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:STANDARD TZNAME:EST TZOFFSETFROM:-0400 TZOFFSETTO:-0500 DTSTART:20251102T060000 END:STANDARD BEGIN:STANDARD TZNAME:EST TZOFFSETFROM:-0400 TZOFFSETTO:-0500 DTSTART:20241103T060000 END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:6a332aa89a189 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:20260617T231552Z END:VEVENT BEGIN:VEVENT UID:6a332aa89b562 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:20260617T231552Z END:VEVENT BEGIN:VEVENT UID:6a332aa89bac9 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:20260617T231552Z END:VEVENT BEGIN:VEVENT UID:6a332aa89bf9c 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:20260617T231552Z END:VEVENT BEGIN:VEVENT UID:6a332aa89c453 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:20260617T231552Z END:VEVENT BEGIN:VEVENT UID:6a332aa89c8a6 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:20260617T231552Z END:VEVENT BEGIN:VEVENT UID:6a332aa89cd3a 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:20260617T231552Z END:VEVENT END:VCALENDAR