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:6a3320beb4264 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb553a 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb5b60 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb60d9 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb65e5 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb6b63 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb7085 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb75ba 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb7aa7 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb7eec 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb84de 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb8a87 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb8f69 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb947e 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb9917 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beb9dfb 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beba244 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320beba63c 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebaad2 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebaf82 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebb3ab 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebb8f0 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebbe07 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebc308 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebc7c3 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebccb6 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebd0b0 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebd494 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebd8d2 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebdda0 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebe1aa 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebe5d9 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebea7d 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebee8a 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebf2ee 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebf79b 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebfbe6 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebfd95 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bebff12 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bec0076 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bec01d6 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bec0361 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bec04ca 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bec0641 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:20260617T223334Z END:VEVENT BEGIN:VEVENT UID:6a3320bec07ae 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:20260617T223334Z END:VEVENT END:VCALENDAR