeep on i ts data federation features. We will also provide insights\ninto how Spice .ai OSS fits into the broader landscape of data and AI\ninfrastructure and how it can be extended to solve complex\,\nproduction-grade problems in m odern applications. \n\n— \n\nLuke Kim is a co-founder and CEO of Spic e AI. He is Australian\, which\nis kind of cool. \n\nThis talk is part of the Database Building Blocks Seminar Series\n\nIn Person and Zoom Partici pation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cd965 DTSTART;TZID=America/New_York:20241021T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241021T163000 URL:https://www.ri.cmu.edu/event/building-scalable-visual-intelligence-from -represention-to-understanding-and-generation/ LOCATION:Newell-Simon 3305 SUMMARY:VASC Seminar - Saining Xie CLASS:PUBLIC DESCRIPTION:Speaker: SAINING XIE\, Assistant Professor of Computer Science\ , Courant\nInstitute of Mathematical Sciences\, New York University\n\nTal k Title: Building Scalable Visual Intelligence: From Represention\nto Unde rstanding and Generation\n\nIn this talk\, we will dive into our recent wo rk on vision-centric\ngenerative AI\, focusing on how it helps with unders tanding and\ncreating visual content like images and videos. We'll cover t he latest\nadvances\, including multimodal large language models for visua l\nunderstanding and diffusion transformers for visual generation. We'll\n explore how these two areas are closely connected\, along with the\nchalle nges and opportunities in building powerful and scalable visual\nintellige nce. Plus\, we'll look at why these developments matter\, both\nin practic al applications and as key steps toward creating robust\nvisual intelligen ce that can better understand and interact with the\nsensory-rich world ar ound us. \n\n— \n\nSaining Xie is an Assistant Professor of Computer S cience at the\nCourant Institute of Mathematical Sciences at New York Univ ersity and\nis affiliated with NYU Center for Data Science. He is also a v isiting\nfaculty researcher at Google DeepMind. Before joining NYU in 2023 \, he\nwas a research scientist at FAIR\, Meta. In 2018\, he received his Ph.D.\ndegree in computer science from the University of California San\nD iego. He works in computer vision and machine learning\, with a\nparticula r interest in scalable visual representation learning for\nmultimodal unde rstanding and generation. His work has been recognized\nwith the Marr Priz e honorable mention\, CVPR best paper finalists and\nan Amazon research aw ard. \n\nThe VASC Seminar is sponsored in part by Meta Reality Labs Pitts burgh\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cdd3b DTSTART;TZID=America/New_York:20241021T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241021T130000 URL:https://www.cylab.cmu.edu/events/2024/10/21-seminar-bao.html LOCATION:Mehrabian Collaborative Innovation Center\, Panther Hollow Room 41 05 SUMMARY:CyLab Seminar - Tiffany Bao CLASS:PUBLIC DESCRIPTION:Speaker: TIFFANY BAO\, Assistant Professor\, School of Computin g and\nAugmented Intelligence\, Associate Director of Research Acceleratio n\,\nGlobal Security Initiative\, \, Center for Cybersecurity & Trusted\nF oundations \, Arizona State University\,\n\nTalk Title: ABC of Harnessing Hacker Insights for Vulnerability\nDiscovery: Automation\, Broadening\, an d Correction\n\nAs defenses in software and operating systems become incre asingly\nsophisticated\, the task of identifying vulnerabilities in modern \napplications and systems has evolved into a pursuit reserved for an\neli te group of highly skilled hackers. These experts bring extensive\nexperie nce\, specialized expertise\, and sharp intelligence to their\nwork. Howev er\, the scalability of human efforts is inherently limited\,\nerror-prone \, and often restricted to specific areas of expertise. \n\nIn this talk\ , I will share insights from my experiences in harnessing\nthe unique pers pectives of these hackers. I will explore scientific\napproaches to automa te the discovery process\, enhance accuracy\, and\nexpand the range of exp ertise\, ultimately increasing our ability to\ndetect vulnerabilities in r eal-world applications. \n\n— \n\nTiffany Bao is an Assistant Professo r at Arizona State University\,\nwhere she focuses on various aspects of s oftware security. Her\nresearch explores innovative techniques\, measureme nts\, and strategies\nfor discovering\, assessing\, and defending against software\nvulnerabilities. Currently\, she is a member of Shellphish Suppo rt\nSyndicate\, a finalist for the AIxCC competition. As a member of the\n Order Of Overflow\, she organized the DEFCOM CTF from 2018 to 2021.\nTiffa ny earned her Ph.D. from Carnegie Mellon University. She is an\nCyLab alum na as well as a formal soda person. \n\nFaculty Host: Fei Fang \n\nIn Pe rson and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ce131 DTSTART;TZID=America/New_York:20241021T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241021T123000 LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Accessibility Seminar - Mina Huh CLASS:PUBLIC DESCRIPTION:Speaker: MINA HUH\n\nGuest Speaker\n\nMina Huh\, Ph.D. Student\ , Department of Computer Science\, University of\nTexas Austin Mina is a P h.D. student at UT Austin working with Prof.\nAmy Pavel on AI for Accessib le Creativity. She will present her\nprojects exploring the use of generat ive AI to make video editing\n(AVscript - CHI'23)\, image generation (GenA ssist - UIST'23 Best Paper\nAward)\, and web design (DesignChecker - UIST' 24) more accessible. \n\nIn Person and Zoom Participation.  See announce ment.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ce42b DTSTART;TZID=America/New_York:20241021T081500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241023T170000 URL:https://2024.iavvc.org LOCATION:Cohon University Center\, Carnegie Mellon University SUMMARY:IAVVC2024: IEEE International Automated Vehicle Validation Conferen ce CLASS:PUBLIC DESCRIPTION:IAVVC 2024 focuses on all aspects related to research\, develop ment\,\nand applications of vehicle and infrastructure connectivity – bo th\nwith a focus on vehicle electrification and vehicle automation.     \n\nThis year's conference theme:  Large Deployment and Large Models\n\nA mong the Distinguished Speakers\n\nMykel Kochenderfer\,  Director\, SAIL- Toyota Center for AI Research\, \nStanford UniversityScott Moura\, Direct or\, PATH\,  University of\nCalifornia\, BerkeleyHenry Liu\, Director\, M City\, University of\nMichiganConrad Tucker\, Director\, CMU-Africa\, Carn egie Mellon\nUniversityRobert Hampshire\, Deputy Assistant Secretary for R esearch\nand Technology and Chief Science Officer\, Department of\nTranspo rtationPhil Koopman\, Department of Electrical and Computer\nEngineering\, Carnegie Mellon UniversityNat Beuse\, Chief Safety\nOfficer\, Aurora Inno vationH. Eric Tseng\,  University Professor\,\nMember of U.S. National Ac ademy of Engineering\, The University of\nTexas at Arlington\n\nConference Chairs\n\nDing Zhao\, General Chair\, Carnegie Mellon University \n\nRaj Rajkumar\, General Chair\, Carnegie Mellon University \n\nJohn Dolan\, T echnical Program Committee\, Carnegie Mellon University\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ce7d3 DTSTART;TZID=America/New_York:20241018T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241018T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Remote Access - Zoom Webinar SUMMARY:AI Institute for Societal Decision Making Seminar - Tom Manzini CLASS:PUBLIC DESCRIPTION:Speaker: TOM MANZINI\, Ph.D. Student\, Machine Learning & Robot ics for\nDisaster Response\, Department of Computer Science and Engineerin g\,\nTexas A&M University\n\nTalk Title: CRASAR-U-DROIDs: A Large Scale Be nchmark Dataset for\nBuilding Alignment and Damage Assessment in Georectif ied sUAS Imagery\n\nREGISTER\n\n→ Link information provided upon registr ation. In Person and Zoom\nParticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ceabf DTSTART;TZID=America/New_York:20241016T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241016T203000 URL:https://community.cmu.edu/customquickevents?id=a3pUk00000021Xt&custom=t rue LOCATION:Kōbo Pizza at Redhook Brewlab\, 714 E Pike Street Seattle\, WA 98 122 SUMMARY:School of Computer Science Seattle Alumni Mixer CLASS:PUBLIC DESCRIPTION:Join us for an alumni mixer in Seattle's Capitol Hill neighborh ood! \n\nThe event will include catering by Kōbo Pizza and drinks by Red hook\nBrewlab. \n\nSCS legend\, Mark Stehlik - University Teaching Profes sor\, Assistant\nDean for Outreach and Director of the CSD Undergraduate P rogram\, and\nCo-Founder\, CMU CS Academy - will be in attendance. \n\nDo n't miss this opportunity to catch up with old friends and meet new\nones.  \n\nREGISTER    \n\n → Registration will close at 5:00 pm - PDT on Friday\, October 11\,\n2025    \n\n→ Space is limited so RSVP today! \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cede9 DTSTART;TZID=America/New_York:20241015T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241015T190000 URL:https://www.carnegielibrary.org/event/cmu-and-clp-presents-ai-panel-dis cussion/ LOCATION:South Wing Reading Room\, Carnegie Library\, 4400 Forbes Avenue SUMMARY:AI Institute for Societal Decision Making: Public Speaker Series CLASS:PUBLIC DESCRIPTION:Speaker: Workforce and Economic Development and AI Panel Discus sion\n\nJoin us for a panel discussion with CMU’s NSF-funded AI Institu te\nfor Societal Decision Making (AI-SDM) to discuss the impact of\nArtif icial Intelligence (AI) and Generative AI on industries across\nthe Pittsb urgh region. \n\nREGISTER  →  suggested for in-person attendance\n\nV IRTUAL REGISTRATION  → attendees must register in advance receive\nthe meeting link\n\nParticipation is free.\n\nEvent Website and Registration\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cf127 DTSTART;TZID=America/New_York:20241015T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241015T153000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-mengchieh-jeremy- lee LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Proposal - Meng-Chieh (Jeremy) Lee CLASS:PUBLIC DESCRIPTION:Speaker: MENG-CHIEH (JEREMY) LEE\, Ph.D. Student\, Computer Sci ence\nDepartment\, Carnegie Mellon University\n\nTalk Title: Explainable M ining of Graphs and Time Series: Algorithms\nand Applications\n\nGiven a s ocial network\, how can we predict the connections between\nusers and dete rmine whether they are based on shared hobbies or common\nfriends? Similar ly\, how can we identify anomalies in time series data\nand explain why th ey are suspicious? Although recent machine learning\nmodels with improved performance are being developed\, they are often\nblack-box methods that a re difficult to interpret. This leads us to\nexplainable artificial intell igence (XAI)\, which offers valuable\ninsights through its explanations.   \n\nIn this thesis proposal\, we introduce carefully designed explaina ble\nmethods tailored for graph and time series data\, with diverse\nappli cations. Each method we proposed is either inherently\nexplainable\, or pr ovides explanations for the data or decisions it\nmakes.  \n\nThesis Com mittee \n\nChristos Faloutsos (Co-Chair)\n\nLeman Akoglu (Co-Chair)\n\nGe offrey Gordon\n\nNina Mishra (Amazon)\n\nAdditional Information\n\nIn Pers on and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cf4ea DTSTART;TZID=America/New_York:20241015T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241015T130000 URL:http://www.cs.cmu.edu/~aiseminar/ LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Artificial Intelligence Seminar - Xiangxiang Xu CLASS:PUBLIC DESCRIPTION:Speaker: XIANGXIANG XU\, Postdoctoral Associate\, Department of \nElectrical Engineering and Computer Science\, Massachusetts Institute\no f Technology\n\nTalk Title: Dependence Induced Representation Learning\n\n Despite the vast progress in deep learning practice\, theoretical\nunderst andings of learned feature representations remain limited. In\nthis talk\, we discuss three fundamental questions from a unified\nstatistical perspe ctive:\n\n What representations carry useful information? How are\nrepre sentations learned from distinct algorithms related? Can we\nseparate rep resentation learning from solving specific tasks?\n\nIn particular\, we fo rmalize representations that extract statistical\ndependence from data\, t ermed dependence-induced representations. We\nprove that representations a re dependence-induced if and only if they\ncan be learned from specific fe atures defined by\nHirschfeld–Gebelein–Rényi (HGR) maximal correlatio n. This\nseparation theorem signifies the key role of HGR features in\nrep resentation learning and enables a modular design of learning\nalgorithms. We further introduce the algorithm design for learning HGR\nfeatures and demonstrate how their mathematical structures enable them\nto simultaneous ly achieve several design objectives\, including minimal\nsufficiency (Tis hby's information bottleneck)\, information\nmaximization\, enforcing unco rrelated features (VICReg)\, and encoding\ninformation at different granul arities (Matryoshka representation\nlearning). We demonstrate that based o n HGR features\, we can obtain\nvarious representations learned by existin g practices\, including\ncross-entropy or hinge loss minimization\, non-ne gative feature\nlearning\, neural density ratio estimators\, and their reg ularized\nvariants. Our development also provides a statistical interpreta tion\nof the neural collapse phenomenon observed in deep classifiers. We\n conclude the talk by discussing the implications of our analyses\,\ninclud ing hyperparameter tuning during inference. \n\n— \n\nXiangxiang Xu re ceived the B.Eng. and Ph.D. degrees in electronic\nengineering from Tsingh ua University\, Beijing\, China\, in 2014 and\n2020\, respectively. He is a postdoctoral associate in the Department\nof EECS at MIT. His research f ocuses on information theory\,\nstatistical learning\, representation lear ning\, and their applications\nin understanding and developing learning al gorithms. He is a recipient\nof the 2016 IEEE PES Student Prize Paper Awar d in Honor of T. Burke\nHayes and the 2024 ITA (Information Theory and App lications) Workshop\nSand Award. \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cf978 DTSTART;TZID=America/New_York:20241014T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241014T163000 URL:https://www.ri.cmu.edu/event/high-fidelity-neural-radiance-fields/ LOCATION:Newell-Simon 3305 SUMMARY:Vision and Autonomous Systems Seminar - Christian Richardt CLASS:PUBLIC DESCRIPTION:Speaker: CHRISTIAN RICHARDT\, Research Scientist Lead\, Meta Re ality\nLabs Research\, Pittsburgh\n\nTalk Title: High-Fidelity Neural Radi ance Fields\n\nI will present three recent projects that focus on high-fi delity\nneural radiance fields for walkable VR spaces: \n\nVR-NeRF (SIGG RAPH Asia 2023) is an end-to-end system for the\nhigh-fidelity capture\, m odel reconstruction\, and real-time rendering\nof walkable spaces in virtu al reality using neural radiance fields. To\nthis end\, we designed and bu ilt a custom multi-camera rig to densely\ncapture walkable spaces in high fidelity and with multi-view high\ndynamic range images in unprecedented q uality and density. To\nrepresent highly detailed scenes\, we introduce a novel perceptual\ncolor space for learning accurate HDR appearance\, and a n efficient\nmip-mapping mechanism for level-of-detail rendering with\nant i-aliasing. Our multi-GPU renderer enables high-fidelity volume\nrendering at the full VR resolution of dual 2K×2K at 36 Hz on our\ncustom demo mac hine.  \n\nHybridNeRF (CVPR 2024 Highlight) leverages the strengths of\ nNeRF-style volumetric rendering and SDF-style surface representations\nby rendering most objects as surfaces while modeling the (typically)\nsmall fraction of challenging regions volumetrically. We evaluate\nHybridNeRF ag ainst the challenging Eyeful Tower dataset along with\nother commonly us ed view synthesis datasets. When compared to\nstate-of-the-art baselines\, including recent rasterization-based\napproaches\, HybridNeRF improves er ror rates by 15–30% while\nachieving real-time framerates (at least 36 F PS) for virtual-reality\nresolutions (2K×2K). \n\nSpecNeRF (CVPR 2024 H ighlight) proposes a learnable Gaussian\ndirectional encoding to better mo del view-dependent effects under\nnear-field lighting conditions. Importan tly\, our new directional\nencoding captures the spatially-varying nature of near-field lighting\nand emulates the behavior of prefiltered environme nt maps. As a\nresult\, it enables the efficient evaluation of preconvolve d specular\ncolor at any 3D location with varying roughness coefficients. We\nfurther introduce a data-driven geometry prior that helps alleviate\nt he shape radiance ambiguity in reflection modeling. \n\n— \n\nChristia n Richardt is a Research Scientist at Meta Reality Labs\nResearch in Pitts burgh\, PA. His research combines insights from\nvision\, graphics and per ception to reconstruct visual information from\nimages and videos\, to cre ate high-quality visual experiences with a\nfocus on VR experiences. Chris tian was previously an Associate\nProfessor and EPSRC-UKRI Innovation Fell ow in the Visual Computing\nGroup and the CAMERA Centre at the University of Bath\, UK. Before\nthat\, he was a postdoc at the Intel Visual Computin g Institute at\nSaarland University and Max-Planck-Institut für Informati k in\nSaarbrücken\, Germany. Previously\, he was a postdoc in the REVES t eam\nat Inria Sophia Antipolis\, France. \n\nChristian graduated with a P hD and BA from the University of Cambridge\nin 2012 and 2007\, respectivel y. His doctoral research investigated the\nfull life cycle of RGBD videos: from their acquisition\, via filtering\nand processing\, to the evaluatio n of stereoscopic display.\n\nThe VASC Seminar is sponsored in part by Met a Reality Labs Pittsburgh\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91cfebd DTSTART;TZID=America/New_York:20241014T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241014T200000 URL:https://aihorizonspgh.com/ LOCATION:Bakery Square SUMMARY:AI Horizons Pittsburgh Summit CLASS:PUBLIC DESCRIPTION:Speaker: Human-First AI Innovations for a Better World\n\nHome to the world’s #1 AI school\, Carnegie Mellon\, this inaugural\nsummit s howcases how human-first AI can be commercialized to benefit\nsociety – with profit as a byproduct. \n\nREGISTER \n\nExplore the transformative potential of AI for societal good in a city\nglobally renowned for its hum ility\, grit\, and technological\nleadership. Share Your AI Innovations An d catalyze change - Are you a\nthought leader or innovator with groundbrea king insights on the\npositive potential of AI?  We invite you to share y our expertise at\nthe AI Horizons Pittsburgh Summit. Topics of interest in clude AI for\nPeace in Our Time\, AI for a Greener World\, AI to Close the Digital\nDivide\, and AI for Better Public Policy. Join us in shaping the future\nof AI for the betterment of humanity. \n\nSpeakers Include\n\nFa rnam Jahanian  —  President\, Carnegie Mellon UniversityJosh\nShapiro   —  48th Governor of the Commonwealth of PennsylvaniaRich\nFitzgerald — Executive Director\, Southwestern Pennaylvania\nCommissionBruce Katz   —  Director\, Nowak Metro Finance Lab\, Drexel\nUniversityAndrew Moo re  —  Lovelace AILaurie Segall  — \nFounder\, Mostly Human Media | Former CNN Senior Tech\nCorrespondentMeredith Meyer Grelli  —  Direc tor\, Project Olympus\,\nSchool of Computer Science\, and Tepper School of Business\, Carnegie\nMellon UniversityTuomas Sandholm — Angel Jordan Un iversity Professor\nof Computer\, Carnegie Mellon UniversityAmeet Talwalke r —  Associate\nProfessor\, Machine Learning Department\, Carnegie Mell on\nUniversityRussell Schwartz —  Professor and Head\, Ray and Stephani e\nLane Department of Computational Biology\, Carnegie Mellon University\n \nand many other distinguished speakers. \n\nSponsored by:  Carnegie Mel lon University\, Governor Josh Shapiro\,\nDuolingo\, Google\, Black Tech N ation\, Partnership to Advance\nResponsible Technology\, University of Pit tsburgh\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d0295 DTSTART;TZID=America/New_York:20241011T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241011T143000 URL:https://www.cmu.edu/leadership/president/lecture-series/index.html LOCATION:Rangos Ballroom\, Cohon University Center SUMMARY:President's Lecture Series - Dr. Sethuraman Dr. Sethuraman \"Panch\ "\nPanchanathanPanchanathan CLASS:PUBLIC DESCRIPTION:Speaker: SETHURAMAN \"PANCH\" PANCHANATHAN\, Director\, U.S. Na tional\nScience Foundation\n\nDr. Sethuraman \"Panch\" Panchanathan is the Senior Vice-President for\nthe Office of Knowledge Enterprise Development (OKED) at ASU and is\nresponsible for advancing Research\, Entrepreneursh ip\, and Economic\nDevelopment activities at ASU. He is a Foundation Chair in Computing\nand Informatics in the School of Computing\, Informatics\, and Decision\nSystems Engineering\, and the Director of the Center for Cog nitive\nUbiquitous Computing (CUbiC) at ASU. He is a Professor of Computer \nScience and Engineering\, an Affiliate Professor in the School of\nElect rical\, Computer\, and Energy Engineering. He was also an Adjunct\nProfess or at the University of Arizona\, College of Medicine\, Phoenix.\nPanch wa s the Founding Director of the School of Computing and\nInformatics and wa s instrumental in founding the Biomedical\nInformatics Department at ASU. He was also the Chair of the Computer\nScience and Engineering Department. \n\nPanch is research interests are in the areas of designing ubiquitous\n computing environments for enhancing quality of life for individuals\nwith disabilities\; human-centered multimedia computing\; genomic signal\nproc essing\; analysis and recognition of face\, gait\, emotions\, events\nand activities\; media processor designs\; and haptic (touch-based) user\ninte rfaces. His research work has been funded by several federal and\nstate ag encies\, foundations and industry. He has published over 430\nresearch pap ers in refereed journals and conferences. He has mentored\nover 100 gradua te students\, post-docs\, research engineers\, and\nresearch scientists wh o occupy leading positions in academia and\nindustry worldwide. Panch co-f ounded two start-up companies MotionEase\nInc.\, and RehabDev LLC.\, focus ed on rehabilitative applications.\nCUbiC's flagship project\, iCARE won t he Arizona Governor's Innovator\nof the Year Academia Award in 2004. His t eam has also won other\nprestigious awards such as the Microsoft Imagine W orld Cup. \n\nPanch was the editor-in-chief of IEEE Multimedia Magazine a nd is on\nthe editorial board of twelve other professional journals. He ha s been\na guest editor of special issues and chair of many conferences. He is\na member of the executive committee of the Council of Research\nProgr ams and Graduate Education (CRPGE) of the Association of Public\nand Land Grant Universities (APLU). Panch was a member of the US-India\nBusiness Co uncil team that was part of President Obama's executive\nmission to India for exploring innovation and economic development\nopportunities. He was a member of the Electronic Health Steering\nCommittee appointed by the Gove rnor of Arizona to define the roadmap\nfor the future of e-health. He was also part of a select team of\nexperts invited by the Governor on her visi t to Canada to explore\nacademic\, industrial and governmental partnership s. Panch is a Fellow\nof the National Academy of Inventors\, Canadian Acad emy of Engineering\n(CAE)\, the Institute of Electrical and Electronics En gineers (IEEE)\nand the Society of Optical Engineering (SPIE).  He was ap pointed to\nthe National Science Board in 2014. \n\nPanch has a Bachelor' s degree in Physics and Electrical Communication\nEngineering from the Uni versity of Madras and Indian Institute of\nScience\, respectively. He has a Master's degree in Electrical\nEngineering from the Indian Institute of Technology\, Madras and a\nPh.D. in Electrical and Computer Engineering fr om University of\nOttawa\, Canada. \n\nREGISTER  →Event registrations close October 10 at 11:59 PM EDT\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d07c9 DTSTART;TZID=America/New_York:20241011T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241011T144500 URL:https://www.hcii.cmu.edu/news/event/2024/10/hcii-seminar-series-james-l anday LOCATION:Mauldin Auditorium\, Newell-Simon 1305 and Livestream SUMMARY:Human-Computer Interaction Institute Seminar - James Landay CLASS:PUBLIC DESCRIPTION:Speaker: JAMES LANDAY\, Professor of Computer Science\, and\, T he Anand\nRajaraman and Venky Harinarayan Professor\, School of Engineerin g\,\nStanford University\n\nTalk Title: “AI For Good” Isn’t Good Eno ugh: A Call for\nHuman-Centered AI\n\nAI for Good initiatives recognize th e potential impacts of AI systems\non humans and societies.  However\, si mply recognizing these impacts\nis not enough. To be truly Human-Centered\ , AI development must be\nuser-centered\, community-centered\, and societa lly-centered.\nUser-centered design integrates techniques that consider th e needs and\nabilities of end users\, while also improving designs through iterative\nuser testing. \n\nCommunity-centered design engages communiti es in the early stages of\ndesign through participatory techniques. Societ ally-centered design\nforecasts and mediates potential impacts on a societ al level\nthroughout a project.  Successful Human-Centered AI requires th e\nearly engagement of multidisciplinary teams beyond technologists\,\ninc luding experts in design\, the social sciences and humanities\, and\ndomai ns of interest such as medicine or law\, as well as community\nmembers. In this talk I will elaborate on my argument for an authentic\nHuman-Centere d AI. \n\n— \n\nJames Landay is a Professor of Computer Science and th e Anand\nRajaraman and Venky Harinarayan Professor in the School of Engine ering\nat Stanford University. He co-founded and is Co-Director of the\nSt anford Institute for Human-Centered Artificial Intelligence (HAI).\nLanday previously was a tenured faculty member at Cornell Tech\, the\nUniversity of Washington\, and UC Berkeley. \n\nHe was also Director of Intel Labs Seattle and co-founder of NetRaker.\nWhile on sabbatical at Microsoft Rese arch Asia in Beijing\, he taught\nfor one semester at Tsinghua University.   Landay received his BS in\nEECS from UC Berkeley\, and MS and PhD in Co mputer Science from\nCarnegie Mellon University. He is a member of the ACM SIGCHI Academy\,\nan ACM Fellow\, and was awarded the ACM SIGCHI Lifetime Research Award.\nHe served on the NSF CISE Advisory Committee for six yea rs. \n\nFaculty Host:  Jeff Bigham \n\nIn Person\, Zoom (Internal to CM U) and Livestream (All Guests)\nParticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d0c3d DTSTART;TZID=America/New_York:20241011T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241011T120000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Wean Hall 1327 SUMMARY:AI-SDM - Student Discussion Session - Melanece Wesley CLASS:PUBLIC DESCRIPTION:The AI Institute for Societal Decision Making (AI-SDM) —  wh ich\nbrings together AI and social sciences researchers to develop\nhuman- centric AI for societal good — hosts student-led brainstorming\ndiscussi on sessions. Our goal is to foster interdisciplinary\ncollaboration and ge nerate ideas on how AI can help solve societal\nproblems\, particularly fr om an angle of decision-making. \n\nThis week\, we welcome Melanece Wesle y\, Ph\,D Candidate in Social Work\,\nHoward University who will discuss A I Tools for Academic Research. \n\n— \n\nMelanece Wesley is Ph.D. stud ent and adjunct professor in Social Work\nat Howard University\, focusing on mental health equity\, health\ndisparities\, and the ethical applicatio n of Artificial Intelligence\n(AI). She holds Master and Bachelor of Socia l Work degrees from the\nUniversity of South Florida. Wesley has a strong background in\nbehavioral health advocacy and education. As a Behavioral H ealth\nAmbassador for the Substance Abuse and Mental Health Services\nAdmi nistration (SAMHSA)\, she played a pivotal role in designing and\ndeliveri ng educational programs that empowered undergraduate students\nat Historic ally Black Colleges and Universities (HBCUs) to pursue\ncareers in behavio ral health. She also served as a behavioral health\nmanager providing ther apeutic services for a monthly caseload of over\n80 patients. Her research interests lie at the intersection of AI\,\nethics\, and equity\, particul arly within Afro-Latine identity and\nmental health. As a member of the AI Institute for Societal\nDecision-Making\, Wesley contributes to research on AI risk\ncommunication for disaster management. She is actively involve d in\nscholarly work\, including a systematic review on the experiences of \nAfro-Latinos at HBCUs. A recognized leader\, she is also a member of\nth e AI-SDM Student Leadership Council\, Howard Universities Graduate\nResear ch Committee\, and has recently become a graduate fellow for the\nAmerican Association for Hispanics in Higher Education. \n\nIn Person and Zoom Pa rticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d10ac DTSTART;TZID=America/New_York:20241010T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241010T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates 8115 and Zoom SUMMARY:Crypto Seminar - Barak Nehoran CLASS:PUBLIC DESCRIPTION:Speaker: BARAK NEHORAN\, PhD Candidate\, Department of Computer Science\,\nPrinceton University\n\nTalk Title: Oracle Separation Between Quantum Commitments and Quantum\nOne-wayness\n\nWe show that there exists a unitary quantum oracle relative to which\nquantum commitments exist but no (efficiently verifiable) one-way\nstate generators exist. Both have bee n widely considered candidates\nfor replacing one-way functions as the min imal assumption for\ncryptography—the weakest cryptographic assumption i mplied by all of\ncomputational cryptography. Recent work has shown that c ommitments can\nbe constructed from one-way state generators\, but the oth er direction\nhas remained open. Our results rule out any black-box constr uction\,\nand thus settle this crucial open problem\, suggesting that quan tum\ncommitments (as well as its equivalency class of EFI pairs\, quantum\ noblivious transfer\, and secure quantum multiparty computation) appear\nt o be strictly weakest among all known cryptographic primitives. \n\nBased on joint work with John Bostanci and Boyang Chen \n\nReference Paper \n \nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d1421 DTSTART;TZID=America/New_York:20241010T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241010T160000 URL:https://aco.math.cmu.edu/abs-24-25/oct10.html LOCATION:Wean 8220 SUMMARY:ACO Seminar - Zongchen Chen CLASS:PUBLIC DESCRIPTION:Speaker: ZONGCHEN CHEN\, Assistant Professor\, School of Comput er\nScience\, Georgia Tech.\n\nTalk Title: Entropy Contractions in Markov Chains: Half-Step\,\nFull-Step and Continuous-Time\n\nGiven a transition m atrix (i.e.\, a row-stochastic matrix) one can\ndefine either a discrete-t ime Markov chain (with the multi-step\ntransition matrix given by the matr ix power) or a continuous-time\nMarkov process (roughly\, the update times are distributed as a Poisson\npoint process). A common way to analyze the speed of convergence of\nMarkov chains is to study the contraction of the relative entropy\n(i.e.\, the Kullback-Leibler divergence). Such entropy contractions are\ncharacterized by the strong data processing inequality f or\ndiscrete-time Markov chains\, or the modified log-Sobolev inequality\n for continuous-time Markov processes. \n\nIn several previous works these two notions of entropy contraction\nwere claimed to be equivalent to each other\, in the sense that the\nrates of contraction differ by universal c onstant factors. We disprove\nthis and related conjectures\, and summarize known comparisons among\ndifferent notions of entropy contraction. In par ticular\, we show that:\n(a) entropy contraction of a continuous-time Mark ov process can be\narbitrarily faster than its discrete-time counterpart\; (b) entropy\ncontraction of an (m+1)-step transition matrix can be arbitr arily\nfaster than the m-step version. \n\nJoint work with Pietro Caputo\ , Yuzhou Gu and Yury Polyanskiy. \n\n4:00 pm → Tea & Cookies in Wean 82 20 sponsored by Jane Street (bring\nyour own mug if possible).  \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d1845 DTSTART;TZID=America/New_York:20241010T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241010T150000 URL:https://csd.cmu.edu/calendar/speaking-skills-talk-carlos-martin LOCATION:Newell-Simon 4305 SUMMARY:Speaking Skills Talk - Carlos Martin CLASS:PUBLIC DESCRIPTION:Speaker: CARLOS MARTIN\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Finding Mixed-Strategy Nash Equilibria of Black-Box\nContinuous-Action Games Using Randomized Pol icy Networks and\nJoint-Perturbation Simultaneous Pseudo-Gradients\n\nWe s tudy the problem of computing an approximate Nash equilibrium of a\ncontin uous-action game without access to gradients of the utility\nfunction. Thi s problem arises for multi-agent reinforcement learning\nsettings where th e environment is treated as a black box. It also\narises for games whose p ayoffs are discontinuous with respect to\nplayers' actions\, such as aucti ons. To tackle this problem\, we apply\nzeroth-order optimization techniqu es that combine smoothed gradient\nestimators with equilibrium-finding dyn amics.\n\nWe model players' strategies using randomized policy networks. T hese\nnetworks take noise in addition to observations as input\, and can\n flexibly approximate arbitrary observation-dependent continuous-action\ndi stributions. This representation power is crucial for tackling\ncontinuous -action games that lack pure-strategy equilibria.\n\nIn addition\, we intr oduce a new training technique that reduces the\nnumber of utility functio n evaluations per iteration from linear to\nconstant in the number of play ers. It achieves this by performing a\nsingle joint perturbation of all pl ayers' strategies\, rather than\nperturbing each one individually. This yi elds dramatic benefits for\nmany-player games in which each evaluation inc urs costs in wall time\,\nmemory\, money\, or other resources.\n\nWe apply our technique to various auctions and resource-allocation\ngames. Experim ents show that our technique quickly finds high-quality\napproximate Nash equilibria of these games. \n\nThis is joint work with Tuomas Sandholm.\n \nPresented in Partial Fulfillment of the CSD Speaking Skills\nRequirement \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d1c6d DTSTART;TZID=America/New_York:20241010T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241010T180000 URL:https://na.eventscloud.com/website/65731/ LOCATION:Simmons Auditorium\, Tepper Building SUMMARY:AI Institutes Exposition and Engagement Showcase CLASS:PUBLIC DESCRIPTION:Speaker: National AI Research Institutes\n\nAI Institutes Expo Day 2024 is your chance to see the forefront of AI\nresearch! Take a peek into cutting-edge work in multiple fields and\nlearn how AI is transformi ng and improving our world.\n\nDiscover the latest in AI research at 25 In stitutes that are part of\nthe National AI Research Institutes program\, i ncluding demonstrations\nof how AI is being used to tackle agricultural ch allenges\, enhance\neducation\, and much more.\n\nConnect with 25 AI Insti tutes and partners and engage with their work\nInteractive displays\, visu al aids\, and manipulatives\, and thought\nleaders from the community Netw ork with 200+ attendees\, researchers\,\nand industry leaders10:00 am-12:3 0 pm → Expo\n12:30 pm-4:30 pm → Welcome\, Hot Topics and Panels\n\nThi s event is affordable and open to the public. REGISTER  |  AI\nInstitut es\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d1fc4 DTSTART;TZID=America/New_York:20241009T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241009T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20241009.html LOCATION:Traffic21 Classroom\, Gates Hillman 6501 (new location) SUMMARY:Theory Lunch Seminar - Zongrui Zou CLASS:PUBLIC DESCRIPTION:Speaker: ZONGRUI ZOU\, Nanjing University\n\nTalk Title: Differ entially Private Multiway and k-Cut\nMathJax.Hub.Config({\ntex2jax: {\ninl ineMath: [ ['$'\,'$'] ]\,\nprocessEscapes: true\n}\n})\;\n\nIn this talk\, we show how to address the challenge of differential\nprivacy in the cont ext of graph cuts\, specifically focusing on the\nminimum $k$-cut and mult iway cut problems. We introduce\nedge-differentially private algorithms th at achieve nearly optimal\nperformance for these problems. For the multiwa y cut problem\, we first\nprovide a private algorithm with a multiplicativ e approximation ratio\nthat matches the state-of-the-art non-private algor ithm. We then\npresent a tight information-theoretic lower bound on the ad ditive\nerror\, demonstrating that our algorithm on weighted graphs is\nne ar-optimal for constant $k$. For the minimum $k$-cut problem\, our\nalgori thms leverage a known bound on the number of approximate\n$k$-cuts\, resul ting in a private algorithm with optimal additive error\n$O(k\\log n)$ for fixed privacy parameter. We also establish a\ninformation-theoretic lower bound that matches this additive error.\nAdditionally\, we give an effici ent private algorithm for $k$-cut even\nfor non-constant $k$\, including a polynomial-time 2-approximation with\nan additive error of $\\tilde{O}(k^ {1.5})$.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d2382 DTSTART;TZID=America/New_York:20241007T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241007T173000 URL:https://db.cs.cmu.edu/events/building-blocks-paradedb-philippe-noel LOCATION:Remote Access - Zoom SUMMARY:Database Building Blocks Seminar - Philippe Noël CLASS:PUBLIC DESCRIPTION:Speaker: PHILIPPE NOËL\, co-founder and Chief Executive Office r\,\nParadeDB\n\nTalk Title: ParadeDB - Postgres for Search and Analytics\ n\nParadeDB is Postgres for search and analytics. It is an alternative to\ nElasticsearch built on Postgres. It offers state-of-the-art full-text\nan d vector search capabilities\, as well as fast aggregations inside\nPostgr es. ParadeDB is built in Rust via Postgres extensions on top of\ndatabase building blocks like Tantivy\, DuckDB\, and Apache DataFusion.\nIt is comp atible with every officially supported PGDG Postgres\nversion. In this tal k\, we'll discuss how we extended Postgres with\nthese building blocks and dive into the technical details of building\nPostgres extensions that int eract with Postgres storage\, planning and\nquery execution. \n\n— \n\ nPhilippe Noël is the co-founder and CEO of ParadeDB\, Postgres for\nsear ch and analytics. \n\nThis talk is part of the Database Building Blocks S eminar\n\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d2713 DTSTART;TZID=America/New_York:20241007T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241007T163000 URL:https://www.ri.cmu.edu/event/reconstructing-everything/ LOCATION:Newell-Simon 3305 and Zoom SUMMARY:VASC Seminar - Noah Snavely CLASS:PUBLIC DESCRIPTION:Speaker: NOAH SNAVELY\, Professor\, Computer Science Department \, Cornell\nUniversity and Cornell Tech\, and\, Research Scientist\, Googl e DeepMind\n\nTalk Title: Reconstructing Everything\n\nThe presentation wi ll be about a long-running\, perhaps quixotic effort\nto reconstruct all o f the world's structures in 3D from Internet\nphotos\, why this is challen ging\, and why this effort might be useful\nin the era of generative AI.  \n\n— \n\nNoah Snavely is a Professor in the Computer Science Departmen t at\nCornell University and Cornell Tech\, and a research scientist at\nG oogle DeepMind in NYC. Noah's research interests are in computer\nvision a nd graphics\, in particular in recovering structure from large\nphoto coll ections for use in understanding and visualizing the world\naround us. Noa h is the recipient of a PECASE\, a Microsoft New Faculty\nFellowship\, an Alfred P. Sloan Fellowship\, a SIGGRAPH Significant New\nResearcher Award\ , and is a Fellow of the ACM. The VASC Seminar is\nsponsored in part by Me ta Reality Labs Pittsburgh\n\n In Person and Zoom Participation.  See an nouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d2a9c DTSTART;TZID=America/New_York:20241007T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241007T130000 URL:https://www.cylab.cmu.edu/events/2024/10/07-seminar-bertino.html LOCATION:Newell-Simon 3305 and Zoom SUMMARY:CyLab Seminar - Elisa Bertino CLASS:PUBLIC DESCRIPTION:Speaker: ELISA BERTINO\, Samuel Conte Distinguished Professor o f\nComputer Science\, Purdue University\n\nTalk Title: Applying Machine Le arning to Securing Cellular Networks\n\nCellular network security is more critical than ever\, given the\nincreased complexity of these networks and the numbers of applications\nthat depend on them\, including telehealth\, remote education\,\nubiquitous robotics and autonomous vehicles\, smart c ities\, and\nIndustry 4.0. \n\nIn order to devise more effective defenses \, a recent trend is to\nleverage machine learning (ML) techniques\, which have become\napplicable because of today advanced capabilities for collec ting data\nas well high-performance computing systems for training of ML m odels.\nRecent large language models (LLMs) are also opening new interesti ng\ndirections for security applications. \n\nIn this talk\, I will first present a comprehensive threat analysis in\nthe context of 5G cellular ne tworks to give a concrete example of the\nmagnitude of the problem of cell ular network security. Then\, I will\npresent two specific applications of ML techniques for the security of\ncellular networks. \n\nThe first appl ication focuses on the use of natural language\nprocessing techniques to t he problem of detecting inconsistencies in\nthe \"natural language specifi cations\" of cellular network protocols.\nThe second application addresses the design of an anomaly detection\nsystem able to detect the presence of malicious base stations and\ndetermine the type of attack. Then I'll conc lude with a discussion on\nresearch directions. \n\n— \n\nElisa Bertin o is a Samuel Conte Distinguished Professor of Computer\nScience at Purdue University. She serves as Director of the Purdue\nCyberspace Security Lab (Cyber2Slab). Prior to joining Purdue\, she was\na professor and departme nt head at the Department of Computer Science\nand Communication of the Un iversity of Milan. She has been a visiting\nresearcher at the IBM Research Laboratory in San Jose (now Almaden)\,\nat Rutgers University\, at Telcor dia Technologies. She has also held\nvisiting professor positions at the S ingapore National University and\nthe Singapore Management University. Her recent research focuses on\nsecurity and privacy of cellular networks and IoT systems\, and on edge\nanalytics for cybersecurity. \n\nElisa Bertin o is a Fellow member of IEEE\, ACM\, and AAAS. She received\nthe 2002 IEEE Computer Society Technical Achievement Award for “For\noutstanding cont ributions to database systems and database security\nand advanced data man agement systems”\, the 2005 IEEE Computer\nSociety Tsutomu Kanai Award f or “Pioneering and innovative research\ncontributions to secure distribu ted systems”\, the 2019-2020 ACM\nAthena Lecturer Award\, and the 2021 I EEE 2021 Innovation in Societal\nInfrastructure Award. She received an Hon orary Doctorate from Aalborg\nUniversity in 2021 and an Honorary Research Doctorate in Computer\nScience from the University of Salerno in 2023. She is currently\nserving as ACM Vice-president. \n\nFaculty Host: Lorrie Cr anor \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d2fad DTSTART;TZID=America/New_York:20241004T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241004T120000 URL:https://www.cmu.edu/flame/ LOCATION:Tepper School 1403 SUMMARY:CMU Foundation and Language Model Seminar - Maarten Sap CLASS:PUBLIC DESCRIPTION:Speaker: MAARTEN SAP\, Assistant Professor\, Language Technolog ies\nDepartment\, Carnegie Mellon University\, and \, Research Scientist\, \nAllen Institute for AI\n\nTalk Title: Artificial Social Intelligence? On the challenges of\nSocially Aware and Ethically informed LLMs\n\nModern A I systems such as LLMs are pervasive and helpful\, but do they\nreally hav e the social intelligence to seamlessly and safely engage in\ninteractions with humans? \n\nIn this talk\, I will delve into the limits of social i ntelligence of\nLLMs and how we can measure and anticipate their risks. Fi rst\, I will\nintroduce Sotopia\, a new social simulation environment to e valuate the\ninteraction abilities of LLMs as social AI agents\, showing t hat even\ntoday's most powerful models struggle to achieve social goals in \ninteractions. \n\nThen\, I will shift to how LLMs pose new ethical chal lenges in their\ninteractions with users. Specifically\, through their lan guage modality\nand possible expressions of uncertainty\, we show that LLM s tend to\nexpress overconfidence in their answers even when incorrect\, w hich\nusers tend to over-rely on. \n\nFinally\, I will introduce Particip AI\, a new framework to anticipate\nfuture AI use cases and dilemmas. Thro ugh our framework\, we show that\nlay users can help us anticipate the ben efits and harms of allowing or\nnot allowing an AI use case\, paving the w ay for more democratic\napproaches to AI design\, development\, and govern ance. I will conclude\nwith some thoughts on future directions towards soc ially aware and\nethically informed AI. \n\n— \n\nMaarten Sap is an as sistant professor in Carnegie Mellon University's\nLanguage Technologies D epartment (CMU LTI)\, and a part-time research\nscientist at the Allen Ins titute for AI. His research focuses on\nmaking NLP systems socially intell igent\, and understanding social\ninequality and bias in language. He has presented his work in top-tier\nNLP and AI conferences\, receiving paper a wards or nominations at EMNLP\n2023\, ACL 2023\, FAccT 2023\, WeCNLP 2020\ , and ACL 2019. \n\nHis research has been covered in the press\, includin g the New York\nTimes\, Forbes\, Fortune\, Vox\, and more. Before joining CMU\, he was a\npostdoc/young investigator at the Allen Institute for AI ( AI2) on\nproject MOSAIC. He received his PhD from the University of\nWashi ngton's Paul G. Allen School of Computer Science & Engineering\nwhere he w as advised by Yejin Choi and Noah Smith.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d342d DTSTART;TZID=America/New_York:20241003T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241003T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room and Zoom SUMMARY:Crypto Seminar - Mahimna Kelkar CLASS:PUBLIC DESCRIPTION:Speaker: MAHIMNA KELKAR\, Ph.D. Student\, Department of Compute r\nScience\, College of Computing and Information Science\, Cornell\nUnive rsity\n\nTalk Title: Complete Knowledge: Preventing Encumbrance of\nCrypto graphic Secrets\n\nMost cryptographic protocols model a player’s knowled ge of secrets\nin a simple way. Informally\, the player knows a secret in the sense\nthat she can directly furnish it as a (private) input to a prot ocol\,\ne.g.\, to digitally sign a message.\n\nThe growing availability of Trusted Execution Environments (TEEs) and\nsecure multiparty computation\ , however\, undermines this model of\nknowledge. Such tools can encumber a secret sk and permit a chosen\nplayer to access sk conditionally\, withou t actually knowing sk. By\npermitting selective access to sk by an adversa ry\, encumbrance of\nsecrets can enable vote-selling in cryptographic voti ng schemes\,\nillegal sale of credentials for online services\, and erosio n of\ndeniability in anonymous messaging systems.\n\nUnfortunately\, exist ing proof-of-knowledge protocols fail to\ndemonstrate that a secret is une ncumbered. We therefore introduce and\nformalize a new notion called compl ete knowledge (CK). A proof (or\nargument) of CK shows that a prover does not just know a secret\, but\nalso has fully unencumbered knowledge\, i.e. \, unrestricted ability to\nuse the secret.\n\nWe introduce two practical CK schemes that use special-purpose\nhardware\, specifically TEEs and off- the-shelf mining ASICs. We prove\nthe security of these schemes and explor e their practical deployment\nwith a complete\, end-to-end prototype with smart-contract verification\nthat supports both. We show how CK can addres s encumbrance attacks\nidentified in previous work. Finally\, we introduce two new\napplications enabled by CK that involve proving ownership of\nbl ockchain assets.\n\nThis is joint work with Kushal Babel\, Philip Daian\,  James Austgen\,\nVitalik Buterin\, and Ari Juels.Reference PaperIn Perso n and Zoom\nParticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d383f DTSTART;TZID=America/New_York:20241003T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241003T160000 URL:https://aco.math.cmu.edu/abs-24-25/oct3.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Chris Eur CLASS:PUBLIC DESCRIPTION:Speaker: CHRIS EUR\, Assistant Professor\, Department of Mathem atical\nSciences\, Carnegie Mellon University\n\nTalk Title: Using Hodge –Riemann relations\n\nWe explain how one can use ideas from algebraic ge ometry to establish\ninequalities for combinatorial invariants.  We do so in two case\nstudies: Tutte polynomials of graphs and \"capacity polynomi als\" of\ncolorful matchings.  \n\nNo algebraic geometry background will be required.\n\n4:00 pm → Tea & Cookies to follow in Wean Hall 6220\, s ponsored by\nJane Street                \n\n → Bring you r own mug if you have one.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d3b5c DTSTART;TZID=America/New_York:20241003T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241003T123000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-junting-hsieh LOCATION:Blellock-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Doctoral Thesis Proposal - Jun-Ting Hsieh CLASS:PUBLIC DESCRIPTION:Speaker: JUN-TING HSIEH\, Ph.D. Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Spectral Algorithms fo r Optimization beyond the\nAverage-Case\n\nSpectral algorithms involve usi ng the eigenvalues and eigenvectors of\nmatrices derived from the input fo r algorithm design and analysis.\nThese techniques have achieved remarkabl e success across a wide range\nof computational problems. In particular\, through the study of random\nmatrices\, spectral methods have been widely used to solve problems in\naverage-case complexity\, where the input is dr awn from some random\nmodel. \n\nIn this proposal\, we explore the robust ness of spectral algorithms for\nhybrids between worst-case and average-ca se input models. Many\nexisting algorithms tend to “overfit” to the sp ecific randomness\nof the input – they fail even with slight perturbatio ns of the input\nmodels. First\, we show that spectral algorithms succeed in both\nrefuting semirandom constraint satisfaction problems (CSPs) and\n solving semirandom planted instances\, generalizing results previously\nkn own only for fully random CSPs. Second\, we demonstrate how upper\nbounds on the second eigenvalue of the adjacency matrix suffice for\nfinding larg e independent sets in 3-colorable graphs\, extending\nexisting results for random 3-colorable graphs.\n\nFinally\, for future work\, we aim to exten d such ideas to worst-case\ninstances by understanding the minimal assumpt ions necessary for\nalgorithmic success. Moreover\, we will explore the li mitations of\nspectral algorithms and investigate formal connections betwe en\nspectral algorithms and low-degree polynomials of the input. \n\nThes is Committee \n\nPravesh K. Kothari (Carnegie Mellon University/Princeton University)\n\nRyan O'Donnell\n\nJason Li\n\nVenkatesan Guruswami (Univer sity of California\, Berkeley)\n\nDavid Steurer (ETH\, Zürich)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d3fcf DTSTART;TZID=America/New_York:20241002T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241002T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20241002.html LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Lunch Seminar - Jiatu Li CLASS:PUBLIC DESCRIPTION:Speaker: JIATU LI\, Ph.D. Student\, Theory Group\, Computer Sci ence and\nArtificial Intelligence Laboratory\, Massachusetts Institute of\ nTechnology\n\nTalk Title: Yao's lemma is all you need (for derandomizatio n)\n\nThis work revisits the study of two classical technical tools in\nth eoretical computer science: Yao's transformation of distinguishers\nto nex t-bit predictors (FOCS 1982)\, and the \"reconstruction paradigm\"\nin pse udorandomness (e.g.\, as in Nisan and Wigderson\, JCSS 1994).\nRecent work s of Pyne\, Raz\, and Zhan (FOCS 2023) and Doron\, Pyne\, and\nTell (STOC 2024) showed that both of these tools can be derandomized\nin the specific context of read-once branching programs (ROBPs)\, but\nleft open the ques tion of derandomizing them in more general settings.\n\nOur main contribut ions give appealing evidence that derandomization of\nthe two tools is pos sible in general settings\, show surprisingly\nstrong consequences of such derandomization\, and reveal several new\nsettings where such derandomiza tion is unconditionally possible for\nalgorithms stronger than ROBPs (with useful consequences). In\nparticular\, we prove that deterministically so lving the search problem\nof producing a transformation from distinguisher s to next-bit\npredictors (i.e. derandomizing Yao's lemma) if equivalent t o generic\nderandomization prBPP = prP. \n\nThis is a joint work with Ed ward Pyne (MIT) and Roei Tell (University\nof Toronto) to appear in FOCS 2 024. \n\n→  Full version of the paper.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d43c0 DTSTART;TZID=America/New_York:20241001T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241001T130000 URL:https://db.cs.cmu.edu/events/db-seminar-json-relational-duality-converg ing-the-worlds-of-objects-documents-and-relational/ LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 SUMMARY:Database Seminar - Tirthankar Lahiri CLASS:PUBLIC DESCRIPTION:Speaker: TIRTHANKAR LAHIRI\, Senior Vice President\, Mission-Cr itical\nData and AI Engines\, Oracle\n\nTalk Title: JSON Relational Dualit y: Converging the worlds of Objects\,\nDocuments\, and Relational\n\nThe \ "Object-Relational Impedance Mismatch\" has been a multi-decade\nproblem f or developers\, and past solutions have all had various\ntradeoffs that ha ve compromised efficiency or consistency.  JSON\nRelational Duality is a breakthrough capability that combines the best\naspects of the Document mo del and the Relational models without the\ndrawbacks of either model. \n\ nThis session will provide an overview and deep dive into the inner\nworki ngs of JSON Relational Duality. We will also discuss some of the\nbenefits of being able to easily access data in terms of Application\nObjects\, wh ile retaining the full power of the Relational Model (which\nincludes supp ort for JSON documents as well) as the underlying storage\nmodel. \n\n—  \n\nTirthankar Lahiri is Senior Vice President of Mission-Critical Data\ nand AI Engines at Oracle. He has a B.Tech in Computer Science from IIT\nK haragpur and an MS in Electrical Engineering from Stanford\nUniversity\, a nd holds 66 patents. \n\nIn Person and Zoom Participation.  See announce ment.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d4753 DTSTART;TZID=America/New_York:20241213T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241213T170000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Student Review (DSR) - General Meeting CLASS:PUBLIC DESCRIPTION:Speaker: Doctoral Student Review (DSR) Talk Title: Doctoral Stu dent\nReview (DSR) General Meeting\n\nGeneral Meeting\n\n(Years 1-3)\n\nSe e email notifications.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d4a17 DTSTART;TZID=America/New_York:20241212T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241212T170000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Student Review (DSR) - PM Research Areas CLASS:PUBLIC DESCRIPTION:Speaker: Doctoral Student Review (DSR) Talk Title: Doctoral Stu dent\nReview (DSR) AI/Graphics/Theory\n\nAI/Graphics/Theory\n\n(Years 4-N) \n\nSee email notifications.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d4cd3 DTSTART;TZID=America/New_York:20241212T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241212T120000 LOCATION:Gates and Hillman Centers SUMMARY:Doctoral Student Review (DSR) F24 - AM Research Areas CLASS:PUBLIC DESCRIPTION:Speaker: Doctoral Student Review (DSR) Talk Title: Doctoral Stu dent\nReview (DSR) PL/Security/Systems\n\nPL/Security/Systems \n\n(Years 4-N)\n\nSee email notifications.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d4f89 DTSTART;TZID=America/New_York:20241210T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241210T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting - December 2024 CLASS:PUBLIC DESCRIPTION:Talk Title: CSD Faculty Meeting\n\nSee email notification.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d520f DTSTART;TZID=America/New_York:20241113T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241113T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting - November 2024 CLASS:PUBLIC DESCRIPTION:Talk Title: CSD Faculty Meeting\n\nSee email notification. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d54a0 DTSTART;TZID=America/New_York:20241008T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20241008T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting - October 2024 CLASS:PUBLIC DESCRIPTION:Talk Title: CSD Faculty Meeting\n\nSee email notification. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d573b DTSTART;TZID=America/New_York:20240911T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240911T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting - Sepember 2024 CLASS:PUBLIC DESCRIPTION:Talk Title: CSD Faculty Meeting\n\nSee email notification.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d59ef DTSTART;TZID=America/New_York:20240930T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240930T173000 URL:https://db.cs.cmu.edu/events/building-blocks-apache-datafusion-comet-an dy-grove LOCATION:Remote Access - Zoom SUMMARY:Database Seminar - Andy Grove CLASS:PUBLIC DESCRIPTION:Speaker: ANDY GROVE\, Apache Arrow \, and\, Apache DataFusion P MC Member\n\nTalk Title: Accelerating Apache Spark workloads with Apache D ataFusion\nComet\n\nApache Spark is one of the most widely-used distribute d data analysis\nframeworks. However\, its JVM-based and row-oriented quer y execution\nengine limits Spark’s performance and scalability. \n\nIn this talk\, we will introduce DataFusion Comet\, an accelerator for\nApach e Spark designed to improve the efficiency of Spark queries by\ntranslatin g them into native queries that leverage Apache Arrow and\nApache DataFusi on. We will explore the core architecture of Comet and\nexplain how Spark plans are translated into native plans and talk\nabout some of the challen ges of providing Spark compatibility. \n\n— \n\nAndy Grove is an Apach e Arrow & Apache DataFusion PMC Member and the\noriginal creator of Apache DataFusion. \n\nThis talk is part of the Database Building Blocks Semina r\n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d5d51 DTSTART;TZID=America/New_York:20240927T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240927T220000 URL:https://www.cmu.edu/family/events/family-weekend/index.html LOCATION:Various Locations and Events on Campus SUMMARY:Fall '24 Family Weekend! CLASS:PUBLIC DESCRIPTION:Family Weekend brings together undergraduate and graduate stude nts\,\nfamilies and friends for several days to experience the vibrancy of \nthe Carnegie Mellon campus! Attending Family Weekend is free\, however\n some events require registration or purchased tickets. \n\nNOTE: →  Pl ease include your CMU student in your family’s\nregistration. →  The full event schedule can be found on the CMU\nEvents App.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d6090 DTSTART;TZID=America/New_York:20240927T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240927T130000 URL:https://www.cylab.cmu.edu/events/2024/09/27-seminar-esteves-ver%C3%ADss imo.html LOCATION:Newell-Simon 4305 and Zoom SUMMARY:CyLab Seminar - Paulo Esteves-Veríssimo CLASS:PUBLIC DESCRIPTION:Speaker: PAULO ESTEVES-VERÍSSIMO\, Paulo Esteves-Veríssimo\, Professor\nof Computer Science\, and\, Director of the Resilient Computing and\nCybersecurity Center (RC3)\, King Abdullah University of Science and \nTechnology (KAUST)\n\nTalk Title: Safe and Secure AI/ML-driven Autonomou s Vehicles? Not\nanywhere near yet...\n\nCurrently\, there is a substantia l push towards autonomous vehicles\n(AV) in the market. However\, largely or autonomous vehicles\, though\nusing extensive fault-tolerance e.g.\, in x-by-wire functions\, are\nstill not quite safe from an accidental faults ’ perspective\, leave\nalone malicious faults. \n\nThe growing number o f recent AV architectures hinging on the\nadvancements in AI/ML models\, g iven their functional power\, must be\nput in context with an equally sign ificant number of related serious\nor fatal accidents. I advance by presen ting important misconceptions\nand pitfalls originating from the use of in appropriate AI/ML\ntechniques in the AV area\, which may be the cause of s erious\naccidents. \n\nFurthermore\, AV present an even greater threat su rface to combined\naccidental faults and malicious attacks. These pitfalls have been very\nslowly recognized by car makers\, with potentially harmin g results. \n\nHowever\, simultaneously securing unavoidable properties o f Safety and\nSecurity is indeed a hard problem. I raise a bit of the curt ain on how\nto break this chicken and egg dilemma—closing the safety-sec urity\ngap—presenting some recent solution avenues based on cyber\nresil ience\, a core subject of my research. \n\nNamely\, I will discuss how in -car architectures might evolve to\nachieve fault and intrusion resilience of ECUs at system level\, as\nwell as reconciling the data-level stochast ic nature of AI/ML\nparadigms with the determinism of driving control theo ry at system\nlevel\, leveraging the best from both worlds: trustworthines s and\nintelligence. \n\n— \n\nPaulo Esteves-Veríssimo is a professor of Computer Science at KAUST\,\nfounding director of the Resilient Comput ing and Cybersecurity Center\n\, and research fellow at the University of Luxembourg (LU). He is\nFellow of IEEE and of ACM\, author of over 200 pee r-refereed\npublications and co-author of 5 books. He is currently interes ted in\nresilient computing\, in areas like: SDN-based infrastructures\;\n autonomous vehicles\; distributed control systems\; digital health and\nge nomics\; or blockchain and cryptocurrencies. \n\nFaculty Host: William Sa nders In Person and Zoom Participation.  See\nannouncement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d6577 DTSTART;TZID=America/New_York:20240926T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240926T160000 URL:https://aco.math.cmu.edu/abs-24-25/sep26.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Colin Tang CLASS:PUBLIC DESCRIPTION:Speaker: COLIN TANG\, Ph.D. Student\, Department of Mathematica l\nSciences\, Carnegie Mellon University\n\nTalk Title: Simplex slicing: a n asymptotically-sharp lower bound\n\nWe show that for the regular n-simpl ex\, the 1-codimensional central\nslice that's parallel to a facet will ac hieve the minimum area (up to\na 1-o(1) factor) among all 1-codimensional central slices\, thus\nimproving the previous best known lower bound (Brze zinski 2013) by a\nfactor of 2sqrt(3)/e≈1.27. In addition to the standar d technique of\ninterpreting geometric problems as problems about probabil ity\ndistributions and standard Fourier-analytic techniques\, we rely on a \nnew idea\, mainly *changing the contour of integration* of a\nmeromorphi c function. \n\n4:00 pm → Tea and Cookies to Follow in the Math Lounge (bring your\nown mug if you have one)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d68f0 DTSTART;TZID=America/New_York:20240925T170000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240925T180000 URL:https://www.cs.cmu.edu/scs-career-center/career-center-events LOCATION:Newell-Simon 3002 SUMMARY:Info Session: Computer Science MS (MSCS) and 5th Year Research MS\n Programs CLASS:PUBLIC DESCRIPTION:Speaker: PETER STEENKISTE and DAVID ECKHARDTAre you interested in\ncontinuing your education and obtaining a Computer Science Master's\nD egree? \nYes?  Attend this information session! \nProfessor Peter Steen kiste\, Professor and Program Director\, will\ndiscuss The 5th-Year Master of Science Program The Fifth Year\nMaster’s Program is a 12-month resea rch-focused program for\ngraduates of SCS bachelor’s programs.  Student s complete four\nPh.D.-level classes while working on thesis research with a research\nadvisor in SCS.  The program has application deadlines in Oc tober and\nJanuary\; admitted students can start in either of the two seme sters\nfollowing the semester in which they were admitted.Professor David\ nEckhardt\, Teaching Professor and Program Co-DIrector\, will discuss The\ nMaster of Science in Computer Science (MSCS) Program The Master of\nScien ce in Computer Science (MSCS) Program offers students with a\nBachelor’s degree the opportunity to improve their training with\nadvanced study in Computer Science (research is optional). We cater to\nstudents with basic analytic skills and a strong aptitude for\nmathematics\, programming and l ogical reasoning.  Students with SCS\nundergraduate minors are welcome to apply.  Application deadlines are\nDecember and July\, for students plan ning to start in August and\nJanuary.Additional Information 5th-Year Progr am Handbook MSCS Program\nHandbook\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d6c82 DTSTART;TZID=America/New_York:20240925T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240925T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20240925.html LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Joint Theory Lunch Seminar / Speaking Skills Talk - William He CLASS:PUBLIC DESCRIPTION:Speaker: WILLIAM HE\, Ph.D. Student\, Computer Science Departme nt\,\nCarnegie Mellon University\n\nTalk Title: Pseudorandomness Propertie s of Random Reversible Circuits\n\nMotivated by cryptography\, quantum inf ormation theory\, circuit\ncomplexity\, and derandomization\, there has be en significant recent\nprogress in the study of random permutations resemb ling a completely\nrandom permutation of n-bit strings. Of particular inte rest is the\ncase where the measure of \"resemblance\" is approximation of the k-th\nmoments of the matrix representations. Such random permutations are\nknown as approximate k-wise independent permutations. In this talk I \nwill discuss some recent results that show that small random\nreversible circuits compute approximate k-wise independent\npermutations:   \n\ni) We show that a random reversible circuit with Õ(nk) gates computes\na co nstant-approximate k-wise independent permutation. This result\nimplies a generalization of Shannon's circuit lower bound argument.\n  \n\nii) We show that a random reversible circuit with Õ(nk2) layers of\n1D-local gat es arranged in a brickwork architecture computes a exp(–\nnk)-approximat e k-wise independent permutation\; connections to block\nciphers exist. \ n\nBased on joint works with William Gay (CMU)\, Lucas Gretta (Berkeley)\, \nNicholas Kocurek (CMU)\, Ryan O'Donnell (CMU)\, Angelos Pelecanos\n(Berk eley). \n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nR equirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d7062 DTSTART;TZID=America/New_York:20240924T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240925T170000 URL:https://www.cylab.cmu.edu/events/2024/09/24-partners-conference.html LOCATION:CMU Campus SUMMARY:CyLab Partners Conference CLASS:PUBLIC DESCRIPTION:Attendees can look forward to brief talks\, poster sessions\, a nd ample\nopportunities for both students and faculty to connect with our\ npartners.\nStay tuned for additional event details as the date approaches !\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d730b DTSTART;TZID=America/New_York:20240923T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240923T163000 URL:https://www.ri.cmu.edu/event/stochastic-graphics-primitives/ LOCATION:Newell-Simon 3305 SUMMARY:Vision and Autonomous Systems Seminar - Bailey Miller CLASS:PUBLIC DESCRIPTION:Speaker: BAILEY MILLER\, Ph.D. Candidate\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Stochastic Graphics P rimitivesFor decades computer\ngraphics has successfully leveraged stochas ticity to enable both\nexpressive volumetric representations of participat ing media like\nclouds and efficient Monte Carlo rendering of large scale\ , complex\nscenes. In this talk\, we’ll explore how these complementary forms of\nstochasticity (representational and algorithmic) may be applied more\ngenerally across computer graphics and vision. In the first part of\ nthe talk\, I’ll discuss our work on rendering probabilistic\nrepresenta tions of 3D geometry\, which explains the connection between\nclassical vo lume rendering and more recent techniques like NeRF. For\nthe second part of the talk\, I’ll discuss our work on Monte Carlo\nsimulation where we ’ve developed accelerated random walk techniques\nfor physics simulation that are analogous to Monte Carlo rendering for\nlight transport.\n—Bai ley Miller is a PhD candidate in the Computer Science Department\nat Carne gie Mellon University where he is advised by Ioannis\nGkioulekas. He works on theory and core algorithms for stochastic\ngraphics primitives which a re leveraged in applications across both\ncomputer graphics and vision. He received his Bachelors in Mathematics\nand Computer Science from Dartmout h College in 2018 where he had the\nprivilege of working with Wojciech Jar osz. During his PhD\, Bailey has\ninterned with Adobe research\, the Explo ratory Design Group at Apple\,\nand the High-Fidelity Physics Research tea m at NVIDIA. He is a\nrecipient of the NSF Graduate Research Fellowship\, the NVIDIA Graduate\nResearch Fellowship\, a Best Paper award at SIGGRAPH 2024\, and a Best\nStudent Paper Honorable Mention award at CVPR 2024.The VASC Seminar is\nsponsored in part by Meta Reality Labs Pittsburgh\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d76d6 DTSTART;TZID=America/New_York:20240923T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240923T140000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-oral-defense-shuqi-dai LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Shuqi Dai CLASS:PUBLIC DESCRIPTION:Speaker: SHUQI DAI\, Ph.D. Candidate \, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Towards Artificial Music ians: Empowering Individual Music\nExpression In Composition\, Performance \, and Synthesis Through Machine\nLearning\n\nRecent advances in music tec hnology and generative AI have\nrevolutionized music creation\, transformi ng how we interact with music\nin various aspects of life. However\, achie ving high musicality and\ncustomizing music to individual preferences rema in significant\nchallenges. This thesis addresses five fundamental proble ms in\ncurrent AI-driven music understanding and creation: (1) multimodal\ nmusic representation\, (2) highly complex and logical music structure\,\n (3) stylistic and personalization controls\, (4) data scarcity and\ncopyri ght\, and (5) ethical concerns. This work integrates music\ndomain knowle dge with machine learning to effectively overcome\nthese obstacles\, by focusing on a practical application: creating\nvirtual musicians or \"re-c reating\" existing musicians.\n\nFirst\, guided by music expertise\, I int roduce novel algorithms that\nanalyze music data to identify and explore p rinciples underlying music\nexpression\, with a focus on music repetition and structure hierarchy.\nNext\, these principles are applied across three levels of music\ncreation: symbolic composition\, expressive performance control\, and\naudio synthesis. For symbolic composition\, both statistica l machine\nlearning and deep learning techniques are employed to compose\n melodies\, harmonies\, and bass lines that imitate specific music styles\n given examples. Expressive performance control\, highly crucial in\nmusic creativity but often ignored\, is realized through diffusion\nmodels that generate timing\, pitch\, dynamics\, and singing techniques.\nAudio synthe sis is demonstrated through singing synthesis\, which\ninvolves generating vocals from scratch and transferring vocal\ntimbres\, including zero-shot and cross-domain synthesis and conversion\nof unseen speech reference. Th ese approaches converge to model music\nexpression across multimodal music representations.\n\nThis thesis emphasizes individual music preference an d stylistic\nmodeling\, offering various controls for composition\, perfor mance\, and\nsynthesis. In symbolic composition\, controls range from micr o-level\nelements such as rhythm patterns and melodic contour\, to macro-l evel\nfeatures like song style\, structure\, and harmony. In singing\nperf ormance and synthesis\, controls include language\, style genre\, and\nsin ging techniques\, with zero-shot capabilities to customize specific\nvocal timbres.\n\nExperiments validate the effectiveness of the models\, demons trating\ncompetitive performance to human music. Ethical and legal\nconcer ns are also discussed. Finally\, I highlight potential\napplications for advancing these technologies in areas like music\ntherapy\, education\, hu man-computer interactive performance systems\,\nand the development of wor ld music theory.\n\nThesis Committee\n\nRoger B. Dannenberg (Chair)\n\nChr is Donahue\n\nJunyan Zhu\n\nJulius O. Smith (Stanford University)\n\nGus G uangyu Xia (Mohamed bin Zayed University of Artificial\nIntelligence)\n\nI n Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d7c34 DTSTART;TZID=America/New_York:20240920T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240920T180000 URL:https://www.cmu.edu/dietrich/statistics-datascience/stamps/events/index .html#launch LOCATION:Simmons Auditorium B\, Tepper Building and Remote Access SUMMARY:STAtistical Methods for the Physical Sciences Research Center Launc h -\nKeynote Talks CLASS:PUBLIC DESCRIPTION:Speaker: KYLE CRANMER and AMY BRAVERMANTalk Title: Two Talks\n\ n►  Dr. KYLE CRANMER      \n\n —  The intersection of statisti cs\, machine learning\, and the\nphysical sciences All models are wrong\, but some are useful” is a\nfamous quip by George Box recognizing that st atisticians commonly\nemploy models that are simplifications of the real w orld. Similarly\,\nphysicists are notorious for invoking ‘spherical cows ’ that\ncapture the essence of a system that is conducive to reasoning o r\npencil-and-paper calculations. While this approach has its place\, the\ ncutting-edge of the physical sciences are characterized by highly\ndetail ed simulations with many components interacting through some\nunderlying m echanistic model. Moreover\, physical scientists are often\nmost intereste d in inferring aspects of the mechanistic model itself\,\nbut this involve s very challenging inverse problems. In contrast\, the\nfield of machine l earning generally eschews interpretable\, generative\nmodels in favor of b lack-box models and an optimization perspective. \nI will describe how ma chine learning techniques\, despite their\nblack-box nature\, are empoweri ng a revolution in principled\nstatistical inference for the physical scie nces. \n\nKyle Cranmer is the David R. Anderson Director of the UW-Madiso n Data\nScience Institute and a Professor of Physics with courtesy\nappoin tments in Statistics and Computer Science. He is also the Editor\nin Chief of the journal Machine Learning Science and Technology.\nCranmer was a Pr ofessor of Physics and Data Science at NYU from 2007\n– 2022. He obtaine d his Ph.D. in Physics from the University of\nWisconsin-Madison in 2005. He was awarded the Presidential Early\nCareer Award for Science and Engine ering in 2007\, the National Science\nFoundation's Career Award in 2009\, and became a Fellow of the American\nPhysical Society in 2021 for his work at the Large Hadron Collider.\nProfessor Cranmer developed a framework th at enables collaborative\nstatistical modeling\, which was used extensivel y for the discovery of\nthe Higgs boson in 2012. His current interests are at the intersection\nof physics\, statistics\, and machine learning. \n\ n►  Dr. AMY BRAVERMAN      \n\n —  Statistical Challenges for the Next Generation of NASA's Earth\nObserving Satellites Remote sensing d ata sets produced by NASA and\nother space agencies are a vast resource fo r the study of climate\nchange and the physical processes that drive it. H owever\, no remote\nsensing instrument actually observes these processes d irectly\;\ninstruments collect electromagnetic spectra aggregated over\ntw o-dimensional ground footprints or three-dimensional voxels (or\nsometimes just at a single point location). Inference on physical\nstate based on t hese spectra occurs via complex\, computationally\nintensive ground data p rocessing algorithms. As we transition from the\nEarth Observing System (E OS\, circa 1999-2025) to the new Earth System\nObservatory (ESO\, circa 20 26) data volumes will explode. For example\,\nthe Surface Biology and Geol ogy (SBG) mission to be launched later\nthis decade\, will acquire million s of highly multivariate spectra per\nsecond over the global land surface of the Earth\, at 30-60 meter\nresolution. This rich content of this vast information asset will\nremain mostly impenetrable unless new methods appr opriate for these\ndata and the questions they address\, emerge. In this t alk\, I will give\nan overview the SBG mission\, its scientific and societ al goals\, and\nthree types of statistical challenges we face: science ana lysis\, data\nprocessing\, and uncertainty quantification. I will also des cribe our\ncurrent thinking on how to address these challenges\, and discu ss\nopportunities for collaboration. \n\nDr. Amy Braverman is a Senior Re search Scientist at the Jet Propulsion\nLaboratory\, California Institute of Technology. She holds a Ph.D.  in\nStatistics from UCLA\, and came to JPL as a post-doctoral scholar in\n1999. Prior to graduate school\, she wa s a Research Director at\nMicronomics\, Inc. in Los Angeles where she led teams preparing\nexhibits for complex civil litigation. Dr. Braverman work ed on various\nNASA missions in various capacities over her 25 years at th e Lab\,\nfirst in designing data reduction methods for massive remote sens ing\ndata sets\, and later expanding to address general statistical\nmetho dology and applications issues related to remote sensing. In 2012\nshe beg an working intensely on uncertainty quantification (UQ)\, and\nhas develop ed practical methods for UQ in high-throughput\, operational\ninverse prob lems of interest to NASA and JPL. She is now serving as\nthe Chair of the SIAM Activity Group on Uncertainty Quantification\,\naiming to bridge the gap between traditional math-based UQ and\nstatistics. Dr. Braverman is a Fellow of the American Statistical\nAssociation\, and is the recipient of the NASA Exceptional Public\nService Medal for her efforts to bring rigoro us UQ to the NASA science\nenterprise. She especially enjoys working with post-docs\, graduate\nstudents\, and academic colleagues to solve new stat istical research\nproblems relevant to Earth and Space sciences. \n\nREGI STER \n\n→  Coffee & Refreshments at 3:30 pm \n\n→   Reception fol lowing the event.\n\nSTAMPS – Many problems in the physical sciences sha re common\nstatistical challenges including heterogeneous data from multip le\nprobes\, uncertainty quantification\, ill-posed inverse problems\,\nsp atio-temporal data and complex simulations. In 2018\, a group of\nfaculty and students at CMU started the STAMPS research group to\ndevelop new stat istical and machine learning methodology tailored to\nthe unique challenge s that arise across multiple areas in the physical\nsciences. STAMPS provi des foundational methodology in statistics\, data\nscience\, machine learn ing and artificial intelligence for two distinct\nbranches of physical sci ence: (i) Astronomy and Particle Physics\, and\n(ii) Climate and Environme ntal Science\, which include applications in\ne.g. Oceanography\, Meteorol ogy\, and Remote Sensing. STAMPS has become\na vibrant forum for interdisc iplinary exchange at the intersection of\nstatistics and the physical scie nces and will become a CMU Research\nCenter in Fall 2024.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d83e3 DTSTART;TZID=America/New_York:20240920T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240920T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Gates Hillman 8102 and Zoom SUMMARY:AI Institute for Societal Decision Making Seminar - Howie Choset CLASS:PUBLIC DESCRIPTION:Speaker: HOWIE CHOSET\, Kavčić-Moura Professor of Computer Sc ience\,\nRobotics Institute\, Carnegie Mellon University\,\n\nTalk Title: Making Large Dimensional Problems Small Again\n\nMotion is all around us. Motion is particularly interesting when it\nhas many degrees of freedom. T his talk covers the design\, sensing\, and\nplanning for snake\, multi-age nt and modular robot high DOF systems.\nThus far\, each system requires di fferent fundamentals – geometric\nmechanics for snake robot locomotion\, deferred planning and ergodic\nsearch for multi-agent systems\, and novel generator and discriminator\nnetworks for modular robots – which will b e covered in this talk.\nWhile no grand unifying theory combines these app roaches\, they all\nshare one aspect in common: reduce complex high dimens ional problems\ninto low dimensional ones. In pursuit of this investigatio n in\nreduction\, my group has created several embedded systems - actuator s\nand edge sensors - to build and deploy robots that stress-tests the\nco re assumptions in the theory and demonstrates efficacy for\napplications o f national importance. These applications include\nminimally invasive surg ery\, urban search and rescue\, manufacturing\,\nassembly in low-Earth orb it\, maintenance of municipal infrastructure\,\nand agile recycling. This talk will discuss these confined space\napplications\, and if time permits \, the five spin off companies\, and\none manufacturing institute\, that m y colleagues and I co-founded to\ncommercialize the core technologies cove red in m this talk. —\n\nHowie Choset is a Professor of Robotics at Carn egie Mellon University\nwhere he serves as the co-director of the Biorobot ics Lab and as\ndirector of the Robotics Major. He received his undergradu ate degrees\nin Computer Science and Business from the University of Penns ylvania\nin 1990. Choset received his Masters and PhD from Caltech in 1991 and\n1996. Choset's research group reduces complicated high-dimensional\n problems found in robotics to low-dimensional simpler ones for design\,\na nalysis\, and planning. Motivated by applications in confined spaces\,\nCh oset has created a comprehensive program in modular\, high DOF\, and\nmult i- robot systems\, which has led to basic research in mechanism\ndesign\, path planning\, motion planning\, and estimation. This work has\nbeen supp orted by both industry and government\; DOD support includes\ntwo MURIs\, one of which Choset was the CO-PI\, a young investigator\naward\, and mult i-PI awards for modular systems. Choset's group has\nproduced over 60 jour nal papers (including 2 in Science and one in\nProceedings of the National Academies of Sceince)\, 180 conference\npapers and 15 patents. Choset's w ork has also been recognized by\nseveral best paper awards and nominations at ICRA\, IROS and other\nrobotics meetings. Choset's research program ha s made contributions to\nchallenging and strategically significant problem s in diverse areas\nsuch as surgery\, manufacturing\, infrastructure inspe ction\, and search\nand rescue. In addition to publications\, this work ha s led to Choset\,\nalong with his students\, to form several companies inc luding\nMedrobotics\, for surgical systems\, Hebi Robotics\, for modular r obots\,\nand Bito Robotics for autonomous guided vehicles. Recently\, Chos et.s\nsurgical snake robot cleared the FDA and has been in use in the US a nd\nEurope since. Choset also leads multi-PI projects centered on\nmanufac turing: (1) automating the programming of robots for auto-body\npainting\; (2) the development of mobile manipulators for agile and\nflexible fixtur e-free manufacturing of large structures in aerospace\,\nand (3) the creat ion of a data-robot ecosystem for rapid manufacturing\nin the commercial e lectronics industry. This year\, Choset co-lead the\nformation of the Adva nced Robotics for Manufacturing Institute\, which\nis $250MM national inst itute advancing both technology development and\neducation for robotics in manufacturing. Finally\, Choset is a founding\nEditor of the journal Scie nce Robotics and is currently serving on the\neditorial board of IJRR. \n \nIn 2002 the MIT Technology Review elected Choset as one of its top 100\n innovators in the world under 35. In 2014\, Popular Science selected\nChos et's medical robotics work as the Best of What's New in Health\nCare. In 2 005\, MIT Press published a textbook\, lead authored by\nChoset\, entitled \"Principles of Robot Motion.\" Finally\, Choset's\nstudents have won bes t paper awards at the RIA in 1999 and ICRA in\n2003\; his group's work has been nominated for best papers at ICRA in\n1997\, IROS in 2003\, 2007\, a nd 2011\, and CLAWAR in 2012\; won best paper\nat IEEE Bio Rob in 2006\, S SRR 2012 and 2015\; won best video at ISMICS\n2006 and ICRA 2011\; and was nominated for best video in ICRA 2012. \n\nREGISTER → upon registering \, you will receive a confirmation email\ncontaining information about joi ning the meeting.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d8aae DTSTART;TZID=America/New_York:20240920T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240920T150000 URL:https://www.cs.cmu.edu/~pop/seminar/ LOCATION:Gates and Hillman Centers SUMMARY:Principles of Programming Seminar - Cameron Freer CLASS:PUBLIC DESCRIPTION:Speaker: CAMERON FREER\, Research Scientist\, Massachusetts Ins titute of\nTechnology\n\nTalk Title: Computability and Symmetry in Probabi listic Programming\n\nWe consider the computable content of several key th eorems in\nprobability theory\, and discuss their implications for the des ign of\nprobabilistic programming languages.\n\nA random variable is said to be exchangeable when its distribution\ndoes not depend on the ordering of the underlying elements. Sequences\,\narrays\, graphs\, and other data structures satisfying this symmetry\ncondition are models for homogeneous data sets and serve as building\nblocks in Bayesian nonparametric statisti cs. Representation theorems\nby de Finetti\, Aldous\, Hoover\, and Kallenb erg show that\nexchangeability gives rise to conditional independence. We establish\nboth positive and negative computable versions of these results \, and\nexplore the consequences for sequential and parallel implementatio ns\nof exchangeable objects in code.\n\nBayes’ theorem describes conditi onal probabilities\, and is\nfundamental in probabilistic inference. We sh ow that not every\ncomputable joint distribution admits a computable condi tional\ndistribution\, and examine the implications for automating Bayesia n\ninference.\n\nThis talk is based on joint work with Nathanael Ackerman\ , Jeremy\nAvigad\, Daniel Roy\, and Jason Rute.\n\n—\n\nCameron Freer is a Research Scientist in the Department of Brain and\nCognitive Sciences a t the Massachusetts Institute of Technology and a\nmember of the MIT Proba bilistic Computing Project. His research\nexplores interactions of randomn ess and computation\, and focuses on\nthe foundations of probabilistic com puting\, efficient samplers and\ntesting methods for probabilistic inferen ce\, and the mathematics of\nrandom structures. Cameron received his PhD i n Mathematics from\nHarvard University advised by Gerald Sacks\, and has h eld positions at\nthe University of Hawaii\, Keio University\, and several industry\nresearch labs.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d8f24 DTSTART;TZID=America/New_York:20240920T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240920T130000 URL:https://csd.cmu.edu/calendar/doctoral-speaking-skills-talk-afonso-tinoc o LOCATION:Mehrabian Collaborative Innovation Center 1301 SUMMARY:Doctoral Speaking Skills Talk - Afonso Tinoco CLASS:PUBLIC DESCRIPTION:Speaker: AFONSO TINOCO\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Oblivious Maps\n\nImagi ne that a privacy-conscious client would like to query a\nkey-value store residing on an untrusted server equipped with a secure\nprocessor. To prot ect the privacy of the client's queries as well as\nthe database\, one app roach is to implement an oblivious map inside a\nsecure enclave. Indeed\, earlier works demonstrated  numerous\napplications of an enclaved-based o blivious map\, including private\ncontact discovery\, key transparency\, a nd secure outsourced\ndatabases. \n\nOur work is motivated by the observa tion that the previous enclave\nimplementations of oblivious algorithms ar e sub-optimal both\nasymptotically and concretely. We make the key observa tion that for\nenclave applications\, the number of page swaps should be a primary\nperformance metric. We therefore adopt techniques from the\nexte rnal-memory algorithms  literature\, and we are the first to\nimplement s uch algorithms inside hardware enclaves. We also devise\nasymptotically be tter algorithms for ensuring a strong notion of\nobliviousness that resist s cache-timing attacks. \n\nWe complement our algorithmic improvements wi th various concrete\noptimizations that save constant factors in practice. The resulting\nsystem\, called Enigmap\, achieves 15x speedup over Signal 's linear scan\nimplementation\, and 53 speedup over the  prior best obli vious\nalgorithm implementation\, at a realistic database size of 256 mill ion\nand a batch size of 1000. The speedup is asymptotical in nature and\n will be even greater as Signal's user base grows. \n\nPresented in Partia l Fulfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d932b DTSTART;TZID=America/New_York:20240919T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240919T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch-Skees Conference Room\, Gates Hillman 8115 and Zoom SUMMARY:Crypto Seminar - Yi Tang CLASS:PUBLIC DESCRIPTION:Speaker: YI TANG\, Ph.D. Student\, Electrical Engineering and C omputer\nScience Department\, University of Michigan\n\nTalk Title: Crypta nalysis of Lattice-Based Sequentiality Assumptions\nand Proofs of Sequenti al Work\n\nThis work completely breaks the sequentiality assumption (and b road\ngeneralizations thereof) underlying the candidate lattice-based proo f\nof sequential work (PoSW) recently proposed by Lai and Malavolta at\nCR YPTO 2023. In addition\, it breaks an essentially identical variant\nof th e PoSW\, which differs from the original in only an arbitrary\nchoice that is immaterial to the design and security proof (under the\nfalsified assu mption). This suggests that whatever security the\noriginal PoSW may have is fragile\, and further motivates the search\nfor a construction based on a sound lattice-based assumption.\n\nSpecifically\, for sequentiality par ameter T and SIS parameters n\, q\, m\n= n*log(q)\, the attack on the sequ entiality assumption finds a\nsolution of quasipolynomial norm mlog T (or norm O(sqrt(m))log T with\nhigh probability) in only logarithmic Õn\,q(l og T) depth\; this\nstrongly falsifies the assumption that finding such a solution\nrequires depth *linear* in T. (The Õ notation hides polylogari thmic\nfactors in the variables appearing in its subscript.) Alternatively \,\nthe attack finds a solution of polynomial norm m1/ε in depth\nÕn\,q (Tε)\, for any constant ε > 0. Similarly\, the attack on the\n(slightly modified) PoSW constructs a valid proof in \npolylogarithmic  Õn\,q(lo g2 T) depth\, thus strongly falsifying the\nexpectation that doing so requ ires linear sequential work.\n\nJoint work with Chris Peikert. Reference p aper\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d97cf DTSTART;TZID=America/New_York:20240919T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240919T160000 URL:https://aco.math.cmu.edu/abs-24-25/sep19.html LOCATION:Wean Hall 8220 SUMMARY:ACO Seminar - Quentin Dubroff CLASS:PUBLIC DESCRIPTION:Speaker: QUENTIN DUBROFF\, Postdoctoral Associate\, Department of\nMathematical Sciences\, Carnegie Mellon University\n\nTalk Title: Thre sholds for graph containment in G_{n\,p} and coupon\ncollectorsThe first g oal of this talk is to describe some recent\nprogress (in joint work with Jeff Kahn and Jinyoung Park) on the\n\"Second\" Kahn-Kalai Conjecture (KKC 2)\, the original conjecture on\ngraph containment in Gn\,p that motivated what is now the Park-Pham\nTheorem (PPT). KKC2 says that pc (H)\, the thr eshold for containing a\ngraph H in Gn\,p\, satisfies pc (H) = O(pE log n) \, where pE is the\nsmallest p such that the expected number of copies of any subgraph of\nH is at least one. In other words\, for this class of pro blems\, the\nexpectation threshold q in PPT can be replaced by the smaller pE. We\nshow that q < O(pE log2 n) (implying pc(H) = O(pE log3 n) via PPT ).\nTime-permitting\, the second portion of the talk will discuss some\nho pes for and failed attempts at sharpening PPT and KKC2.\n4:00 pm → Tea a nd Cookies\, Wean 6220 (bring your own mug if you have\none).\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91d9c12 DTSTART;TZID=America/New_York:20240919T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240919T140000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:TCS Hall 358 and Zoom SUMMARY:Joint Computational Social Science / AI Institute for Societal\nDec ision Making Seminar CLASS:PUBLIC DESCRIPTION:Speaker: DIYI YANG\, Assistant professor\, Computer Science Dep artment\,\nStanford University\n\nTalk Title: Building Human-Centered NLP for Social Impact\n\nLarge language models have revolutionized the way hum ans interact with\nAI systems\, transforming a wide range of fields and di sciplines. The\nbenefits and promises of LLMs are accompanied by an increa se in\nevidence and concern about its negative aspects.  In this talk\, w e\ndiscuss how to build socially responsible LLMs for social impact from\n a human-centered perspective.  The first half presents a\nparticipatory design approach to develop dialect-inclusive language\ntools and adaptatio n techniques for low-resourced language and\ndialect\, and further introdu ces a distilled voice assistant using\ncross-modal context distillation t o enable positive speech\ninteraction.  The second half looks at skill tr aining with LLMs by\ndemonstrating how we use LLMs to empower novice thera pists through\nsimulated practice and deliberative feedback. We conclude\ nby discussing how human-centered LLMs can empower individuals and\nfoste r positive change. \n—\nDiyi Yang is an assistant professor in the Comp uter Science Department\nat Stanford University.  Her research focuses on human-centered\nnatural language processing and computational social scie nce.  She is\na recipient of a Microsoft Research Faculty Fellowship (202 1)\, NSF\nCareer Award (2022) an ONR Young Investigator Award (2023) and a Sloan\nResearch Fellowship (2024).  Her work has received multiple paper \nawards or nominations at top NLP and HCI conferences.\nIn Person and Zoo m Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91da00c DTSTART;TZID=America/New_York:20240919T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240919T130000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-aditi-kabra LOCATION:Gordon Bell Conference Room\, Gates Hillman 5117 SUMMARY:Doctoral Thesis Proposal - Aditi Kabra CLASS:PUBLIC DESCRIPTION:Speaker: ADITI KABRA\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Verified Control Envelope Synthesis for Hybrid Systems\n\nMany cyber-physical systems\, such as tra ins\, planes\, and self-driving\ncars\, are safety-critical but difficult to reason about. Formal\nverification can provide strong safety guarantees \, but most industrial\ncontrollers are too complex to formally verify. Sa fe control envelopes\ncharacterize families of safe controllers and are us ed to monitor\nuntrusted controllers on verifiable abstractions of control systems\nthat isolate the parts relevant to safety without the full compl exity\nof a specific control implementation\, at runtime. They can put com plex\ncontrollers\, even when machine learning based\, within the reach of \nformal guarantees. But correct control envelopes are still hard to\ndesi gn because the control engineer needs to identify correct control\nconditi ons that tell the controller what to do right now to stay safe\nat all tim es in the future by anticipating the behavior of the system\nover complex dynamics and an uncountably infinite state space. \n\nThis thesis propose s to provide synthesis techniques to automatically\nsynthesize provably co rrect control conditions\, greatly reducing the\nmanual effort required fo r control envelope design. It aims to scale\nsynthesis to complexity of re al-world systems. The input of the\nsynthesis tool is a sketch of the cont rol envelope in a hybrid system\nshowing what kind of control behavior is physically possible. The tool\nfills in the blanks of the sketch by synthe sizing control conditions\nusing hybrid system game theory. The output is a provably correct\nsymbolic control envelope. Existing controller synthes is techniques do\nnot solve control envelope synthesis because control env elopes have\nthe higher-order constraint of permitting as many valid contr ol\nsolutions as possible. \n\nCompleted work provides the algorithm CESA R (Control Envelope\nSynthesis via Angelic Refinement) which solves a clas s of problems\nwhere a set of systematic game refinements allows automatic control\nenvelope synthesis. Proposed work generalizes synthesis to a bro ad\nclass of systems (characterized by admitting a natural representation\ nin differential game logic) and develops a system that allows users to\np rovide the human intuition based insights that\, together with\nautomated reasoning\, can complete the control envelope synthesis\nprocess in more c omplex cases. \n\nThesis Committee\n\nAndré Platzer (Co-chair\, Carnegie Mellon University/Karlsruhe\nInstitute of Technology)\n\nStefan Mitsch (C o-chair\, Carnegie Mellon University/DePaul University)\n\nEunsuk Kang\n\n Armando Solar-Lezama (Massachusetts Institute of Technology)\n\n \n\nAdd itional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91da4fa DTSTART;TZID=America/New_York:20240918T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240918T130000 URL:https://www.cs.cmu.edu/~theorylunch/abstractsHTML/20240918.html LOCATION:Newell-Simon 3305 (new location) SUMMARY:Theory Lunch Seminar - Grand Schoenebeck CLASS:PUBLIC DESCRIPTION:Speaker: GRANT SCHOENEBECK\, Associate Professor\, School of\nI nformation\, University of Michigan\n\nTalk Title: Eliciting Informative T ext Evaluations with Large Language\nModelsIn a wide variety of contexts i ncluding peer grading\, peer\nreview\, and crowd-sourcing (e.g. evaluating LLM outputs) we would like\nto design mechanisms which reward agents for producing high quality\nresponses. Unfortunately\, computing rewards by co mparing to ground\ntruth or gold standard is often cumbersome\, costly\, o r impossible.\nInstead we would like to compare agent reports.Peer predict ion\nmechanisms motivate high-quality feedback with provable guarantees.\n However\, current methods only apply to rather simple reports\, like\nmult iple-choice or scalar numbers. We aim to broaden these techniques\nto the larger domain of text-based reports\, drawing on the recent\ndevelopments in large language models. This greatly increases the\napplicability of pee r prediction mechanisms as textual feedback is the\nnorm in a large variet y of feedback channels: peer reviews\, e-commerce\ncustomer reviews\, and comments on social media. We introduce two\nmechanisms\, the Generative Pe er Prediction Mechanism (GPPM) and the\nGenerative Synopsis Peer Predictio n Mechanism (GSPPM). These\nmechanisms utilize LLMs as predictors\, mappin g from one agent’s\nreport to a prediction of her peer’s report. Theor etically\, we show\nthat when the LLM prediction is sufficiently accurate\ , our mechanisms\ncan incentivize high effort and truth-telling as an (app roximate)\nBayesian Nash equilibrium. Empirically\, we confirm the efficac y of our\nmechanisms through experiments conducted on two real datasets: t he\nYelp review dataset and the ICLR OpenReview dataset. We highlight the\ nresults that on the ICLR dataset\, our mechanisms can differentiate\nthre e quality levels — human written reviews\, GPT-4-generated\nreviews\, an d GPT-3.5-generated reviews in terms of expected scores.\nAdditionally\, G SPPM penalizes LLM-generated reviews more effectively\nthan GPPM.\n\nNo ba ckground is required\, and the talk will introduce the relevant\ntheoretic al background of peer prediction mechanisms.\n—\n\nGrant Schoenebeck is an associate professor at the University of\nMichigan in the School of Inf ormation.  His work spans diverse areas\nin theoretical computer science but has recently focused on combining\nideas from theoretical computer sci ence\, machine learning\, and\neconomics (e.g game theory\, mechanism desi gn\, and information design)\nto develop and analyze systems for eliciting and aggregating\ninformation from of diverse group of agents with varying information\,\ninterests\, and abilities.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91daa00 DTSTART;TZID=America/New_York:20240917T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240917T190000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:South Wing Reading Room\, Carnegie Library of Pittsburgh\, 4400 Fo rbes\nAve\, Pittsburgh\, PA 15213 SUMMARY:AI-SDM Public Speaker Series - Roni Rosenfeld CLASS:PUBLIC DESCRIPTION:Speaker: RONI ROSENFELD\, Professor of Machine Learning\, Langu age\nTechnologies\, Computer Science\, and Computational Biology\, School of\nComputer Science\, Carnegie Mellon University\n\nTalk Title: What is A I and Gen AI?\n\nJoin in for this informative session with CMU’s NSF-fun ded AI\nInstitute for Societal Decision Making (AI-SDM) as they share an\n overview of Artificial Intelligence (AI) and Generative AI. If you\nhave e ver wondered about the advances of computer intelligence and the\ncurrent trends\, this program is for you. \n\nPresenter Roni Rosenfeld from Car negie Mellon University’s Machine\nLearning Department\,  will introduc e attendees to the world of AI\nduring this engaging session. \n\nREGISTR ATION   |  Additional Information \n\n⇒  This program is a livestre am event that will be recorded for\nlater access. \n\n⇒  Attendees are welcome to join in-person\, which takes place in\nthe South Wing Reading Room\, CLP-Main Library\, or virtually by\nregistering and using the the l ink provided in your RSVP confirmation\nemail. \n\n⇒  Free and Open to the Public\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dadff DTSTART;TZID=America/New_York:20240916T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240916T163000 URL:https://www.ri.cmu.edu/event/instant-visual-3d-worlds-through-split-loh mann-displays/ LOCATION:Newell-Simon 3305 SUMMARY:Vision and Autonomous Systems Seminar - Yingsi Qin CLASS:PUBLIC DESCRIPTION:Speaker: YINGSI QIN\, EDWARD LU\, and MOSAM DABHI\, PhD Candida tes\,\nCarnegie Mellon University\n\nTalk Title: Three TalksSplit-Lohmann displays provide a novel approach\nto creating instant visual 3D worlds th at support realistic eye\naccommodation. Unlike commercially available VR headsets that show\ncontent at a fixed depth\, the proposed display can op tically place\neach pixel region to a different depth\, instantly creating \neye-tracking-free 3D worlds without using time-multiplexing. This\nenabl es real-time streaming of 3D content over a large depth range at\nhigh spa tial resolution\, offering an exciting step towards a more\nimmersive real -time 3D experience. We demonstrate the technology's\ncapabilities through a lab prototype\, showcasing high-quality visuals\nacross various static\ , dynamic\, and interactive 3D scenes.\n—\nYingsi Qin is a PhD candidate in Electrical and Computer Engineering\nat Carnegie Mellon University\, a dvised by Aswin C. Sakaranarayanan and\nMatthew P. O'Toole. Her research focuses on designing and building\nnext-generation computational 3D displa ys for Virtual\, Augmented\, and\nMixed Reality. The interdisciplinary wor k involves a fusion of\ncomputer vision\, optics\, signal processing\, and machine\nlearning. Yingsi received the Best Paper Award at SIGGRAPH 2023 and\nthe Best Demo Award at ICCP 2023. Yingsi holds a B.S. in Computer\nS cience from Columbia University and a B.A. in Physics from Colgate\nUniver sity. She was a research intern at Meta Reality Labs in the\nDisplay Syste ms Research team (2024) and Snap Research in the\nComputational Imaging te am (2020). She was also a software engineering\nintern at Google Search (2 019).\nThe VASC Seminar is sponsored in part by Meta Reality Labs Pittsbur gh\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91db227 DTSTART;TZID=America/New_York:20240916T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240916T130000 URL:https://www.cylab.cmu.edu/events/2024/09/16-seminar-savage.html LOCATION:Newell-Simon 4305 and Zoom (new location) SUMMARY:CyLab Seminar - Stefan Savage CLASS:PUBLIC DESCRIPTION:Speaker: STEFAN SAVAGE\, Professor\, Department of Computer Sci ence and\nEngineering\, University of California\, San Diego\n\nTalk Title : Inside-out and backwards: a retrospective look at how\nmeasurement resea rch really happens\n\nMeasurement research is commonly presented in finish ed form\, with each\nfinding leading clearly and cleanly to the next\, as natural and\nexpected as night following day. This narrative\, whether in writing or\nspeech\, exists to support the research's rhetorical goal — framing\nthe problem and the measurements in a way that highlights the ins ight\nbeing presented. However\, this is not how such research actually\nh appens. Measurement research does not spring into existence like\nBotticel li's Venus\, fully formed and without blemish\, but is\ninevitably a compl ex combination of semi-random exploration\, hard\nwork\, serendipity and a variety of exogenous factors that defy any\nclean and clear generalizatio n. Researchers new to the process thus\ncan despair that the \"dumpster fi re\" that they see in their own\nresearch process will never produce the p erfection seen in others'\npaper\, not realizing that in fact even the mos t celebrated research\nhas similarly chaotic underpinnings. In this talk\, I will explore this\nreality by pulling back the curtain on my own experi ence and exploring\nthe unfiltered history of how a variety of my more suc cessful research\nprojects actually happened and the lessons that I tried to extract\nfrom those experiences.\n\n— \n\nStefan Savage is a profess or of Computer Science and Engineering at\nthe University of California\, San Diego. He received his Ph.D. in\nComputer Science and Engineering from the University of Washington and\na B.S. in Applied History from Carnegie Mellon University. He\ncurrently serves as the co-director for UCSD's Cen ter for Network\nSystems (CNS). Savage is known for his work on network se curity and\nreliability\, on cybercrime economics and defense\, and on the empirical\nmeasurement of cybersecurity and cyberinfrastructure. He is a member\nof the National Academy of Engineering and the American Academy of \nArts and Sciences\, a MacArthur Fellow\, an ACM Fellow\, and is the\nrec ipient of ACM's Prize in Computing and AAAS' Golden Goose award. He\ncurre ntly holds the Irwin and Joan Jacobs Chair in Information and\nComputer Sc ience\, but is a fairly down-to-earth guy and only writes\nabout himself i n the third person when asked. \n\nFaculty Host:  Justine Sherry \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91db794 DTSTART;TZID=America/New_York:20240913T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240913T120000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Wean Hall 1327 and Zoom SUMMARY:AI Institute for Societal Decision Making - Student Brainstorming\n Session CLASS:PUBLIC DESCRIPTION:The AI-SDM students meet regularly to participate in informal\n discussion sessions that delve into cutting-edge AI topics. These\nregular meetings throughout the semester provide a platform for\nstudents to deep en understanding of specific areas and broaden\nknowledge by exploring cro ss-cutting connections between various AI\ndisciplines. \n\nDiscussions a re a breeding ground for collaboration\, innovative\nthinking\, and proble m-solving from the ground up. They provide a\nstress-free forum for exchan ging ideas\, brainstorming new approaches\nto challenges\, and fostering l asting connections within the AI-SDM\ncommunity in an environment distinct from a traditional seminar. \n\nRSVP Requested.\n\nIn Person and Zoom Pa rticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dbaf2 DTSTART;TZID=America/New_York:20240912T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240912T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Blelloch Skees Conference Room\, Gates Hillman 8115 SUMMARY:Crypto Seminar - Quang Dao CLASS:PUBLIC DESCRIPTION:Speaker: QUANG DAO\, Ph.D. Student\, Computer Science Departmen t\,\nCarnegie Mellon University\,\n\nTalk Title: Non-Interactive Zero-Know ledge from LPN and MQ\n\nOver the past few decades\, we have seen a prolif eration of advanced\ncryptographic primitives with lossy or homomorphic pr operties built\nfrom various assumptions such as Quadratic Residuosity\, D ecisional\nDiffie-Hellman\, and Learning with Errors. These primitives imp ly hard\nproblems in the complexity class SZK (statistical zero-knowledge) \; as\na consequence\, they can only be based on assumptions that are brok en\nin BPPSZK. This poses a barrier for building advanced primitives from\ ncode-based assumptions\, as the only known such assumption is Learning\nP arity with Noise (LPN) with an extremely low noise rate log2(n)/n\,\nwhich is broken in quasi=polynomial time. \n\nIn this work\, we propose a new code-based assumption: Dense-Sparse\nLPN\, that falls in the complexity cl ass BPPSZK and is conjectured to\nbe secure against subexponential time ad versaries. Our assumption is a\nvariant of LPN that is inspired by McEliec e’s cryptosystem and\nrandom k-XOR in average-case complexity. Roughly\, the assumption\nstates that (T⋅M\, s⋅T⋅M + e) is indistinguishable from (T⋅M\,\nu)\, for a random (dense) matrix T\, random sparse matrix M \, and sparse\nnoise vector e drawn from the Bernoulli distribution with i nverse\npolynomial noise probability. We leverage our assumption to build\ nlossy trapdoor functions (Peikert-Waters STOC 08). This gives the\nfirst post-quantum alternative to the lattice-based construction in\nthe origina l paper. Lossy trapdoor functions\, being a fundamental\ncryptographic too l\, are known to enable a broad spectrum of both lossy\nand non-lossy cryp tographic primitives\; our construction thus implies\nthese primitives in a generic manner. In particular\, we achieve\ncollision-resistant hash fun ctions with plausible subexponential\nsecurity\, improving over a prior co nstruction from LPN with noise rate\nlog2(n)/n that is only quasi-polynomi ally secure. \n\nThis is joint work with Aayush Jain. \n\nReference Pape r\n\nIn Person and Zoom Participation. See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dbfb5 DTSTART;TZID=America/New_York:20240912T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240912T160000 URL:https://aco.math.cmu.edu/abs-24-25/sep12.html LOCATION:Wean 8220 SUMMARY:ACO Seminar - Imre Leader\, Univ. of Cambridge CLASS:PUBLIC DESCRIPTION:Speaker: IMRE LEADER\, Professor of Pure Mathematics\, Centre f or\nMathematical Sciences\, University of Cambridge\n\nTalk Title: Turan d ensities for daisies and hypercubes\n\nThe Turan problem for hypercubes as ks: how few vertices of the\nn-dimensional cube can we take so that they m eet every d-dimensional\nsubcube? A longstanding conjecture states that th e best one can do\n(asymptotically) is 1/(d+1) of all vertices\, by taking every (d+1)th\nlayer of the cube. In this talk we will explain the connec tion to\nTuran questions for `daisy’ hypergraphs\, and present a disproo f of\nthe conjecture. T\n\nhis is joint work with David Ellis and Maria Iv an.\n\nTea and Cookies to follow talk.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dc319 DTSTART;TZID=America/New_York:20240912T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240912T120000 URL:https://www.cmu.edu/swartz-center-for-entrepreneurship/events/james-r-s wartz-leadership-series/index.html LOCATION:Swartz Center for Entrepreneurship\, Tepper Building SUMMARY:James R. Swartz Entrepreneurial Leadership Series Talk /Joint SCS\n Distinguished Industry Talk CLASS:PUBLIC DESCRIPTION:Speaker: ALFRED SPECTOR\, Visiting Scholar\, Electrical & Compu ter\nScience Department\, Massachusetts Institute of Technology\, and\, Se nior\nAdvisor\, Blackstone\n\nTalk Title: Beyond Models – Applying AI an d Data Science Effectively\n\nApplying artificial intelligence and data sc ience effectively requires\na considerably broader focus than just data an d machine learning.\nBased on the speaker and his co-authors' recent book\ , Data Science in\nContext (and an associated MIT Course)\, this presentat ion distills\nthese additional challenges into a rubric and illustrates it s\napplication with a number of examples. Beyond the rubric\, the\npresent ation also presents useful frameworks for making the complex\ntrade-offs t hat are present and growing.  While the talk should have\npractical value to those applying and regulating AI and DS\, it also\nillustrates contemp orary research challenges. \n\n— \n\nDr. Alfred Spector is a Visiting Scholar at MIT and a Senior Advisor\nat Blackstone. His career has led him from innovation in large scale\,\nnetworked computing systems to broad en gineering and research\nleadership. Recently\, he co-authored a Cambridge University Press\ntextbook\, “Data Science in Context: Foundations\, Cha llenges\,\nOpportunities.” \n\nPreviously\, Dr. Spector was CTO and Hea d of Engineering at Two Sigma\nInvestments. Before that\, he spent eight y ears as VP of Research and\nSpecial Initiatives at Google\, and he held va rious senior-level\npositions at IBM\, including as global VP of Services and Software\nResearch and global CTO of IBM’s Software Business. Earlie r in his\ncareer\, he founded Transarc Corporation\, a pioneer in distribu ted\ntransaction processing and wide-area file systems.  He was also a\nt enured professor at Carnegie Mellon University\, and he led its\nInformati on Technology Center in 1998 and 1999. \n\nSpector was a Hertz Fellow at Stanford and is also a Fellow of both\nthe ACM and the IEEE. He is a membe r of the National Academy of\nEngineering and the American Academy of Arts and Sciences. Dr. Spector\nwon the 2001 IEEE Kanai Award for Distributed Computing and the 2016\nACM Software Systems Award. In 2018-19\, Dr. Spect or lectured widely as\na Phi Beta Kappa Scholar (for example\, on the grow ing importance of\ncomputer science across all disciplines based on the ev ocative phrase\,\n“CS+X”). He has been a member of the ACM Turing Awar d Committee\nand has done national service through chairing the NSF’s CI SE\nAdvisory Board and his membership on the Army and Defense Science\nBoa rds. Dr. Spector obtained a Ph.D. in computer science from Stanford\nand a B.A. in applied math from Harvard.\n\n REGISTRATION\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dc80e DTSTART;TZID=America/New_York:20240911T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240911T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Reddy Conference Room\, Gates Hillmsan 4405 SUMMARY:Theory Lunch Seminar - Tim Hsieh CLASS:PUBLIC DESCRIPTION:Speaker: TIM HSIEH\, Ph.D. Student\, Computer Science Departmen t\,\nCarnegie Mellon University\n\nTalk Title: Rounding Large Independent Sets on ExpandersIn this talk\,\nwe will present a new approach for approx imating large independent\nsets when the input graph is a one-sided spectr al expander – that\nis\, the uniform random walk matrix of the graph has the second\neigenvalue bounded away from 1. Consequently\, we obtain a po lynomial\ntime algorithm to find linear-sized independent sets in one-side d\nexpanders that are almost 3-colorable or are promised to contain an\nin dependent set of size (1/2 - ε)n. Somewhat surprisingly\, we observe\ntha t the analogous task of finding a linear-sized independent set in\nalmost 4-colorable one-sided expanders (even when the second\neigenvalue is o(1)) is NP-hard\, assuming the Unique Games Conjecture.\n\nOur rounding builds on the method of simulating multiple samples from\na pseudodistribution i ntroduced by Bafna et. al. for rounding Unique\nGames instances. The key t o our analysis is a new clustering property\nof large independent sets in expanding graphs — every large\nindependent set has a larger-than-expect ed intersection with some\nmember of a small list— and its formalization in the low-degree\nsum-of-squares proof system.\n\nBased on joint work wi th Mitali Bafna and Pravesh K. Kothari.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dcc0e DTSTART;TZID=America/New_York:20240910T154500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240910T163000 URL:https://events.cmu.edu/event/10256-welcome-reception-for-new-faculty LOCATION:Simmons Auditorium\, Tepper Building SUMMARY:CMU Welcome Reception for New Faculty and Fall Faculty Reception CLASS:PUBLIC DESCRIPTION:Talk Title: New Faculty and Fall Faculty ReceptionPresident Far nam\nJahanian and Provost Jim Garrett \n      invite you and a guest to the Fall Faculty Welcome and\nReception New Faculty Welcome |  3:45-4 :30 pm All Faculty\nReception   |  4:30-6:00 pmNew CMU faculty members are invited to\njoin us as we welcome you and fellow new faculty who have joined\nCarnegie Mellon University this year.\nRSVP Required → by Friday \, September 6.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dcf0a DTSTART;TZID=America/New_York:20240910T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240910T163000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Machine Learning/Duolingo Seminar - Nika Haghtalab CLASS:PUBLIC DESCRIPTION:Speaker: NIKA HAGHTALAB\, Assistant Professor\, Department of E lectrical\nEngineering and Computer Sciences\, Co-director\, Center for th e\nTheoretical Foundations of Learning\, Inference\, Information\,\nIntell igence\, Mathematics and Microeconomics \, University of\nCalifornia\, Ber keley\,\n\nTalk Title: Multiobjective learning: A unified framework for\nr obustness\, fairness\, and collaboration in machine learning\n\nPervasive needs for robustness\, multi-agent collaboration\, and\nfairness have moti vated the design of new methods in research and\ndevelopment. However\, th ese methods remain largely stylized\, lacking a\nfoundational perspective and provable performance. In this talk\, I\nwill introduce and highlight t he importance\nof multi-objective learning as a unifying paradigm for  addressing\nthese needs. This paradigm aims to optimize complex and unst ructured\nobjectives from only a small amount of sampled data. I will also \ndiscuss how the multi-objective learning paradigm relates to the\nclassi cal and modern considerations in machine learning broadly\,\nintroduce tec hnical tools with versatile provable guarantees\, and\nempirical evidence for its performance on a range of important\nbenchmarks. \n\n— \n\nNik a Haghtalab is an Assistant Professor in the Department of\nElectrical E ngineering and Computer Sciences at UC Berkeley. She works\nbroadly on the theoretical aspects of machine learning and algorithmic\neconomics. Prof.  Haghtalab's work builds theoretical foundations for\nensuring both the p erformance of learning algorithms in the presence\nof everyday economic fo rces and the integrity of the social and\neconomic forces that are born ou t of the use of machine learning\nsystems. \n\nShe received her Ph.D. fro m the Computer Science Department of\nCarnegie Mellon University\, where h er thesis won the CMU School of\nComputer Science Dissertation Award (ACM nomination) and the SIGecom\nDissertation Honorable Mention. She is a co-f ounder of Learning Theory\nAlliance (LeT-All). Among her honors are an NSF CAREER award\, Sloan\nfellowship\, Schmidt Sciences AI2050 fellowship\, NeurIPS and ICAPS\nbest paper awards\, an EC exemplary in AI track award\, and several\nindustry awards and fellowships. \n\nIn Person and Zoom Par ticipation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dd371 DTSTART;TZID=America/New_York:20240909T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240909T163000 URL:https://www.ri.cmu.edu/event/unlocking-magic-personalization-of-diffusi on-models-for-novel-applications/ LOCATION:Newell-Simon 3305 SUMMARY:Vision and Autonomous Systems Seminar - Nathaniel Ruiz CLASS:PUBLIC DESCRIPTION:Speaker: NATANIEL RUIZ\, Research Scientist\, Google\n\nTalk Ti tle: Unlocking Magic: Personalization of Diffusion Models for\nNovel Appli cations\n\nSince the recent advent of text-to-image diffusion models for\n high-quality realistic image generation\, a plethora of creative\napplicat ions have suddenly become within reach. I will present my\nwork at Google where I have attempted to unlock magical applications\nby proposing simpl e techniques that act on these large text-to-image\ndiffusion models. Part icularly\, a large class of these applications\ncan be unlocked using pers onalization by finetuning\, starting with our\npopular work on DreamBooth  where we can learn a subject's appearance\nand generate that subject in different contexts and with different\nsemantic modifications. My presenta tion will include a deeper dive\ninto our recent works ZipLoRA\, RealFil l\, RB-Modulation and our\nlatest work Magic Insert. \n\n— \n\nNata niel Ruiz is a Research Scientist at Google and the lead author of\nDreamB ooth\, which was selected for a Best Paper Award at CVPR 2023.\nHis main r esearch interests revolve around generative models\, and he\nhas authored other works in the areas of controllability and\npersonalization of diffus ion models\, including StyleDrop\, ZipLoRA\, and\nHyperDreamBooth. He obta ined his PhD from Boston University\, his\nMaster's from Georgia Tech\, an d his Bachelor's from École\nPolytechnique in Paris. Prior to joining Goo gle\, he also interned at\nApple\, Amazon\, and NEC Labs. \n\nThe VASC Se minar is sponsored in part by Meta Reality Labs Pittsburgh\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dd758 DTSTART;TZID=America/New_York:20240909T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240909T130000 LOCATION:Panther Hollow Room 4105\, Mehrabian Collaborative Innovation Cent er SUMMARY:CyLab Seminar - Sanghamitra Dutta CLASS:PUBLIC DESCRIPTION:Speaker: SANGHAMITRA DUTTA\, Assistant Professor\, Department o f\nElectrical and Computer Engineering\, University of Maryland College\nP ark\n\nTalk Title: Information-Theoretic Methods for Explainability in\nHi gh-Stakes Applications\n\nHow do we ensure that the machine learning algor ithms in high-stakes\napplications are explainable\, fair\, and lawful? To wards addressing\nthis urgent question\, this talk provides strategies for explainability\nthat are deep-rooted in information theory and probabilit y. \n\nIn the first part of the talk\, I will discuss an emerging problem in\nexplainability\, also called robust counterfactual explanations: how do\nwe guide a rejected applicant to receive a favorable model outcome\nwh ile also being robust to model multiplicity (Rashomon Effect)? We\npropose strategies to provide counterfactual explanations that remain\nrobust und er model changes with probabilistic guarantees. In the\nsecond part of the talk\, I will introduce an information theoretic\ntool called Partial Inf ormation Decomposition and discuss its role in\nfairness and explainabilit y problems. \n\n— \n\nSanghamitra Dutta is an assistant professor in t he Department of\nElectrical and Computer Engineering at the University of Maryland\nCollege Park since Fall 2022. She is also affiliated with the C enter\nfor Machine Learning  (CML) at UMIACS. Prior to joining UMD\, she was\na senior research associate at JPMorgan Chase AI Research New York in \nthe Explainable AI Centre of Excellence (XAI CoE). She received her\nPh. D. and Masters's from Carnegie Mellon University and B. Tech. from\nIIT Kh aragpur\, all in Electrical and Computer Engineering. \n\nHer research in terests broadly revolve around reliable and trustworthy\nmachine learning where she brings in novel foundational perspectives\ndeep-rooted in inform ation theory\, statistics\, causality\, and\noptimization. In her prior wo rk\, she has also examined problems in\nreliable computing for large-scale distributed machine learning\, using\ntools from coding theory (an emergi ng area called “coded\ncomputing”). She is a recipient of the 2024 NSF CAREER Award\, 2023\nJP Morgan Faculty Award\, 2023 Northrop Grumman Seed Grant\, 2022 Simons\nInstitute Fellowship for Causality\, 2021 AG Milnes Outstanding Thesis\nAward from CMU and 2019 K&L Gates Presidential Fellows hip in Ethics\nand Computational Technologies. She has also pursued summer research\ninternships at IBM Research and Dataminr. \n\nFaculty Host: Pu lkit Grover \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ddc7f DTSTART;TZID=America/New_York:20240907T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240907T130000 URL:https://www.cs.cmu.edu/scs-career-center/career-center-events SUMMARY:GRAY SWAN: Jailbreaking Championship 2024 CLASS:PUBLIC DESCRIPTION:Talk Title: $40\,000 in Bounties: Jailbreaking Championship 202 4\n\nGray Swan AI is hosting a jailbreaking championship\, offering $40\,0 00\nin bounties! No coding required! \n\nThe final showdown between Hacke rs and LLMs! Simply jailbreak any\nmodel in our chat interface to earn you r share of $40\,000 in bounties!\nThe championship begins at 10:00 AM PT o n September 7 and will\nconclude when at least K (TBD) participants have s uccessfully\njailbroken each model. The timer for all models will start\ns imultaneously at exactly 10:00 AM PT for everyone.\n\n REGISTER  → reg istration now open! |   Learn more! \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ddfe8 DTSTART;TZID=America/New_York:20240906T154500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240906T174500 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-elisaweta-masserova LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Thesis Oral Defense - Elisaweta Masserova CLASS:PUBLIC DESCRIPTION:Speaker: ELISAWETA MASSEROVA\, Ph.D. Candidate\, Computer Scien ce\nDepartment\, Carnegie Mellon University\n\nTalk Title: Distributed Cry ptography as a Service\n\nToday’s world is undeniably data-driven. The e xplosion of the\nInternet has generated vast volumes of data\, and the adv ent of machine\nlearning has unlocked captivating applications that thrive on this\ndata. In such a world\, it is evident that the ability to store\ ,\ntransmit\, and process data securely is paramount. Distributing trust\n is one of the fundamental cryptographic principles that enable such\nsecur ity\, and it is at the core of key cryptographic tools such as\nmulti-part y computation (MPC) and randomness generation. As the demand\nfor secure a nd reliable cryptographic solutions grows\, there is\nincreasing interest in offering such distributed protocols as a\nservice. \n\nThese services are typically expected to run continuously for long\nperiods of time\, req uiring significant resource commitments from all\nparticipating parties. O ne approach to mitigate this issue is to\ndesign distributed cryptographic protocols that are stateless. With\nsuch protocols\, parties can contribu te to the execution of a\ndistributed cryptographic protocol by participat ing only for a short\ntime\, without committing to a long-term computation . \n\nIn this work\, we study such mostly stateless protocols. We start b y\nintroducing a blockchain-based MPC protocol which does not require\npar ties to be online at the same time and requires no interaction\nbetween th e participants. We construct this protocol in the blockchain\nmodel and un der the assumption of what we call Conditional Storage and\nRetrieval (CSa R) systems. In our next step\, we eliminate the CSaR\nrequirement and desi gn a stateless MPC protocol without relying on\nthis assumption. \n\nMore concretely\, we focus on the recently introduced You Only Speak\nOnce (YO SO) paradigm. In this model participating parties are allowed\nto send onl y a single message\; i.e.\, they speak only once. We improve\nthe state of the art in YOSO MPC by designing a protocol with better\ncommunication co mplexity than the currently known solutions. Then\, we\nfocus on improving the efficiency of special-purpose YOSO MPC\nprotocols. Specifically\, we consider the task of distributed\nrandomness generation\, and design a sui te of protocols\, each balancing\ndifferent trade-offs in terms of underly ing assumptions\, efficiency\,\nand corruption threshold.\n\nThesis Commit tee \n\nBryan Parno (Co-chair)\n\nVipul Goyal (Co-chair)\n\nElaine Shi\n\ nAntigoni Polychroniadou (J.P. Morgan AI Research)\n\nTal Rabin (Universit y of Pennsylvania  / Amazon Web Services)\n\nIn Person and Zoom Participa tion.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91de4f1 DTSTART;TZID=America/New_York:20240906T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240906T120000 URL:https://www.cmu.edu/flame/ LOCATION:Tepper Building 1403 SUMMARY:CMU Foundation and Language Model Center (FLAME) Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91de6b1 DTSTART;TZID=America/New_York:20240906T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240906T170000 URL:https://sites.google.com/andrew.cmu.edu/cmu-workshop-on-cryptography-s/ home LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:CMU Workshop on Cryptography CLASS:PUBLIC DESCRIPTION:Made possible thanks to support from CYLAB and Stellar Foundati on. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91de960 DTSTART;TZID=America/New_York:20240905T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240905T173000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:Carnegie Mellon Graphics Colloquium - Ravi Ramamoorthi CLASS:PUBLIC DESCRIPTION:Speaker: RAVI RAMAMOORTHI\, Ronald L. Graham Professor of Compu ter\nScience\, Founding Director\, UCSD Center for Visual Computing\,\nDep artment of Computer Science and Engineering\, University of\nCalifornia\, San Diego\n\nTalk Title: Sampling and Signal-Processing for High-Dimension al Visual\nAppearance in Computer Graphics and Vision\n\nMany problems in computer graphics and vision\, such as acquiring\nimages of a scene to ena ble synthesis of novel views from many\ndirections for virtual reality\, c omputing realistic images by\nintegrating lighting from many different inc ident directions across a\nrange of scene pixels and viewing angles\, or a cquiring and modeling\nthe appearance of realistic materials like fur or s kin\, require\nsampling and signal-processing on high-dimensional visual a ppearance\nspaces involving changes in lighting\, viewpoint\, spatial loca tion and\nother parameters.  Over my career\, my group has developed a nu mber of\nnovel mathematical and signal-processing tools to address these\n challenges\, significantly reducing the cost of acquisition and\ncomputati on.  In this talk\, we describe significant theoretical and\npractical ad vances in real-time high quality precomputed rendering\,\nMonte Carlo rend ering with orders of magnitude fewer samples\, and\nrealistic novel view s ynthesis.  In all cases\, the methods are now\nwidely deployed in product ion\, and we discuss new computational and\nsignal-processing tools we hav e developed\, including reflection as\nconvolution\, sheared and multiple axis-aligned filtering\, plenoptic\nlight field sampling and neural radian ce fields.  \n\n— \n\nRavi Ramamoorthi is the Ronald L. Graham Profes sor of Computer Science\nat UCSD and founding director of the UC San Diego Center for Visual\nComputing.  He earlier held tenured faculty positions at UC Berkeley\nand Columbia University\, in all of which he played a key leadership\nrole in building multi-faculty research groups recognized as leaders\nin computer vision and graphics.  He has authored more than 200\ nrefereed publications in computer graphics and vision\, including 100+\nA CM SIGGRAPH/TOG papers. He has consulted with Pixar and startups in\ncomp utational imaging\, and currently holds a part-time appointment as\na Dist inguished Research Scientist at NVIDIA.  \n\nProf. Ramamoorthi has recei ved about twenty major honors including the\nACM SIGGRAPH Significant New Researcher Award for his research in\ncomputer graphics\, and the Presiden tial Early Career Award for\nScientists and Engineers for his work on phys ics-based computer\nvision.  He is a fellow of IEEE\, ACM and the SIGGRAP H Academy\,\nreceived two inaugural Frontiers of Science Awards\, and has twice been\nhonored with the edX Prize certificate for exceptional contrib utions\nin online teaching and learning.  He has graduated more than 30\n postdoctoral and Ph.D. students\, whose theses have been recognized by\nth e ACM Dissertation Award honorable mention\, the SIGGRAPH outstanding\ndis sertation award and the UCSD Chancellor's Dissertation Medal.  \n\nFacul ty Host:  Ioannis Gkioulekas \n\nThe Carnegie Mellon Graphics Colloquium is hosted by the Carnegie\nMellon Graphics Lab and generously supported b y Meta and Adobe.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91deec9 DTSTART;TZID=America/New_York:20240905T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240905T160000 URL:https://aco.math.cmu.edu/abs-24-25/sep5.html LOCATION:Wean 8220 SUMMARY:The ACO Seminar - Robert Krueger CLASS:PUBLIC DESCRIPTION:Speaker: ROBERT KRUEGER\, NSF Postdoctoral Researcher\, Departm ent of\nMathematical Sciences\, Carnegie Mellon University\n\nTalk Title: Lipschitz functions on weak expanders\n\nGiven a connected finite graph G\ , an integer-valued function f on V(G)\nis called M-Lipschitz if the value of f changes by at most M along the\nedges of G. In 2013\, Peled\, Samoti j\, and Yehudayoff showed that random\nM-Lipschitz functions on sufficient ly good \"expander\" graphs typically\nexhibit small fluctuations\, giving sharp bounds on the typical range\nof such functions\, assuming M is not too large. We prove that the same\nconclusion holds under a relaxed expans ion condition and for larger M\,\nusing a combination of Sapozhenko's grap h container methods and\nentropy methods. In this talk\, I aim to discuss our result and some\ncontext\, some elements of the proof\, and some open problems. \n\nThis is joint work with Lina Li and Jinyoung Park.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91df229 DTSTART;TZID=America/New_York:20240904T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240904T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Lunch Seminar - Richard Peng CLASS:PUBLIC DESCRIPTION:Speaker: RICHARD PENG\, Associate Professor\, Computer Science\ nDepartment\, Carnegie Mellon University\n\nTalk Title: Survey on Recent P rogresses in Dynamic Matchings\nMathJax.Hub.Config({\ntex2jax: {\ninlineMa th: [ ['$'\,'$'] ]\,\nprocessEscapes: true\n}\n})\;\n\nThis talk discusses the dynamic matching problem\, which seeks to\nmaintain a large matching / small vertex cover in a graph undergoing\nedge insertions/deletions. It covers recent progresses on maintaining\n$(1 + \\epsilon)$-approximations in unweighted bipartite graphs. The\ntwo main topics of focus are sparsifi cation and MWU based reductions\nto vertex-dynamic cases with larger error s\, and connections with\nRusza-Szemeredi graphs and the online matrix-vec tor multiplication\nconjecture in such settings.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91df582 DTSTART;TZID=America/New_York:20240904T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240904T110000 URL:https://www.cmu.edu/dietrich/philosophy/hott/seminars/index.html LOCATION:Baker Hall 150 (special date/time) SUMMARY:Philosophy - Homotopy Type Theory Seminar CLASS:PUBLIC DESCRIPTION:Speaker: JON STERLING\, Associate Professor in Logical Foundati ons and\nFormal Methods\, Department of Computer Science and Technology\,\ nUniversity of Cambridge\, and\, Bye-Fellow\, Clare College\n\nTalk Title: Hofmann–Streicher lifting of fibred categories  In\n1997\, Hofmann and Streicher introduced an explicit technique to lift a\nGrothendieck univer se 𝓤 from 𝐒𝐞𝐭 into the category of\n𝐒𝐞𝐭-valued preshe aves on a 𝓤-small category 𝓑. More\nrecently\, Awodey presented an e legant functorial analysis of this\nconstruction in terms of the ‘catego rical nerve’\, the right\nadjoint to the functor that takes a presheaf t o its category of\nelements\; in particular\, applying the categorical ner ve to the\nuniversal 𝓤-small discrete fibration gives the generic famil y of\n𝓤’s Hofmann–Streicher lifting. \n\n  Although Awodey has in vestigated Hofmann–Streicher lifting in\nterms of a 1-functor 𝐂𝐚 𝐭→𝐏𝐫(𝓑)\, his analysis can\nbe extended to a 2-functor 𝐂 𝐚𝐭→𝐅𝐢𝐛(𝓑) that is\nobserved by Weber to be right 2-adj oint to the 2-functor that takes a\nfibred category to its total category (i.e. the oplax colimit of the\ncorresponding diagram of categories under straightening). A\ngeneralised form of Hofmann–Streicher lifting that ca n be applied to\ncategories other than universes is then obtained by conju gating this\nright 2-adjoint with duality involutions. \n\n  In joint wo rk with Daniel Gratzer and Andrew Slattery\, we have\nconstructed a relati ve version of the 2-functorial Hofmann–Streicher\nlifting: given a fibra tion p:𝓐→𝓑\, we have a 2-functor\nΔ[p]:𝐅𝐢𝐛(𝓑)→𝐅 𝐢𝐛(𝓐) which is not base change\nbut rather (we conjecture) right pseudo-adjoint to the 2-functor\nΣ[p]:𝐅𝐢𝐛(𝓑)→𝐅𝐢𝐛( 𝓐) that sends a fibration\nq:𝓔→𝓐 to the composite fibration p ∘q:𝓔→𝓑. A relative\nversion of Hofmann–Streicher lifting could give a more regular\ntheory to the practice of computing internal lifting s of lifted\nuniverses.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91df9a2 DTSTART;TZID=America/New_York:20240904T070000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240905T160000 LOCATION:David L. Lawrence Convention Center\, Pittsburgh PA 15222 SUMMARY:2024 Secure and Trustworthy Cyberspace Principal Investigators’\n Meeting CLASS:PUBLIC DESCRIPTION:The NSF SaTC PI Meeting is a biennial forum of the SaTC researc h\ncommunity with leading experts from academia\, industry\, and Federal\n agencies\, who will come together to discuss game changing challenges\nres ulting from the global adoption of cyberspace by: Reviewing new\ndevelopme nts in SaTC fundamental ideas and concepts that minimize the\nmisuses of c ybersecurity Discussing ways to bolster education and\ntraining in cyberse rcurity Identifying new\, emerging applications\nTransitioning promising r esearch into practiceThe National Science\nFoundation established the SaTC program to protect cyber-systems\n(including host machines\, the internet and other cyber-infrastructure)\nfrom malicious behavior\, while preservi ng privacy and promoting\nusability. The SaTC program supports a broad spe ctrum of innovative\nresearch that will improve the resilience of individu al hosts\,\nnetworked systems\, hardware\, software\, applications and cri tical\ninfrastructure from malicious cyber-attacks while preserving privac y\nand promoting effective and safe usability. SaTC recognizes that this\n is not only a problem of developing trustworthy computing technology\,\nbu t also of understanding the economic\, social\, and behavioral factors\nth at affect its use and deployment. Addressing this problem requires\nmulti- disciplinary expertise in human\, statistical\, mathematical\,\ncomputatio nal\, and computer sciences and ultimately the transition of\nnew concepts and technologies to practice\, as well will require the\nexpertise and re sources from a wide range of disciplines\, including\ncomputer science\, e ngineering\, economics mathematics\, and behavioral\nsciences. Forerunners of the SaTC program include the Trustworthy\nComputing\, Cyber Trust\, an d Trusted Computing programs.\nThe program of the meeting includes invited keynotes and PI talks\,\npanels\, and breakout discussions about SaTC res earch-related projects\nfunded by NSF and other Federal agencies. The prog ram is being\norganized by a program committee comprising SaTC community e xperts\nidentified by NSF.\nRegistration InformationHosted by the Carnegie Mellon University's\nCyLab Security and Privacy Institute\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91dfddd DTSTART;TZID=America/New_York:20240903T143000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240903T160000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-mingxun-zhou LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Doctoral Thesis Proposal - Mingxun Zhou CLASS:PUBLIC DESCRIPTION:Speaker: MINGXUN ZHOU\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Practical Private Inform ation Retrieval and Searching with\nSublinear Cost\nMathJax.Hub.Config({\n tex2jax: {\ninlineMath: [ ['$'\,'$']\, [\"\\\\(\"\,\"\\\\)\"] ]\,\nprocess Escapes: true\n}\n})\;\n\nPrivate Information Retrieval (PIR) is a long-st udied cryptographic\nprimitive that allows a user to retrieve information from a public\ndatabase without revealing the query to the service provid er.\nClassically\, PIR was studied in a setting without any preprocessing\ ,\nand this setting is known to have inherent limitations that the total\n computation cost per query must be linear in the size of the database\nto achieve privacy. This limitation is the fundamental barrier for\nscaling P IR to large databases and building practical product systems\nbased on PIR . Pioneered by Beimel\, Ishai and Malkin (Crypto 2000) and\nCorrigan-Gibbs and Kogan (Eurocrypt 2020)\, the preprocessing PIR\nparadigm has been sho wn to overcome the linear computation cost\nbarrier. Nonetheless\, prior w orks on preprocessing PIR remain mostly\nin the theoretical space due to t heir use of heavy cryptographic\nprimitives and/or convoluted algorithmic constructions.\n\nAchieved Results: We proposed a practical single-server PIR\nconstruction in the preprocessing setting\, named Piano (S P 2023)\,\ nthat achieved sublinear query cost based on lightweight cryptographic\npr imitives. Piano achieved $\\tilde O(\\sqrt{n})$ amortized\ncommunication a nd computation per query given a database of size n \,\nand only required $\\tilde O(\\sqrt{n})$ client storage. Notably\, it is\nalso concretely ef ficient – for a 100GB database of 1.6 billion\nentries\, Piano takes onl y 12ms online computation time on a single\nCPU-core. Subsequently\, we im proved the construction and obtained a\nnew practical PIR scheme\, Quarter PIR (Eurocrypt 2024)\, that reduced\nthe online communication cost to $\\t ilde O(n^{1/4})$ per query\, while\nmaintaining competitive practical perf ormances. Proposed Direction:\nNearly all existing PIR constructions are d esigned for point accesses\nin an array-type database\, while many practic al information retrieval\nsystems like search engines do need more advance d access algorithms to\nsupport semantic and similarity queries. We propos e to construct a\nprivate search algorithm that can handle semantic/simila rity queries\nwith sublinear communication and computation query costs bas ed on our\npractical PIR constructions and graph-based Nearest Neighbor Se arch\n(NNS) algorithms. This construction will be the first private search \nalgorithm that achieves sublinear query cost\, and it has the potential\ nto support multi-modal search queries including text\, image\, voice and\ nvideo search.\n\nThesis Committee: \n\nElaine Shi (Co-chair)\n\nGiulia F anti (Co-chair)\n\nBryan Parno\n\nDavid Wu (University of Texas at Austin) \n\nAdditional Information\n\nIn Person and Zoom Participation. See announ cement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e0323 DTSTART;TZID=America/New_York:20240903T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240903T213000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-afonso-tinnoco LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Doctoral Thesis Proposal - Afonso Tinnoco CLASS:PUBLIC DESCRIPTION:Speaker: AFONSO TINOCO\, Ph.D. Student\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Towards Practical and V erifiable Distributed Systems:\nApplications of Oblivious Algorithms\, Gar bled Circuits and Formal\nMethods+\n\nIn my research\, I want to design an d develop practically efficient and\nprovably secure distributed systems. To this end\, I combine applied\ncryptography and formal verification tool s. The goal of this proposal\nis to provide primitives that can be used in distributed systems\, as\nwell as methodologies to verify consensus proto cols\, to ensure the\nsecurity\, efficiency\, and correctness of distribut ed systems.\nSpecifically\, the research focuses on three main areas: prac tical\nimplementations of oblivious algorithms for Trusted Execution\nEnvi ronments (TEEs)\; practical and verified implementations of Garbled\nRAM\; and methodologies to verify safety and liveness of consensus\nprotocols.  \n\nTEEs can be used to offer efficient crash fault nodes with\nconfiden tial computations\; however\, most TEEs implementations leak the\npage-lev el memory access pattern to the host machine where the TEE is\nrunning. Th erefore\, to guarantee confidential computations\, TEE\nprograms need not only correct implementations but also to be memory\ntrace-oblivious. There has been extensive theoretical research in\noblivious algorithms\; howeve r\, there is a gap with practical\nimplementations\, particularly in the T EE setting. In this proposal\, we\naim to close this gap\, providing an ob livious data structure library\nakin to C++'s STL\, and extending it with oblivious graph algorithms. \n\nTEEs require trust in the hardware manufa cturer and a certain level of\nhardware/software integrity. In scenarios w here this isn't possible\,\nGarbled Circuits can be used to provide equiva lent secure processor\nguarantees based on cryptographic assumptions. To a chieve an efficient\nGarbled Circuit processor\, Garbled RAM is necessary\ , and recent\ntheoretical advancements suggest using tristate circuits to implement\nit. In this proposal\, we aim to achieve concretely more effici ent\nGarbled RAM constructions\, as well as provide methodologies to verif y\nthe correctness of tristate circuits. \n\nFinally\, while TEEs and Gar bled RAM aim to achieve secure\ncomputations\, ensuring the correctness of the underlying protocols\nthat use them is critical. We have previously d eveloped a python DSL\nto verify safety properties of distributed system p rotocols. In this\nproposal\, we aim to extend this framework to verify li veness\nproperties of distributed system protocols\, focusing on proof\nau tomation and generalization.  \n\nThesis Committee: \n\nElaine Shi (Co- chair)\n\nRodrigo Rodrigues (Co-chair\, University of Lisbon\, Instituto S uperior\nTécnico)\n\nBryan Parno\n\nPedro Adão (University of Lisbon\, I nstituto Superior Técnico)\n\nJosé Fragoso (University of Lisbon\, Insti tuto Superior Técnico)\n\nAndrew Miller (University of Illinois Urbana-Ch ampaign)\n\n \n\nAdditional Information\n\nIn Person and Zoom Participati on. See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e08a0 DTSTART;TZID=America/New_York:20240828T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240828T170000 URL:http://www.cs.cmu.edu/~theorylunch/ LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Theory Lunch Seminar - Angelos Pelecanos CLASS:PUBLIC DESCRIPTION:Speaker: ANGELOS PELECANOS\, Ph.D. Student in Computer Science\ , The\nDepartment of Electrical Engineering & Computer Sciences\, Universi ty\nof California\, Berkeley\n\nTalk Title: On the t-wise independence of block ciphers\n\nBlock ciphers such as the Advanced Encryption Standard (R ijndael) are\nused extensively in practice\, yet our understanding of thei r security\ncontinues to be highly incomplete. We will present some result s\nproving the security of block ciphers against important and\nwell-studi ed classes of attacks. In particular\, we will introduce a\nconcrete block cipher construction paradigm called the\nsubstitution-permutation network (SPN) and study its (almost) t-wise\nindependence as a relevant and meani ngful property that captures\nsecurity. We will also survey some recent t- wise independence results\nthat are not limited to the SPN paradigm.  \n \nBased on works with Tianren Liu\, Stefano Tessaro\, Vinod\nVaikuntanatha n:  https://eprint.iacr.org/2021/507\,\nhttps://eprint.iacr.org/2024/083 and Lucas Gretta\, William He: \nhttps://eprint.iacr.org/2024/847\n\nSpec ial Time and Location\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e0c41 DTSTART;TZID=America/New_York:20240827T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240827T140000 URL:https://csd.cmu.edu/calendar/speaking-skills-talk-victor-akinwande LOCATION:Newell-Simon 4305 SUMMARY:Speaking Skills Talk - Victor Akinwande CLASS:PUBLIC DESCRIPTION:Speaker: VICTOR AKINWANDE\, Ph.D. Student\, Computer Science De partment\,\nCarnegie Mellon University\n\nTalk Title: HyperCLIP: Adapting Vision-Language models with\nHypernetworks\n\nSelf-supervised vision-langu age models trained with contrastive\nobjectives perform  better as one in creases their scale. Typically\,\nthe image encoder in such models are lar ger than the text encoder and\nwe are often able to amortize the inference cost of the text encoder\nby using a predefined set of text-embedding but not with the image\nencoder. This poses a challenge for deploying large v ision-language\nmodels especially in resource-constrained environments. \ n\nIn this talk\, I will present HyperCLIP - a vision-language\narchitectu re that dynamically adapts a small image encoder using a\nhypernetwork. Th is hypernetwork learns to produce a subset of the\nimage encoder parameter s conditioned on the text-embedding\, and the\nentire model (hypernetwork\ , image encoder\, and text encoder) are\ntrained jointly end-end. HyperCLI P increases the zero-shot accuracy of\nSigLIP models with small image enco ders by up to 3% on ImageNet and 5%\non CIFAR-100 with minimal training th roughput overhead. \n\nPresented in Partial Fulfillment of the CSD Speaki ng Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e0fbb DTSTART;TZID=America/New_York:20240826T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240826T203000 SUMMARY:First Day of Fall Term CLASS:PUBLIC DESCRIPTION:Talk Title: First Day of Classes F24\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e11f7 DTSTART;TZID=America/New_York:20240823T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240823T163000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-brian-zhang LOCATION:Gates Hillman 8102 SUMMARY:Doctoral Thesis Proposal - Brian Zhang CLASS:PUBLIC DESCRIPTION:Speaker: BRIAN ZHANG\, Ph.D. Student\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: New Solution Concepts\, A lgorithms\, and Applications for\nExtensive-Form Games: Learning\, Correla tion\, Communication\, and Common\nKnowledge\n\nComputational game theory has led to significant breakthroughs in AI\ndating back to the start of AI as a discipline. These include the\nstrongest AI agents for both recreati onal and practical applications.\nIt has been instrumental in enabling sup erhuman AI from recreational\ngames such as two-player zero-sum games ches s\, go\, and heads-up poker\nto multiplayer games such as six-player poker and Hanabi\, and even in\ngames involving human language such as Diplomac y. It has also\nempowered a growing range of non-recreational applications \, such as\ntrading\, machine learning robustness and safety\, negotiation \, conflict\nresolution\, mechanism design\, information design\, security \, political\ncampaigning\, and self-driving cars. \n\nThis thesis pushes the boundary on computational game theory\,\nespecially in imperfect-info rmation sequential (extensive-form) games\,\nwhich are most prevalent in p ractical applications both in zero-sum\ngames and beyond. We present new t heoretical concepts and frameworks\,\nstate-of-the-art and often provably optimal algorithms for computing\nand learning equilibria\, and new ways t o apply such algorithms to\nreal-world problems\, including problems in ec onomics such as mechanism\nand information design.\n\nThe thesis contains four parts.\n\nI)  We develop new solution concepts and state-of-the-art complexity\nresults for adversarial team games. Among other results\, we c ompute\nexact solutions to several variants of the popular game The\nResis tance: Avalon with up to six players.\n\nII)  We develop an algorithmic f ramework for generalized mechanism\ndesign that covers sequential mechanis m design\, sequential information\ndesign\, optimal correlated equilibria and more for the first time\, and\nreduces them to zero-sum games. This en ables computation using any\nzero-sum game solving technique\, including d eep reinforcement\nlearning.\n\nIII)  We develop the fastest learning alg orithms for minimizing\nregret against linear and low-degree deviations\, which are the\ntightest solution concepts known to be efficiently learnabl e in games.\n\nIV)  We develop new techniques for subgame solving that wo rk even\nwhen the common-knowledge set is too large to work with\, and use these\ntechniques to build the first strong bot—and\, to our knowledge\ ,\ncurrently the best bot—for the game of dark chess. \n\nThesis Commit tee\n\nTuomas Sandholm (Chair)\n\nVincent Conitzer\n\nZico Kolter\n\nKevin Leyton-Brown (University of British Columbia)\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e16fa DTSTART;TZID=America/New_York:20240823T121500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240823T134500 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-arjun-laksmipathy LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Doctoral Thesis Proposal - Arjun Laksmipathy CLASS:PUBLIC DESCRIPTION:Speaker: ARJUN LAKSMIPATHY\, Ph.D. Student\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Contact Areas for D exterous Manipulation\n\nHumans use their hands to effortlessly manipulate objects of\narbitrarily complex geometries and physical properties every day\;\nhowever\, adapting such manipulations to dexterous robots and virtu al\ncharacters is an extremely difficult task. Understanding the ways in\n which humans exploit contact to perform these manipulations has the\npoten tial to greatly advance progress towards this goal. \n\nUnsurprisingly\, research efforts have analyzed contact in the context\nof dexterous manipu lation for decades. We now have numerous metrics\nfor evaluating grasp qua lity in terms of contacts\, efficient means of\ncomputing contact in physi cal simulation\, and countless strategies\nexploiting contact corresponden ces between hands and objects to\nsynthesize grasps and manipulations. But the majority of existing\nworks fundamentally characterize contact in the same way: as points\,\nlines\, or planes of interaction. \n\nBut contact s in the real world are much more complicated. Instead\,\nreal bodies inte rface with one another via areas of contact which\ngreatly vary with the g eometries of contacting surfaces. If we wish to\nmodel the complexities of manipulations as they actually occur\, then\nwe must progress beyond such simplifying assumptions and deal with the\nmessy nature of reality. This thesis aims to do so by presenting\nframeworks and algorithms for the mode ling\, capture\, mutation\, and\nexploitation of contact areas. Our intent ion is to establish the\nfoundations necessary to elevate contact areas to first-class\nprimitives and demonstrate their inherent value across a ran ge of\npractical applications in dexterous manipulation and adjacent\ndoma ins.    \n\nFirst\, we introduce three novel models of contact areas al ongside\noperations supported by each model which are fundamentally design ed to\nrun on real geometries rather than simple primitive shapes. Second\ , we\nintroduce a low cost approach for capturing human contacts from the\ nreal world. Then\, using these models\, we introduce: a new set of\nintui tive artist tools for drafting high quality grasps\, a novel\nmotion retar geting pipeline for dexterous manipulations\, a novel\ncontact-driven cont rol framework for dexterous robot hands\, and two\npractical extensions of our work. The contributions in this thesis are\nnot intended to be the la st words\, but rather important first steps\ndesigned to promote future re search in contact area modeling.\n\nThesis Committee\n\nNancy S. Pollard ( Chair)\n\nJessica K. Hodgins\n\nKeenan Crane\n\nC. Karen Liu (Stanford Uni versity)\n\nAdditional Information\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e1c57 DTSTART;TZID=America/New_York:20240823T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240823T140000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-oral-defense-daniel-linkit -wong LOCATION:Panther Hollow Conference Room\, Mehrabian Collaborative Innovatio n\nCenter 4th Floor SUMMARY:Doctoral Thesis Oral Defense - Daniel Lin-Kit Wong CLASS:PUBLIC DESCRIPTION:Speaker: DANIEL LIN-KIT WONG\, Ph.D. Candidate\, Computer Scien ce\nDepartment\, Carnegie Mellon University\n\nTalk Title: Machine Learnin g for Flash Caching in Bulk Storage Systems\n\nFlash caches are used to re duce peak backend load for\nthroughput-constrained data center services\, reducing the total number\nof backend servers required. Bulk storage syst ems are a large-scale\nexample\, backed by high-capacity but low-throughpu t hard disks\, and\nuse flash caches to provide a cost-effective storage layer underlying\neverything from blobstores to data warehouses. \n\nHow ever\, flash caches must manage their limited write endurance and\nlimit  the flash write rate to avoid premature wear-out. They do so\nvia admissio n policies that filter cache insertions and maximize the\nworkload-reduct ion value of each write. \n\nI evaluate and demonstrate potential uses of ML in place of\ntraditional heuristic cache management policies for flas h caches in\nbulk storage systems. The most successful elements of my res earch are\nembodied in a flash cache system called Baleen\, which uses\nc oordinated ML admission and prefetching to reduce peak backend load.\nAft er learning painful lessons with early ML policy attempts\, I\nexploit a new cache residency model (episodes) to guide model\ntraining. I focus on optimizing an end-to-end metric (Disk-head Time)\nthat measures backend load more accurately than IO or byte miss rate.\nEvaluation using 7-day M eta traces from 7 storage clusters shows\nBaleen reducing Peak Disk-head Time (and backend hard disks required)\nby 12% over state-of-the-art poli cies for a fixed flash write rate. \n\nI present a TCO (total cost of ow nership) formula quantifying\nthe costs of additional flash writes agains t reductions in Peak\nDisk-head Time in terms of flash drives and hard di sks needed.\nBaleen-TCO chooses optimal flash write rates and reduces est imated\nTCO by 17%. \n\nWorkloads change over time\, requiring that cache s adapt to\nmaintain performance. I present a strategy for peak load\nred uction that adapts selectivity to load levels. I evaluated\nworkload drif t and its impact on ML policy performance on 30-day Meta\ntraces. \n\nBa leen is the result of substantial exploration and experimentation\nwith ML  for caching. I present lessons learned from additional\nstrategies consi dered and explain why they saw limited success on our\nworkloads. These  include improvements for ML eviction and more\nadvanced ML models. Code an d traces are available\n\nThesis Committee\n\nGregory R. Ganger (Chair)\n\ nNathan Beckmann\n\nDavid G. Andersen\n\nDaniel S. Berger (Microsoft Resea rch / University of Washington)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e2198 DTSTART;TZID=America/New_York:20240822T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240822T170000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-oral-defense-juncheng-yang LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Doctoral Thesis Oral Defense - Juncheng Yang CLASS:PUBLIC DESCRIPTION:Speaker: JUNCHENG YANG\, Ph.D. Candidate\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Designing Efficient a nd Scalable Cache Management Systems\n\nSoftware caches have been widely d eployed at scale in today's\ncomputing infrastructure to improve data acce ss latency and\nthroughput. These caches consume PBs of DRAM at many compa nies\, which\nnecessitates high efficiency --- achieving the same miss rat io with\nless DRAM consumption. Meanwhile\, modern servers have hundreds o f\ncores\, making scalability a critical requirement for designing\nsoftwa re caches. This thesis explores different approaches to\nimproving the eff iciency and scalability of software caches. \n\nThis thesis has two parts . The first part focuses on system designs\nthat allow caches to store mor e objects in the cache to achieve a low\nmiss ratio. In this part\, I will describe three works. First\, I will\ndiscuss what key-value cache worklo ads at Twitter look like using a\nlarge-scale workload analysis. Second\, drawing on insights from the\nworkload study\, I will describe the design of Segcache\, a TTL-indexed\nsegment-structured key-value cache that quick ly removes expired\nobjects\, provides tiny object metadata\, and enables close-to-linear\nscalability. Third\, I will present C2DN to demonstrate a \nfault-tolerant CDN cache cluster using erasure coding for low-overhead\n redundancy. \n\nThe second part focuses on algorithms that allow the cach e to store\nmore useful objects in the cache\, which is also critical for cache\nefficiency. First\, I will investigate the design of a low-overhead \nlearned cache. Existing caches using machine learning often incur\nsigni ficant storage and computation overheads. I will show that\nlearning on th e group level amortizes overheads and accumulates more\ninformation for be tter learning. While GL-Cache is faster than\nexisting learned caches\, it is still more complex compared to simple\nheuristics. In the following ch apter\, I will discuss two techniques\,\nlazy promotion and quick demotion \, which enable us to design simple\nyet effective eviction algorithms. In the third chapter\, I will\ndiscuss an example using the two techniques\, S3-FIFO\, a new eviction\nalgorithm only composed of FIFO queues. In the last chapter\, I will\npresent SIEVE\, a new eviction algorithm that uses one queue to achieve\nlazy promotion and quick demotion. SIEVE is simpler than LRU\, but\nachieves state-of-the-art efficiency and scalability.\n\nT hesis Committee: \n\nRashmi Vinayak (Chair)\n\nGreg Ganger\n\nPhillip Gib bons\n\nVijay Chidambaram (University of Texas at Austin)\n\nIon Stoica (U niversity of California\, Berkeley)\n\n \n\nn Person and Zoom Participati on.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e2665 DTSTART;TZID=America/New_York:20240821T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240821T130000 URL:https://www.pdl.cmu.edu/talk-series/index.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Talk - Vaibhav Arora CLASS:PUBLIC DESCRIPTION:Speaker: VAIBHAV ARORA\, Principal Member of Technical Staff\,\ nSalesforce\n\nTalk Title: LSM Management and using LSM immutability for d ata\nvirtualization\n\nLSM (Log-Structured Merge) trees are now the bedroc k of many storage\nengines and datastores like RocksDB\, HBase\, Cassandra etc. They\nprovide the ability to avoid random-writes\, and provide immut ability.\nData is organized in multiple-levels that are exponentially incr easing\nin size. Each data mutation writes a new version of an object\, an d\nbackground processes named merge/compaction continuously remove the\nun used versions\, while moving the data across the layers of the LSM\ntree a nd maintain its shape. \n\nThis talk will describe how the immutability o f LSMs can be used to\nprovide data virtualization. Since the underlying d ata in persistence\nin an LSM never changes\, it provides an opportunity t o build a\nvirtualization layer over it. \n\nIn this talk\, we will descr ibe a mechanism to use metadata in the form\nof many to 1 references over data files in the LSM. This metadata can\nthen be used to create constant- time clones of data-sets without\nphysically copying the data. These clone s can then be used for\ntesting\, and experimentation and also for taking back-ups. The cloning\ncan also be used for fast data-migration between mu ltiple datastores\nwith LSM based-storage\, running on a common distribute d storage\nlayer. \n\n— \n\nVaibhav Arora works at Salesforce where he works on various\nlarge-scale data management problems including LSM\, mu lti-tenancy\,\ntransaction processing\, indexing and other access layer related\nchallenges in the database engine. Prior to joining Salesforce\, he\ndid his PhD in Computer Science from University of California\, Santa \nBarbara\, where he worked on problems related to data variety - both in\ nstructure and access. Aside from that\, he has worked on various\ndistrib uted database problems during his time at Amazon\, Microsoft\nResearch\, H P Labs and Yahoo!. Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e2aa5 DTSTART;TZID=America/New_York:20240819T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240819T130000 URL:https://www.cylab.cmu.edu/events/2024/08/29-seminar-solt.html LOCATION:Panther Hollow Conference Room\, Mehrabian Collaborative Innovatio n\nCenter 4105 SUMMARY:CyLab Seminar - Flavien Solt CLASS:PUBLIC DESCRIPTION:Speaker: FLAVIEW SOLT\, Postdoctoral Researcher\, Department of \nInformation Technology and Electrical Engineering\, COMSEC Group\, ETH\n Zürich\n\nTalk Title: Software Inspired Techniques for Digital Hardware S ecurity\n\nWe entered an era where new hardware flourishes at an unprecede nted\npace and with unseen diversity. We are also living in an era where\n security and safety are paramount\, and where the potential impact of a\ns ingle bug can be catastrophic. Hence\, we urgently need foundations to\nde tect as many hardware bugs as possible before their deployment.\n\nHardwar e validation is universally recognized as complex\, expensive\nand tedious . Despite genuine best efforts\, the last decade has shown\nthat the indus try is incapable of producing non-trivial bug-free\nhardware. What will th en happen with the rise of open-source hardware?\nWithout effective and ea sy-to-adopt solutions for validation\, it is\nhard to believe that the ope n-source hardware community will be able\nto produce safe and secure hardw are\, despite its best intentions.\n\nInterestingly\, the exact same situa tion occurred in the software world\nsome decades ago. Software was plague d with myriads of bugs and\nsecurity issues\, after what the software comm unity developed a\nformidable set of tools and methodologies to detect bug s and security\nissues. Could we adapt some of these tools and methodologi es to\nhardware?\n\nTo answer this question\, we first observe many CPU er rata\, deduce the\nmost promising techniques from software security\, and adapt them. To\nunderstand contemporary CPU bugs\, we build the RemembERR database\nbased on thousands of errata. We deduce two techniques inspired by\nsoftware security that are particularly promising for hardware:\ndynam ic information flow tracking and fuzzing. We introduce CellIFT\,\nthe firs t scalable hardware dynamic information flow tracking\nmechanism and showc ase 4 new architectural or microarchitectural\nsecurity applications. We t hen introduce Cascade\, a black-box CPU\nfuzzer that found dozens of new b ugs and outperforms other fuzzers’\ncoverage. We finally demonstrate MiR TL\, a new class of hardware\nattacks that relies on EDA software bugs\, a nd propose TransFuzz\, a\nfuzzer that produces complex hardware descriptio ns to find such bugs\nin popular open-source EDA software.\n\nAll these co ntributions demonstrate that when properly adapted\,\nsoftware security te chniques can provide effective and easy-to-adopt\nsolutions that will empo wer safer and more secure hardware.\n\n— \n\nFlavien Solt is a postdoct oral researcher in the Department of\nInformation Technology and Electrica l Engineering at ETH Zurich at the\nCOMSEC group. His research mainly focu ses on digital hardware security\nand led to multiple publications in top- tier security or computer\narchitecture venues (e.g.\, S&P\, USENIX Securi ty\, MICRO\, etc.).\n\nFaculty Host: Riccardo Paccagnella \n\nIn Person a nd Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e2ff2 DTSTART;TZID=America/New_York:20240819T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240830T170000 URL:https://www.cs.cmu.edu/~csd-ic/ LOCATION:Gates Hillman Center SUMMARY:Computer Science Department New Ph.D. Introductory Course CLASS:PUBLIC DESCRIPTION:Additional information provided by your program and event manag ers.\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e32a1 DTSTART;TZID=America/New_York:20240814T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240814T130000 URL:https://pdl.cmu.edu/talk-series/2024/081424.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Talks - Javier González & Arun George CLASS:PUBLIC DESCRIPTION:Speaker: JAVIER GONZÁLEZ and ARUN GEORGE\, Javier: Principal E ngineer\,\nArun: Associate Technical Director\, Samsung\n\n► Javier Gonz ález\, Principal Engineer\, Samsung     \n\n—  Enabling High-Capac ity SSDs in the Open-Ecosystem \n\nQLC NAND is gaining mainstream accepta nce in cloud and enterprise\nenvironments. While past NAND technology evol ution resulted in a high\npercentage of turnover (i.e.\, SLC to MLC to TLC )\, the characteristics\nof QLC makes it suitable to cover new use cases i nvolving\nHigh-Capacity SSDs. It has also prompted the opportunity to enha nce\nQLC adoption further. In order to transparently enable these SSDs in\ nexisting applications\, we need changes at the operating system\nlevel.  \n\nIn this talk\, we will present the three main challenges when\nintrodu cing QLC-based High-Capacity SSDs in existing applications:\nScalability\, Endurance and Performance. We will report how we are\npreparing the ecosy stem to address these challenges using open\nstandards and open-source. Sp ecifically\, we will cover our progress on\n(i) enabling and enhancing QLC drives with larger Indirection Units\n(IUs) by supporting larger block si zes\, (ii) increasing QLC endurance\nthrough Flexible Data Placement (FDP) \, and (iii) improving performance\nin mainstream databases by leveraging the atomicity associated with\nlarger IUs. \n\nJavier González leads Sam sung Semiconductor's Global Open-ecoSystem\nTeam (GOST). I take care of ou r ecosystem activities and manage a\ndistributed team of highly talented e ngineers. This includes defining\nour vision\, strategy\, internal / exter nal communication\, and\nday-to-day execution. I am also the founder and s ite manager for\nSamsung Semiconductor Denmark Research (SSDR) - Samsung ’s Memory\nSolutions first R&D center in Europe and fifth worldwide. I a m a Ph.D\nin operating systems with a strong background in technical leade rship\,\nexperimental research\, and Linux Kernel development. My interest s lay\nprimarily in the hardware / software co-design space\, where system s\,\nhardware architecture\, and open-ecosystem meet. I am dedicated to\nd efining safe environments for motivated software engineers to be\ncreative and get things done. I am a contributor to a wide range of\nopen source p rojects including the Linux Kernel\, as well as to the\nNVMe Specification . I am a regular speaker at several top industry and\nacademic conferences . \n\n►  Arun George\,  Associate Technical Director\, Samsung\n       \n\n—  Mitigating the Write Amplification Challenges in CacheLi b using\nFDP \n\nCacheLib is an open source caching engine from Meta to b uild high\nthroughput\, low overhead caching services. It has the built-in ability\nto transparently leverage DRAM and SSD. Flexible Data Placement( FDP)\nusing NVMe is the latest technology in the field of Data Placement i n\nSSDs. This talk explains how FDP helps Cachelib to mitigate the device\ nwrite amplification(WAF) challenges in SSD. We will also discuss how\nFDP can help the Cachelib deployments in the multi-tenant use cases.\nThis co ver how FDP can mitigate the device WAF challenges arising due\nto the dif fering IO patterns of the various tenants. \n\nArun George — Software A rchitect\, Developer and Leader\; Contributing\nto Storage and Memory doma ins. Working at SSIR\, Samsung Research India\nLtd\, Bangalore. Loves deve loping software for Next Generation Storage\nSystems and distributed file systems. Researching and Innovating on\nStorage solutions for Cloud Deploy ments with emphasis on Performance\nand Latencies. Focus on Enterprise Sys tems and Datacenter technologies\nnow\; with focus in Embedded Systems in early part of career. Interests\ninclude Distributed Storage\, Open Source Technologies\, Linux Kernel\,\nCeph\, File Systems\, SPDK etc. Contribute d to cutting-edge storage\ndevice technologies such as SCM\, NVDIMM-C\, NV Me etc.. Contributed to\nSystem Software and Firmware development for many storage devices such\nas SSDs\, Micro-SD\, NAND etc. 11 patents\, 1 Publi cation. \n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e383b DTSTART;TZID=America/New_York:20240812T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240812T173000 URL:https://csd.cmu.edu/calendar/thesis-proposal-runtian-zhai LOCATION:Traffic21 Classroom\, Gates Hillm 6501 and Zoom SUMMARY:Thesis Proposal - Runtian Zhai CLASS:PUBLIC DESCRIPTION:Speaker: RUNTIAN ZHAI\, Ph.D. Student\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Learning Generalizable a nd Transferable Representations\nwith Big Models\n\nMachine learning has s hifted to a new paradigm driven by\nrepresentation learning and foundation models\, big encoders that\nextract useful features from data. This thesi s studies how big models\nlearn good representations\, and especially focu ses on two aspects:\ngeneralization and transferability. For generalizatio n\, the problem is\nwhy big foundation models wouldn’t overfit as classi cal theory\nsuggests.\n\nMy work proves a generalization bound that works for big models\, by\nviewing them as algorithmic models instead of data mo dels. For\ntransferability\, my work focuses on reweighting\, the most pop ular\nclass of methods. This talk will focus on one issue with reweighting \nthat is its sensitivity to outliers\, and propose a solution that\nsigni ficantly improves the performance and stability of reweighting.\nFinally\, I will propose two future work\, feature learning for tabular\ndata\, and combining multiple sources of prior knowledge.\n\nThesis Committee:\n\nPr adeep Ravikumar (Co-chair)\n\nZico Kolter (Co-chair)\n\nAndrej Risteski\n\ nYuandong Tian (Meta AI)\n\nAdditional Information\n\nIn Person and Zoom P articipation. See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e3c0f DTSTART;TZID=America/New_York:20240808T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240808T133000 LOCATION:Gates Hillman 9115 SUMMARY:5th Year Thesis Presentation - Piratach (Bas) Yoovidhya CLASS:PUBLIC DESCRIPTION:Speaker: PIRATACH (BAS) YOOVIDHYA\, Master's Student\, Computer Science\nDepartment\, Carnegie Mellon University\n\nTalk Title: Mechanism s for Efficient Memory Access in Near-Cache\nAccelerators\n\nIn traditiona l computer systems\, data has to move through the memory\nhierarchy for co mputation to take place within the core. This data\nmovement cost has been dominating computer systems’ performance\, and\nwill only get worse ove r time. Many proposals address this problem by\nintroducing architectures that move compute closer to data.  Like\nsome of these proposals\, our ap proach to this places engines within\nthe cache hierarchy\, allowing the c ore to offload work to the caches.\nWhen the engine experiences a cache mi ss\, the requested data could be\nresiding at two different levels within the memory hierarchy. Sending\na request to only one of the two locations could result in a miss\,\nincreasing the miss latency of the engine. Howev er\, sending a request\nto both locations at once also leads to higher ene rgy consumption. \n\nIn this thesis\, we introduce a novel Memory Access Predictor to the\nsystem that assists the engine in sending requests that minimizes\nenergy usage\, while retaining high performance. We evaluate th e\npredictor on various micro-benchmarks\, showing that it is able to\nimp rove the performance by 21%\, and reduces additional energy\nconsumption b y 83% in other applications. \n\nThesis Committee:\n\nNathan Beckmann (Ch air)\n\nPhillip Gibbons\n\n \n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e3ff6 DTSTART;TZID=America/New_York:20240807T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240807T130000 URL:https://pdl.cmu.edu/SDI/index.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Talk - Suhas Jayaram Subramanya CLASS:PUBLIC DESCRIPTION:Speaker: SUHAS JAYARAM SUBRAMANYA\, Ph.D. Student\, Computer Sc ience\nDepartment\, Carnegie Mellon University\n\nTalk Title: Sia: Heterog eneity-aware\, goodput-optimized ML-cluster\nscheduling\n\nLarge GPU clust ers are increasingly becoming more heterogeneous due to\nadvances in GPU d esign and incremental deployment of a mix of GPU\ntypes over time. Deep le arning (DL) training jobs running on these GPU\nclusters can see varying j ob completion times depending on the\nresources allocated by the cluster s cheduler and job hyper-parameters\nconfigured by users at the time of job submission. Sia is a cluster\nscheduler that (1) efficiently assigns heter ogeneous GPU resources to\nelastic resource-adaptive DL training jobs\, an d (2) configures the job\nhyper-parameters to maintain high training effic iency for all running\njobs without sacrificing the quality of trained mod els.\n\nWe will discuss challenges in optimizing resource-adaptivity for d eep\nlearning training (DLT) jobs on large clusters with many GPU types\,\ nand introduce a new scheduling formulation that efficiently matches\nDLT jobs and their configurations to GPU types and counts\, while\nadapting to changes in cluster load and job mix over time. On job\ntraces derived fro m real datacenters\, Sia improves job completion\ntimes by 30-93% while us ing 12-60% fewer GPU hours. Furthermore\, its\nscheduling policy is quick to evaluate and easily scales to GPU\nclusters with many GPU types and 100 0s of GPUs.\n\n—\n\nSuhas Jayaram Subramanya is a final-year PhD student in the CS\nDepartment\, advised by Prof. Greg Ganger. His primary researc h area is\ndeep learning systems.\nZoom Participation.  See announcement. \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e43cf DTSTART;TZID=America/New_York:20240807T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240807T120000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-travis-hance LOCATION:Reddy Conference Room\, Gates HIllman 4405 and Zoom SUMMARY:Thesis Oral Defense - Travis Hance CLASS:PUBLIC DESCRIPTION:Speaker: TRAVIS HANCE\, Ph.D. Candidate\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Verifying Concurrent S ystems Code\n\nConcurrent software is notoriously difficult to write corre ctly\, so to\nincrease confidence in it\, it is often desirable to apply f ormal\nverification techniques. One technique that is especially promising \nfor verifying concurrent software is concurrent separation logic\n(CSL)\ , which uses reasoning principles based on resource ownership.\nHowever\, even with CSL\, verifying complex systems at scale (e.g.\,\nthose with 100 0s of lines of code) remains challenging. The reasons it\nremains challeng ing include:\n\nThe manual proof effort required by many existing CSL fram eworks.The\ninherent complexity of the target systems. Sophisticated syste ms may\nhave custom\, low-level synchronization logic\, which may be deepl y\nintertwined with domain logic\, in the interest of performance.\n\nWe p osit that a promising way to overcome (1) is\, rather than using\nCSL dire ctly\, to use an ownership type system such as Rust's\, taking\nadvantage of its sophisticated but efficient type-checking algorithms.\nTo demonstra te this\, we develop a full methodology\, from theory to\nimplementation\, based around this core idea\, showing that we can\nrecover the rich reaso ning principles of CSL in this setting. In\nparticular\, we show that this methodology is rich enough to support\nthe verification of inherently com plex systems as in (2).\n\nThesis Committee:\n\nBryan Parno (Chair)\n\nDav e Andersen\n\nFrank Pfenning\n\nDerek Dreyer (Max Planck Institute for Sof tware Systems) \n\nIn Person and Zoom Participation.  See announcement.\ n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e4814 DTSTART;TZID=America/New_York:20240806T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240806T163000 URL:https://csd.cmu.edu/calendar/doctoral-thesis-proposal-suhas-jayaram-sub ramanya LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Doctoral Thesis Proposal - Suhas Jayaram Subramanya CLASS:PUBLIC DESCRIPTION:Speaker: SUHAS JAYARAM SUBRAMANYA\, Ph.D. Student\, Computer Sc ience\nDepartment\, Carnegie Mellon University\n\nTalk Title: Efficient jo b-resource co-adaptivity for deep learning\nworkloads on large heterogeneo us GPU clusters\n\nThe training performance of a deep learning (DL) traini ng job is\ndetermined by the number\, type and arrangement of the allocate d GPU\nresources\, and the job parameters (like batch size and learning ra te)\nused for execution. Modern clusters for DL training contain tens of\n thousands of GPUs of many types\, and a cluster scheduler allocates\nGPUs to training jobs to maximize collective training progress in the\ncluster. Existing DL cluster schedulers cannot handle the large space\nof adaptivi ty choices (i.e.\, combined space of GPU allocations and job\nparameters) for large\, heterogeneous GPU clusters — many are not\nheterogeneity-awa re\, few are adaptivity-aware\, and none scale to large\nclusters without sacrificing allocation fidelity and cluster\nefficiency. \n\nIn this thes is\, we introduce (a) a scheduler to facilitate efficient\njob-resource ad aptivity for DL training jobs on large heterogeneous\nGPU clusters\, and ( b) a method to scale optimization-based scheduling\nto much larger cluster sizes without sacrificing allocation fidelity\nand resource efficiency. O ur adaptivity-aware scheduler\, Sia\, uses GPU\nresources judiciously to l earn a job's training performance across\ndifferent GPU types\, and contin uously co-optimizes the GPU allocation\nand job execution parameters to ma ximize cluster-wide training\nprogress in heterogeneous GPU clusters. We t hen scale Sia to large\ncluster sizes by modeling the scheduling policy as a continuous\noptimization problem. We show that it is possible to augmen t the\ninterface between a scheduler and the optimization problem solver t o\nefficiently track changes to the scheduling problem arising from\nchang ing cluster conditions like job arrivals\, departures and phase\nchanges. We develop a prototype solver with the augmented interface\nfor the Sia sc heduling policy that can efficiently recover allocations\nfor very large c lusters. As an additional contribution\, we observe\nthat many other resou rce-allocation problems can also be formulated as\ncontinuous optimization problems and can be solved both quickly and\nefficiently using our propos ed solver. \n\nThesis Committee:\n\nGreg Ganger (Chair)\n\nZhihao Jia\n\n Virginia Smith\n\nAmar Phanishayee (Meta)\n\n \n\nAdditional Information\ n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e4cad DTSTART;TZID=America/New_York:20240805T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240805T143000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:5th Year Thesis Presentation - Lawrence Chen CLASS:PUBLIC DESCRIPTION:Speaker: LAWRENCE CHEN\, Master's Student\, Computer Science De partment\,\nCarnegie Mellon University\n\nTalk Title: Spline-FRIDA: Enhanc ing Robot Painting with Human\nBrushstroke\n\nA painting is more than just a picture on a wall\; a painting is a\nprocess comprised of many intenti onal brush strokes\, leading to a\nperformance far richer than the final o utput. The shapes of individual\nstrokes are an important component of a p ainting's style. This is\nespecially true for sparse sketches\, where indi vidual strokes are\nlikely to be visible. Prior work in modeling brush str oke trajectories\neither does not work with real-world robotics or is not flexible\nenough to capture the complexity of human-made brush strokes. In this\nwork\, we aim to develop a robotic drawing agent with controllable\ nstroke-level style based on human trajectories. \n\nTo achieve this\, we develop a framework to collect brush trajectories\nfrom human artists on a real canvas. We model these trajectories with\nan autoencoder. Finally\, we incorporate the autoencoder into the\nplanning pipeline in the FRIDA r obotic painting system. We find that\,\noff-the-shelf\, FRIDA's brush stro ke renderer struggles or fails to\nlearn the complex trajectories from the human demonstration data\,\nespecially with narrow brushes or markers. We present a novel brush\nstroke renderer that is capable of generalizing to complex\, human-made\nbrush strokes while maintaining a small Sim2Real ga p. Our code is open\nsourced along with the dataset of drawing trajectorie s collected from\npeople using real-world drawing tools. \n\nThesis Commi ttee: \n\nJean Oh (Chair)\n\nJim McCann \n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e50b1 DTSTART;TZID=America/New_York:20240731T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240731T170000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-mikhail-khodak SUMMARY:Thesis Oral Defense - Mikhail Khodak CLASS:PUBLIC DESCRIPTION:Speaker: MIKHAIL KHODAK\, Ph.D. Candidate\, Computer Science De partment\,\nCarnegie Mellon University\n\nTalk Title: The Learning of Algo rithms and Architectures\n\nHow should we design the algorithms we run and the architectures we\nlearn? Several high-impact areas of computing have begun to automate\nthese procedures using machine learning (ML)\, reducing the need for\nhuman effort by using our expanding amount of data and comp ute. We use\nideas from ML\, algorithm design\, and optimization to advanc e our\nunderstanding of these areas of data-driven computing—meta-learni ng\,\nalgorithms with predictions\, and architecture search—and to\ntran slate the resulting methodologies into state-of-the-art\nimplementations.\ n\nIn meta-learning\, which uses ML itself to improve ML algorithms by\nle arning across many learning tasks\, we introduce ARUBA\, a framework\nfor designing and analyzing meta-learning methods. Our analysis yields\nthe fi rst guarantees for gradient-based meta-learning\, showing how\nsuch method s improve performance based upon quantifiable measures of\nsimilarity betw een learning tasks. We use ARUBA to extend the\npractical impact of meta-l earning to new areas of ML\, including to\nlearning with partial feedback and to federated learning.We build upon\nthe success of ARUBA by taking it s core approach—the optimization of\nsurrogate loss functions approximat ing algorithmic objectives—and\nextending it beyond learning algorithms to show learning guarantees\nfor algorithms with predictions\, which are a lgorithms that take\nadvantage of learned predictions about their instance s\; in particular\,\nwe show the first learning-theoretic guarantees for p redictions that\ndepend on the instance the algorithm is run on\, a crucia l property for\npractical applications. We apply our framework while intro ducing\nalgorithms with predictions to new areas such as scientific comput ing\,\nwhere we design learning algorithms that\, under natural structural \nassumptions\, can learn to make instance-optimal predictions.Lastly\, we \naddress the problem of finding neural network architectures to train\non specific learning tasks\, or architecture search\, where we make\nprogres s towards understanding the optimization and generalization\nproperties of weight-sharing\, a dominant heuristic used throughout the\nfield. We then extend weight-sharing to design new search spaces based\naround neural op erations that allow for the automated discovery of\ntruly novel architectu res from data\; the culmination of this effort is\nDASH\, a method that ef ficiently finds architectures that outperform\nhuman expert-designed neura l architectures on the majority of diverse\ntasks we test.\n\nThesis Commi ttee:\n\nMaria-Florina Balcan (Co-Chair)\n\nAmeet Talwalkar (Co-Chair)\n\n Tom Mitchell\n\nPeter Bartlett (University of California\, Berkeley)\n\nPi otr Indyk (Massachusetts Institute of Technology)\n\nAlexander Smola (Boso n AI)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e558e DTSTART;TZID=America/New_York:20240731T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240731T130000 URL:https://pdl.cmu.edu/talk-series/2024/073124.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talk Series - Jacob Baskin CLASS:PUBLIC DESCRIPTION:Speaker: JACOB BASKIN\, Software Engineer\, Jane Street\n\nTalk Title: Superstore: What We Learned Building a Data Warehouse\n\nIn 2022\, Jane Street decided to build an on-premises data warehouse\,\ncalled Supe rstore\, which launched in 2023 and stores about 2PB of\ndata. While we us ed existing software for most of the heavy lifting\,\nsome of our design d ecisions were a bit more customized. \n\nIn this talk\, I will give a bri ef architecture overview of Superstore\nand discuss the choices we made\, how they worked in practice\, and what\nwe could or should have done diffe rently. Is Parquet the storage\nformat of the future? Does data locality m atter? How do you\nefficiently handle arbitrarily wide data sets with a fi xed amount of\nRAM? Our opinions on all these questions have changed signi ficantly in\nthe past year.\n\n — \n\nJacob Baskin is a software engin eer at Jane Street. His previous jobs\nhave included CTO and co-founder of Coord\, an urban transportation\nstartup\, and software engineer at Googl e. His focus is on managing\ndata effectively at the application level at scales ranging from \"big\"\nto \"artisanal small-batch\". He graduated fr om Brown University with a\nB.A. in Computer Science. Zoom Participation.   See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e595f DTSTART;TZID=America/New_York:20240729T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240729T170000 LOCATION:Gates Hillman 9115 SUMMARY:5th Year Thesis Presentation - Trevor Leong CLASS:PUBLIC DESCRIPTION:Speaker: TREVOR LEONG\, Master's Student\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Improvements to the e valuation of ML models in FHE\n\nFHE (Fully Homomorphic encryption) is a c lass of encryption schemes\nthat enables computation over ciphertexts that correspond to\noperations on the encrypted plaintexts. This useful proper ty allows\nclients to outsource computation to servers without revealing w hat\ntheir inputs were. A notable application of FHE is in private Machine \nLearning as a Service (MLaaS)\, where clients can submit data to a\nserv er-hosted machine learning model and receive processed results\,\nall whil e preserving data confidentiality.\n\nHowever\, the practical implementati on of FHE in evaluating machine\nlearning models remains challenging. A si gnificant hurdle is the\nrestricted set of operations permissible under FH E. This is compounded\nby the fact that each FHE operation incurs a signif icant performance\noverhead when compared to its plaintext counterpart. As a result\,\ncomputation of nonlinear functions like Softmax requires comp lex\npolynomial approximations and even basic operations such as Matrix\nm ultiplication require a significant amount of time to evaluate.\n\nThis th esis seeks to mitigate these performance constraints associated\nwith FHE in machine learning model evaluation. Initially\, I introduce\na modificat ion to the HELR algorithm\, enhancing its efficiency when\ndeployed on a h ardware accelerator\, achieving an 8x reduction in\nlatency. Then\, I prop ose a novel application of a softmax\napproximation for evaluation in FHE that leads to a 4x reduction in\nlatency. Lastly\, I describe a procedure for evaluating the embedding\nlayer on the server without the client learn ing the model’s\nembedding matrix at minimal computation overhead.\n\nTh esis Committee:\n\nWenting Zheng (Advisor)\n\nLujo Bauer\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e5d8d DTSTART;TZID=America/New_York:20240729T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240729T160000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-jatin-arora LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Thesis Oral Defense - Jatin Arora CLASS:PUBLIC DESCRIPTION:Speaker: JATIN ARORA\, Ph.D. Candidate\, Computer Science Depar tment\,\nCarnegie Mellon University\n\nTalk Title: Provably Efficient Cosc heduling of Computation and Data\nthrough Disentanglement\n\nBecause of it s many desirable properties\, such as its ability to\ncontrol effects and thus potentially disastrous race conditions\,\nfunctional programming offe rs a viable approach to programming modern\nmulticore computers. This has  led to  the past decade several\nparallel functional languages\, typica lly based on dialects of ML and\nHaskell\, have been developed. These lang uages\, however\, have\ntraditionally underperformed compared to procedura l languages (such as\nC and Java).The primary reason for this underperform ance has been the\nlack of scalable memory management techniques capable of matching the\nincreased demand of memory in parallel execution.\n\nIn t his thesis\, we propose provably efficient techniques for memory\nmanageme nt of parallel functional programs. The key idea behind our\ntechniques is to coschedule the parallel computation with its data\,\nenabling the memo ry manager to exploit the disentanglement\nhypothesis---the idea that para llel tasks of a program largely execute\nindependently and avoid side-effe cting data that may be accessed by\nothers--for efficiency. We implement t hese techniques in the MPL\ncompiler for parallel ML and our experimental results show that the\ntechniques can marry the safety benefits of functio nal programming\nwith performance.\n\nThesis Committee:\n\nUmut A. Acar (C hair)\n\nGuy E. Blelloch\n\nRobert Harper\n\nRustan Leino (Amazon)\n\n \n \nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e6184 DTSTART;TZID=America/New_York:20240729T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240729T120000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-yue-niu LOCATION:Gates HIllman 8102 SUMMARY:Thesis Oral Defense - Yue Niu CLASS:PUBLIC DESCRIPTION:Speaker: YUE NIU\, Ph.D. Candidate\, Computer Science Departmen t\,\nCarnegie Mellon University\n\nTalk Title: Cost-sensitive Programming\ , Verification\, and Semantics\n\nComputational cost is a fundamental aspe ct of the behavior of computer\nprograms. However\, existing program verif ication techniques do not\nsimultaneously provide both faithful representa tion of cost structure\nand a way to reason about the pure functional mean ing of\ncost-instrumented programs. \n\nThis thesis introduces a logical framework for integrating\ncost-sensitive and functional program verificat ion and semantics by\nmeans of the internal modal type theory of presheaf categories\, an\napproach to programming language semantics first introduc ed by\nSterling and Harper in the context of program modules and data\nabs traction. I demonstrate that a range of common algorithms can be\nformulat ed and formally verified to meet both their functional and\ncost specifica tions within the framework. Lastly\, I extend the logical\nframework and u se it as a metalanguage for studying the cost semantics\nof programming la nguages\, culminating in an internal cost-sensitive\ncomputational adequac y theorem for PCF that relates the denotational\nand operational cost sema ntics in the style of Plotkin. \n\nThesis Committee: \n\nRobert Harper ( Chair)\n\nJan Hoffmann\n\nSteve Brookes\n\nJon Sterling (University of Cam bridge)\n\nNeel Krishnaswami (University of Cambridge)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e655c DTSTART;TZID=America/New_York:20240726T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240726T110000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-ranysha-ware LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Thesis Oral Defense - Ranysha Ware CLASS:PUBLIC DESCRIPTION:Speaker: RANYSHA WARE\, Ph.D. Candidate\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Battle for Bandwidth: On The Deployability of New\nCongestion Control Algorithms\n\nThe Internet has become the central source of information and\ncommunication in modern society. Congestion control algorithms (CCAs)\nare critical for the stabi lity of the Internet: ensuring that users\nare able to fairly and efficien tly share the network. Over the past 30\nyears\, researchers and Internet content providers have proposed and\ndeployed dozens of new CCAs designed to keep up with the growing\ndemands of faster networks\, diverse applicat ions\, and mobile users.\nWithout tools to understand this growing heterog eneity in CCAs\ndeployed in the Internet\, the fairness of the Internet is at stake.\n\nTowards understanding this growing heterogeneity\, we develo p\nCCAnalyzer\, a tool to determine what CCA a particular web service\ndep loys\, outperforming previous classifiers in accuracy and\nefficiency. Wit h CCAnalyzer\, we show that new CCAs\, both known and\nunknown\, have wide spread deployment in the Internet today\, including a\nrecently proposed C CA by Google: BBRv1. Next\, we develop the first\nmodel of BBRv1\, and pro ve BBRv1 can be very unfair to legacy\nloss-based CCAs\, an alarming findi ng given the prolific deployment of\nBBRv1.\n\nConsequently\, we argue the need for a better methodology for\ndetermining if a new CCA is safe to de ploy in the Internet today. We\ndescribe how the typical methodology testi ng for equal-rate fairness\n(every user gets the same bandwidth) is both a n unachievable goal and\nultimately\, not the right threshold for determin ing if a new CCA is\nsafe to deploy alongside others. Instead of equal-rat e fairness\, we\npropose a new metric we call\, harm\, and argue for a har m-based\nthreshold. Lastly we present RayGen\, a novel framework for evalu ating\ninteractions between heterogeneous CCAs. RayGen uses a genetic\nalg orithm to efficiently explore the large state space of possible\nworkloads and network settings when two CCAs compete. With a small\nbudget of exper iments\, RayGen finds more harmful scenarios than a\nparameter sweep and r andom search.\n\nThesis Committee: \n\nJustine Sherry (Co-Chair)\n\nSrini vasan Seshan (Co-Chair)\n\nTheophilus A. Benson\n\nJim Kurose (University of Massachusetts Amherst)\n\n \n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e6a03 DTSTART;TZID=America/New_York:20240724T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240724T130000 URL:https://pdl.cmu.edu/talk-series/2024/072424.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Talk - Matt Butrovich CLASS:PUBLIC DESCRIPTION:Speaker: MATT BUTROVICH\, Post-Doctoral Researcher\, Computer S cience\nDepartment\, Carnegie Mellon University\n\nTalk Title: On Embeddin g Database Management System Logic in Operating\nSystems via Restricted Pr ogramming Environments\n\nThe rise in computer storage and network perform ance means that disk\nI/O and network communication are often no longer bo ttlenecks in\ndatabase management systems (DBMSs). Instead\, the overheads associated\nwith operating system (OS) services (e.g.\, system calls\, th read\nscheduling\, and data movement from kernel-space) limit query\nproce ssing responsiveness. User-space applications can elide these\noverheads w ith a kernel-bypass design. However\, extracting benefits\nfrom kernel-byp ass frameworks is challenging\, and the libraries are\nincompatible with s tandard deployment and debugging tools. \n\nThis talk presents an alterna tive in user-bypass: a design that\nextends OS behavior for DBMS-specific features\, including\nobservability\, networking\, and query execution. Hi storically\, DBMS\ndevelopers avoid kernel extensions for safety and secur ity reasons\,\nbut recent improvements in OS extensibility present new opp ortunities.\nWith user-bypass\, developers write safe\, event-driven progr ams to push\nDBMS logic into the kernel and avoid user-space overheads. Th ere are\ntwo ways to to invoke user-bypass logic: (1) when a DBMS in user- space\ninvokes these programs\, user-bypass provides behavior similar to a new\nOS system call\, albeit without kernel modifications. In contrast\, (2)\nwhen an OS thread or interrupt triggers these programs in\nkernel-spa ce\, user-bypass inserts DBMS logic into the kernel stack. \n\nFirst\, we present a framework that employs user-bypass to collect\ntraining data fo r self-driving DBMSs efficiently. User-bypass programs\nreduce the number of round trips to kernel-space to retrieve\nperformance counters and other system metrics. Next\, we present a\ndatabase proxy that applies user-byp ass to support features like\nconnection pooling and workload replication while reducing data\ncopying and user-space thread scheduling. User-bypass programs embed\nDBMS network protocol logic in multiple layers of the OS network\nstack\, applying DBMS proxy logic in a kernel-space fast path. La stly\,\nwe present an embedded DBMS for future user-bypass applications. W e\ndiscuss the design decisions\, environment challenges\, and performance \ncharacteristics of a DBMS that offers ACID transactions over\nmulti-vers ioned data in kernel-space. We also explore applications of\nthis user-byp ass DBMS and compare them to modern user-space systems. \n\nThe technique s proposed in this talk show user-bypass benefits across\nmultiple DBMS de sign disciplines and provide a template for future\nDBMS and OS co-design.  \n\n— \n\nMatt Butrovich is a recent Ph.D. graduate (now postdoctoral \nresearcher) from the Computer Science Department at Carnegie Mellon\nUni versity\, researching database management systems (DBMSs). He\nexplores op erating system and DBMS co-design opportunities via safe\nkernel extension mechanisms like eBPF. He is incredibly fortunate to\nbe advised by Andy P avlo\, and is a member of the Database Group\n(CMU-DB)\, Parallel Data Lab (PDL)\, and Systems\, Networking\, and\nPerformance (SNAP) Lab. \n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e6f28 DTSTART;TZID=America/New_York:20240722T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240722T150000 URL:https://csd.cmu.edu/calendar/5th-year-thesis-SEO-2024-07-22 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:5th Year Thesis Presentation - William Seo___ CLASS:PUBLIC DESCRIPTION:Speaker: WILLIAM SEO\, Master's Student\, Computer Science Depa rtment\,\nCarnegie Mellon University\n\nTalk Title: Level Aware Bootstrapp ing Placement for Fully Homomorphic\nEncryption Using MaxSAT\n\nFully Homo morphic Encryption (FHE) is a cryptographic technique that\nallows computa tions to be performed on encrypted data without having\nto decrypt it. Thi s property preserves data privacy\, enabling a wide\nrange of applications in fields such as cloud computing\, secure data\nanalysis\, and privacy-p reserving machine learning. \n\nIn FHE computations\, each ciphertext can only handle a limited number\nof operations before the accumulation of no ise makes decryption\nimpossible. To resolve this\, the bootstrapping oper ation must be used\nto reset the noise in the ciphertext. Bootstrapping is a\ncomputationally expensive computation\, and it influences the costs of \nother operations. Consequently\, the strategic placement of\nbootstrappi ng operations is a critical aspect of FHE performance. \n\nThis thesis in troduces Saturn\, a novel method for automatically\ndetermining the placem ent of bootstrapping operations which minimizes\nprogram runtime. Given a directed acyclic graph (DAG) representing an\nFHE computation\, Saturn lev erages the Maximum Satisfiability (MaxSAT)\noptimization problem to find t he most efficient bootstrapping\nplacement. Additionally\, we propose two methods for reducing the\ncomplexity of the input computational DAG\, sign ificantly decreasing\nthe solve time of our MaxSAT formulation. Saturn's e ffectiveness is\nevaluated on various deep learning models\, demonstrating its potential\nto enhance FHE performance.\n\nThesis Committee:\n\nWentin g Zheng (Chair)\n\nFraser Brown\n\nAdditional Information\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e731c DTSTART;TZID=America/New_York:20240719T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240719T160000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-jenny-lin LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Thesis Oral Defense - Jenny Lin CLASS:PUBLIC DESCRIPTION:Speaker: JENNY LIN\, Ph.D. Candidate\, Computer Science Departm ent\,\nCarnegie Mellon University\n\nTalk Title: Formalizing Object Equiva lence in Machine Knitting\n\nCorrectness is a desirable property for any p rogram\, whether that\nprogram computes an equation\, controls a machine\, or interprets data.\nDefining what it means for a program to be correct c an be surprisingly\nnuanced\, however\, especially when that program is us ed to create a\nphysical object. We can reframe this problem by treating c orrectness\nas a question of equivalence. Given some target object\, is th e result\nof a fabrication process equivalent to the target object? Howeve r\,\nthis now requires that we answer the still complicated question of\nw hat it means for two objects to be equivalent. In order to do so\, we\nnot only need a precise definition of object meaning\, but also a\nstrong und erstanding of how we create and interact with the objects\naround us. \n\ nIn this thesis I tackle this problem of meaning and equivalence for\nmach ine knitting programs. Knitting is the act of taking a few strands\nof yar n and deforming them into interlocking loops forming a stable\nstructure. While knitting machines are capable of quickly fabricating\na vast array o f structures with controllable material properties\, the\ncomplexity of bo th the machine control process and the resulting\nphysical object makes tr anslating between the two incredibly\ndifficult. This gap prevents existin g programing and design tools from\naccessing the full breadth of its fabr ication possibilities. \n\nTo address this\, I formally characterize the complete space of machine\nknitting programming. I begin by introducing fe nced tangles\, a novel\nmathematical object designed to match intuition ab out knit object\nmeaning\, to define semantics for Knitout\, which is a lo w-level\nlanguage for controlling v-bed knitting machines. The underlying\ nprogram meaning is then used to reason about the correctness of a set\nof practical program transformations. This semantic function is used\nas gui dance for developing Instruction Graphs\, which are an\nintermediate repre sentation of knit objects. Unlike existing knit\nobject representations\, Instruction Graphs can capture the full range\nof machine knittable object s and can be verified as machine knittable\nusing three easy to check grap h embedding properties. Finally\, I\ndiscuss how fabrication constraints m ay enable an algebraic approach\nto computing machine knitting program equ ivalence.   \n\nThesis Committee:\n\nJames McCann (Chair)\n\nJan Hoffma nn\n\nScott Hudson\n\nAdriana Schulz (University of Washington)\n\nIn Pers on and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e7803 DTSTART;TZID=America/New_York:20240718T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240718T160000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-shiva-kaul LOCATION:Mauldin Auditorium\, Newell-Simon 1305 SUMMARY:Thesis Oral Defense - Shiva Kaul CLASS:PUBLIC DESCRIPTION:Speaker: SHIVA KAUL\, Ph.D. Candidate\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Classical Improvements t o Modern Machine Learning\n\nThe following two dilemmas of modern foundati on models concern\nstatistical accuracy and computational efficiency\, res pectively:\n\nCan language models be trusted to rigorously answer importan t\nscientific questions? (Specifically\, causal questions from\nevidence-b ased medicine which are currently answered through\nmeta-analysis)Can Tran sformers and RNNs be replaced by faster\nstate-space models (which are lin ear across time / sequence length)\nwithout sacrificing expressive power?\ n\nI present solutions to both. For (1)\, I adapt conformal prediction to\ nmeta-analysis\, which may be thought of as a regression from treatment\na nd population features (e.g. \"800mg of amiodarone for atrial\nfibrillatio n patients\") to treatment effect (e.g. \"60% chance of\nreversion to norm al heart rhythm\"). By using conformal prediction to\nsafely incorporate u ntrusted data (i.e. observational studies and\nother background informatio n)\, this complex regression problem can be\nsatisfactorily addressed even with a small number of randomized\ncontrolled trials. The main technical challenges are computationally\nsimplifying full conformal prediction (whi ch is necessary due to the\nsmall number of trials) and handling noisy obs ervations (due to the\nlimited number of participants in each trial). \n\ nFor (2)\, I present a general scheme by which nonlinearity across time\nc an be replaced by nonlinearity along depth. This involves stacking\nlinear systems with interposed local corrections. This scheme is fast\,\nconstru ctive\, involves no additional parameters\, provably converges\neven in th e worst case\, and empirically exhibits fast convergence. It\ncan be used to practically develop fast sequence models and to\ntheoretically understa nd the power of depth. \n\nBoth of these solutions exemplify a broader th esis of developing\nsyntheses between classical machine learning technique s (such as\nmeta-analytic averaging or linear dynamical systems) and moder n\napproaches (such as deep nonlinear regression or Transformers). The\nhi gh-level goal is to combine the safety and tractability of classical\nappr oaches with the accuracy of modern ones through a close (and\nsometimes su rprising) examination of their technical relationship. \n\nThesis Committ ee: \n\nGeoffrey Gordon (Chair)\n\nZachary Lipton\n\nAditi Raghunathan\n\ nRyan Tibshirani (University of California\, Berkeley)\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e7cd3 DTSTART;TZID=America/New_York:20240717T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240717T140000 URL:https://pdl.cmu.edu/talk-series/2024/071724.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talks - Dimitrios Skarlatos & Amarn ath\nJolad CLASS:PUBLIC DESCRIPTION:Speaker: DIMITRIOS SKARLATOS and AMARNATH JOLAD\n\nTalk One\n\n  ► DIMITRIOS SKARLATOS\n\n  Assistant Professor\, Computer Science De partment\, Carnegie Mellon\nUniversity\n\n  — Perspective: A Principle d Framework for Pliable and Secure\nSpeculation in Operating Systems\n\nTr ansient execution attacks present an unprecedented threat to\ncomputing sy stems. Protecting the operating system (OS) is\nexceptionally challenging because a transient execution gadget in the\nOS can potentially leak the e ntire memory.\n\nIn this talk\, I will present Perspective: a principled f ramework for\nbuilding pliable and secure speculative execution defenses f or the OS.\nPerspective offers a pliable interface that allows the OS to\n communicate its security requirements to hardware defenses\, enabling\ntai lored protection against transient execution attacks with little\nperforma nce overhead.\n\nThe design of Perspective is driven by a taxonomy of tran sient\nexecution attacks in the OS kernel: (i) active transient execution\ nattacks\, where the attacker process exploits its own kernel thread to\ns peculatively execute a transient execution gadget in the kernel\, and\n(ii ) passive transient execution attacks\, where the attacker coerces\nthe vi ctim process’s kernel thread to execute a transient execution\ngadget.\n \nBased on the taxonomy\, Perspective introduces Data Speculation Views\n( DSVs) and Instruction Speculation Views (ISVs)\, to mitigate active\nand p assive attacks\, respectively. DSVs define the ownership of kernel\ndata b y a given execution context and block any speculative access to\ndata outs ide the DSV. ISVs define the set of kernel functions that can\nbe speculat ively executed by a given execution context. Any\ntransmitter instructions —whose execution could leak secrets\, such as\nload instructions—that belong to kernel functions outside the ISVs\nare blocked from speculative execution. ISVs open up new opportunities\nof (i) swiftly patching gadgets in the OS\, (ii) reducing the surface\nof passive attacks\, and (iii) spe eding up the process of auditing\ntransient execution gadgets in the OS.\n \nAdditional Reading: Perspective - ISCA'24\n\nDimitrios Skarlatos is an a ssistant professor in the Computer Science\nDepartment at Carnegie Mellon University. His research bridges\ncomputer architecture and operating syst ems\, focusing on performance\,\nsecurity\, and scalability. He has receiv ed several awards for his\ncross-cutting research including the NSF CAREER award\, four Meta\nFaculty Awards in systems\, AI\, and security\, the jo int ACM SIGARCH &\nIEEE CS TCCA Outstanding Dissertation award\, the David J. Kuck\nOutstanding Ph.D. Thesis Award\, an ISCA Best Paper Award\, two ASPLOS\nBest Paper Awards\, and four IEEE MICRO Top Picks. Dimitrios has\n released several open-source frameworks\, with some of his work\nupstreame d in Linux\, adopted by Android\, and deployed in production at\nMeta acro ss millions of servers.\n\nTalk Two\n\n►  AMARNATH JOLAD\n\n   Archit ect\, Oracle Database Product Development\, Oracle\n\n   — Breaking Ba rriers: Successful Strategies for Integrating RDMA\nin the Enterprise\n\nR emote Direct Memory Access (RDMA) enables direct access to\napplication me mory across distributed compute or storage nodes in\nserver clusters. By l everaging standards such as InfiniBand and RoCE\n(RDMA over Converged Ethe rnet)\, organizations can boost efficiency\,\nreduce server count\, and im prove network performance. RDMA minimizes\nCPU overhead\, enhances network efficiency\, and supports scalability.\nThis technology forms the foundat ion of the Oracle Exadata platform\,\nan enterprise database platform desi gned to manage Oracle Database\nworkloads of any size and importance. For instance\, RDMA plays a\ncrucial role in the Oracle Cache Fusion cluster d atabase architecture.\n\nHowever\, integrating RDMA into mission-critical enterprise platforms\npresents unique challenges\, including memory scalab ility (handling\nmulti-terabytes per node)\, scaling connections across th ousands of\nprocesses\, and ensuring secure RDMA fabric in multi-tenant\ne nvironments. This discussion examines these challenges and explores\ninnov ative solutions that enable the adoption of RDMA technology in\nthe Oracle Exadata enterprise database platform.\n\nAmarnath Jolad is an Architect i n Oracle Database product development.\nHis research interests include hig h performance networks\n(RDMA/InfiniBand/RoCE)\, interconnects (PCIe/CXL)\ , accelerators\, and\ndistributed systems. His current research is focused on efficient and\nscalable parallel programs in a distributed computing e nvironment.\nPrior to joining Oracle\, he worked on clustered data protect ion\nsystems\, storage systems and networking (Fibre Channel/iSCSI/SRP). H e\nholds a Master's degree in Telecom and Software Engineering.\n\nZoom Pa rticipation.  See announcement.\n\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e835f DTSTART;TZID=America/New_York:20240715T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240715T170000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-haithem-turki LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Thesis Oral Defense - Haithem Turki CLASS:PUBLIC DESCRIPTION:Speaker: HAITHEM TURKI\, Ph.D. Candidate\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: Towards City-Scale Ne ural Rendering\n\nAdvances in neural rendering techniques have led to sign ificant\nprogress towards photo-realistic novel view synthesis. When combi ned\nwith increases in data processing and compute capability\, this\nprom ises to unlock numerous VR applications\, including virtual\ntelepresence\ , search and rescue\, and autonomous driving. Large-scale\nvirtual reality \, long the domain of science fiction\, feels markedly\nmore tangible. \n \nThis thesis explores the frontier of large-scale neural rendering by\nbu ilding upon Neural Radiance Fields (NeRFs)\, a family of methods\nattracti ng attention due to their state-of-the-art rendering quality\nand conceptu al simplicity. Since its inception\, at least 3\,000 papers\nhave been pro posed in less than three years by research groups across\nthe world across numerous use cases. However\, many shortcomings\nremain. The first is sca le itself. Only a handful of existing methods\ncapture scenes larger than a room. Those that do only handle static\nreconstruction\, which limits th eir applicability. Another is speed\, as\nrendering falls below interactiv e thresholds. Current acceleration\nmethods remain too slow or degrade qua lity at high resolution. Quality\nis a third issue\, as NeRF assumes ideal viewpoint conditions that are\nunrealistic in practice and degrades when they are violated. \n\nWe first explore scaling within the context of sta tic reconstruction.\nWe design a sparse network structure that specializes parameters to\ndifferent regions of the scene that can be trained in para llel\,\nallowing us to scale linearly as we increase model capacity (vs\nq uadratically in the original NeRF)\, and reconstruct urban-scale\nenvironm ents orders of magnitude larger than prior work. We then\naddress dynamic reconstruction of entire cities\, and build the largest\ndynamic NeRF repr esentation to date. To accelerate rendering\, we\nimprove sampling efficie ncy through a hybrid surface-volume\nrepresentation that encourages the mo del to represent as much of the\nworld as possible through surfaces (which require few samples per ray)\nwhile maintaining the freedom to render tra nsparency and finer details\n(which pure surface representations struggle to capture). We finally\npropose a fast anti-aliasing method that greatly improves rendering\nquality when training with data collected from freefor m camera\ntrajectories. Importantly\, our method incurs a minimal performa nce\noverhead and is compatible with the scale and speed improvements\npre viously mentioned.\n\nThesis Committee:\n\nDeva Ramanan (Chair)\n\nShubham Tulsiani\n\nJessica K. Hodgins\n\nMartial Hebert\n\nJonathan T. Barron (G oogle DeepMind)\n\nIn Person and Zoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e8848 DTSTART;TZID=America/New_York:20240704T000000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240704T235900 SUMMARY:Independence Day Holiday Observance 2024 CLASS:PUBLIC DESCRIPTION:The Fourth of July observation is a federal holiday in the Unit ed\nStates\, commemorating the Declaration of Independence ratification by \nthe Second Continental Congress on July 4\, 1776 — establishing the\nU nited States of America.\nNormal campus class schedules and office hours r esume Monday\, July 8.\n \n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e8acb DTSTART;TZID=America/New_York:20240701T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240701T130000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-peter-manohar LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Thesis Oral Defense - Peter Manohar CLASS:PUBLIC DESCRIPTION:Speaker: PETER MANOHAR\, Ph.D. Candidate\, Computer Science Dep artment\,\nCarnegie Mellon University\n\nTalk Title: New Spectral Techniqu es in Algorithms\, Combinatorics\, and\nCoding Theory: The Kikuchi Matrix Method\n\nIn this thesis\, we present a new method to solve algorithmic an d\ncombinatorial problems by (1) reducing them to bounding the maximum\,\n over x in {-1\,1}n\, of homogeneous degree-q multilinear polynomials\,\nan d then (2) bounding the maximum value attained by these polynomials\nby an alyzing the spectral properties of appropriately chosen induced\nsubgraphs of Cayley graphs on the hypercube (and related variants)\ncalled \"Kikuch i matrices\". We will present the following applications\nof this method.\ n\nDesigning algorithms for refuting/solving semirandom and smoothed\ninst ances of constraint satisfaction problems\;Proving Feige's\nconjectured hy pergraph Moore bound on the extremal girth vs. density\ntrade-off for hype rgraphs\;Proving a cubic lower bound for 3-query\nlocally decodable codes and an exponential lower bound for 3-query\nlocally correctable codes.\n\n Thesis Committee: \n\nVenkatesan Guruswami (Co-Chair\, Carnegie Mellon Un iversity /\nUniversity of California\, Berkeley)\n\nPravesh K. Kothari (Co -Chair\, Carnegie Mellon University / Princeton\nUniversity)\n\nRyan O’D onnell\n\nUriel Feige (Weizmann Institute)\n\nIn Person and Zoom Participa tion.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e8e87 DTSTART;TZID=America/New_York:20240628T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240628T160000 URL:https://csd.cmu.edu/calendar/thesis-oral-defense-david-kahn LOCATION:Mauldin Auditorium\, Newell-Simon 1305 SUMMARY:Thesis Oral Defense - David Kahn CLASS:PUBLIC DESCRIPTION:Speaker: DAVID KAHN\, Ph.D. Candidate\, Computer Science Depart ment\,\nCarnegie Mellon University\n\nTalk Title: Leveraging Linearity to Improve Automatic Amortized\nResource Analysis\n\nAfter correctness\, the most important properties of programs concern\ntheir resource requirements \, like how much time they take to run or\nhow much memory they need. It i s therefore desirable to automate the\nderivation of a program’s costs. One successful approach to such\nautomatic derivation is the type system k nown as Automatic Amortized\nResource Analysis (AARA). AARA finds polynomi al bounds on resource\nusage by using its types to apply the physicist’s method of\namortized cost analysis. Type inference in AARA can be reduced to\nlinear programming\, thereby automating resource analysis. This balan ce\nof expressive bounds and efficient analysis has brought AARA success.\ n\nUnfortunately\, deriving a program’s resource usage can be difficult\ n— in fact it is generally not computable. Thus\, despite AARA’s\nsucc ess\, it is not surprising that there are many natural program\npatterns t hat it cannot analyze well. Sometimes AARA finds loose\nresource bounds\, other times it finds bounds slowly\, and sometimes it\ncannot find any bou nds at all. \n\nThis thesis addresses such shortcomings by developing a v ariety of\nupgrades to the AARA type system that allow the efficient deriv ation\nof tight resource bounds for more programs. The key theme underlyin g\nthese upgrades is the leveraging of linear reasoning principles. These\ nideas integrate well with AARA because AARA exists in the intersection\no f various forms of linearity: the linear flavor its type system\, the\nlin ear relations of its cost bound templates\, and the linear\nphysicality be hind the physicist’s method of amortized cost\nanalysis.\n\nThis work fi rst upgrades the type system with remainder contexts to\nbetter reason abo ut reusable resources like memory. Then the class of\nAARA’s bounding fu nctions is enlarged to include\, e.g.\, exponential\nbounds. This class of functions is further enlarged to be\nmultivariate\, allowing dependence o n products of data structure sizes.\nNext\, this work provides a more effi cient\, matrix-based approach to\ninferring the cost-free AARA types neede d for\, e.g.\, non-tail\nrecursion. Finally the physicist’s method of am ortized cost analysis\nis refined into the quantum physicist’s method\, which provides an\nautomatable framework for reasoning about resource real location\, while\nalso allowing resource bounds to depend on data structur e height. \n\nThesis Committee:\n\nJan Hoffmann (Chair)\n\nFrank Pfenning \n\nStephanie Balzer\n\nThomas Reps (University of Wisconsin)\n\nIn Person and Zoom Participation. See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e938a DTSTART;TZID=America/New_York:20240626T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240626T203000 URL:https://community.cmu.edu/customquickevents?id=a3pUk0000000uIL&custom=t rue LOCATION:895 Villa Street Mountain View\, CA 94041 SUMMARY:2024 SCS Alumni Mixer - South Bay CLASS:PUBLIC DESCRIPTION:Speaker: MARK STEHLIK\n\nJoin us for an SCS alumni mixer with h ors d'oeuvres\, drinks and games.\nAs a bonus\, SCS legend\, Mark Stehli k\, University Teaching Professor\,\nAssistant Dean for Outreach\, and Dir ector of CSD Undergraduate\nPrograms\, and Co-Founder\, CMU CS Academy\, w ill be in attendance. Don't\nmiss this opportunity to catch up with old fr iends and meet new ones!\nSpace is limited! \n\nRSVP  ⇒  Registration opens Monday\, June 3\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e967e DTSTART;TZID=America/New_York:20240626T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240626T140000 URL:https://pdl.cmu.edu/talk-series/2024/062624.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talks - Shasank Chavan and Greg Gan ger CLASS:PUBLIC DESCRIPTION:Speaker: Two Talks: SHASANK CHAVAN and GREG GANGER\n\n►  SHA SANK CHAVAN \n\n      Vice President\, Data In-Memory and AI Techno logies\, Oracle\n      \n\n—  Leveraging Generative AI with Oracl e AI Vector Search\n\nAI Vector Search in Oracle 23ai is a new\, transform ative way to\nintelligently search through your unstructured business data \nefficiently\, and accurately\, by using AI techniques to match on the\ns emantics\, or meaning\, of the underlying data. With the inclusion of a\nn ew VECTOR datatype\, new approximate search indexes\, and new SQL\noperato rs and extensions\, enterprise companies can quickly and easily\nleverage AI Vector Search to build modern\, generative-ai applications\nwith just a few lines of SQL! And with this simplicity comes power\, as\nAI Vector Se arch is fully integrated with Oracle’s enterprise-grade\nfunctionality\, such as transactions\, RAC\, and Exadata. This talk will\ndive into the m echanics of AI Vector Search\, ensuring a solid\nunderstanding of its impl ementation and benefits. \n\nShasank Chavan is the Vice President of the Data\, In-Memory and AI\nTechnologies group at Oracle. He leads an organiz ation of brilliant\nengineers working on the nexus between AI systems and modern\ndatabases. His team is currently hyper-focussed on developing the\ nnext-generation\, AI-centric data storage engine\, designed for\nin-memor y OLTP\, Analytics and Vector Search capabilities to power the\nAI and Gen erative AI revolution to come. Shasank earned his BS/MS in\nComputer Scien ce at the University of California\, San Diego. He has\naccumulated 50+ pa tents over a span of 25 years working on systems\nsoftware technology. \n \n►  GREG GANGER\n\n      Jatras Professor of Electrical and Compu ter Engineering\,\nCarnegie Mellon University\n\n      Director\, Pa rallel Data Laboratory (PDL)\n\n—  Cluster Storage Systems Need Declara tive IO Interfaces \n\nStorage systems continue to be built around decade s-old imperative\ninterfaces\, like read/write and get/put. Although this low-level\ninterface can be used for any framework or application\, it can lead to\nsignificant IO inefficiencies\, especially in cases (e.g.\, data \nmaintenance tasks like compaction\, integrity checks\, rebalancing\,\net c.)\, for which caches tend to be least effective. Although not a new\nfac t\, IO efficiency is reaching emergency status\, as the IOPS/TB (or\nBW/TB ) available from each storage device in large-scale cluster\nstorage drops with each increase in device capacity...new approaches\nare needed to mor e efficiently use the IOPS/TB available. It's time to\naugment cluster sto rage with declarative interfaces\, whereby data\nmaintenance tasks and dat a management applications can register need\nfor sets of data items and al low the storage system to orchestrate the\ncorresponding IO. So\, rather t han converting order-flexible and\ntime-flexible needs into an arbitrary o rdering of \"do this now\"\nimperative IO\, the flexibility can be exposed to and exploited by the\nstorage system. With this flexibility\, signific ant opportunities arise\nfor eliminating redundant IO (e.g.\, data read fo r an integrity check\ncould also be used for rebalancing)\, smoothing IO b ursts\, and\ncoelescing IOs. This talk will describe the declarative IO co ncept\,\nargue for their importance\, talk about our early exploration int o\nthem\, and invite discussion and collaboration.   \n\nGreg Ganger is the Jatras Professor of ECE and CS (by courtesy) at\nCarnegie Mellon Unive rsity (CMU). Since 2001\, he has also served as\nthe Director of CMU's Par allel Data Laboratory (PDL) research center\nfocused on data storage and p rocessing systems. He has broad research\ninterests in computer systems\, including storage/file systems\, cloud\ncomputing\, ML systems\, distribut ed systems\, and operating systems. He\nearned his collegiate degrees from the University of Michigan and did\na postdoc at MIT before joining CMU. He still loves playing\nbasketball..he's lost a step but developed a sweet 3-point shot. And\,\nno\, the surfing pictures are not photoshopped. \n\ nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e9c1d DTSTART;TZID=America/New_York:20240625T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240625T203000 URL:https://community.cmu.edu/customquickevents?id=a3pUk0000001Lon&custom=t rue LOCATION:7 Warriors Way Suite 206\, San Francisco\, CA 94158 SUMMARY:2024 SCS Alumni Mixer - San Francisco CLASS:PUBLIC DESCRIPTION:Speaker: MARK STEHLIK\n\nJoin us for an SCS alumni mixer with h ors d'oeuvres\, drinks and games.\nAs a bonus\, SCS legend\, Mark Stehli k\, University Teaching Professor\,\nAssistant Dean for Outreach\, and Dir ector of CSD Undergraduate\nPrograms\, and Co-Founder\, CMU CS Academy\, w ill be in attendance. Don't\nmiss this opportunity to catch up with old fr iends and meet new ones!\nSpace is limited! \n\nRSVP  ⇒  Registration opens Monday\, June 3\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91e9f38 DTSTART;TZID=America/New_York:20240620T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240620T150000 LOCATION:17:30 pm Spanish Standard Time | 1:30 pm Eastern Standard Time SUMMARY:16th Edition: Frontiers of Knowledge Award Ceremony CLASS:PUBLIC DESCRIPTION:Speaker: Honoring TAKEO KANADE\n\nThe Frontiers of Knowledge A wards\, established in 2008\, recognize and\nreward contributions of singu lar impact in science\, art and the\nhumanities\, privileging those that s ignificantly expand the frontiers\nof the known world\, open up new fields \, or emerge from the interaction\nof various disciplinary areas. \n\nSCS Founders University Professor Takeo Kanade will receive the BBVA\n(Banco Bilbao Vizcaya Argentaria) Foundation's Frontiers of\nKnowledge Award  in Information and Communication Technologies (2023)\nfor developing mathe matical foundations that underlie the current\ncapabilities of computers a nd robots to comprehend and interpret\nvisual images and scenes. \n\nThe event will be streamed live on the BBVA Foundation website. \n\nAdditiona l Details\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ea291 DTSTART;TZID=America/New_York:20240618T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240618T160000 URL:https://csd.cmu.edu/calendar/speaking-skills-talk-christoper-canel LOCATION:Gates Hillman 8102 SUMMARY:Speaking Skills Talk - Christoper Canel CLASS:PUBLIC DESCRIPTION:Speaker: CHRISTOPHER CANEL\, Ph.D. Student\, Computer Science\n Department\, Carnegie Mellon University\n\nTalk Title: Measuring and Mitig ating Incast Bursts in Modern\nDatacenters\n\nIn datacenter networks\, com mon many-to-one traffic patterns known as\nincast are challenging because they violate the basic premise of\nbandwidth stability on which Transmissi on Control Protocol (TCP)\ncongestion control is built\, overwhelming shal low switch buffers and\ncausing packet loss and high application latency.   \n\nFirst\, to understand why these challenges remain despite decades of\nresearch on datacenter transport\, this talk presents an in-depth\ninv estigation into high-degree incasts both in production workloads at\na maj or datacenter operator and in simulation. In addition to\ncharacterizing t he bursty nature of these incasts and their impacts on\nthe network\, our findings demonstrate the shortcomings of widely\ndeployed window-based con gestion control techniques used to address\nincast problems. \n\nSecond\, we show how congestion control can leverage existing features\nin TCP to mitigate high-degree incast without a new protocol design.\nDrawing on sim ple receiver-assisted congestion control techniques\, we\ndemonstrate that tuning an incast receiver's TCP advertised window\n(RWND) provides the co ordination necessary to help independent senders\nreduce the volume of in- flight data to keep queues short and prevent\npacket loss.  \n\nFurtherm ore\, we find that hosts associated with a specific application\nor servic e exhibit similar and predictable incast traffic properties\nacross hours\ , enabling us to avoid computationally expensive online\nadaptation and in stead adopt a static RWND tuning approach with low\noverheads. We draw on three years of production deployment experience\nto demonstrate that RWND tuning for incast bursts reduces\nretransmissions by 3.8x\, leading to a 2 —8x reduction in application\ntail latency. \n\nPresented in Partial Fu lfillment of the CSD Speaking Skills\nRequirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ea6b3 DTSTART;TZID=America/New_York:20240617T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240617T120000 URL:https://csd.cmu.edu/calendar/speaking-skills-talk-yue-niu LOCATION:Gates Hillman 8102 SUMMARY:Speaking Skills Talk - Yue Niu CLASS:PUBLIC DESCRIPTION:Speaker: YUE NIU\, Ph.D. Student\, Computer Science Department\ , Carnegie\nMellon University\n\nTalk Title: A type-theoretic view of info rmation flow\n\nInformation dependency is a fundamental notion in the theo ry of\nprogramming languages. Examples of dependencies include resource us age\nand program modules\, where the computational cost of a function\ndep ends on the value of inputs and the \"shape\" of runtime program\ncomponen ts is determined by static type information. \n\nIn the aforementioned ex amples dependency is asymmetric --- the value\nof a function does not depe nd on the cost of inputs\, and the shape of\na module does not depend on r untime components. The most prototypical\nsuch asymmetry arises in informa tion flow security\, where the value of\na high security bit may depend on a low security bit\, but not vice\nversa. A phase distinction is a type-t heoretic way of organizing the\ninteractions amongst programs existing at different security levels. \n\nThis talk broaches the logic and geometry of phase distinctions and\nindicates applications to information flow secu rity. \n\nPresented in Partial Fulfillment of the CSD Speaking Skills\nRe quirement\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91eaa33 DTSTART;TZID=America/New_York:20240617T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240617T150000 URL:http://www.cs.cmu.edu/~ai-teachers/index.html LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:Crash Course in AI for Teachers CLASS:PUBLIC DESCRIPTION:Carnegie Mellon University School of Computer Science is offeri ng a\nweek-long workshop\, Crash Course in Artificial Intelligence along w ith\na Learning Material Design Session to support high school educators\n looking to gain familiarity with AI and to offer AI-related\neducational a ctivities to their students.\nOur week-long workshop will cover an introdu ction to a broad range of\nartificial intelligence topics including: Machi ne Learning Deep\nLearning Search Algorithms Information Retrieval Recomme nder\nSystemsEach topic will include a hands-on activity to help you conne ct\nthe topic to a real-world application. On the last two days\, we will\ nform teams of CMU faculty and teachers to work together to create\nAI-rel ated activities for your classrooms. We will work directly with\nyou to de sign and build short demonstrations you can use in your\nclassrooms. Examp les include machine learning\, search and navigation\,\nand face recogniti on. We will also have faculty available to help\ndesign lecture slide cont ent and talking points for you to use in your\nclassrooms.\nTeachers will gain: hands-on skills to implement and apply AI\nalgorithms\, access to al l of the written materials and activities\npresented in the course\, addit ional activities and content that they\nhelped design to be tailored to th eir classrooms and their students’\ninterests\, a professional network o f Pittsburgh area STEM teachers and\nCarnegie Mellon faculty who can act a s resources and collaborators for\nfuture projects.We will be available to you throughout the school\nyear. CMU faculty will follow up several times during the school year\nto determine if the materials were sufficient and if you need\nadditional support.\nCONTACT US TO REGISTER\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91eadf3 DTSTART;TZID=America/New_York:20240613T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240613T180000 URL:https://csd.cmu.edu/calendar/thesis-oral-DERY-2024-06-13 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Thesis Oral Defense - Lucio Mwinmaarong Dery CLASS:PUBLIC DESCRIPTION:Speaker: LUCIO MWINMAARONG DERY\, Ph.D. Candidate\, Computer Sc ience\nDepartment\, Carnegie Mellon University\n\nTalk Title: On Resource Efficient Transfer Learning via End Task Aware\nTraining\n\nIn Transfer le arning\, performance on a desired end task (or tasks) is\nimproved by expl oiting \"knowledge\" from other tasks. The technique has\nbecome a critica l workhorse driving many of the advances in machine\nlearning. The current formula is relatively simple -- train a large\nmodel on large amounts of data from the transfer task(s)\; then apply\nthe learned model either zero -shot or adapted to the desired\ndownstream task(s). \n\nThis thesis reco gnizes that these powerful models are not developed\nin-vacuo but rather r equire non-trivial resources to train and deploy.\nAs such\, there are a w ide range of salient problems and communities of\nresearchers that the sta tus-quo leaves behind. In the first part of\nthis thesis\, we will focus o n the training time problem of\ndata-efficient transfer learning. We will begin by making a case for\nexploiting advanced knowledge of the desired d ownstream task(s) —\nwhich is commonly the case in many ML settings — to inform different\ndimensions of transfer learning. We dub this end task aware transfer\nlearning. Next\, we will present a set of novel end task aware\noptimization algorithms that bias the learning trajectory towards\n data-efficient solutions with strong generalization on the end task.\nWe w ill close this part by providing an automated approach to\nconstructing an d searching over task-relevant transfer objectives when\nonly end task dat a is available and in limited amounts. \n\nWe will proceed to develop alg orithms for compute and memory efficient\ntransfer learning. Our goal will be to deliver a small and efficient\nyet performant task specific model f or deployment seeded from a large\,\ngeneralist model that has already bee n pre-trained on a transfer task\n(or set of tasks). Focusing on structure d pruning for making models\nsmaller\, we will investigate pruning under t wo resource constrained\nsettings:\n\nlimited task data\, where we will ex ploit extra transfer tasks to learn\npruning structures that\, at the same task performance\, lead to more\ncompute and memory efficient modelssetti ngs of limited memory\, where\nmany of the classical pruning techniques br eak down because they\nrequire gradient-based optimization which can have prohibitive memory\noverhead.\n\nThesis Committee: \n\nGraham Neubig (Co- Chair)\n\nAmeet Talwalkar (Co-Chair)\n\nZico Kolter\n\nLuke Zettlemoyer (U niversity of Washington / Meta)\n\nMarc'Aurelio Ranzato  (Google DeepMind )\n\nIn Person and Zoom Participation. See announcement.\n\nMeeting ID:  797 670 0891\n\nPasscode:  ldm-thesis\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91eb2d6 DTSTART;TZID=America/New_York:20240612T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240612T160000 LOCATION:Mehrabian Collaborative Innovation Center 2101 SUMMARY:CyLab Summer Reading Group CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91eb451 DTSTART;TZID=America/New_York:20240612T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240612T130000 URL:https://pdl.cmu.edu/talk-series/2024/061224.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talk - PAT HELLAND CLASS:PUBLIC DESCRIPTION:Speaker: PAT HELLAND\, Salesforce\n\nTalk Title: Scalable OLTP in the Cloud: What's the BIG DEAL?\n\nThe pursuit of scalable OLTP systems has been the holy grail of my\ncareer. Because OLTP systems are typically split into applications and\ndatabases\, the isolation semantics provided by the DB and used by the\napp have a major impact on the scalability of the OLTP system as a\nwhole. The isolation semantics are a BIG DEAL! \n\n This thought experiment explores the asymptotic limits to scale for\nOLTP systems. An OLTP (OnLine Transaction Processing) system is a\ndomain-speci fic application using a RCSI (READ COMMIT- TED SNAPSHOT\nISOLATION) SQL da tabase to provide transactions across many concurrent\nusers. This interfa ce provides the contractual BIG DEAL between OLTP\ndatabases and OLTP appl ications. \n\nFocusing on the BIG DEAL\, shows today’s popular database s\nunnecessarily limit scale. Similarly\, we identify common app pat-\nter ns that inhibit scale. We can reimagine the way we build both\ndatabases a nd applications to empower scale. All while complying with\nthe establishe d SQL and RCSI interface (i.e.\, the BIG DEAL).  \n\nPerhaps\, this can provoke discussions within the database community\nleading to new opportun ities for OLTP systems. To me\, that would be a\nbig deal! \n\n— \n\nP at Helland has been building distributed systems\, database systems\,\nhig h-performance messaging systems\, and multiprocessors since 1978\,\nshortl y after dropping out of UC Irvine without a bachelor's degree.\nThat hasn' t stopped him from having a passion for academics and\npublication. From 1 982 to 1990\, Pat was the chief architect for TMF\n(Transaction Monitoring Facility)\, the transaction logging and\nrecovery systems for NonStop SQL \, a message-based fault-tolerant\nsystem providing high-availability solu tions for business critical\nsolutions. In 1991\, he moved to HaL Computer s where he was chief\narchitect for the Mercury Interconnect Architecture\ , a cache-coherent\nnon-uniform memory architecture multiprocessor. \n\nI n 1994\, Pat moved to Microsoft to help the company develop a business\npr oviding enterprise software solutions. He was chief architect for\nMTS (Mi crosoft Transaction Server) and DTC (Distributed Transaction\nCoordinator) . Starting in 2000\, Pat began the SQL Service Broker\nproject\, a high-pe rformance transactional exactly-once in-order\nmessage processing and app execution engine built deeply into\nMicrosoft SQL Server 2005. From 2005-2 007\, he worked at Amazon on\nscalable enterprise solutions\, scale-out us er facing services\,\nintegrating product catalog feeds from millions of s ellers\, and\nhighly-available eventually consistent storage. From 2007 t o 2011\,\nPat was back at Microsoft working on a number of projects includ ing\nStructured Streams in Cosmos. Structured streams kept metadata within \nthe \"big data\" streams that were typically 10s of terabytes in size.\n This metadata allowed affinitized placement within the cluster as well\nas efficient joins across multiple streams. On launch\, this doubled\nthe wo rk performed within the 250PB store. Pat also did the initial\ndesign for Baja\, the distributed transaction support for a distributed\nevent-proces sing engine implemented as an LSM atop structured streams\nproviding trans actional updates targeting the ingestion of \"the entire\nweb in one table \" with changes visible in seconds. Starting in 2012\,\nPat has worked at Salesforce on database technology running within\ncloud environments. \n\ nHis current interests include latency bounding of online\nenterprise-grad e transaction systems in the face of jitter\, the\nmanagement of metastabi lity in complex environments\, and zero-downtime\nupgrades to databases an d stateful applications. In his spare time\,\nPat regularly writes for ACM Queue\, Communications of the ACM\, and\nvarious conferences. He has been deeply involved in the organization\nof the HPTS (High Performance Transa ctions Systems) workshop since\n1985.  His blog is at pathelland.substack .com and he parsimoniously\ntweets with the handle @pathelland. \n\nZoom Participation.  See announcement.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91eb9f7 DTSTART;TZID=America/New_York:20240705T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240705T200000 URL:https://www.cmu.edu/hr/benefits/time-away/holidays.html SUMMARY:Community Appreciation Day Summer 2024 CLASS:PUBLIC DESCRIPTION:Talk Title: Community Appreciation Day M24\n\nLimited campus op erations will be open on these days\, such as Dining\nServices\, Universit y Libraries and the Cohon Center\; the hours of\noperation will be noted o n relevant university websites. Summer\nclasses scheduled on July 5 will p roceed as planned.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ebcb0 DTSTART;TZID=America/New_York:20240605T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240605T130000 URL:https://www.pdl.cmu.edu/talk-series/index.shtml LOCATION:Remote Access - Zoom SUMMARY:Parallel Data Laboratory Summer Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ebe5b DTSTART;TZID=America/New_York:20240605T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240605T120000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Thesis Oral Defense - Dravyansh Sharma CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ebfbe DTSTART;TZID=America/New_York:20240604T091500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240604T091500 URL:https://www.cs.cmu.edu/scs-ece-career-center/career-center-events SUMMARY:6th Learning Theory Alliance Mentorship workshop CLASS:PUBLIC DESCRIPTION:Speaker: For upper-level undergraduate and all-level graduate\n students\, and postdoctoral researchers interested in theoretical\ncompute r science and machine learningWe are pleased to invite you to\nthe 6th Le arning Theory Alliance Mentorship workshop\, to be held on\n4-5 June 2024. The workshop is free and fully virtual.\nThe workshop is intended for up per-level undergraduate and all-level\ngraduate students as well as postdo ctoral researchers who are\ninterested in theoretical computer science and machine learning. No\nprior research experience in the field is expected\ , and some sessions\nmay be of interest to researchers in adjacent fields. We have several\nplanned events including: A “how-to” talk on how to be a good\ncollaborator (discussing what healthy collaborations do and don ’t\nlook like\, setting expectations\, transitioning from junior to seni or\ncollaborator roles). A “how-to” talk on how to do theory research\ n(covering topics such as formulating research questions and theory\nprobl ems\, breaking a larger problem into smaller toy problems\, and\nday-to-da y best practices). A panel discussion on time management (for\nexample\, m aintaining a balance between learning and solving\, work-life\nbalance\, d eciding how many projects to work on). A social hour with\nmentoring table s.Our lineup includes Shuchi Chawla (UT)\, Adam Groce\n(Reed College)\, Zh iyi Huang (University of Hong Kong)\, Varun Kanade\n(Oxford)\, Po-Ling Loh (University of Cambridge)\, Audra McMillan\n(Apple)\, Ankur Moitra (MIT)\ , Devi Parikh (Georgia Tech)\, Aaditya\nRamdas (CMU)\, and Steven Wu (CMU) .\nA short application form is required to participate with an\napplicat ion deadline of Tuesday\, 28 May 2024.\nStudents with backgrounds that are underrepresented or underserved in\nrelated fields are especially encoura ged to apply. We are trying our\nbest to accommodate all time zones.   M ore information (including\nthe schedule) can be found on the event’s we bsite.\nThis workshop is part of our broader community-building initiative \ncalled the Learning Theory Alliance. Learn more...\nTo connect with fell ow participants and stay in touch for more\nannouncements\, we encourage e veryone to join the LeT-All slack.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ec380 DTSTART;TZID=America/New_York:20240530T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240530T170000 LOCATION:Reddy Conference Room\, Gates HIllman 4405 and Zoom SUMMARY:Thesis Oral Defense - Praneeth Kacham CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ec508 DTSTART;TZID=America/New_York:20240529T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240529T153000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Thesis Proposal - Eric Sturzinger CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ec69f DTSTART;TZID=America/New_York:20240528T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240528T130000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Thesis Proposal - Mingjie Sun CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ec7f4 DTSTART;TZID=America/New_York:20240521T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240521T133000 URL:https://www.cylab.cmu.edu/events/2024/05/21-seminar-alashoor.html LOCATION:Mehrabian Collaborative Innovation Center\, Panther Hollow Room 41 05 SUMMARY:CyLab Seminar - TAWFIQ ALASHOOR CLASS:PUBLIC DESCRIPTION:Speaker: TAWFIQ ALASHOOR\, Assistant Professor of Operations\,\ nInformation and Technology\, IESE Business School\, University of\nNavarr a\n\nTalk Title: The General Relativity of Privacy: From Spacetime to\nCon textime\n\nIn this seminar\, taking cues from Einstein’s theory of gener al\nrelativity\, Alashoor will introduce the “General Relativity of\nPri vacy\,” offering a new way to see privacy today. He will touch on\nfound ational terms in the multidisciplinary domain of privacy\,\nstarting with the basic principles of privacy and security\, and\nconcluding with an exp loration of data practices and their\nimplications at the individual\, org anizational\, and country level. \n\nBuilding on this foundation\, he wil l draw from several studies to\nshare empirical findings that spotlight pr ivacy manipulation\ntechniques influencing decisions to share or protect p ersonal data.\nThese techniques might not only reveal personal information assets\,\nbut they could also\, directly or indirectly\, expose organizat ional\ninformation assets\, posing significant cybersecurity risks across\ nvarious levels. By the end of the seminar\, attendees will be equipped\nw ith valuable knowledge and soft skills\, such as privacy awareness\,\nthat will aid them in making informed privacy decisions\, thereby\nhelping to mitigate potential cybersecurity risks\n\n— \n\nTawfiq Alashoor is an A ssistant Professor at IESE Business School\,\nUniversity of Navarra (Spain ). Alashoor is a researcher\, educator\, and\nconsultant with more than 12 years of experience in the cybersecurity\,\nIT\, and banking industry. He holds B.S.\, M.S.\, and Ph.D. degrees from\nKing Fahd University of Petro leum and Minerals (KFUPM\, Saudi Arabia)\,\nPenn State University (U.S.)\, and Georgia State University (U.S.)\,\nrespectively. He has completed a p ostdoc program at the University of\nNotre Dame (U.S.) and has taught mana gerial cybersecurity at KFUPM\,\nUniversity of Notre Dame\, and Copenhagen Business School to MBA and\nundergraduate students. \n\nAlashoor’s mai n research focuses on privacy decision making and\ncybersecurity. His rese arch has been published in globally recognized\npeer-reviewed journals and conferences\, such as Information Systems\nResearch and Academy of Manage ment. He received several academic\nawards\, and he serves as a reviewer f or several FT50 journals\nincluding other editorial roles in the privacy r esearch community.\n“It is seriously time for privacy education\, traini ng\, and awareness\n(PETA) programs” is the title of his recent book cha pter. \n\nAlashoor is a Co-founder and the Publications Director at the S audi\nAssociation for Information Systems (AIS) and a member of the AIS\nS pecial Interest Group on Information Security and Privacy.\nAlashoor’s w ork has been featured in the media and industry reports\,\nand he has been a guest on several podcasts. \n\nThe CyLab seminar is open only to partn ers and Carnegie Mellon\nUniversity faculty\, students\, and staff.\n DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ecc7e DTSTART;TZID=America/New_York:20240520T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240520T150000 URL:https://www.cs.cmu.edu/~pop/seminar/ LOCATION:McWilliams Classroom\, Gates Hillman 4303 SUMMARY:Principles of Programming Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ece22 DTSTART;TZID=America/New_York:20240516T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240516T103000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ecf78 DTSTART;TZID=America/New_York:20240516T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240516T123000 LOCATION:Cohon University Center SUMMARY:Global Accessibility Awareness Day at CMU CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ed0d4 DTSTART;TZID=America/New_York:20240516T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240516T130000 URL:https://www.cmu.edu/computing/dao/news/2024/april/gaad2024.html LOCATION:Cohon University Center SUMMARY:Global Accessibility Awareness Day CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ed23d DTSTART;TZID=America/New_York:20240515T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240515T133000 LOCATION:Wiegand Gymnasium and the Tartan Pavilion\, Resnik House / Legacy Plaza SUMMARY:CMU Community Picnic CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ed382 DTSTART;TZID=America/New_York:20240514T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240514T133000 LOCATION:Mehrabian Collaborative Innovation Center\, Panther Hollow Room 41 05 SUMMARY:CyLab Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ed4c5 DTSTART;TZID=America/New_York:20240512T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240512T133000 LOCATION:Carnegie Music Hall\, 4400 Forbes Avenue SUMMARY:SCS Commencement: Undergraduate Diploma Ceremony 2024 CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ed619 DTSTART;TZID=America/New_York:20240511T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240511T110000 LOCATION:Carnegie Music Hall\, 4400 Forbes Avenue SUMMARY:SCS Commencement: Ph.D. Diploma Ceremony 2024 CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ed75c DTSTART;TZID=America/New_York:20240511T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240511T140000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:SCS Honors Distribution CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ed89a DTSTART;TZID=America/New_York:20240510T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240510T170000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:SCS Honors Distribution CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ed9ea DTSTART;TZID=America/New_York:20240510T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240510T113000 LOCATION:Soldiers and Sailors Memorial Hall\, 4141 Fifth Avenue SUMMARY:SCS Commencement: Master's Diploma Ceremony 2024 CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91edb2b DTSTART;TZID=America/New_York:20240508T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240508T150000 LOCATION:Gates Hillman 8102 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91edc73 DTSTART;TZID=America/New_York:20240507T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240507T163000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Computer Science 5th Years Master's Thesis Presentation CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91eddc3 DTSTART;TZID=America/New_York:20240507T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240507T110000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91edf02 DTSTART;TZID=America/New_York:20240506T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240506T170000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ee03f DTSTART;TZID=America/New_York:20240506T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240506T130000 URL:https://www.cylab.cmu.edu/events/2024/05/06-seminar-telang.html LOCATION:Panther Hollow Room\, Mehrabian Collaborative Innovation Center 45 01 SUMMARY:Crypto Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ee191 DTSTART;TZID=America/New_York:20240506T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240506T120000 LOCATION:Newell-Simon 3305 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ee342 DTSTART;TZID=America/New_York:20240506T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240506T110000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Computer Science 5th Years Master's Thesis Presentation CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ee493 DTSTART;TZID=America/New_York:20240503T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240503T120000 LOCATION:McWilliams Classroom\, Gates Hillman 4303 and Zoom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ee5e2 DTSTART;TZID=America/New_York:20240502T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240502T170000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ee725 DTSTART;TZID=America/New_York:20240502T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240502T133000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Computer Science Thesis Proposal CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91eed85 DTSTART;TZID=America/New_York:20240502T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240502T120000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91eef2a DTSTART;TZID=America/New_York:20240502T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240502T100000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Computer Science 5th Year Master's Thesis Presentation CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ef096 DTSTART;TZID=America/New_York:20240501T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240501T160000 LOCATION:Gordon Bell Conference Room\, Gates Hillman 5117 and Zoom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ef1f7 DTSTART;TZID=America/New_York:20240501T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240501T150000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ef356 DTSTART;TZID=America/New_York:20240501T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240501T153000 LOCATION:McWilliams Classroom\, Gates HIllman 4303 and Zoom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ef4fb DTSTART;TZID=America/New_York:20240501T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240501T120000 LOCATION:Wean Hall 5328 SUMMARY:Computer Science 5th Year Master's Thesis Presentation CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ef654 DTSTART;TZID=America/New_York:20240430T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240430T163000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Mehrabian Collaborative Innovation Center 2201 and Zoom SUMMARY:Crypto Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ef7c1 DTSTART;TZID=America/New_York:20240430T133000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240430T153000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and ZOom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91ef95e DTSTART;TZID=America/New_York:20240430T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240430T130000 URL:http://www.cs.cmu.edu/~aiseminar/ LOCATION:Newell-Simon 3305 and Zoom SUMMARY:Artificial Intelligence Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91efaa8 DTSTART;TZID=America/New_York:20240430T103000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240430T113000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Computer Science 5th Year Master's Thesis Presentation CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91efc26 DTSTART;TZID=America/New_York:20240430T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240430T113000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Computer Science Thesis Proposal CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91efd97 DTSTART;TZID=America/New_York:20240429T143000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240429T153000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Computer Science 5th Year Masters Thesis Presentation CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91efee7 DTSTART;TZID=America/New_York:20240429T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240429T140000 LOCATION:Gates HIllman 9115 SUMMARY:Computer Science 5th Year Master of Science Thesis Presentation CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f0045 DTSTART;TZID=America/New_York:20240429T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240429T140000 LOCATION:McWilliams Classroom\, Gates Hillman 4303 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f0194 DTSTART;TZID=America/New_York:20240429T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240429T130000 URL:https://www.cylab.cmu.edu/events/2024/04/29-seminar-boneh.html LOCATION:Simmons Auditorium A \, Tepper Building and Youtube SUMMARY:CyLab Blockchain Distinguished Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f0302 DTSTART;TZID=America/New_York:20240429T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240429T130000 LOCATION:Gates Hillman 8102 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f048b DTSTART;TZID=America/New_York:20240429T113000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240429T123000 LOCATION:Gates Hillman 7501 SUMMARY:Computer Science 5th Year Master of Science Thesis Presentation CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f05dc DTSTART;TZID=America/New_York:20240426T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240426T210000 LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:15-322/622 Computer Music Concert CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f0729 DTSTART;TZID=America/New_York:20240426T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240426T163000 URL:https://www.cmu.edu/ai-sdm/events/index.html LOCATION:Newell-Simon 4305 and Zoom SUMMARY:AI Institute for Societal Decision Making Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f088a DTSTART;TZID=America/New_York:20240426T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240426T170000 LOCATION:Newell-Simon Hall 3002 SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f09d4 DTSTART;TZID=America/New_York:20240426T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240426T150000 LOCATION:Gates Hillman 6121 and Zoom SUMMARY:AI Institute for Societal Decision Making - Student Session CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f0b21 DTSTART;TZID=America/New_York:20240426T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240426T120000 LOCATION:Wean Hall 5415 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f0c9e DTSTART;TZID=America/New_York:20240426T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240426T100000 LOCATION:Gordon Bell Conference Room\, Gates Hillman 5117 SUMMARY:Computer Science Thesis Proposal CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f0deb DTSTART;TZID=America/New_York:20240425T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240425T180000 LOCATION:ASA Conference Room\, Gates Hillman 6115 SUMMARY:School of Computer Science Professorships Celebration CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f0f4b DTSTART;TZID=America/New_York:20240425T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240425T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Crypto Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f10b6 DTSTART;TZID=America/New_York:20240425T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240425T160000 URL:https://aco.math.cmu.edu/abs-23-24/apr25.html LOCATION:Wean Hall 8220 SUMMARY:Algorithms\, Combinatorics and Optimization Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f1204 DTSTART;TZID=America/New_York:20240424T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240424T180000 LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f135d DTSTART;TZID=America/New_York:20240424T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240424T150000 LOCATION:Gates Hillman 8102 SUMMARY:Joint Theory Seminar / Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f151c DTSTART;TZID=America/New_York:20240424T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240424T135000 LOCATION:Remote Access - Zoom SUMMARY:15-445 Class Speaker CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f166b DTSTART;TZID=America/New_York:20240424T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240424T133000 LOCATION:Smith Hall 236 SUMMARY:Computer Imaging Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f17d6 DTSTART;TZID=America/New_York:20240424T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240424T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Gates Hillman 8102 SUMMARY:Theory Lunch Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f192b DTSTART;TZID=America/New_York:20240424T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240424T100000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 SUMMARY:Computer Science 5th Years Master's Thesis Presentation - POSTPONED CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f1a70 DTSTART;TZID=America/New_York:20240423T123000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240423T133000 URL:https://privacy.cs.cmu.edu/masters/seminar/index.html LOCATION:Group Viewing Hamburg Hall 1002 and Zoom SUMMARY:Privacy Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f1bc3 DTSTART;TZID=America/New_York:20240422T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240422T143000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 and Zoom SUMMARY:Computer Science Thesis Proposal CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f1d07 DTSTART;TZID=America/New_York:20240422T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240422T120000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f1e60 DTSTART;TZID=America/New_York:20240419T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240419T130000 LOCATION:Gates HIllman 9115 SUMMARY:Special Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f1fe4 DTSTART;TZID=America/New_York:20240419T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240419T130000 LOCATION:Gates Hillman 8102 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f212c DTSTART;TZID=America/New_York:20240418T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240418T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Crypto Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f226e DTSTART;TZID=America/New_York:20240418T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240418T160000 URL:https://www.cs.cmu.edu/~pop/seminar/2024-04-18-Michel/ LOCATION:Newell-Simon 3305 SUMMARY:Principles of Programming Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f23e0 DTSTART;TZID=America/New_York:20240418T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240418T120000 LOCATION:Newell-Simon 4305 and Zoom SUMMARY:SCS Faculty Candidate CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f253d DTSTART;TZID=America/New_York:20240417T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240417T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Gates Hillman 8102 SUMMARY:Joint Theory Lunch Seminar / Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f268b DTSTART;TZID=America/New_York:20240417T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240417T130000 LOCATION:Wean Hall 3701 and Zoom SUMMARY:Joint Carnegie Mellon Electricity Industry Center and CyLab Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f27f1 DTSTART;TZID=America/New_York:20240416T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240416T170000 URL:http://www.cs.cmu.edu/~aiseminar/ LOCATION:Newell-Simon 3305 and Zoom SUMMARY:Joint Artificial Intelligence Seminar / Computer Science Speaking\n Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f2938 DTSTART;TZID=America/New_York:20240416T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240416T140000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f2a84 DTSTART;TZID=America/New_York:20240416T083000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240417T180000 URL:https://www.cylab.cmu.edu/research/blockchain/secure-blockchain-summit/ index.html LOCATION:Simmons Auditorium\, Tepper Building SUMMARY:CMU Secure Blockchain Summit CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f2be3 DTSTART;TZID=America/New_York:20240416T080000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240417T200000 URL:https://www.cylab.cmu.edu/research/blockchain/secure-blockchain-summit/ index.html LOCATION:Tepper Building SUMMARY:Second Annual Carnegie Mellon Blockchain Conference CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f2d29 DTSTART;TZID=America/New_York:20240415T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240415T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Mehrabian Collaborative Innovation Center 2201 SUMMARY:Crypto Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f2e6c DTSTART;TZID=America/New_York:20240415T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240415T140000 LOCATION:Reddy Conference Room\, Gates Hillman 4405 SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f2fc7 DTSTART;TZID=America/New_York:20240412T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240412T140000 LOCATION:Gates Hillman 9115 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f310a DTSTART;TZID=America/New_York:20240412T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240412T130000 URL:https://www.cylab.cmu.edu/events/2024/04/12-seminar-kongs.html LOCATION:Hamburg Hall A301 and Zoom SUMMARY:CyLab Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f324d DTSTART;TZID=America/New_York:20240412T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240412T120000 LOCATION:Newell-Simon 3002 SUMMARY:Principles of Programming Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f33d4 DTSTART;TZID=America/New_York:20240411T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240411T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Gates Hillman 8102 and Zoom SUMMARY:Crypto Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f351c DTSTART;TZID=America/New_York:20240411T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240411T170000 LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f3689 DTSTART;TZID=America/New_York:20240411T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240411T150000 LOCATION:Gates Hillman 7501 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f37ea DTSTART;TZID=America/New_York:20240411T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240411T130000 URL:https://pdl.cmu.edu/SDI/2024/041124.html LOCATION:Newell-Simon 3305 SUMMARY:Systems Design and Implementation Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f3930 DTSTART;TZID=America/New_York:20240410T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240410T130000 URL:https://www.cs.cmu.edu/~theorylunch/ LOCATION:Gates Hillman 8102 SUMMARY:Joint Computer Science Speaking Skills Talk / Theory Lunch Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f3aa2 DTSTART;TZID=America/New_York:20240410T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240410T130000 URL:https://pdl.cmu.edu/SDI/2024/041024.html LOCATION:ASA Conference Room\, Gates Hillman 6115 and Zoom SUMMARY:Special Systems Design and Implementation Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f3c0f DTSTART;TZID=America/New_York:20240409T110000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240409T120000 LOCATION:Traffic21 Classroom\, Gates Hillman 6121 SUMMARY:Computer Science Speaking Skills Talk CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f3d5d DTSTART;TZID=America/New_York:20240405T180000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240405T180000 URL:http://sigbovik.org/2024/ LOCATION:Rashid Auditorium\, Gates Hillman 4401 SUMMARY:SIGBOVIK 2024 CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f3ec9 DTSTART;TZID=America/New_York:20240405T163000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240405T173000 URL:https://sites.google.com/view/crypto-seminar/home LOCATION:Gates Hillman 7501 and Zoom SUMMARY:Crypto Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f4021 DTSTART;TZID=America/New_York:20240405T141500 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240405T151500 URL:https://www.cmu.edu/dietrich/statistics-datascience/events/index.html LOCATION:Doherty Hall A302 SUMMARY:Statistics and Data Science Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef91f4167 DTSTART;TZID=America/New_York:20240405T140000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240405T160000 SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9200034 DTSTART;TZID=America/New_York:20240405T100000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240405T113000 LOCATION:Traffic21 Classroom\, Gates Hillman 6501 and Zoom SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef920017e DTSTART;TZID=America/New_York:20240404T150000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240404T160000 URL:https://aco.math.cmu.edu/abs-23-24/apr04.html LOCATION:Wean Hall 8220 SUMMARY:Algorithms\, Combinatorics and Optimization Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef92002c4 DTSTART;TZID=America/New_York:20240404T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240404T140000 LOCATION:Newell-Simon SUMMARY:SCS Faculty Candidate CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9200423 DTSTART;TZID=America/New_York:20240403T160000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240403T180000 LOCATION:Gates and HIllman Centers SUMMARY:Computer Science Thesis Oral CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9200585 DTSTART;TZID=America/New_York:20240402T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240402T130000 URL:http://www.cs.cmu.edu/~aiseminar/ LOCATION:ASA Conference Room\, Gates HIllman 6115 and Zoom SUMMARY:Artificial Intelligence Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef92006e2 DTSTART;TZID=America/New_York:20240401T120000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240401T130000 URL:https://www.cylab.cmu.edu/events/2024/04/01-seminar-boyle.html LOCATION:Panther Hollow Room 4105\, 4th Floor\, Mehrabian Collaborative\nIn novation Center and Livestream SUMMARY:CyLab Blockchain Seminar CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9200859 DTSTART;TZID=America/New_York:20240401T090000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240401T103000 LOCATION:Gates and HIllman Centers SUMMARY:Computer Science Thesis Proposal CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef92009ab DTSTART;TZID=America/New_York:20240510T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240510T170000 LOCATION:Gates and HIllman Centers SUMMARY:Doctoral Student Review (DSR) General Meeting CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9200af4 DTSTART;TZID=America/New_York:20240509T130000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240509T170000 LOCATION:Gates and HIllman Centers SUMMARY:Doctoral Student Review (DSR) PL/Security/Systems CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9200c4f DTSTART;TZID=America/New_York:20240509T093000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240509T120000 LOCATION:Gates and HIllman Centers SUMMARY:Doctoral Student Review (DSR) AI/Grapics/Theory CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT BEGIN:VEVENT UID:6a09ef9200d9a DTSTART;TZID=America/New_York:20240409T153000 SEQUENCE:0 TRANSP:TRANSPARENT DTEND;TZID=America/New_York:20240409T170000 LOCATION:Gates and Hillman Centers SUMMARY:CSD Faculty Meeting CLASS:PUBLIC DTSTAMP:20260517T164050Z END:VEVENT END:VCALENDAR