BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Halıcıoğlu Data Science Institute - UC San Diego - ECPv6.16.2//NONSGML v1.0//EN CALSCALE:GREGORIAN METHOD:PUBLISH X-ORIGINAL-URL:https://datascience.ucsd.edu X-WR-CALDESC:Events for Halıcıoğlu Data Science Institute - UC San Diego REFRESH-INTERVAL;VALUE=DURATION:PT1H X-Robots-Tag:noindex X-PUBLISHED-TTL:PT1H BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:20230312T100000 END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:20231105T090000 END:STANDARD BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:20240310T100000 END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:20241103T090000 END:STANDARD BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:20250309T100000 END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:20251102T090000 END:STANDARD BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:20260308T100000 END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:20261101T090000 END:STANDARD BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:20270314T100000 END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:20271107T090000 END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20260330T140000 DTEND;TZID=America/Los_Angeles:20260330T150000 DTSTAMP:20260531T014033 CREATED:20260206T212001Z LAST-MODIFIED:20260312T171805Z UID:10000543-1774879200-1774882800@datascience.ucsd.edu SUMMARY:Pietro Perona Distinguished Lecture DESCRIPTION:Talk Information\nSpeaker: Pietro Perona\nDate & Time: Monday March 30th\, 2pm\nLocation: HDSI Multipurpose Room 123\n\n\nTITLE Responsible AI – A case for causal reasoning\n\nABSTRACT “As Artificial Intelligence  (AI) finds increasing applications in industry and society\, responsible deployment demands that we measure and correct algorithmic biases vis-a-vis protected attributes such as sex\, age and ethnicity. State of the art methods for measuring algorithmic bias rely on test sets that are collected in the wild and are then annotated for the protected attributes. Such methods are therefore observational and yield correlational observations. I will argue that\, in order to obtain useful information to discover and correct biases we need causal information which is only available if we use experimental methods. I will show that modern generative models offer a promising starting point to develop experimental testing methods. I will review our recent work in face synthesis and demonstrate its application to the study of algorithmic bias in gender classification\, face recognition\, and social judgment of faces.”\n\nBIO Pietro Perona is the Allan E. Puckett Professor of Electrical Engineering at the California Institute of Technology. He is known for his research in computer vision and is the director of the Caltech Computational Vision Group. Professor Perona’s research focuses on vision: how do we see and how can we build machines that see. Professor Perona is currently interested in visual recognition\, more specifically visual categorization. He is studying how machines can learn to recognize frogs\, cars\, faces and trees with minimal human supervision\, and how machines can learn from human experts. His project `Visipedia’ has produced two smart device apps (iNaturalist and Merlin Bird ID) that anyone can use to recognize the species of plants and animals from a photograph. In collaboration with Professors Anderson and Dickinson\, professor Perona is building vision systems and statistical techniques for measuring actions and activities in fruit flies and mice. This enables geneticists and neuroethologists to investigate the relationship between genes\, brains and behavior. Professor Perona is also interested in studying how humans perform visual tasks\, such as searching and recognizing image content. One of his recent projects studies how to harness the visual ability of thousands of people on the web. URL:https://datascience.ucsd.edu/event/pietro-perona-distinguished-lecture/ CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20260304T140000 DTEND;TZID=America/Los_Angeles:20260304T150000 DTSTAMP:20260531T014033 CREATED:20260226T194755Z LAST-MODIFIED:20260226T194755Z UID:10000550-1772632800-1772636400@datascience.ucsd.edu SUMMARY:HDSI Seminar - Eddy Keming Chen - AI Meets Philosophy of Science: Towards a Foundation of AI DESCRIPTION:Speaker:Eddy Keming Chen\nDate & Time: Wednesday March 4nd\, 2:00pm\nLocation: HDSI Multipurpose Room 123 \nTitle:  AI Meets Philosophy of Science: Towards a Foundation of AI\n\n\nAbstract: \n Why do simple learning rules yield AI systems that generalize far beyond their training data? I argue that this reflects an abundance of learnable structure in nature\, and that this abundance motivates Nomic Liberalism\, a conception of laws developed from Minimal Primitivism (Chen and Goldstein\, 2022). On this view\, laws can be simple\, predictive\, representation-relative\, existing at many scales and in domains far beyond fundamental physics. Simplicity functions as a nomic razor: an epistemic guide to discovering laws\, rather than a general guide to truth (Chen\, 2024\, 2025).\n\nI show how puzzling phenomena in machine learning—such as double descent\, scaling laws\, and the emergence of AGI from simple objectives such as next-token prediction (Chen\, Belkin\, Bergen\, and Danks\, 2026)—can be understood as learning systems discovering such liberal laws\, often in domains traditionally thought to lack lawful structure\, and in coordinate systems quite unlike familiar human concepts. AI success thus provides concrete evidence for this expanded conception of lawhood. Yet abundance has limits. Recent results in quantum foundations establish in-principle constraints on learning: in high-dimensional quantum systems\, nearly all quantum states are observationally indistinguishable—a limit no learning algorithm can overcome (Chen and Tumulka\, 2025). A satisfactory theory of induction must therefore explain both why learning works so well and why it must sometimes fail.\n\n\n\n\n\nSpeaker Bio: ​”I am an associate professor of philosophy\, a faculty affiliate of the Halıcıoğlu Data Science Institute and the Chinese studies program at the University of California\, San Diego\, and ​a fellow of the John Bell Institute for the Foundations of Physics.\n\nMy primary research interests are foundations of AI\, philosophy of physics\, philosophy of science\, and metaphysics. I am also interested in philosophy of mind\, decision theory\, formal epistemology\, philosophy of mathematics\, philosophy of religion\, and Chinese philosophy. I have a side interest in using films to popularize philosophical ideas. Currently\, I’m co-writing a screenplay about a time-travel romance (inspired by a fascinating article in the SEP).” URL:https://datascience.ucsd.edu/event/hdsi-seminar-eddy-keming-chen-ai-meets-philosophy-of-science-towards-a-foundation-of-ai/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20260303T140000 DTEND;TZID=America/Los_Angeles:20260303T150000 DTSTAMP:20260531T014033 CREATED:20260206T211743Z LAST-MODIFIED:20260226T195702Z UID:10000542-1772546400-1772550000@datascience.ucsd.edu SUMMARY:HDSI Seminar - Jessilyn Dunn - The Digital Physiome: Wearables for Disease Detection and Monitoring DESCRIPTION:Speaker: Jessilyn Dunn\nDate & Time: Monday March 2nd\, 2:00pm\nLocation: HDSI Multipurpose Room 123\nZoom Link: http://bit.ly/HDSI-Seminars\n\n\n\nTitle:  The Digital Physiome: Wearables for Disease Detection and Monitoring\n\n\nAbstract: Digital health is rapidly expanding due to surging healthcare costs\, deteriorating health outcomes\, and the growing prevalence and accessibility of mobile health and wearable technologies. Recent technological advancements make it possible to closely and continuously monitor individuals using multiple measurement modalities in real time. We are collecting and integrating such wearables data with clinical information to gain a more precise understanding of health and disease and develop actionable\, predictive health models for improving outcomes. We are simultaneously developing open source data science and machine learning tools for the digital health community\, including the Digital Biomarker Discovery Pipeline (DBDP)\, to facilitate the use of mobile device data in healthcare.\n\n\n\n\nSpeaker Bio: Jessilyn Dunn\, PhD\, is an Associate Professor of Biomedical Engineering and Biostatistics & Bioinformatics at Duke University. She directs the BIG IDEAs Lab\, which is focused on digital health innovation\, wearable sensors\, and the development and validation of AI-driven digital biomarkers. Dr. Dunn is the Principal Investigator of research initiatives funded by the NIH\, NSF\, and FDA which are developing digital biomarkers of conditions ranging from pre- and type 2 diabetes to influenza-like illness to Opioid Use Disorder. She sits on the Google Consumer Health Advisory Panel and is a recipient of the NSF CAREER Award and the IEEE EMBS Early Career Achievement Award for her leadership and innovation across engineering and medicine. URL:https://datascience.ucsd.edu/event/jessilyn-dunn-seminar/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20260302T140000 DTEND;TZID=America/Los_Angeles:20260302T150000 DTSTAMP:20260531T014033 CREATED:20260206T211446Z LAST-MODIFIED:20260219T170241Z UID:10000541-1772460000-1772463600@datascience.ucsd.edu SUMMARY:HDSI Seminar - Emily Aiken - Empirically understanding the relative value of prediction in allocation DESCRIPTION:Speaker: Emily Aiken\nDate & Time: Monday March 2nd\, 2:00pm\nLocation: HDSI Multipurpose Room 123\n  \nTalk Title:\nEmpirically understanding the relative value of prediction in allocation \nAbstract:\nPublic institutions increasingly use prediction to allocate scarce resources. From a design perspective\, better predictions compete with other investments\, such as expanding capacity or improving treatment quality. Here\, the big question is not how to solve a specific allocation problem\, but rather which problem to solve. In this talk\, I will introduce an empirical toolkit\nto help planners form principled answers to this question and quantify the bottom-line welfare impact of investments in prediction versus other policy levers such as expanding capacity and improving treatment quality. I will then apply the framework to two real-world case studies on predictive methods for allocating humanitarian aid in Bangladesh and Ethiopia. Related papers: this and this. \nSpeaker bio: \nEmily is an assistant professor jointly appointed in HDSI and the School of Global Policy and Strategy. Her research interests are at the intersection of data science and development economics\, with a focus on analyzing large digital traces to inform the design of social protection and humanitarian aid programs. Her work centers on the implementation and evaluation of data-driven allocation systems for aid delivery in low-income countries. \nPrior to joining UC San Diego\, Emily was a postdoctoral scholar at Carnegie Mellon University Africa. She received her PhD from the UC Berkeley School of Information\, and holds an MS in computer science from UC Berkeley and a BA in computer science from Harvard University. URL:https://datascience.ucsd.edu/event/emily-aiken-seminar/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20260225T140000 DTEND;TZID=America/Los_Angeles:20260225T150000 DTSTAMP:20260531T014033 CREATED:20260219T165539Z LAST-MODIFIED:20260219T165539Z UID:10000549-1772028000-1772031600@datascience.ucsd.edu SUMMARY:HDSI Seminar - Zhijian Liu - Efficient AI in the Era of Large Models DESCRIPTION:Speaker: Zhijian Liu\nDate & Time: Wednesday February 25th\, 2:00pm\nLocation: HDSI Multipurpose Room 123 \nTalk Title:\nEfficient AI in the Era of Large Models \nAbstract:\nLarge foundation models now deliver remarkable capabilities in understanding\, reasoning\, and acting across digital and physical domains. Yet their computational cost has become the primary bottleneck to scaling and real-world deployment. As models expand in size\, context length\, and modality\, efficiency is no longer optional. It is a first-order challenge in algorithms and systems design. \nIn this talk\, I present advances along three complementary directions: parallel decoding\, quantized reasoning\, and structured sparsity. I show how diffusion-based parallel drafting reduces autoregressive latency through speculative decoding; how 4-bit quantization remains near-lossless even under demanding reasoning workloads; and how structured sparsity accelerates long-context inference and long-form reasoning. I conclude with case studies demonstrating how these principles enable efficient multimodal and physical AI systems. Together\, these results show that advances in algorithms and systems design are essential to making large models faster\, more affordable\, and deployable at scale. \nSpeaker Bio: Zhijian Liu is an assistant professor at UCSD. Previously\, he received his Ph.D. and S.M. from MIT and his B.Eng. from Shanghai Jiao Tong University. His research focuses on efficient machine learning and systems. He was selected as the recipient of the Qualcomm Innovation Fellowship. He was also recognized as a Rising Star in ML and Systems by MLCommons and a Rising Star in Data Science by UChicago and UCSD. URL:https://datascience.ucsd.edu/event/hdsi-seminar-zhijian-liu-efficient-ai-in-the-era-of-large-models/ LOCATION:Halıcıoğlu Data Science Institute Room 123\, 3234 Matthews Ln\, La Jolla\, CA 92093\, USA CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20260211T110000 DTEND;TZID=America/Los_Angeles:20260211T123000 DTSTAMP:20260531T014033 CREATED:20251002T164405Z LAST-MODIFIED:20251002T164405Z UID:10000531-1770807600-1770813000@datascience.ucsd.edu SUMMARY:HDSI/TILOS Seminar - Maarten de Hoop DESCRIPTION:Talk Details TBA URL:https://datascience.ucsd.edu/event/hdsi-tilos-seminar-maarten-de-hoop/ CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20260209T140000 DTEND;TZID=America/Los_Angeles:20260209T150000 DTSTAMP:20260531T014033 CREATED:20260116T194400Z LAST-MODIFIED:20260123T223602Z UID:10000536-1770645600-1770649200@datascience.ucsd.edu SUMMARY:HDSI Seminar Series - Xiuyuan Cheng - Wasserstein-regularized learning via neural transport maps: from gradient flow to minimax DESCRIPTION:Speaker: Xiuyuan Cheng\nDate & Time: Monday February 9th\, 2:00pm\nLocation: HDSI Multipurpose Room 123\n\n\n\nTalk Title: Wasserstein-regularized learning via neural transport maps: from gradient flow to minimax \n\nAbstract: Wasserstein regularization has become a common tool in learning problems over distributions\, while its practical computation remains challenging\, especially in high dimensions. This talk presents a framework for Wasserstein-regularized learning based on directly parameterizing transport maps with neural networks. Rather than working with dual potentials or entropic relaxations\, we explicitly model the transport map\, which enables principled optimization in Wasserstein space as well as scalable out-of-sample evaluation. One example is the implementation of Wasserstein-2 proximal steps via flow networks that realize the Jordan–Kinderlehrer–Otto (JKO) scheme\, yielding a variational interpretation of flow-based generative models. The main focus of the talk is a different regime motivated by distributionally robust optimization (DRO)\, where transport is chosen adversarially through a minimax objective. In this setting\, we establish convergence guarantees for gradient descent–ascent (GDA) dynamics of the Wasserstein minimax problem\, and show how a neural transport map learned from a sample-based matching loss enables direct generation from the worst-case distribution. Together\, these examples highlight both the opportunities and challenges of incorporating Wasserstein regularization into generative modeling\, suggesting that the transport-map formulation offers a flexible and promising way to leverage Wasserstein geometry in learning problems over distributions.\n\nSpeaker Bio: Prof Xiuyuan Cheng currently is Professor of Mathematics at Duke University\, previously as Associate Professor of Mathematics at Trinity College of Arts & Sciences. Prof. Cheng works as an applied analyst\, developing theoretical and computational techniques to solve problems in high-dimensional statistics\, signal processing and machine learning. They are a past recipient of the National Science Foundation CAREER Award as well as the Sloan Research Fellowship-Mathematics. URL:https://datascience.ucsd.edu/event/hdsi-seminar-series-xiuyuan-cheng/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20260205T150000 DTEND;TZID=America/Los_Angeles:20260205T160000 DTSTAMP:20260531T014033 CREATED:20260116T200658Z LAST-MODIFIED:20260116T200658Z UID:10000537-1770303600-1770307200@datascience.ucsd.edu SUMMARY:HDSI Seminar Series - Stephen Pizer - A geometric model for anatomic objects targeted for statistical analysis DESCRIPTION:Title:\nA geometric model for anatomic objects targeted for statistical analysis\nSpeaker:\nStephen Pizer\, Kenan Professor of Computer Science\, University of North Carolina at Chapel Hill \nAbstract:\nA novel representation called the evolutionary s-rep represents anatomic objects by encompassing an unusually rich collection of geometric features\, image intensity based\nfeatures\, genetic features\, etc.  It can be used for shape-based classification of anatomic objects\, hypothesis testing on inter-class shape variations\, and object segmentation\, using web-\nresident software. The evolutionary s-rep and example applications will be discussed. This new representation provides better statistics than alternatives because its across-object positional correspondences are based on geometry not just on the object boundary but also in its interior\, on second-order and not just positional (0th order) features\, and on finding for each case a\ndiffeomorphism from a basic\, ellipsoidal object. \nSpeaker Bio\n“My PhD dissertation\, in 1967\, was the first in medical image computing\, and I have been teaching\, writing\, and doing research in that area continuously since my first summer job at Massachusetts General Hospital in 1962 and since my joining the UNC Computer Science faculty in 1967. I have had many collaborations across the UNC Medical School\, in some of whose departments I have had adjunct appointments. I led the committee that led to the formation of the Biomedical Research Imaging Center at UNC. 61 PhDs have been produced with me as principal advisor. \nMy early research emphases were on image quality restoration and on 2D and 3D display\, as well as related models of human vision. I helped advise Charles Metz’s dissertation in that area\, and we collaborated for some years\, including a paper that was an early component of the EM algorithm. However\, for the last 4.5 decades my focus has been on geometric models of anatomic objects\nespecially suited for statistics\, statistical analyses of these\, and the applications of that in diagnosis\, treatment planning\, and object segmentation and registration\, with clinical targets all over the body and with many image sources. Major medical targets were radiation oncology\, neuroscience\, and recently colonoscopy. The form of model I have developed is skeletal\, and in particular what we call the evolutionary s-rep\, with its advantages over object boundary based models of also locally capturing interior properties such as object curvature and cross-object width and of providing an object-relative coordinate system important in accessing image intensities as they are used for segmentation and registration. Towards diagnosis and other statistical objectives\, my statistics professor colleague JS Marron and I have made important contributions in methods of statistics of shape that recognize that object geometry cannot be directly analyzed by Euclidean methods because abstractly it resides on a curved manifold. Most especially\, the method called Principal Nested Spheres (CPNS)\, allowing statistical analysis of directional data\, was developed in our laboratory. Collaborations with Kitware\, Inc. in software development and tutorials related to shape analysis in the salt.slicer.org toolkit including s- reps uses\, especially by Jared Vicory\, have been and continue to be important. Successes of a variety of types for statistics on s-reps include the commercial success of the company\, Morphormics\, now part of\nAccuray\, that we spun off and whose main product at that time\, built upon statistics of skeletal models\, focused on segmentation of male pelvic organs from CT for radiation treatment planning. Other work showed registrations and segmentations of mobile structures across medical imaging modalities. Our work over the last decade or so shows that our latest form of s-reps that evolve from ellipsoids while according to rich geometric properties of the object interior and boundary yield notable improvements over boundary point distribution models and models based on smooth deformations of means for object representation in classification\, hypothesis testing\, and production of correspondence across a population of neuroanatomic objects.” URL:https://datascience.ucsd.edu/event/hdsi-seminar-series-stephen-pizer-a-geometric-model-for-anatomic-objects-targeted-for-statistical-analysis/ LOCATION:Computer Science & Engineering Building (CSE)\, Room 1242\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar ATTACH;FMTTYPE=image/png:https://datascience.ucsd.edu/wp-content/uploads/2026/01/steve-p.png END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20260113T140000 DTEND;TZID=America/Los_Angeles:20260113T150000 DTSTAMP:20260531T014033 CREATED:20260107T193320Z LAST-MODIFIED:20260107T225249Z UID:10000535-1768312800-1768316400@datascience.ucsd.edu SUMMARY:HDSI Seminar Series - Ann Kennedy - Latent state inference from neural dynamics and behavior DESCRIPTION:Talk Time: Mon Jan 12th\, 2026 | 2:00pm \nLocation: Data Science Building 1st Floor\, Room 123 \nTalk Abstract\nAs we interact with the world around us\, we experience a constant stream of sensory inputs\, and must generate a constant stream of behavioral actions. What makes brains more than simple input-output machines is their capacity to integrate sensory inputs with our internal motivations and drives to produce behavior that is flexible and adaptive. How can we uncover evidence of this integration from observed neural activity? In this talk\, I will present our lab’s recent work inferring the structure and dynamics of latent motivational states from the observed actions and neural activity of freely behaving mice. We draw on classical normative models of animal behavior from ethology\, showing how these models can help us analyze and interpret experimental data to understand how the circuit architecture of the brain gives rise to the algorithms of survival. \nSpeaker Biography\nAnn Kennedy is a theoretical neuroscientist interested in understanding how the structure of the nervous system gives rise to its function. Originally from Virginia\, they studied Biomedical Engineering at Johns Hopkins and Neuroscience at Columbia University\, where they earned my PhD in the lab of Larry Abbott in the Center for Theoretical Neuroscience. Dr Kennedy pursued postdoctoral training in the lab of David Anderson at Caltech\, and opened their lab at Northwestern University in 2020\, moving to Scripps Research in 2024.\nThe Kennedy Lab studies the underlying neuroscience and brain structures that give rise to fundamental behaviors related to fear\, survival and social interactions. By better understanding the neural activities that guide our decision-making and behavior\, Dr. Kennedy’s work aims to reveal insights about the guiding principles of behavior\, as well as what happens in cases of dysfunctions—for example\, social dysfunction or excessive fear and anxiety. URL:https://datascience.ucsd.edu/event/hdsi-seminar-series-ann-kennedy/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20251208T140000 DTEND;TZID=America/Los_Angeles:20251208T153000 DTSTAMP:20260531T014033 CREATED:20251002T164243Z LAST-MODIFIED:20251208T220555Z UID:10000530-1765202400-1765207800@datascience.ucsd.edu SUMMARY:HDSI Distinguished Seminar - Ila Fiete DESCRIPTION:Speaker: Ila Fiete\nDate & Time: Mon Dec 8th\, 2:00pm\nNEW LOCATION:  Jacobs Hall\, Qualcomm Conference Center B\, 1st Floor \nAbstract: Modular and hierarchical structures are ubiquitous in the brain\, and the decomposition of tasks into invariant subfactors of variation are arguably the basis for robust\, efficient learning\, compositional generalization\, and lack of forgetting. In this talk\, I will describe simple mechanisms for the emergent self-organization of structures for such computation\, arising from local competition and fault-tolerance constraints. I will show how these principles are implemented in brains\, leading to qualitative and topologically robust predictions about brain organization. I will discuss how these biological principles can be abstracted to drive modularity in artificial neural networks\, and show how modular organization can serve as an inductive bias for world structure learning\, including the discovery of vanishingly few modular solutions to modular problems in the space of all possible solutions. \nBio: Ila Fiete is a professor of brain and cognitive sciences\, associate member of the McGovern Institute\, and director of the K. Lisa Yang ICoN Center at MIT. Fiete earned a BS in mathematics and physics at the University of Michigan\, obtaining her PhD in physics at Harvard University in 2004. She conducted her postdoctoral work at the Kavli Institute for Theoretical Physics at the University of California\, Santa Barbara while she was also a visiting member of the Center for Theoretical Biophysics at the University of California\, San Diego. Fiete subsequently spent two years at Caltech as a Broad Fellow in brain circuitry\, then joined the faculty of the University of Texas at Austin before coming to MIT in 2019. URL:https://datascience.ucsd.edu/event/hdsi-seminar-ila-fiete/ LOCATION:Qualcomm Conference Room at JSOE\, Jacobs Hall\, 9736 Engineers Ln\, La Jolla\, San Diego\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20251203T110000 DTEND;TZID=America/Los_Angeles:20251203T123000 DTSTAMP:20260531T014033 CREATED:20251002T164101Z LAST-MODIFIED:20251002T164101Z UID:10000529-1764759600-1764765000@datascience.ucsd.edu SUMMARY:HDSI/TILOS Seminar - Jeremy Schwartz DESCRIPTION:Talk Details TBA URL:https://datascience.ucsd.edu/event/hdsi-tilos-seminar-jeremy-schwartz/ CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20251117T140000 DTEND;TZID=America/Los_Angeles:20251117T153000 DTSTAMP:20260531T014033 CREATED:20251002T163926Z LAST-MODIFIED:20251029T222512Z UID:10000528-1763388000-1763393400@datascience.ucsd.edu SUMMARY:HDSI Distingushed Seminar - Yisong Yue DESCRIPTION:TITLE Design\, Measure\, Interpret: Foundation Models in the Scientific Loop \nABSTRACT “As foundation models become powerful scientific priors\, a central question emerges: how can they drive the full cycle of discovery—from designing experiments to interpreting results? This talk presents a probabilistic framework that unites experiment design and inverse problems under a common foundation model. I will highlight recent progress in adaptive design algorithms and diffusion-based inversion methods\, and discuss how these ideas point toward an AI-driven ecosystem for science.” \nBIO Yisong Yue is a Professor of Computing and Mathematical Sciences at the California Institute of Technology. He was previously a research scientist at Disney Research. Before that\, he was a postdoctoral researcher in the Machine Learning Department and the iLab at Carnegie Mellon University. He received a Ph.D. from Cornell University and a B.S. from the University of Illinois at Urbana-Champaign. Yisong previously served as Senior Program Chair of ICLR 2024 and General Chair of ICLR 2025\, and currently serves on the ICLR board. Yisong’s research interests are centered around machine learning and artificial intelligence\, particularly in getting AI to work in high-stakes and high-expertise domains. To that end\, his research agenda spans both fundamental and applied pursuits\, from novel learning frameworks all the way to deployment in autonomous driving on public roads. His work has been recognized with multiple paper awards and nominations\, including in robotics\, computer vision\, sports analytics\, machine learning for health\, and information retrieval. During his time in industry\, Yisong worked on machine learning approaches to behavior modeling and motion planning for autonomous driving. URL:https://datascience.ucsd.edu/event/hdsi-seminar-yisong-yue/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20251112T110000 DTEND;TZID=America/Los_Angeles:20251112T120000 DTSTAMP:20260531T014033 CREATED:20251106T170549Z LAST-MODIFIED:20251106T170549Z UID:10000534-1762945200-1762948800@datascience.ucsd.edu SUMMARY:TILOS-HDSI Seminar: Adam Oberman - AI Safety Theory: The Missing Middle Ground DESCRIPTION:The next TILOS-HDSI seminar will be Wednesday\, November 12 at 11am PST with Adam Oberman (McGill University). The title is AI Safety Theory: The Missing Middle Ground. \nTalk Information \nSpeaker: Adam Oberman (McGill University) \nDate & Time: Wednesday\, November 12 @ 11am PST \nVenue: HDSI 123 \nAbstract: Over the past few years\, the capabilities of generative artificial intelligence (AI) systems have advanced rapidly. Along with the benefits of AI\, there is also a risk of harm. In order to benefit from AI while mitigating the risks\, we need a grounded theoretical framework. \nThe current AI safety theory\, which predates generative AI\, is insufficient. Most theoretical AI safety results tend to reason absolutely: a system is a system is “aligned” or “mis-aligned”\, “honest” or “dishonest”. But in practice safety is probabilistic\, not absolute. The missing middle ground is a quantitative or relative theory of safety — a way to reason formally about degrees of safety. Such a theory is required for defining safety and harms\, and is essential for technical solutions as well as for making good policy decisions. \nIn this talk I will: \n\nReview current AI risks (from misuse\, from lack of reliability\, and systemic risks to the economy) as well as important future risks (lack of control).\nReview theoretical predictions of bad AI behavior and discuss experiments which demonstrate that they can occur in current LLMs.\nExplain why technical and theoretical safety solutions are valuable\, even by contributors outside of the major labs.\nDiscuss some gaps in the theory and present some open problems which could address the gaps.\n\nBio: Adam Oberman is a Full Professor of Mathematics and Statistics at McGill University\, a Canada CIFAR AI Chair\, and an Associate Member of Mila. He is a research collaborator at LawZero\, Yoshua Bengio’s AI Safety Institute. He has been researching AI safety since 2024. His research spans generative models\, reinforcement learning\, optimization\, calibration\, and robustness. Earlier in his career\, he made significant contributions to optimal transport and nonlinear partial differential equations. He earned degrees from the University of Toronto and the University of Chicago\, and previously held faculty and postdoctoral positions at Simon Fraser University and the University of Texas at Austin. URL:https://datascience.ucsd.edu/event/tilos-hdsi-seminar-adam-oberman-ai-safety-theory-the-missing-middle-ground/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:HDSI Event,Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20251110T140000 DTEND;TZID=America/Los_Angeles:20251110T143000 DTSTAMP:20260531T014033 CREATED:20251002T163809Z LAST-MODIFIED:20251104T223610Z UID:10000527-1762783200-1762785000@datascience.ucsd.edu SUMMARY:HDSI Seminar - Tianhao Wang DESCRIPTION:Speaker: Tianhao Wang\n\nDate & Time: Monday Nov 10th\, 2pm\nLocation: HDSI Multipurpose Room 123\, 1st floor \n\n\nTitle: Adaptive Optimizers: From Structured Preconditioners to Adaptive Geometry\n\nAbstract: Adaptive optimizers such as Adam and Shampoo are workhorses of modern machine learning\, enabling efficient training of large-scale models across architectures and domains. In this talk\, we will present a unified framework for adaptive optimizers with structured preconditioners\, encompassing a variety of existing methods and introducing new ones. Our analysis reveals the fundamental interplay between preconditioner structures and loss geometries\, highlighting in particular that more adaptivity is not always helpful. Furthermore\, the dominance of adaptive methods has recently been challenged by the surprising effectiveness of simpler normalized steepest descent (NSD)–type methods such as Muon\, while a consensus has emerged that both families of methods succeed by exploiting the non-Euclidean geometry of the loss landscape. Building on the proposed framework\, we show that the convergence of adaptive optimizers is governed by a notion of adaptive smoothness\, which contrasts with the standard smoothness assumption leveraged by NSD. In addition\, although adaptive smoothness is a stronger condition\, it enables acceleration via Nesterov momentum\, which cannot be achieved under the standard smoothness assumption in non-Euclidean settings. Finally\, we develop a notion of adaptive gradient variance that parallels adaptive smoothness and yields qualitatively improved guarantees compared to those based on standard gradient variance.\n\n\nSpeaker Bio: Tianhao Wang is an Assistant Professor at the Halıcıoğlu Data Science Institute\, University of California\, San Diego. Prior to UCSD\, he was a Research Assistant Professor at Toyota Technological Institute at Chicago. He received his PhD from the Department of Statistics and Data Science at Yale University in 2024. His research focuses on theoretical foundations at the intersection of deep learning\, optimization\, and statistics.\nMore info is available on Professor Wang’s website: https://tiiao.github.io/ URL:https://datascience.ucsd.edu/event/hdsi-seminar-tianhao-wang/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20251007T140000 DTEND;TZID=America/Los_Angeles:20251007T150000 DTSTAMP:20260531T014033 CREATED:20250929T215834Z LAST-MODIFIED:20250929T215856Z UID:10000524-1759845600-1759849200@datascience.ucsd.edu SUMMARY:HDSI Seminar - Dragan Radulović - A New Paradigm for the Analysis of Large Text Files DESCRIPTION:Speaker:Dragan Radulović\nDate & Time: Tuesday Oct 7th\, 2pm\nLocation: HDSI Multipurpose Room 123\, 1st floor\n\n \nTitle:  A New Paradigm for the Analysis of Large Text Files\n \nTalk Abstract: The problem is as follows: a large text file containing information on thousands of individuals serves as the input. The output is a simple yes-or-no prediction. For example\, the algorithm might receive a new patient’s file and must provide a prognosis—yes or no—for a given disease (or treatment\, or test\, etc.). I have developed a rather unusual (and quite peculiar) method for doing this. The algorithm has been successfully used for several years by an (undisclosed) American professional sports team. It has never been published\, and until recently\, I was not even allowed to share it with anyone. Now that the confidentiality clause in my agreement has expired\, I am free to share it with the world.\n \nSpeaker Bio: Dragan Radulović is a mathematician specializing in probability on Banach spaces\, empirical processes\, and copula functions. Parallel to this more theoretical career\, Dragan also explores applications—particularly in data analysis. He was the principal mathematician at the successful startup Quantiva (Princeton\, 1999–2003)\, where he designed a suite of algorithms tailored to detecting anomalies in internet traffic. From 2002 to 2011\, he worked on problems in molecular biology. In this area\, he was the first author of several high-profile papers (Nature Genetics\, PLOS Biology\, Cancer Informatics). His key innovation was a novel algorithm that analyzes mass spectrometry data to provide protein quantification—something that was not possible at the time.\n\n\n\n\nMore recently\, he worked as a contractor for the Chicago Blackhawks\, a professional hockey team. There he designed a suite of algorithms that processed large numerical and textual datasets collected by scouts and hockey professionals. The output of these algorithms was predictive modeling of players’ future performances. \nDragan Radulović is also an author. His first book\, On the Road Again (2018)\, recounts his road trip through Iran and Afghanistan. His second book\, Why Does Math Work? (Cambridge\, 2023)\, received praise in the Notices of the American Mathematical Society: \n“If you have wondered about the philosophical underpinnings of mathematics\, this book is for you. It contains insightful queries for a mathematician to ponder and could definitely be the start of some enlightening conversations\, perhaps in a departmental book club or seminar course. I found myself enjoying the many tangents (pun intended!) and digressions in this wonderfully unique and well-articulated book.” —Emily J. Olson\, Notices of the American Mathematical Society \nDragan has had stints at Princeton University and Yale University. He later moved to South Florida\, where he surfs\, writes\, and does mathematics. URL:https://datascience.ucsd.edu/event/hdsi-seminar-dragan-radulovic-a-new-paradigm-for-the-analysis-of-large-text-files/ LOCATION:Halıcıoğlu Data Science Institute Room 123\, 3234 Matthews Ln\, La Jolla\, CA 92093\, USA CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20250929T140000 DTEND;TZID=America/Los_Angeles:20250929T153000 DTSTAMP:20260531T014033 CREATED:20250918T200519Z LAST-MODIFIED:20250918T205524Z UID:10000523-1759154400-1759159800@datascience.ucsd.edu SUMMARY:HDSI Seminar - Regina Liu - Fusion Learning: Fusing Inferences from Diverse Data Sources DESCRIPTION:HDSI will be hosting its first Seminar Series speaker of the academic year at the end of this month. Regina Liu (Rutgers University) will be giving a talk Monday Sept 29th at 2pm in the HDSI Multipurpose room\, 1st floor Room 123. \n\nSpeaker: Regina Liu\nDate & Time: Monday Sept 29th\, 2pm\nLocation: HDSI Multipurpose Room 123\, 1st floor \n\nTalk Title: Fusion Learning: Fusing Inferences from Diverse Data Sources\n\nAbstract: \nAdvanced data acquisition technology has greatly increased the accessibility of complex inferences\, based on summary statistics or sample data\, from diverse data sources. Fusion learning refers to combining complex inferences from multiple sources to yield a more effective overall. We focus on the tasks: 1) Whether/When to combine inferences? 2) How to combine inferences efficiently? 3) How to combine inferences to enhance an individual study\, thus named i-Fusion?\n\nWe present a general framework for nonparametric and efficient fusion learning. The main tool underlying this framework is the new notion of depth confidence distribution (depth-CD)\, developed by combining data depth\, bootstrap and confidence distributions. We show that a depth-CD is an omnibus form of confidence regions\, whose contours of level sets shrink toward the true parameter value\, and thus an all-encompassing inferential tool. The approach is efficient\, general and robust\, and readily applies to heterogeneous studies covering a broad range of complex settings. The approach is demonstrated with an aviation safety analysis application in tracking aircraft landing performance and a zero-event studies in clinical trials with non- estimable parameters. \n\nKey words: confidence distribution\, data depth\, fusion learning\, heterogeneous studies\n\nSpeaker Bio:\nRegina Liu is Distinguished Professor\, Rutgers University. Her research areas include data depth\, resampling\, nonparametric statistics\, confidence distribution\, and fusion learning. Aside from theoretical and methodological research\, she has long collaborated with the FAA on aviation safety research projects on process control\, text mining and risk management. \nShe is an elected fellow of the Institute of Mathematical Statistics (IMS) and the American Statistical Association (ASA). She is the recipient of 2021 Noether Distinguished Scholar Award (ASA)\, 2024 Elizabeth Scott Award (Committee of Presidents of Statistical Societies (COPSS))\, and the IMS 2025 Neyman Award & Lecture. She has served as Co-Editor for the Journal of the American Statistical Association and as Associate Editor for several journals. She was elected President of the Institute of Mathematical Statistics (IMS)\, 2020-2021. URL:https://datascience.ucsd.edu/event/hdsi-seminar-reginia-liu-fusion-learning-fusing-inferences-from-diverse-data-sources/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20250528T130000 DTEND;TZID=America/Los_Angeles:20250528T143000 DTSTAMP:20260531T014033 CREATED:20250519T164145Z LAST-MODIFIED:20250519T164145Z UID:10000521-1748437200-1748442600@datascience.ucsd.edu SUMMARY:Shuang Hao - Empowering and Strengthening Security in the AI Era DESCRIPTION:When Wednesday May 28th\, 1pm\nWhere: HDSI 1st Floor Multipurpose Room 123 \n\nTitle: Empowering and Strengthening Security in the AI Era\n \nAbstract: Revolutionary advances in artificial intelligence (AI) techniques have led to promising applications and widespread deployment accessible to users. However\, AI techniques are increasingly being abused by cybercriminals\, such as creating synthetic content for scams or injecting malicious instances into services. It is imperative to cultivate systematic analysis and defenses against security threats in the era of AI.\nIn this talk\, I will describe my research on developing empirical-theoretical approaches to address AI abuses and attacks. First\, I will introduce the approaches of leveraging user intelligence to characterize and detect AI-generated face images\, enabling human-AI collaboration to strengthen security of generative AI. Second\, I will describe the analysis of attacks exploiting machine unlearning in the AI ecosystem\, and quantify model degradation and risks in unlearning scenarios. My research builds systematic approaches and principled solutions to advance AI security. \n \nBio: Shuang Hao is an Associate Professor of Computer Science at the University of Texas at Dallas. He obtained his Ph.D. from the Georgia Institute of Technology\, and he was a postdoctoral scholar at the University of California\, Santa Barbara before joining UT Dallas. His research interests are in security and its intersection with AI\, data science\, and user behavior analysis. His current research focuses on designing data-driven approaches to advance security in the AI ecosystem. He has published extensively in top-tier security conferences including S&P\, USENIX Security\, CCS\, and NDSS. He has received multiple awards and recognitions\, including an NSF CAREER Award\, an IETF Applied Networking Research Prize\, a DSN Best Paper Award\, an IMC Best Paper Award Runner-up\, two-time CSAW Best Security Paper Award Finalist\, and a Yahoo! Key Scientific Challenges Program Award. His work has been featured in media outlets such as MIT Technology Review\, Slashdot\, Fortune\, CNN\, and The Wall Street Journal. More about his research can be found at https://www.utdallas.edu/~shao/ URL:https://datascience.ucsd.edu/event/shuang-hao-empowering-and-strengthening-security-in-the-ai-era/ LOCATION:Halıcıoğlu Data Science Institute Room 123\, 3234 Matthews Ln\, La Jolla\, CA 92093\, USA CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20250423T130000 DTEND;TZID=America/Los_Angeles:20250423T143000 DTSTAMP:20260531T014033 CREATED:20250404T162241Z LAST-MODIFIED:20250404T162241Z UID:10000517-1745413200-1745418600@datascience.ucsd.edu SUMMARY:LIVED EXPERIENCE RESEARCH SUMMIT DESCRIPTION:Location: HDSI MPR \nTo watch via zoom please contact: hdsiassistant@ucsd.edu \nFormerly Incarcerated professor speaking on his lived experience in research. Researchers from\nSmarr Lab and MOSAIC lab. \nNoel Vest\, PhD\, is an Assistant Professor at the Boston University School of Public Health. His research interests include mental health\, substance use disorders\, and addiction recovery. As a formerly incarcerated scholar and a person in long-term recovery\, Dr. Vest is an advocate for social justice issues and public policy concerning substance use disorder recovery and prison reentry. He completed his PhD in Experimental Psychology from Washington State University and did his postdoctoral fellowship at Stanford University. \n  URL:https://datascience.ucsd.edu/event/lived-experience-research-summit/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Guest Lecture,Seminar ATTACH;FMTTYPE=image/jpeg:https://datascience.ucsd.edu/wp-content/uploads/2025/04/TUS_HS_HDSI_Collaboration_V4_Flyer-scaled.jpg END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20250411T140000 DTEND;TZID=America/Los_Angeles:20250411T150000 DTSTAMP:20260531T014033 CREATED:20250408T201744Z LAST-MODIFIED:20250408T201744Z UID:10000518-1744380000-1744383600@datascience.ucsd.edu SUMMARY: DESCRIPTION:Seminar Information\n\nSeminar Date\nApril 11\, 2025 – 2:00 PM\n\n\n\nLocation\nThe FUNG Auditorium – PFBH\n\n  \n\n\n\n\n\n\n\n\n\n  \n\nAbstract\n\nPersonal and population health applications built on top of large-scale mobile sensor data and computing platforms have a great potential to impact the way we diagnose diseases\, track\, and manage our health. However\, the existing sensing mechanisms often fail to accurately capture and infer syndromic signatures that are indicative of anomalies in internal physiological and behavioral processes at an earlier stage. A mobile sensing system that can harness early syndromic signals at an individual or a community level can pave the way to effective just-in-moment intervention\, early screening\, and prevention. \nIn this talk\, I will present our recent and ongoing research to demonstrate how physiological time series data harnessed from on-body wearable systems can be used for modeling opioid use/administration\, affective states including craving\, pain\, stress and euphoria\, and opioid misuse.  I will talk about different approaches (attention based approach and Large Language Model based approach) to fuse multimodal physiological biomarker data\, behavioral data\, clinical health record data\, demographic data as well as symptom data. I will highlight how integration of pharmacological knowledge such as Pharmacokinetics of a specific substance can help neural networks to better generalize and learn opioid related physiological events better than a purely data driven approach. \nChronic opioid use induces neuroplastic changes in brain circuits\, causing predictable changes in different states including heightened stress\, increased pain\, and intense cravings. The fluctuations or changes in stress\, pain and craving states carries telltale signal for opioid misuse risk. In the last part of the talk\, I will present how the heart rate variability data from a wearable wristband can be used to predict momentary pain\, stress and craving state trajectories with a personalized hierarchical deep learning model\, obviating the need for obtrusive ecological momentary assessments throughout the day. We adopt a nonlinear dynamical systems approach with different features including persistence entropy to extract subtle trends from the moment-by-moment fluctuations or changes in pain\, stress and craving states. Our analysis reveals a hidden counter-intuitive association between high entropy or lack of predictability (i.e.\, chaos) in the momentary pain\, stress and craving states with the decrease in opioid misuse risk. Leveraging Chaos Theory\, the entropy-based nonlinear dynamical features can be used to train a deep learning based approach for accurate opioid misuse risk assessment. \n\n\n\nSpeaker Bio\n\nTauhidur Rahman is an Assistant Professor in the Halıcıoğlu Data Science Institute and Computer Science and Engineering at the University of California San Diego where he directs the Mobile Sensing and Ubiquitous Computing Laboratory (MOSAIC Lab). His current research focuses on building novel ubiquitous and mobile health sensing technologies that capture observable low-level physical signals in the form of an acoustic and electromagnetic wave from our bodies and surrounding environments and map them to relevant biological and behavioral measurements. Some of his notable accomplishments include a Google Research Scholar Award 2023\, a Google Ph.D. fellowship in 2016 in mobile computing\, a finalist position in Qualcomm innovation fellowship in 2015\, Outstanding Teaching Award 2015 from Cornell University\, one best paper award in ACM Digital Health 2016\, one best paper honorable mention award in ACM Ubicomp 2015 and a distinguished paper award from ACM IMWUT in 2021. Tauhidur received his B.S. in Electrical and Electronic Engineering from the Bangladesh University of Engineering and Technology\, his M.S. in Electrical Engineering from the University of Texas at Dallas and PhD in Information Science from Cornell University. He has a long track-record working with large-scale multi-modal and multi-rate sensor data\, especially in the application areas of digital epidemiology\, substance use disorder\, mental health and sleep. His work has been featured in several US-based and International media outlets including Wall Street Journal\, MIT Technology Review\, NewScientist\, Public Television for Western New England\, Daily Mail (UK) and Hindustan Times (India). His laboratory has been funded by NSF\, NIH\, DARPA and industry grants. \n  URL:https://datascience.ucsd.edu/event/34697/ LOCATION:Powell-Focht Bioengineering Hall (PFBH)\, FUNG Auditorium CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20250407T130000 DTEND;TZID=America/Los_Angeles:20250407T150000 DTSTAMP:20260531T014033 CREATED:20250401T200812Z LAST-MODIFIED:20250401T200812Z UID:10000516-1744030800-1744038000@datascience.ucsd.edu SUMMARY:HDSI Seminar - Xiaofei Shi- Continuous-time Reinforcement Learning with Forward-Backward Stochastic Differential Equations DESCRIPTION:When Monday April 7th 1:00pm\nWhere: HDSI 1st Floor Multipurpose Room 123\nTitle: Continuous-time Reinforcement Learning with Forward-Backward Stochastic Differential Equations \nAbstract:\nIn this talk we introduce a mathematical formulation of reinforcement learning problem with a system of forward-backward stochastic differential equations (FBSDEs). With the Deep FBSDE Solver proposed by Han\, Jentzen\, and E (2018)\, deep architecture for FBSDE systems shows great success in continuous-time stochastic control problems. In our work\, we show how to further leverage the FBSDE formulation to solve traditionally intractable equilibrium problems in finance. We present a general computational framework for solving continuous-time financial market equilibria under minimal modeling assumptions while incorporating realistic financial frictions\, such as trading costs\, and supporting multiple interacting agents. Inspired by generative adversarial networks (GANs)\, our approach employs a novel generative deep reinforcement learning framework with a decoupling feedback system embedded in the adversarial training loop\, which we term as the reinforcement link. This architecture stabilizes the generator by integrating the information from the discriminator. Our theoretically guided feedback mechanism enables the decoupling of the equilibrium system\, overcoming challenges that hinder conventional numerical algorithms. \nBio: Professor Xiaofei Shi is an Assistant Professor in the Department of Statistical Sciences at the University of Toronto. Before joining U of T\, they worked as a Term Assistant Professor at Columbia University. Professor Shi obtained their PhD in Mathematical Finance at Carnegie Mellon University\, under the supervision of Prof. Johannes Muhle-Karbe. They are mainly interested in stochastic optimization and stochastic differential equations with applications to mathematical finance and have also worked on various topics in data science\, including crowdsourcing\, dimensionality reduction\, and sparse recovery. URL:https://datascience.ucsd.edu/event/hdsi-seminar-xiaofei-shi-continuous-time-reinforcement-learning-with-forward-backward-stochastic-differential-equations/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20250331T110000 DTEND;TZID=America/Los_Angeles:20250331T120000 DTSTAMP:20260531T014033 CREATED:20250318T195914Z LAST-MODIFIED:20250318T195914Z UID:10000513-1743418800-1743422400@datascience.ucsd.edu SUMMARY:Seminar - Jeremy Bernstein - Metrized Deep Learning DESCRIPTION:Jeremy Bernstein\n\nMIT CSAIL\n \n\n\nMonday\, March 31\n11:00 AM – 12:00 PM (PST) \nCSE 1242\n\nTitle: Metrized Deep Learning\n\n\nAbstract:\nWe build neural networks in a modular and programmatic way using software libraries like PyTorch and JAX. But optimization theory has not caught up to the flexibility of this paradigm\, and practical advances in neural net optimization are largely driven by heuristics. In this talk\, I will argue that to treat deep learning rigorously\, we must build our optimization theory programmatically and in lockstep with the neural network itself. To instantiate this idea we propose the “modular norm”\, which is a norm on the weight space of general neural architectures. The modular norm is constructed by stitching together norms on individual tensor spaces as the architecture is constructed. The modular norm has several applications: automatic Lipschitz certificates for general architectures in both weights and inputs; automatic learning rate transfer across scale; and most recently we built the duality theory for the modular norm\, leading to fast optimizers like “Muon”\, which set speed records for training transformers. We are building the theory of the modular norm into a software library called Modula to ease the development and deployment of metrized deep learning algorithms—you can find out more at https://modula.systems/.\n\n\n\nBiosketch:\n\nJeremy Bernstein is a postdoc in CSAIL at MIT advised by Phillip Isola. His goal is to uncover the computational and statistical laws of natural and artificial intelligence\, and thereby design learning systems that are more efficient\, more automatic and more useful in practice. He has a PhD in Computation & Neural Systems from Caltech and Bachelor’s and Master’s degrees in Physics from the University of Cambridge. He was a recipient of the NVIDIA graduate fellowship. URL:https://datascience.ucsd.edu/event/seminar-jeremy-bernstein-metrized-deep-learning/ LOCATION:Computer Science & Engineering Building (CSE)\, Room 1242\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Guest Lecture,Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20250317T110000 DTEND;TZID=America/Los_Angeles:20250317T120000 DTSTAMP:20260531T014033 CREATED:20250313T162327Z LAST-MODIFIED:20250313T162327Z UID:10000512-1742209200-1742212800@datascience.ucsd.edu SUMMARY:Seminar: Deep Learning Theory in the Age of Generative AI - Sadhika Malladi DESCRIPTION:Monday\, March 17\n11:00 AM – 12:00 PM (PST) \nCSE 1242  \nTitle: Deep Learning Theory in the Age of Generative AI \nAbstract:\nModern deep learning has achieved remarkable results\, but the design of training methodologies largely relies on guess-and-check approaches. Thorough empirical studies of recent massive language models (LMs) is prohibitively expensive\, underscoring the need for theoretical insights\, but classical ML theory struggles to describe modern training paradigms. I present a novel approach to developing prescriptive theoretical results that can directly translate to improved training methodologies for LMs. My research has yielded actionable improvements in model training across the LM development pipeline — for example\, my theory motivates the design of MeZO\, a fine-tuning algorithm that reduces memory usage by up to 12x and halves the number of GPU-hours required. Throughout the talk\, to underscore the prescriptiveness of my theoretical insights\, I will demonstrate the success of these theory-motivated algorithms on novel empirical settings published after the theory. \nBiosketch:\n\nSadhika Malladi is a final-year PhD student in Computer Science at Princeton University advised by Sanjeev Arora. Her research advances deep learning theory to capture modern-day training settings\, yielding practical training improvements and meaningful insights into model behavior. She has co-organized multiple workshops\, including Mathematical and Empirical Understanding of Foundation Models at ICLR 2024 and Mathematics for Modern Machine Learning (M3L) at NeurIPS 2024. She was named a 2025 Siebel Scholar. URL:https://datascience.ucsd.edu/event/seminar-deep-learning-theory-in-the-age-of-generative-ai-sadhika-malladi/ LOCATION:CSE 1242 CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20250221T140000 DTEND;TZID=America/Los_Angeles:20250221T153000 DTSTAMP:20260531T014033 CREATED:20250219T201049Z LAST-MODIFIED:20250219T201049Z UID:10000509-1740146400-1740151800@datascience.ucsd.edu SUMMARY:HDSI Seminar - Victor Minces - The Sound of Data DESCRIPTION:When Friday\, February 21st\nWhere: HDSI MPR 123\n\n\n\n\n\n\n\n\nTitle: The Sound of Data\n\nSpeaker: Victor Minces\n\nAbstract: In this talk\, Dr. Minces will give an overview of his career and how it led to the development of Listening to Waves\, a program that creates playful activities and web applications that connect music with science through data visualization and sonification. He will demonstrate how to use the applications created by his team to create surprising sounds and how the applications can help people understand the science of waves\, signal processing\, and music. He will discuss the impact of his program on children’s attitudes toward science and the education system. Further\, he will demonstrate new projects for sonifying data\, such as ‘the talking hand\,’ an application transforming hand movements into phonemes.\n\nBio: Dr. Minces is a neuroscientist of music\, sound artist\, performer\, and developer of educational programs centered on the STEM of music. He studied fine arts and physics at the University of Buenos Aires and obtained his Ph.D. in Computational Neurobiology at the University of California\, San Diego\, in Andrea Chiba’s laboratory. He is now a research scientist in the Department of Cognitive Science. He has studied how large neural networks in the brain encode sensory information and how the brain processes musical rhythm. He has created Listening to Waves\, a widely adopted program that develops web applications and activities for people to learn about the science of sound through playful exploration. URL:https://datascience.ucsd.edu/event/hdsi-seminar-victor-minces-the-sound-of-data/ LOCATION:Halıcıoğlu Data Science Institute Room 123\, 3234 Matthews Ln\, La Jolla\, CA 92093\, USA CATEGORIES:Guest Lecture,Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20250205T140000 DTEND;TZID=America/Los_Angeles:20250205T153000 DTSTAMP:20260531T014033 CREATED:20250130T190217Z LAST-MODIFIED:20250130T190217Z UID:10000508-1738764000-1738769400@datascience.ucsd.edu SUMMARY:HDSI Seminar - Hongzhe Li DESCRIPTION:When Wednesday Feb 5th 2:00pm\nWhere: Computer Science & Engineering (CSE) 1st floor\, Seminar Room 1242 \nTitle: Fréchet Regression of Random Objects on Vector Covariates and Its applications for Single Cell RNA-seq Data Analysis \nAbstract: \nPopulation-level single-cell RNA-seq data captures gene expression profiles across thousands of cells from each individual in a sizable cohort. This data facilitates the construction of cell-type- and individual-specific gene co-expression networks by estimating covariance matrices. Investigating how these co-expression networks relate to individual-level covariates provides critical insights into the interplay between molecular processes and biological or clinical traits. This talk introduces Fréchet regression\, modeling covariance matrices as outcomes and vector covariates as predictors\, using the Wasserstein distance between covariance matrices as a metric instead of the Euclidean distance. A test statistic is proposed based on the Fréchet mean and covariate-weighted Fréchet mean\, with its asymptotic null distribution derived. Analysis of large-scale single-cell RNA-seq data reveals an association between the co-expression network of genes in the nutrient-sensing pathway and age\, highlighting perturbations in gene co-expression networks with aging. \nAdditionally\, a robust local Fréchet regression approach\, leveraging neural unbalanced optimal transport\, is briefly discussed to explore how cells are temporally organized during the differentiation of human embryonic stem cells into embryoid bodies. \nBio: Bio: Hongzhe Li (Lee) is Perelman Professor of Biostatistics\, Epidemiology and Informatics and Vice Chair of Research Integration at the Perelman School of Medicine at the University of Pennsylvania (Penn). He is also Director of Center for Statistics in Biomedical Big Data and a faculty member in the graduate groups of Genomics and Computational Biology and Computational and Applied Mathematics at Penn. Dr Li also has a secondary appointment in the Department of Statistics at the Wharton School. His research has been focused on developing powerful statistical and computational methods for analysis of large-scale genetic\, genomics and metagenomics data. URL:https://datascience.ucsd.edu/event/hdsi-seminar-hongzhe-li/ LOCATION:Computer Science & Engineering Building (CSE)\, Room 1242\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20241106T140000 DTEND;TZID=America/Los_Angeles:20241106T150000 DTSTAMP:20260531T014033 CREATED:20241112T204134Z LAST-MODIFIED:20241112T204134Z UID:10000506-1730901600-1730905200@datascience.ucsd.edu SUMMARY:Revisiting Scalarization in Multi-Task Learning | Prof. Han Zhao DESCRIPTION:Title: Revisiting Scalarization in Multi-Task Learning \nAbstract: Linear scalarization\, i.e.\, combining all loss functions by a weighted sum\, has been the default choice in the literature of multi-task learning (MTL) since its inception. In recent years\, there has been a surge of interest in developing Specialized Multi-Task Optimizers (SMTOs) that treat MTL as a multi-objective optimization problem. However\, it remains open whether there is a fundamental advantage of SMTOs over scalarization. In fact\, heated debates exist in the community comparing these two types of algorithms\, mostly from an empirical perspective. In this talk\, I will revisit scalarization from a theoretical perspective. I will be focusing on linear MTL models and studying whether scalarization is capable of fully exploring the Pareto front. Our findings reveal that\, in contrast to recent works that claimed empirical advantages of scalarization\, scalarization is inherently incapable of full exploration\, especially for those Pareto optimal solutions that strike the balanced trade-offs between multiple tasks. More concretely\, when the model is under-parametrized\, we reveal a multi-surface structure of the feasible region and identify necessary and sufficient conditions for full exploration. This leads to the conclusion that scalarization is in general incapable of tracing out the Pareto front. Our theoretical results provide a more intuitive explanation of why scalarization fails beyond non-convexity. I will conclude the talk by briefly discussing the extension of our results to general nonlinear neural networks.\nBio: Dr. Han Zhao is an Assistant Professor of Computer Science and\, by courtesy\, of Electric and Computer Engineering at the University of Illinois Urbana-Champaign (UIUC). He is also an Amazon Visiting Academic at Amazon AI. Dr. Zhao earned his Ph.D. degree in machine learning from Carnegie Mellon University. His research interest is centered around trustworthy machine learning\, with a focus on algorithmic fairness\, robust generalization under distribution shifts and model interpretability. He has been named a Kavli Fellow of the National Academy of Sciences and has been selected for the AAAI New Faculty Highlights program. His research has been recognized through a Google Research Scholar Award\, an Amazon Research Award\, and a Meta Research Award. URL:https://datascience.ucsd.edu/event/revisiting-scalarization-in-multi-task-learning-prof-han-zhao/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20241030T130000 DTEND;TZID=America/Los_Angeles:20241030T143000 DTSTAMP:20260531T014033 CREATED:20241029T173302Z LAST-MODIFIED:20241029T173302Z UID:10000504-1730293200-1730298600@datascience.ucsd.edu SUMMARY:HDSI Seminar - Maksim Kitsak -Modeling and Inference of Complementarity Mechanisms in Networks. DESCRIPTION:Talk Information:\nWhen Wednesday Oct 30th 1:00pm\nWhere: HDSI MPR 123\nZoom Info: http://bit.ly/HDSI-Seminars \nTitle: Modeling and Inference of Complementarity Mechanisms in Networks. \nAbstract: “In many networks\, including networks of protein-protein interactions\, interdisciplinary collaboration networks\, and semantic networks\, connections are established between nodes with complementary rather than similar properties. What is complementarity?\nThe Oxford Dictionary asserts that “”two people or things that are complementary are different but together form a useful or attractive combination of skills\, qualities or physical features.”” Sadly\, our understanding of complementarity in networks does not\ngo far beyond definition. While complementarity is abundant in networks\, we lack mathematical intuition and quantitative methods to study complementarity mechanisms in these systems. Instead\, we routinely retreat to using available off-the-shelf methods developed in the first place for similarity-driven networks. \nIn my talk\, I will discuss my group’s recent achievements in the analysis of complementarity mechanisms in networks. I will first explain why existing similarity-based inference and learning methods are not readily applicable to systems where complementarity between interacting nodes plays a significant role. I will then deduce\, starting with the definition by the Oxford Dictionary\, a general complementarity framework for networks capable of describing any matching relations and containing both similarity and antitheses relations as special cases. Using the general framework\, I will formulate a minimal null model to learn complementarity embeddings of real networks via maximum-likelihood estimation. I will demonstrate how complementarity embeddings can be used to infer both complementary and similar nodes in a network\, enabling network inference tasks\, such as link prediction and community detection. I will conclude my talk with an outlook on the interplay of similarity and complementarity in the formation of networks\, arguing for a careful re-evaluation of existing similarity-inspired methods.” \nBio: “Maksim Kitsak is an Associate Professor of the Electrical Engineering\, Mathematics\, and Computer Science faculty of the Delft University of Technology\, the Netherlands. Prof. Kitsak has been working at the intersection of Network Theory\, Machine Learning\, and Statistical Physics. Prof. Kitsak is particularly interested in the fundamental principles behind non-Euclidean network embeddings and novel applications of network embeddings in communication and biological networks. His research is often published in prestigious journals\, such as Nature and Science Families. Prof. Kitsak gratefully acknowledges the financial support of the National Science Foundation (NSF\, USA)\, Army Research Office (ARO\, USA)\, and the Dutch Research Council (NWO\, NL).” URL:https://datascience.ucsd.edu/event/hdsi-seminar-maksim-kitsak-modeling-and-inference-of-complementarity-mechanisms-in-networks/ LOCATION:Halıcıoğlu Data Science Institute\, 3234 Matthews Ln\, La Jolla\, CA 92093\, USA Room 123 CATEGORIES:Seminar ATTACH;FMTTYPE=image/png:https://datascience.ucsd.edu/wp-content/uploads/2024/01/cropped-HDSI-UCSD-Image-e1712856546428.png END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20241024T100000 DTEND;TZID=America/Los_Angeles:20241024T110000 DTSTAMP:20260531T014033 CREATED:20241023T181207Z LAST-MODIFIED:20241023T182044Z UID:10000503-1729764000-1729767600@datascience.ucsd.edu SUMMARY:Deep Learning: a Non-parametric Statistical Viewpoint DESCRIPTION:ABSTRACT \nThe advent of deep learning has completely revolutionized how we perceive data to obtain superhuman performance across all fields of modern science. However\, despite the remarkable empirical successes of deep learners\, the theoretical guarantees for their statistical accuracy remain rather pessimistic. In particular\, the data distributions on which deep learners are generally applied\, such as natural images\, are often hypothesized to have an intrinsic low-dimensional structure in a typically high-dimensional feature space. However\, this is often not reflected in the derived rates in the state-of-the-art analyses. This talk aims to bridge the gap between the theory and practice of deep learning from a statistical perspective. We demonstrate that deep learners exhibit a convergence rate determined solely by the intrinsic dimensionality of the data\, rather than its nominal high-dimensional feature representation. Our work not only provides practical guidelines for selecting suitable network architectures but also connects the theoretical analyses of these models to established convergence rates in optimal transport and non-parametric statistics literature. In particular\, we derive the sharpest convergence rates for various learning scenarios\, including Generative Adversarial Networks (GANs)\, Wasserstein Autoencoders (WAEs)\, federated learning\, Bi-directional GANs\, and general deep supervised learners. Furthermore\, we introduce a novel measure\, called the entropic dimension\, to characterize the intrinsic dimension of probability measures and achieve the sharpest known approximation results for neural networks employing Rectified Linear Unit (ReLU) activation\, improving upon classical benchmarks. \nBIOGRAPHY \nSaptarshi Chakraborty is a fifth-year Ph.D. student in Statistics at the University of California\, Berkeley\, advised by Prof. Peter Bartlett. Prior to joining Berkeley\, he earned his M.Stat and B. Stat (Hons.) degrees in Statistics from the Indian Statistical Institute (ISI)\, Kolkata\, India. He is primarily interested in the theoretical and methodological foundations of machine learning\, especially\, deep learning theory\, unsupervised learning\, dimensionality reduction\, optimal transport\, and optimization. \nZOOM LINK: https://ucsd.zoom.us/j/93363424503 URL:https://datascience.ucsd.edu/event/deep-learning-a-non-parametric-statistical-viewpoint/ LOCATION:Atkinson Hall\, Fourth Floor CATEGORIES:Seminar ATTACH;FMTTYPE=image/jpeg:https://datascience.ucsd.edu/wp-content/uploads/2024/10/Saptarshi-Chakraborty-EnCORE-Flyer.jpg END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20241021T130000 DTEND;TZID=America/Los_Angeles:20241021T140000 DTSTAMP:20260531T014033 CREATED:20241029T173423Z LAST-MODIFIED:20241029T173423Z UID:10000502-1729515600-1729519200@datascience.ucsd.edu SUMMARY:HDSI Seminar - Generative Social Choice | Ariel Procaccia DESCRIPTION:Talk Information: \nWhen Monday Oct 21st 1:00pm\nWhere: HDSI MPR 123\nZoom Info: http://bit.ly/HDSI-Seminars \nTitle: Generative Social Choice \nAbstract: “The mathematical study of voting\, social choice theory\, has traditionally only been applicable to choices among a few predetermined alternatives\, but not to open-ended decisions such as collectively selecting a textual statement. This limitation is addressed by generative social choice\, a design methodology for open-ended democratic processes that combines the rigor of social choice theory with the capability of large language models to generate text and extrapolate preferences. I’ll introduce a framework that divides the design of AI-augmented democratic processes into two components: first\, proving that the process satisfies representation guarantees when given access to oracle queries; second\, empirically validating that these queries can be approximately implemented using a large language model. I’ll also discuss the application of this framework to the problem of summarizing free-form opinions into a proportionally representative slate of opinion statements. By providing rigorous guarantees\, generative social choice could alleviate concerns about AI-driven democratic innovation and help unlock its potential. \nBio: Ariel Procaccia is Gordon McKay Professor of Computer Science at Harvard University. He works on a broad and dynamic set of problems related to AI\, algorithms\, economics\, and society. He has helped create systems and platforms that are widely used to solve everyday fair division problems\, resettle refugees and select citizens’ assemblies. To make his research accessible to the public\, he regularly writes opinion and exposition pieces for publications such as the Washington Post\, Bloomberg\, Wired and Scientific American. He is a AAAI Fellow (2024) and a recipient of the ACM SIGecom Mid-Career Award (2024)\, Social Choice and Welfare Prize (2020)\, Guggenheim Fellowship (2018)\, IJCAI Computers and Thought Award (2015) and Sloan Research Fellowship (2015). URL:https://datascience.ucsd.edu/event/hdsi-seminar-generative-social-choice-ariel-procaccia/ LOCATION:Halıcıoğlu Data Science Institute Room 123\, 3234 Matthews Ln\, La Jolla\, CA 92093\, USA CATEGORIES:Seminar END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20241015T123000 DTEND;TZID=America/Los_Angeles:20241015T140000 DTSTAMP:20260531T014033 CREATED:20241008T002133Z LAST-MODIFIED:20241008T003755Z UID:10000500-1728995400-1729000800@datascience.ucsd.edu SUMMARY:MathWorks Technical Seminar: AI & Machine Learning in Real-World Systems DESCRIPTION:Beyond their use in traditional data analytics problems\, AI and Machine Learning techniques are changing the way real-world complex systems (like vehicles\, airplanes\, and even industrial production lines) are designed\, tested and fabricated. When building these systems\, engineers rely heavily on modeling and computer-assisted simulations. This seminar showcases how MATLAB and Simulink can help integrate AI models into the component and system level simulation stages of the design process\, and how they can help deploy the resulting control algorithms into the real-world systems. We will use an example where deep-learning and machine learning can be used to create a Virtual Sensor algorithm. Join us for a complementary seminar to see how this is done! Highlights include: \n\nDesigning and training machine learning components with Statistics and Machine Learning Toolbox\nDesigning and training deep learning components with Deep Learning Toolbox\nImporting trained TensorFlow models into MATLAB\nIntegrating machine learning and deep learning models into Simulink for system-level simulation\nGenerating library-free C code and performing PIL tests\nLSTM Model Compression using projection\n\nThis is the first in a 3 part seminar series in partnership with HDSI\, JSOE\, and SIO for this academic year. \nPlease note that this is open to all faculty\, staff\, and students (all grade & majors). \nRegister Here \n*Lunch is included in registration. URL:https://datascience.ucsd.edu/event/mathworks-technical-seminar-ai-machine-learning-in-real-world-systems/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar ATTACH;FMTTYPE=image/png:https://datascience.ucsd.edu/wp-content/uploads/2024/10/10.2024-Mathworks-UCSD-Technical-Seminar-Series.png END:VEVENT BEGIN:VEVENT DTSTART;TZID=America/Los_Angeles:20240724T100000 DTEND;TZID=America/Los_Angeles:20240724T110000 DTSTAMP:20260531T014033 CREATED:20240717T171141Z LAST-MODIFIED:20240717T171141Z UID:10000489-1721815200-1721818800@datascience.ucsd.edu SUMMARY:HDSI/TILOS Seminar | Rob Nowak | What Kinds of Functions do Neural Networks Learn? Theory and Practical Applications DESCRIPTION:When Wednesday July 24th 10:00am *Updated\nWhere: HDSI 123 * Updated\nZoom Info: https://ucsd.zoom.us/j/99334315002 *Updated \nTitle: What Kinds of Functions do Neural Networks Learn?  Theory and Practical Applications \nAbstract:  This talk presents a theory characterizing the types of functions neural networks learn from data. Specifically\, the function space generated by deep ReLU networks consists of compositions of functions from the Banach space of second-order bounded variation in the Radon transform domain. This Banach space includes functions with smooth projections in most directions. A representer theorem associated with this space demonstrates that finite-width neural networks suffice for fitting finite datasets. The theory has several practical applications. First\, it provides a simple and theoretically grounded method for network compression. Second\, it shows that multi-task training can yield significantly different solutions compared to single-task training\, and that multi-task solutions can be related to kernel ridge regressions. Third\, the theory has implications for improving implicit neural representations\, where multi-layer neural networks are used to represent continuous signals\, images\, or 3D scenes. This exploration bridges theoretical insights with practical advancements\, offering a new perspective on neural network capabilities and future research directions.\nBio: Robert Nowak is the Grace Wahba Professor of Data Science and Keith and Jane Nosbusch Professor in Electrical and Computer Engineering at the University of Wisconsin-Madison. His research focuses on machine learning\, optimization\, and signal processing. He serves on the editorial boards of the SIAM Journal on the Mathematics of Data Science and the IEEE Journal on Selected Areas in Information Theory.meeting with him. If… URL:https://datascience.ucsd.edu/event/hdsi-tilos-seminar-rob-nowak-what-kinds-of-functions-do-neural-networks-learn-theory-and-practical-applications/ LOCATION:Halıcıoğlu Data Science Institute (HDSI)\, Room 123\, 3234 Matthews Ln\, La Jolla\, CA\, 92093\, United States CATEGORIES:Seminar ATTACH;FMTTYPE=image/png:https://datascience.ucsd.edu/wp-content/uploads/2023/10/TILOS-Square_HDSI-Website-e1712854679822.png END:VEVENT END:VCALENDAR