BEGIN:VCALENDAR VERSION:2.0 PRODID:-//UC Irvine//CML Seminars//EN CALSCALE:GREGORIAN METHOD:PUBLISH X-WR-CALNAME:CML AI/ML Seminar Series X-WR-TIMEZONE:America/Los_Angeles BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:20070311T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:20071104T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:2021-01-11-florian-wenzel@cml.ics.uci.edu DTSTAMP:20210111T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210111T130000 DTEND;TZID=America/Los_Angeles:20210111T140000 SUMMARY:[CML Seminar] Florian Wenzel: Towards Reliable Deep Learning LOCATION:Online (live stream) DESCRIPTION:Florian Wenzel\, Postdoctoral Researcher\, Google Brain Berlin\ n\nTitle: Towards Reliable Deep Learning\n\nAbstract: Deep learning models are bad at detecting their failure\, tending to make over-confident mista kes especially under distribution shift. We discuss two approaches to reli able deep learning. First\, we focus on Bayesian neural networks and cast doubt on the current understanding of Bayes posteriors in deep networks\, showing that they can be improved significantly through a cold posterior t hat sharply deviates from the Bayesian paradigm\, and discuss hypotheses t hat could explain it. Second\, we discuss ensembles: we show that the dive rsity of predictions can be improved by considering models with different hyperparameters\, and present an efficient method that leverages hyperpara meter diversity within a single model.\n\nhttps://cml.ics.uci.edu/seminars /2021-01-11-florian-wenzel X-ALT-DESC;FMTTYPE=text/html:Florian Wenzel\, Postdoctor al Researcher\, Google Brain Berlin

Title: Towards Reliable Deep Learning

Abstract: Deep learning models are bad at dete cting their failure\, tending to make over-confident mistakes especially u nder distribution shift. We discuss two approaches to reliable deep learni ng. First\, we focus on Bayesian neural networks and cast doubt on the cur rent understanding of Bayes posteriors in deep networks\, showing that the y can be improved significantly through a cold posterior that sharply devi ates from the Bayesian paradigm\, and discuss hypotheses that could explai n it. Second\, we discuss ensembles: we show that the diversity of predict ions can be improved by considering models with different hyperparameters\ , and present an efficient method that leverages hyperparameter diversity within a single model.

https://cml.ics.uci.edu/seminars/2021-01-11-flori an-wenzel URL:https://cml.ics.uci.edu/seminars/2021-01-11-florian-wenzel END:VEVENT BEGIN:VEVENT UID:2021-01-25-yezhou-yang@cml.ics.uci.edu DTSTAMP:20210125T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210125T130000 DTEND;TZID=America/Los_Angeles:20210125T140000 SUMMARY:[CML Seminar] Yezhou Yang: Visual Recognition beyond Appearances\, and its Robotic Applications LOCATION:Online (live stream) DESCRIPTION:Yezhou Yang\, Assistant Professor\, School of Computing\, Infor matics\, and Decision Systems Engineering\, Arizona State University\n\nTi tle: Visual Recognition beyond Appearances\, and its Robotic Applications\ n\nAbstract: The goal of Computer Vision is to develop algorithms to answe r What is Where at When from visual appearance. The speaker recognizes the importance of studying underlying entities and relations beyond visual ap pearance\, following an Active Perception paradigm. This talk presents eff orts ranging from reasoning beyond appearance for visual question answerin g\, image understanding\, and video captioning\, through temporal knowledg e distillation with incremental knowledge transfer\, to their roles in a r obotic visual learning framework via a robotic indoor object search task. The talk also features the Active Perception Group's ongoing projects in a utonomous driving\, AI security\, and healthcare.\n\nhttps://cml.ics.uci.e du/seminars/2021-01-25-yezhou-yang X-ALT-DESC;FMTTYPE=text/html:Yezhou Yang\, Assistant Pro fessor\, School of Computing\, Informatics\, and Decision Systems Engineer ing\, Arizona State University

Title: Visual Recognition bey ond Appearances\, and its Robotic Applications

Abstract: The goal of Computer Vision is to develop algorithms to answer What is Where at When from visual appearance. The speaker recognizes the importance of s tudying underlying entities and relations beyond visual appearance\, follo wing an Active Perception paradigm. This talk presents efforts ranging fro m reasoning beyond appearance for visual question answering\, image unders tanding\, and video captioning\, through temporal knowledge distillation w ith incremental knowledge transfer\, to their roles in a robotic visual le arning framework via a robotic indoor object search task. The talk also fe atures the Active Perception Group's ongoing projects in autonomous drivin g\, AI security\, and healthcare.

https://cml.ics.uci.edu/seminars/2021-01- 25-yezhou-yang URL:https://cml.ics.uci.edu/seminars/2021-01-25-yezhou-yang END:VEVENT BEGIN:VEVENT UID:2021-02-01-joe-marino@cml.ics.uci.edu DTSTAMP:20210201T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210201T130000 DTEND;TZID=America/Los_Angeles:20210201T140000 SUMMARY:[CML Seminar] Joe Marino: Connecting Variational Autoencoders Back to the Brain LOCATION:Online (live stream) DESCRIPTION:Joe Marino\, PhD Student\, Computation and Neural Systems\, Cal ifornia Institute of Technology\n\nTitle: Connecting Variational Autoencod ers Back to the Brain\n\nAbstract: Unsupervised machine learning has recen tly dramatically improved our ability to model and extract structure from data. One such approach is deep latent variable models\, including variati onal autoencoders (VAEs)\, which can be traced back to the Helmholtz machi ne and\, in turn\, ideas from theoretical neuroscience. Neuroscientists ha ve further developed these ideas into a popular theory: predictive coding. Yet the machine learning community remains largely unaware of these conne ctions. In this talk\, I discuss the links between modern deep latent vari able models and predictive coding\, yielding several implications for corr espondences between machine learning and neuroscience\, including the sear ch for backpropagation in the brain.\n\nhttps://cml.ics.uci.edu/seminars/2 021-02-01-joe-marino X-ALT-DESC;FMTTYPE=text/html:Joe Marino\, PhD Student\, Computation and Neural Systems\, California Institute of Technology

Title: Connecting Variational Autoencoders Back to the Brain
Abstract: Unsupervised machine learning has recently dramatically improved our ability to model and extract structure from data. One such a pproach is deep latent variable models\, including variational autoencoder s (VAEs)\, which can be traced back to the Helmholtz machine and\, in turn \, ideas from theoretical neuroscience. Neuroscientists have further devel oped these ideas into a popular theory: predictive coding. Yet the machine learning community remains largely unaware of these connections. In this talk\, I discuss the links between modern deep latent variable models and predictive coding\, yielding several implications for correspondences betw een machine learning and neuroscience\, including the search for backpropa gation in the brain.

https://cml.ics.uci.edu/seminars/2021-02-01-joe-marino< /a> URL:https://cml.ics.uci.edu/seminars/2021-02-01-joe-marino END:VEVENT BEGIN:VEVENT UID:2021-02-08-junkyu-lee@cml.ics.uci.edu DTSTAMP:20210208T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210208T130000 DTEND;TZID=America/Los_Angeles:20210208T140000 SUMMARY:[CML Seminar] Junkyu Lee: Decomposition Bounds for Influence Diagra ms LOCATION:Online (live stream) DESCRIPTION:Junkyu Lee\, AI Planning Group\, IBM Research\n\nTitle: Decompo sition Bounds for Influence Diagrams\n\nAbstract: Influence diagrams (IDs) extend Bayesian networks with decision variables and utility functions to model the interaction between an agent and a system. The standard task is to compute the maximum expected utility (MEU) and optimal policies\, one of the most challenging tasks in graphical models. Computing upper bounds on the MEU is desirable because they can guide search or sampling-based me thods. In this talk\, I present bounding schemes for solving IDs: one exte nds variational decomposition bounds in marginal MAP\, and another is a ne w submodel tree decomposition method. Empirical results show these bounds are orders of magnitude tighter than previous methods.\n\nhttps://cml.ics. uci.edu/seminars/2021-02-08-junkyu-lee X-ALT-DESC;FMTTYPE=text/html:Junkyu Lee\, AI Planning Gr oup\, IBM Research

Title: Decomposition Bounds for Influence Diagrams

Abstract: Influence diagrams (IDs) extend Bayesian networks with decision variables and utility functions to model the inter action between an agent and a system. The standard task is to compute the maximum expected utility (MEU) and optimal policies\, one of the most chal lenging tasks in graphical models. Computing upper bounds on the MEU is de sirable because they can guide search or sampling-based methods. In this t alk\, I present bounding schemes for solving IDs: one extends variational decomposition bounds in marginal MAP\, and another is a new submodel tree decomposition method. Empirical results show these bounds are orders of ma gnitude tighter than previous methods.

https://cml.ics.uci.edu/seminars/2021 -02-08-junkyu-lee URL:https://cml.ics.uci.edu/seminars/2021-02-08-junkyu-lee END:VEVENT BEGIN:VEVENT UID:2021-03-01-robert-logan@cml.ics.uci.edu DTSTAMP:20210301T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210301T130000 DTEND;TZID=America/Los_Angeles:20210301T140000 SUMMARY:[CML Seminar] Robert Logan: Fill in the ___: Prompt-Based Solutions for NLP LOCATION:Online (live stream) DESCRIPTION:Robert Logan\, PhD Student\, Department of Computer Science\, U niversity of California\, Irvine\n\nTitle: Fill in the ___: Prompt-Based S olutions for NLP\n\nAbstract: Recent progress in NLP has been driven by la rge neural language models (e.g.\, GPT-2 and BERT) that are pretrained usi ng self-supervised learning before being finetuned to downstream tasks. In this talk\, we describe the technique of prompting\, which reformulates t asks as fill-in-the-blank questions. We show how prompts can measure the f actual\, linguistic\, and task-specific knowledge contained in language mo dels\, introduce an approach for automatically constructing prompts via gr adient-guided search\, and cover ongoing work investigating whether prompt ing can replace finetuning — with early results showing prompting can be more effective in few-shot scenarios while being substantially more param eter efficient.\n\nhttps://cml.ics.uci.edu/seminars/2021-03-01-robert-loga n X-ALT-DESC;FMTTYPE=text/html:Robert Logan\, PhD Student\ , Department of Computer Science\, University of California\, Irvine
Title: Fill in the ___: Prompt-Based Solutions for NLP

Abstract: Recent progress in NLP has been driven by large neural langu age models (e.g.\, GPT-2 and BERT) that are pretrained using self-supervis ed learning before being finetuned to downstream tasks. In this talk\, we describe the technique of prompting\, which reformulates tasks as fill-in- the-blank questions. We show how prompts can measure the factual\, linguis tic\, and task-specific knowledge contained in language models\, introduce an approach for automatically constructing prompts via gradient-guided se arch\, and cover ongoing work investigating whether prompting can replace finetuning — with early results showing prompting can be more effective in few-shot scenarios while being substantially more parameter efficient. 
https://cml.ics.uci.edu/seminars/2021-03-01-robert-logan URL:https://cml.ics.uci.edu/seminars/2021-03-01-robert-logan END:VEVENT BEGIN:VEVENT UID:2021-04-12-sanmi-koyejo@cml.ics.uci.edu DTSTAMP:20210412T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210412T130000 DTEND;TZID=America/Los_Angeles:20210412T140000 SUMMARY:[CML Seminar] Sanmi Koyejo: The Measurement and Mismeasurement of T rustworthy ML LOCATION:Online (live stream) DESCRIPTION:Sanmi Koyejo\, Assistant Professor\, Department of Computer Sci ence\, University of Illinois at Urbana-Champaign\n\nTitle: The Measuremen t and Mismeasurement of Trustworthy ML\n\nAbstract: Across healthcare\, sc ience\, and engineering\, we increasingly employ machine learning to autom ate decision-making that affects our lives in profound ways. However\, ML can fail\, and reliably measuring such failures is the first step toward b uilding trustworthy learning machines. Consider algorithmic fairness\, whe re widely-deployed fairness metrics can exacerbate group disparities and a re often incompatible. Measurement is also crucial for robustness\, partic ularly in federated learning with error-prone devices. Across ML applicati ons\, the dire consequences of mismeasurement are a recurring theme. This talk outlines emerging strategies for addressing the measurement gap in ML and how this impacts trustworthiness.\n\nhttps://cml.ics.uci.edu/seminars /2021-04-12-sanmi-koyejo X-ALT-DESC;FMTTYPE=text/html:Sanmi Koyejo\, Assistant Pr ofessor\, Department of Computer Science\, University of Illinois at Urban a-Champaign

Title: The Measurement and Mismeasurement of Tru stworthy ML

Abstract: Across healthcare\, science\, and engi neering\, we increasingly employ machine learning to automate decision-mak ing that affects our lives in profound ways. However\, ML can fail\, and r eliably measuring such failures is the first step toward building trustwor thy learning machines. Consider algorithmic fairness\, where widely-deploy ed fairness metrics can exacerbate group disparities and are often incompa tible. Measurement is also crucial for robustness\, particularly in federa ted learning with error-prone devices. Across ML applications\, the dire c onsequences of mismeasurement are a recurring theme. This talk outlines em erging strategies for addressing the measurement gap in ML and how this im pacts trustworthiness.

https://cml.ics.uci.edu/seminars/2021-04-12-sanmi-k oyejo URL:https://cml.ics.uci.edu/seminars/2021-04-12-sanmi-koyejo END:VEVENT BEGIN:VEVENT UID:2021-04-19-kate-crawford@cml.ics.uci.edu DTSTAMP:20210419T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210419T160000 DTEND;TZID=America/Los_Angeles:20210419T170000 SUMMARY:[CML Seminar] Kate Crawford: Towards an Atlas of AI: A Conversation with Kate Crawford LOCATION:Online (live stream) DESCRIPTION:Kate Crawford\, Senior Principal Researcher\, Microsoft Researc h\, New York\n\nTitle: Towards an Atlas of AI: A Conversation with Kate Cr awford\n\nAbstract: Where do the motivating ideas behind Artificial Intell igence come from and what do they imply? What claims to universality or pa rticularity are made by AI systems? How do the movements of ideas\, data\, and materials shape the present and likely futures of AI development? A c onversation with social scientist and AI scholar Kate Crawford (with Paul Dourish and Geof Bowker of the Department of Informatics\, UC Irvine) abou t the intellectual history and geopolitical contexts of contemporary AI re search and practice. (Sponsored by the Steckler Center for Responsible\, E thical\, and Accessible Technology\, CREATE.)\n\nhttps://cml.ics.uci.edu/s eminars/2021-04-19-kate-crawford X-ALT-DESC;FMTTYPE=text/html:Kate Crawford\, Senior Prin cipal Researcher\, Microsoft Research\, New York

Title: Towa rds an Atlas of AI: A Conversation with Kate Crawford

Abstract: Where do the motivating ideas behind Artificial Intelligence come from and what do they imply? What claims to universality or particularity are made by AI systems? How do the movements of ideas\, data\, and materials s hape the present and likely futures of AI development? A conversation with social scientist and AI scholar Kate Crawford (with Paul Dourish and Geof Bowker of the Department of Informatics\, UC Irvine) about the intellectu al history and geopolitical contexts of contemporary AI research and pract ice. (Sponsored by the Steckler Center for Responsible\, Ethical\, and Acc essible Technology\, CREATE.)

https://cml.ics.uci.edu/seminars/2021-04-19 -kate-crawford URL:https://cml.ics.uci.edu/seminars/2021-04-19-kate-crawford END:VEVENT BEGIN:VEVENT UID:2021-04-26-yibo-yang@cml.ics.uci.edu DTSTAMP:20210426T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210426T130000 DTEND;TZID=America/Los_Angeles:20210426T140000 SUMMARY:[CML Seminar] Yibo Yang: Exploring the limits of lossy data compres sion with deep learning LOCATION:Online (live stream) DESCRIPTION:Yibo Yang\, PhD Student\, Department of Computer Science\, Univ ersity of California\, Irvine\n\nTitle: Exploring the limits of lossy data compression with deep learning\n\nAbstract: Probabilistic machine learnin g\, particularly deep learning\, is reshaping data compression. Recent wor k established a close connection between lossy compression and latent vari able models such as variational autoencoders (VAEs). In this talk\, I give an overview of learned data compression and present research addressing s ome of its limitations: algorithmic improvements inspired by variational i nference that push the performance limits of VAE-based lossy compression t o a new state of the art on images\; a new algorithm that compresses the v ariational posteriors of pre-trained latent variable models\; and ongoing work exploring fundamental bounds on lossy compression using stochastic ap proximation.\n\nhttps://cml.ics.uci.edu/seminars/2021-04-26-yibo-yang X-ALT-DESC;FMTTYPE=text/html:Yibo Yang\, PhD Student\, D epartment of Computer Science\, University of California\, Irvine

Title: Exploring the limits of lossy data compression with deep lear ning

Abstract: Probabilistic machine learning\, particularly deep learning\, is reshaping data compression. Recent work established a close connection between lossy compression and latent variable models such as variational autoencoders (VAEs). In this talk\, I give an overview of learned data compression and present research addressing some of its limit ations: algorithmic improvements inspired by variational inference that pu sh the performance limits of VAE-based lossy compression to a new state of the art on images\; a new algorithm that compresses the variational poste riors of pre-trained latent variable models\; and ongoing work exploring f undamental bounds on lossy compression using stochastic approximation.

https: //cml.ics.uci.edu/seminars/2021-04-26-yibo-yang URL:https://cml.ics.uci.edu/seminars/2021-04-26-yibo-yang END:VEVENT BEGIN:VEVENT UID:2021-05-03-levi-lelis@cml.ics.uci.edu DTSTAMP:20210503T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210503T130000 DTEND;TZID=America/Los_Angeles:20210503T140000 SUMMARY:[CML Seminar] Levi Lelis: Policy and Heuristic-Guided Tree Search A lgorithms LOCATION:Online (live stream) DESCRIPTION:Levi Lelis\, Assistant Professor\, Department of Computer Scien ce\, University of Alberta\n\nTitle: Policy and Heuristic-Guided Tree Sear ch Algorithms\n\nAbstract: In this talk I describe two tree search algorit hms that use a policy to guide the search. I start with Levin tree search (LTS)\, a best-first search algorithm with guarantees on the number of nod es it needs to expand\, based on the quality of the policy it employs. I t hen describe Policy-Guided Heuristic Search (PHS)\, which uses both a poli cy and a heuristic function to guide the search\, with guarantees based on the quality of both. Empirical results show that LTS and PHS compare favo rably with A*\, Weighted A*\, Greedy Best-First Search\, and PUCT on singl e-agent shortest-path problems.\n\nhttps://cml.ics.uci.edu/seminars/2021-0 5-03-levi-lelis X-ALT-DESC;FMTTYPE=text/html:Levi Lelis\, Assistant Prof essor\, Department of Computer Science\, University of Alberta

T itle: Policy and Heuristic-Guided Tree Search Algorithms

Abs tract: In this talk I describe two tree search algorithms that use a p olicy to guide the search. I start with Levin tree search (LTS)\, a best-f irst search algorithm with guarantees on the number of nodes it needs to e xpand\, based on the quality of the policy it employs. I then describe Pol icy-Guided Heuristic Search (PHS)\, which uses both a policy and a heurist ic function to guide the search\, with guarantees based on the quality of both. Empirical results show that LTS and PHS compare favorably with A*\, Weighted A*\, Greedy Best-First Search\, and PUCT on single-agent shortest -path problems.

https://cml.ics.uci.edu/seminars/2021-05-03-levi-lelis URL:https://cml.ics.uci.edu/seminars/2021-05-03-levi-lelis END:VEVENT BEGIN:VEVENT UID:2021-05-10-david-alvarez-melis@cml.ics.uci.edu DTSTAMP:20210510T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210510T130000 DTEND;TZID=America/Los_Angeles:20210510T140000 SUMMARY:[CML Seminar] David Alvarez-Melis: Ideal made real: machine learnin g with limited data and interpretable outputs LOCATION:Online (live stream) DESCRIPTION:David Alvarez-Melis\, Postdoctoral Researcher\, Microsoft Resea rch New England\n\nTitle: Ideal made real: machine learning with limited d ata and interpretable outputs\n\nAbstract: Success stories in machine lear ning tend to be concentrated on ideal scenarios where clean labeled data a re abundant and constraints are rare. But machine learning in practice is rarely so pristine. In this talk we explore how to reconcile these along t wo axes: learning with scarce or heterogeneous data\, and making complex m odels interpretable. First\, I present approaches for amplifying datasets based on the theory of Optimal Transport\, with applications in machine tr anslation\, transfer learning\, and dataset shaping. Second\, I present wo rk on designing methods to extract explanations from complex models and a novel framework for interpretable machine learning inspired by the study o f human explanation in the social sciences.\n\nhttps://cml.ics.uci.edu/sem inars/2021-05-10-david-alvarez-melis X-ALT-DESC;FMTTYPE=text/html:David Alvarez-Melis\, Postd octoral Researcher\, Microsoft Research New England

Title: I deal made real: machine learning with limited data and interpretable outpu ts

Abstract: Success stories in machine learning tend to be concentrated on ideal scenarios where clean labeled data are abundant and constraints are rare. But machine learning in practice is rarely so pristi ne. In this talk we explore how to reconcile these along two axes: learnin g with scarce or heterogeneous data\, and making complex models interpreta ble. First\, I present approaches for amplifying datasets based on the the ory of Optimal Transport\, with applications in machine translation\, tran sfer learning\, and dataset shaping. Second\, I present work on designing methods to extract explanations from complex models and a novel framework for interpretable machine learning inspired by the study of human explanat ion in the social sciences.

https://cml.ics.uci.edu/seminars/2021-0 5-10-david-alvarez-melis URL:https://cml.ics.uci.edu/seminars/2021-05-10-david-alvarez-melis END:VEVENT BEGIN:VEVENT UID:2021-05-17-megan-peters@cml.ics.uci.edu DTSTAMP:20210517T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210517T130000 DTEND;TZID=America/Los_Angeles:20210517T140000 SUMMARY:[CML Seminar] Megan Peters: How do we evaluate our own uncertainty? Uncovering our metacognitive computations LOCATION:Online (live stream) DESCRIPTION:Megan Peters\, Assistant Professor\, Department of Cognitive Sc iences\, University of California\, Irvine\n\nTitle: How do we evaluate ou r own uncertainty? Uncovering our metacognitive computations\n\nhttps://cm l.ics.uci.edu/seminars/2021-05-17-megan-peters X-ALT-DESC;FMTTYPE=text/html:Megan Peters\, Assistant Pr ofessor\, Department of Cognitive Sciences\, University of California\, Ir vine

Title: How do we evaluate our own uncertainty? Uncoveri ng our metacognitive computations

https://cml.ics.uci.edu/seminars/2021-05 -17-megan-peters URL:https://cml.ics.uci.edu/seminars/2021-05-17-megan-peters END:VEVENT BEGIN:VEVENT UID:2021-05-24-jing-zhang@cml.ics.uci.edu DTSTAMP:20210524T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20210524T130000 DTEND;TZID=America/Los_Angeles:20210524T140000 SUMMARY:[CML Seminar] Jing Zhang: Effective representation learning to diss ect the gene regulatory grammar LOCATION:Online (live stream) DESCRIPTION:Jing Zhang\, Assistant Professor\, Department of Computer Scien ce\, University of California\, Irvine\n\nTitle: Effective representation learning to dissect the gene regulatory grammar\n\nAbstract: Recent advanc es in sequencing technologies provide unprecedented opportunities to decip her multi-scale gene regulatory grammars at diverse cellular states. I int roduce our computational efforts on cell/gene representation learning to e xtract biologically meaningful information from high-dimensional\, sparse\ , and noisy genomic data. First\, we proposed SAILER\, a deep generative m odel to learn low-dimensional latent cell representations from single-cell epigenetic data that are invariant to confounding factors. Then at the ne twork level\, we developed TopicNet using latent Dirichlet allocation to e xtract latent gene communities and quantify regulatory network rewiring be tween cell states\, applied across 13 cancer types.\n\nhttps://cml.ics.uci .edu/seminars/2021-05-24-jing-zhang X-ALT-DESC;FMTTYPE=text/html:Jing Zhang\, Assistant Prof essor\, Department of Computer Science\, University of California\, Irvine

Title: Effective representation learning to dissect the gen e regulatory grammar

Abstract: Recent advances in sequencing technologies provide unprecedented opportunities to decipher multi-scale gene regulatory grammars at diverse cellular states. I introduce our compu tational efforts on cell/gene representation learning to extract biologica lly meaningful information from high-dimensional\, sparse\, and noisy geno mic data. First\, we proposed SAILER\, a deep generative model to learn lo w-dimensional latent cell representations from single-cell epigenetic data that are invariant to confounding factors. Then at the network level\, we developed TopicNet using latent Dirichlet allocation to extract latent ge ne communities and quantify regulatory network rewiring between cell state s\, applied across 13 cancer types.

https://cml.ics.uci.edu/seminars/2021-05 -24-jing-zhang URL:https://cml.ics.uci.edu/seminars/2021-05-24-jing-zhang END:VEVENT BEGIN:VEVENT UID:2022-01-10-roy-fox@cml.ics.uci.edu DTSTAMP:20220110T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220110T130000 DTEND;TZID=America/Los_Angeles:20220110T140000 SUMMARY:[CML Seminar] Roy Fox: Curiously effective ensemble and double-orac le reinforcement-learning methods LOCATION:Online (live stream) DESCRIPTION:Roy Fox\, Assistant Professor\, Department of Computer Science\ , University of California\, Irvine\n\nTitle: Curiously effective ensemble and double-oracle reinforcement-learning methods\n\nAbstract: Ensemble me thods for reinforcement learning represent model uncertainty and use it to guide exploration and reduce value estimation bias. We present MeanQ\, a very simple ensemble method with improved performance that reduces estimat ion variance enough to operate without a stabilizing target network — cu riously\, it is theoretically almost equivalent to a non-ensemble method i t significantly outperforms. In adversarial environments\, double-oracle ( DO) methods grow a population of policies by iteratively adding best respo nses. We present XDO\, a DO algorithm that exploits the game's sequential structure to exponentially reduce the worst-case population size.\n\nhttps ://cml.ics.uci.edu/seminars/2022-01-10-roy-fox X-ALT-DESC;FMTTYPE=text/html:Roy Fox\, Assistant Profess or\, Department of Computer Science\, University of California\, Irvine

Title: Curiously effective ensemble and double-oracle reinforc ement-learning methods

Abstract: Ensemble methods for reinfo rcement learning represent model uncertainty and use it to guide explorati on and reduce value estimation bias. We present MeanQ\, a very simple ense mble method with improved performance that reduces estimation variance eno ugh to operate without a stabilizing target network — curiously\, it is theoretically almost equivalent to a non-ensemble method it significantly outperforms. In adversarial environments\, double-oracle (DO) methods grow a population of policies by iteratively adding best responses. We present XDO\, a DO algorithm that exploits the game's sequential structure to exp onentially reduce the worst-case population size.

https://cml.ics.uci.edu/semin ars/2022-01-10-roy-fox URL:https://cml.ics.uci.edu/seminars/2022-01-10-roy-fox END:VEVENT BEGIN:VEVENT UID:2022-01-24-ransalu-senanayake@cml.ics.uci.edu DTSTAMP:20220124T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220124T130000 DTEND;TZID=America/Los_Angeles:20220124T140000 SUMMARY:[CML Seminar] Ransalu Senanayake: Propagating Uncertainty from Mode ling into Decision-Making for Trustworthy Autonomy LOCATION:Online (live stream) DESCRIPTION:Ransalu Senanayake\, Postdoctoral Scholar\, Department of Compu ter Science\, Stanford University\n\nTitle: Propagating Uncertainty from M odeling into Decision-Making for Trustworthy Autonomy\n\nAbstract: Autonom ous agents such as self-driving cars have gained the capability to perform individual tasks such as object detection and lane following in simple\, static environments. While advancing robots towards full autonomy\, it is important to minimize deleterious effects on humans and infrastructure. Fo r robots to safely operate in the real world\, it is vital to quantify the multimodal aleatoric and epistemic uncertainty around them and use that u ncertainty for decision-making. In this talk\, I discuss how we can levera ge approximate Bayesian inference\, kernel methods\, and deep neural netwo rks to develop interpretable autonomous systems for high-stakes applicatio ns.\n\nhttps://cml.ics.uci.edu/seminars/2022-01-24-ransalu-senanayake X-ALT-DESC;FMTTYPE=text/html:Ransalu Senanayake\, Postdo ctoral Scholar\, Department of Computer Science\, Stanford University
 Title: Propagating Uncertainty from Modeling into Decision-Makin g for Trustworthy Autonomy

Abstract: Autonomous agents such as self-driving cars have gained the capability to perform individual task s such as object detection and lane following in simple\, static environme nts. While advancing robots towards full autonomy\, it is important to min imize deleterious effects on humans and infrastructure. For robots to safe ly operate in the real world\, it is vital to quantify the multimodal alea toric and epistemic uncertainty around them and use that uncertainty for d ecision-making. In this talk\, I discuss how we can leverage approximate B ayesian inference\, kernel methods\, and deep neural networks to develop i nterpretable autonomous systems for high-stakes applications.

https: //cml.ics.uci.edu/seminars/2022-01-24-ransalu-senanayake URL:https://cml.ics.uci.edu/seminars/2022-01-24-ransalu-senanayake END:VEVENT BEGIN:VEVENT UID:2022-01-31-dylan-slack@cml.ics.uci.edu DTSTAMP:20220131T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220131T130000 DTEND;TZID=America/Los_Angeles:20220131T140000 SUMMARY:[CML Seminar] Dylan Slack: Exposing Shortcomings and Improving the Reliability of Machine Learning Explanations LOCATION:Online (live stream) DESCRIPTION:Dylan Slack\, PhD Student\, Department of Computer Science\, Un iversity of California\, Irvine\n\nTitle: Exposing Shortcomings and Improv ing the Reliability of Machine Learning Explanations\n\nAbstract: For doma in experts to adopt machine learning models in high-stakes settings such a s health care and law\, they must understand and trust model predictions. Researchers have proposed numerous ways to explain complex ML models\, but these approaches suffer from critical drawbacks such as vulnerability to adversarial attacks\, instability\, and inconsistency. This talk describes the shortcomings of explanations\, demonstrates how they are vulnerable t o adversarial attacks\, and presents recent work on explanations that leve rage uncertainty estimates to overcome several critical explanation shortc omings.\n\nhttps://cml.ics.uci.edu/seminars/2022-01-31-dylan-slack X-ALT-DESC;FMTTYPE=text/html:Dylan Slack\, PhD Student\, Department of Computer Science\, University of California\, Irvine

Title: Exposing Shortcomings and Improving the Reliability of Mach ine Learning Explanations

Abstract: For domain experts to ad opt machine learning models in high-stakes settings such as health care an d law\, they must understand and trust model predictions. Researchers have proposed numerous ways to explain complex ML models\, but these approache s suffer from critical drawbacks such as vulnerability to adversarial atta cks\, instability\, and inconsistency. This talk describes the shortcoming s of explanations\, demonstrates how they are vulnerable to adversarial at tacks\, and presents recent work on explanations that leverage uncertainty estimates to overcome several critical explanation shortcomings.

< a href="https://cml.ics.uci.edu/seminars/2022-01-31-dylan-slack">https://c ml.ics.uci.edu/seminars/2022-01-31-dylan-slack URL:https://cml.ics.uci.edu/seminars/2022-01-31-dylan-slack END:VEVENT BEGIN:VEVENT UID:2022-02-07-maja-rudolph@cml.ics.uci.edu DTSTAMP:20220207T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220207T130000 DTEND;TZID=America/Los_Angeles:20220207T140000 SUMMARY:[CML Seminar] Maja Rudolph: Modeling Irregular Time Series with Con tinuous Recurrent Units LOCATION:Online (live stream) DESCRIPTION:Maja Rudolph\, Senior Research Scientist\, Bosch Center for AI\ n\nTitle: Modeling Irregular Time Series with Continuous Recurrent Units\n \nAbstract: Recurrent neural networks (RNNs) are a popular choice for mode ling sequential data but assume constant time-intervals between observatio ns. In many datasets (e.g. medical records) observation times are irregula r and can carry important information. We propose continuous recurrent uni ts (CRUs) — a neural architecture that naturally handles irregular inter vals. The CRU assumes a hidden state that evolves according to a linear st ochastic differential equation\, integrated into an encoder-decoder framew ork via the continuous-discrete Kalman filter in closed form. We find that the CRU can interpolate irregular time series better than methods based o n neural ODEs.\n\nhttps://cml.ics.uci.edu/seminars/2022-02-07-maja-rudolph X-ALT-DESC;FMTTYPE=text/html:Maja Rudolph\, Senior Resea rch Scientist\, Bosch Center for AI

Title: Modeling Irregula r Time Series with Continuous Recurrent Units

Abstract: Recu rrent neural networks (RNNs) are a popular choice for modeling sequential data but assume constant time-intervals between observations. In many data sets (e.g. medical records) observation times are irregular and can carry important information. We propose continuous recurrent units (CRUs) — a neural architecture that naturally handles irregular intervals. The CRU as sumes a hidden state that evolves according to a linear stochastic differe ntial equation\, integrated into an encoder-decoder framework via the cont inuous-discrete Kalman filter in closed form. We find that the CRU can int erpolate irregular time series better than methods based on neural ODEs.
h ttps://cml.ics.uci.edu/seminars/2022-02-07-maja-rudolph URL:https://cml.ics.uci.edu/seminars/2022-02-07-maja-rudolph END:VEVENT BEGIN:VEVENT UID:2022-02-14-ruiqi-gao@cml.ics.uci.edu DTSTAMP:20220214T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220214T130000 DTEND;TZID=America/Los_Angeles:20220214T140000 SUMMARY:[CML Seminar] Ruiqi Gao: Advanced training of energy-based models LOCATION:Online (live stream) DESCRIPTION:Ruiqi Gao\, Research Scientist\, Google Brain\n\nTitle: Advance d training of energy-based models\n\nAbstract: Energy-based models (EBMs) are an appealing class of probabilistic models that can be learned from un labeled data\, but two challenges remain for training them on high-dimensi onal datasets: maximum-likelihood learning requires expensive MCMC samplin g\, and energy potentials learned with non-convergent MCMC can be highly b iased. I present two algorithms to tackle these challenges: (1) Diffusion Recovery Likelihood\, which tractably learns and samples from a sequence o f EBMs trained on increasingly noisy versions of a dataset\, and (2) Flow Contrastive Estimation\, which jointly estimates an EBM and a flow-based m odel via a shared adversarial value function.\n\nhttps://cml.ics.uci.edu/s eminars/2022-02-14-ruiqi-gao X-ALT-DESC;FMTTYPE=text/html:Ruiqi Gao\, Research Scient ist\, Google Brain

Title: Advanced training of energy-based models

Abstract: Energy-based models (EBMs) are an appealing class of probabilistic models that can be learned from unlabeled data\, b ut two challenges remain for training them on high-dimensional datasets: m aximum-likelihood learning requires expensive MCMC sampling\, and energy p otentials learned with non-convergent MCMC can be highly biased. I present two algorithms to tackle these challenges: (1) Diffusion Recovery Likelih ood\, which tractably learns and samples from a sequence of EBMs trained o n increasingly noisy versions of a dataset\, and (2) Flow Contrastive Esti mation\, which jointly estimates an EBM and a flow-based model via a share d adversarial value function.

https://cml.ics.uci.edu/seminars/2022-02-14-rui qi-gao URL:https://cml.ics.uci.edu/seminars/2022-02-14-ruiqi-gao END:VEVENT BEGIN:VEVENT UID:2022-02-28-sunipa-dev@cml.ics.uci.edu DTSTAMP:20220228T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220228T130000 DTEND;TZID=America/Los_Angeles:20220228T140000 SUMMARY:[CML Seminar] Sunipa Dev: Towards Inclusive and Socially Aware Lang uage Technologies LOCATION:Donald Bren Hall 4011 DESCRIPTION:Sunipa Dev\, Research Scientist\, Ethical AI Team\, Google AI\n \nTitle: Towards Inclusive and Socially Aware Language Technologies\n\nAbs tract: Large language models are commonly used across natural language pro cessing and are known for their efficiency as well as their overall lack o f interpretability. Their data-driven approach for emulating human languag e often results in human biases being encoded and even amplified\, potenti ally leading to cyclic propagation of representational and allocational ha rm. In this talk we discuss detecting\, evaluating\, and mitigating biases and associated harms in a holistic\, inclusive\, and culturally-aware man ner. In particular\, we discuss the disparate impact of common language to ols that are not inclusive of all gender identities.\n\nhttps://cml.ics.uc i.edu/seminars/2022-02-28-sunipa-dev X-ALT-DESC;FMTTYPE=text/html:Sunipa Dev\, Research Scien tist\, Ethical AI Team\, Google AI

Title: Towards Inclusive and Socially Aware Language Technologies

Abstract: Large lan guage models are commonly used across natural language processing and are known for their efficiency as well as their overall lack of interpretabili ty. Their data-driven approach for emulating human language often results in human biases being encoded and even amplified\, potentially leading to cyclic propagation of representational and allocational harm. In this talk we discuss detecting\, evaluating\, and mitigating biases and associated harms in a holistic\, inclusive\, and culturally-aware manner. In particul ar\, we discuss the disparate impact of common language tools that are not inclusive of all gender identities.

https://cml.ics.uci.edu/seminars/2022-0 2-28-sunipa-dev URL:https://cml.ics.uci.edu/seminars/2022-02-28-sunipa-dev END:VEVENT BEGIN:VEVENT UID:2022-03-07-mukund-sundararajan@cml.ics.uci.edu DTSTAMP:20220307T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220307T130000 DTEND;TZID=America/Los_Angeles:20220307T140000 SUMMARY:[CML Seminar] Mukund Sundararajan: Analyzing deep neural networks u sing attribution LOCATION:Online (Zoom) DESCRIPTION:Mukund Sundararajan\, Principal Research Scientist\, Google\n\n Title: Analyzing deep neural networks using attribution\n\nAbstract: A tal k on analyzing deep neural networks using attribution methods — includin g integrated gradients\, counterfactual and Shapley-value approaches\, and attributing predictions to training data — to understand and debug mode l behavior (for example\, discovering that a cancer-from-X-rays model was fixating on radiologist markings). (The speaker's original abstract was co mposed in verse.)\n\nhttps://cml.ics.uci.edu/seminars/2022-03-07-mukund-su ndararajan X-ALT-DESC;FMTTYPE=text/html:Mukund Sundararajan\, Princ ipal Research Scientist\, Google

Title: Analyzing deep neura l networks using attribution

Abstract: A talk on analyzing d eep neural networks using attribution methods — including integrated gra dients\, counterfactual and Shapley-value approaches\, and attributing pre dictions to training data — to understand and debug model behavior (for example\, discovering that a cancer-from-X-rays model was fixating on radi ologist markings). (The speaker's original abstract was composed in verse. )

https://cml.ics.uci.edu/seminars/2022-03-07-mukund-sundararajan URL:https://cml.ics.uci.edu/seminars/2022-03-07-mukund-sundararajan END:VEVENT BEGIN:VEVENT UID:2022-05-02-maurizio-filippone-and-ba-hien-tran@cml.ics.uci.edu DTSTAMP:20220502T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220502T130000 DTEND;TZID=America/Los_Angeles:20220502T140000 SUMMARY:[CML Seminar] Maurizio Filippone and Ba-Hien Tran: Functional Prior s for Bayesian Deep Learning LOCATION:Donald Bren Hall 4011 DESCRIPTION:Maurizio Filippone and Ba-Hien Tran\, Associate Professor and P hD Student\, EURECOM\n\nTitle: Functional Priors for Bayesian Deep Learnin g\n\nAbstract: The Bayesian treatment of neural networks dictates that a p rior distribution is specified over their weight and bias parameters\, whi ch is challenging because modern networks have huge numbers of parameters and non-linearities. Gaussian processes instead offer a rigorous non-param etric framework to define priors over the space of functions. In this talk we introduce a robust framework to impose functional priors on modern neu ral networks by minimizing the Wasserstein distance between samples of sto chastic processes\, and extend it to model selection for Bayesian autoenco ders. Coupling these priors with scalable MCMC offers systematically large performance improvements over alternative choices of priors.\n\nhttps://c ml.ics.uci.edu/seminars/2022-05-02-maurizio-filippone-and-ba-hien-tran X-ALT-DESC;FMTTYPE=text/html:Maurizio Filippone and Ba-Hien Tran\, Associate Professor and PhD Student\, EURECOM

Title: Functional Priors for Bayesian Deep Learning

Abstract: T he Bayesian treatment of neural networks dictates that a prior distributio n is specified over their weight and bias parameters\, which is challengin g because modern networks have huge numbers of parameters and non-linearit ies. Gaussian processes instead offer a rigorous non-parametric framework to define priors over the space of functions. In this talk we introduce a robust framework to impose functional priors on modern neural networks by minimizing the Wasserstein distance between samples of stochastic processe s\, and extend it to model selection for Bayesian autoencoders. Coupling t hese priors with scalable MCMC offers systematically large performance imp rovements over alternative choices of priors.

https ://cml.ics.uci.edu/seminars/2022-05-02-maurizio-filippone-and-ba-hien-tran URL:https://cml.ics.uci.edu/seminars/2022-05-02-maurizio-filippone-and-ba-h ien-tran END:VEVENT BEGIN:VEVENT UID:2022-05-09-ties-van-rozendaal@cml.ics.uci.edu DTSTAMP:20220509T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220509T130000 DTEND;TZID=America/Los_Angeles:20220509T140000 SUMMARY:[CML Seminar] Ties van Rozendaal: Instance-adaptive data compressio n: Improving Neural Codecs by Training on the Test Set LOCATION:Donald Bren Hall 4011 DESCRIPTION:Ties van Rozendaal\, Senior Machine Learning Researcher\, Qualc omm AI Research\n\nTitle: Instance-adaptive data compression: Improving Ne ural Codecs by Training on the Test Set\n\nAbstract: Neural data compressi on has been shown to outperform classical methods in terms of rate-distort ion performance. These models are fitted to a training dataset and cannot be expected to optimally compress test data in general\, due to limits on model capacity\, distribution shifts\, and imperfect optimization. Instanc e-adaptive methods take adaptation to the extreme\, adapting the model to a single test instance and signaling the updated model in the bitstream. I n this talk\, we show the potential of different types of instance-adaptiv e methods and discuss the tradeoffs they pose.\n\nhttps://cml.ics.uci.edu/ seminars/2022-05-09-ties-van-rozendaal X-ALT-DESC;FMTTYPE=text/html:Ties van Rozendaal\, Senior Machine Learning Researcher\, Qualcomm AI Research

Title: I nstance-adaptive data compression: Improving Neural Codecs by Training on the Test Set

Abstract: Neural data compression has been show n to outperform classical methods in terms of rate-distortion performance. These models are fitted to a training dataset and cannot be expected to o ptimally compress test data in general\, due to limits on model capacity\, distribution shifts\, and imperfect optimization. Instance-adaptive metho ds take adaptation to the extreme\, adapting the model to a single test in stance and signaling the updated model in the bitstream. In this talk\, we show the potential of different types of instance-adaptive methods and di scuss the tradeoffs they pose.

https://cml.ics.uci.edu/seminars/2022 -05-09-ties-van-rozendaal URL:https://cml.ics.uci.edu/seminars/2022-05-09-ties-van-rozendaal END:VEVENT BEGIN:VEVENT UID:2022-05-16-robin-jia@cml.ics.uci.edu DTSTAMP:20220516T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220516T130000 DTEND;TZID=America/Los_Angeles:20220516T140000 SUMMARY:[CML Seminar] Robin Jia: Out-of-Distribution Evaluation: The How\, the Which\, and the What?! LOCATION:Donald Bren Hall 4011 DESCRIPTION:Robin Jia\, Assistant Professor of Computer Science\, Universit y of Southern California\n\nTitle: Out-of-Distribution Evaluation: The How \, the Which\, and the What?!\n\nAbstract: NLP models have achieved impres sive accuracies on in-distribution benchmarks but are unreliable out-of-di stribution (OOD). In this talk\, I preview my group's ongoing work on eval uating and improving model performance in OOD settings. First\, I propose likelihood splits\, a general-purpose way to create challenging non-i.i.d. benchmarks by measuring generalization to the tail of the data distributi on. Second\, I describe the advantages of neurosymbolic approaches over en d-to-end pretrained models for OOD generalization in visual question answe ring. Finally\, I show how synthesized examples can improve open-set recog nition.\n\nhttps://cml.ics.uci.edu/seminars/2022-05-16-robin-jia X-ALT-DESC;FMTTYPE=text/html:Robin Jia\, Assistant Profe ssor of Computer Science\, University of Southern California

Tit le: Out-of-Distribution Evaluation: The How\, the Which\, and the What ?!

Abstract: NLP models have achieved impressive accuracies on in-distribution benchmarks but are unreliable out-of-distribution (OOD) . In this talk\, I preview my group's ongoing work on evaluating and impro ving model performance in OOD settings. First\, I propose likelihood split s\, a general-purpose way to create challenging non-i.i.d. benchmarks by m easuring generalization to the tail of the data distribution. Second\, I d escribe the advantages of neurosymbolic approaches over end-to-end pretrai ned models for OOD generalization in visual question answering. Finally\, I show how synthesized examples can improve open-set recognition.

< a href="https://cml.ics.uci.edu/seminars/2022-05-16-robin-jia">https://cml .ics.uci.edu/seminars/2022-05-16-robin-jia URL:https://cml.ics.uci.edu/seminars/2022-05-16-robin-jia END:VEVENT BEGIN:VEVENT UID:2022-06-06-bobak-pezeshki@cml.ics.uci.edu DTSTAMP:20220606T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20220606T130000 DTEND;TZID=America/Los_Angeles:20220606T140000 SUMMARY:[CML Seminar] Bobak Pezeshki: AND/OR Branch-and-Bound for Computati onal Protein Design Optimizing K* LOCATION:Donald Bren Hall 4011 DESCRIPTION:Bobak Pezeshki\, PhD Student\, Department of Computer Science\, University of California\, Irvine\n\nTitle: AND/OR Branch-and-Bound for C omputational Protein Design Optimizing K*\n\nAbstract: Computational prote in design (CPD) is the task of creating new proteins to fulfill a desired function. In this talk\, I share work accepted at UAI 2022 based on a new formulation of CPD as a graphical model designed for optimizing subunit bi nding affinity (approximated by a quantity called K*). I relate this to th e task of MMAP\, present the formulation of the problem as a graphical mod el\, and introduce a weighted mini-bucket heuristic for bounding K* and gu iding search. Finally\, I share our algorithm AOBB-K* and modifications th at enhance it\, describing its empirical benefits and limitations.\n\nhttp s://cml.ics.uci.edu/seminars/2022-06-06-bobak-pezeshki X-ALT-DESC;FMTTYPE=text/html:Bobak Pezeshki\, PhD Studen t\, Department of Computer Science\, University of California\, Irvine

Title: AND/OR Branch-and-Bound for Computational Protein Design Optimizing K*

Abstract: Computational protein design (CPD) is the task of creating new proteins to fulfill a desired function. In thi s talk\, I share work accepted at UAI 2022 based on a new formulation of C PD as a graphical model designed for optimizing subunit binding affinity ( approximated by a quantity called K*). I relate this to the task of MMAP\, present the formulation of the problem as a graphical model\, and introdu ce a weighted mini-bucket heuristic for bounding K* and guiding search. Fi nally\, I share our algorithm AOBB-K* and modifications that enhance it\, describing its empirical benefits and limitations.

https://cml.ics.uci.e du/seminars/2022-06-06-bobak-pezeshki URL:https://cml.ics.uci.edu/seminars/2022-06-06-bobak-pezeshki END:VEVENT BEGIN:VEVENT UID:2022-10-10-furong-huang@cml.ics.uci.edu DTSTAMP:20221010T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20221010T130000 DTEND;TZID=America/Los_Angeles:20221010T140000 SUMMARY:[CML Seminar] Furong Huang: Trustworthy Machine Learning in Complex Environments LOCATION:Donald Bren Hall 4011 DESCRIPTION:Furong Huang\, Assistant Professor of Computer Science\, Univer sity of Maryland\n\nTitle: Trustworthy Machine Learning in Complex Environ ments\n\nAbstract: With the burgeoning use of machine learning models\, th ere is a need to rapidly and reliably deploy models in a variety of enviro nments. These trustworthy models must be able to: (i) adapt and generalize to previously unseen worlds although trained on data that only represent a subset of the world\, (ii) allow for non-iid data\, (iii) be resilient t o (adversarial) perturbations\, and (iv) conform to social norms and make ethical decisions. In this talk\, I will cover reinforcement learning algo rithms that achieve fast adaptation by guaranteed knowledge transfer\, pri ncipled methods that measure the vulnerability and improve the robustness of RL agents\, and ethical models that make fair decisions under distribut ion shifts.\n\nhttps://cml.ics.uci.edu/seminars/2022-10-10-furong-huang X-ALT-DESC;FMTTYPE=text/html:Furong Huang\, Assistant Pr ofessor of Computer Science\, University of Maryland

Title: Trustworthy Machine Learning in Complex Environments

Abstract: With the burgeoning use of machine learning models\, there is a need to rapidly and reliably deploy models in a variety of environments. These tr ustworthy models must be able to: (i) adapt and generalize to previously u nseen worlds although trained on data that only represent a subset of the world\, (ii) allow for non-iid data\, (iii) be resilient to (adversarial) perturbations\, and (iv) conform to social norms and make ethical decision s. In this talk\, I will cover reinforcement learning algorithms that achi eve fast adaptation by guaranteed knowledge transfer\, principled methods that measure the vulnerability and improve the robustness of RL agents\, a nd ethical models that make fair decisions under distribution shifts.
 http s://cml.ics.uci.edu/seminars/2022-10-10-furong-huang URL:https://cml.ics.uci.edu/seminars/2022-10-10-furong-huang END:VEVENT BEGIN:VEVENT UID:2022-10-17-bodhisattwa-prasad-majumder@cml.ics.uci.edu DTSTAMP:20221017T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20221017T130000 DTEND;TZID=America/Los_Angeles:20221017T140000 SUMMARY:[CML Seminar] Bodhisattwa Prasad Majumder: Effective\, explainable\ , and equitable predictions in NLP models with world knowledge and convers ations LOCATION:Donald Bren Hall 4011 DESCRIPTION:Bodhisattwa Prasad Majumder\, PhD Student\, Department of Compu ter Science and Engineering\, University of California\, San Diego\n\nTitl e: Effective\, explainable\, and equitable predictions in NLP models with world knowledge and conversations\n\nAbstract: The use of artificial intel ligence in knowledge-seeking applications has shown remarkable effectivene ss\, but increasing demand for interaction and accessibility requires the underlying components to be grounded in up-to-date real-world context. In this talk\, I discuss methods to effectively inject up-to-date knowledge i nto an existing dialog model without additional training\, the role of bac kground knowledge in generating faithful natural language explanations\, a nd a conversational framework to address subjectivity—balancing task per formance and bias mitigation for fair interpretable predictions.\n\nhttps: //cml.ics.uci.edu/seminars/2022-10-17-bodhisattwa-prasad-majumder X-ALT-DESC;FMTTYPE=text/html:Bodhisattwa Prasad Majumder \, PhD Student\, Department of Computer Science and Engineering\, Universi ty of California\, San Diego

Title: Effective\, explainable\ , and equitable predictions in NLP models with world knowledge and convers ations

Abstract: The use of artificial intelligence in knowl edge-seeking applications has shown remarkable effectiveness\, but increas ing demand for interaction and accessibility requires the underlying compo nents to be grounded in up-to-date real-world context. In this talk\, I di scuss methods to effectively inject up-to-date knowledge into an existing dialog model without additional training\, the role of background knowledg e in generating faithful natural language explanations\, and a conversatio nal framework to address subjectivity—balancing task performance and bia s mitigation for fair interpretable predictions.

https://cm l.ics.uci.edu/seminars/2022-10-17-bodhisattwa-prasad-majumder URL:https://cml.ics.uci.edu/seminars/2022-10-17-bodhisattwa-prasad-majumder END:VEVENT BEGIN:VEVENT UID:2022-10-24-mark-steyvers@cml.ics.uci.edu DTSTAMP:20221024T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20221024T130000 DTEND;TZID=America/Los_Angeles:20221024T140000 SUMMARY:[CML Seminar] Mark Steyvers: Human-AI collaboration LOCATION:Donald Bren Hall 4011 DESCRIPTION:Mark Steyvers\, Professor of Cognitive Sciences\, University of California\, Irvine\n\nTitle: Human-AI collaboration\n\nAbstract: Artific ial intelligence and machine learning models are being increasingly deploy ed in real-world applications where there is strong motivation to develop hybrid systems in which humans and AI algorithms work together. I will dis cuss a Bayesian framework for statistically combining predictions from hum ans and machines while accounting for the unique ways human and algorithmi c confidence is expressed\, recent work on AI-assisted decision making and estimating individual AI-reliance policies\, and the question of machine theory of mind — how humans and machines can efficiently form mental mod els of each other.\n\nhttps://cml.ics.uci.edu/seminars/2022-10-24-mark-ste yvers X-ALT-DESC;FMTTYPE=text/html:Mark Steyvers\, Professor o f Cognitive Sciences\, University of California\, Irvine

Title: Human-AI collaboration

Abstract: Artificial intelligence and machine learning models are being increasingly deployed in real-world applications where there is strong motivation to develop hybrid systems i n which humans and AI algorithms work together. I will discuss a Bayesian framework for statistically combining predictions from humans and machines while accounting for the unique ways human and algorithmic confidence is expressed\, recent work on AI-assisted decision making and estimating indi vidual AI-reliance policies\, and the question of machine theory of mind — how humans and machines can efficiently form mental models of each oth er.

https://cml.ics.uci.edu/seminars/2022-10-24-mark-steyvers< /html> URL:https://cml.ics.uci.edu/seminars/2022-10-24-mark-steyvers END:VEVENT BEGIN:VEVENT UID:2022-10-31-alex-boyd@cml.ics.uci.edu DTSTAMP:20221031T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20221031T130000 DTEND;TZID=America/Los_Angeles:20221031T140000 SUMMARY:[CML Seminar] Alex Boyd: Predictive Querying for Autoregressive Neu ral Sequence Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Alex Boyd\, PhD Student\, Department of Statistics\, University of California\, Irvine\n\nTitle: Predictive Querying for Autoregressive N eural Sequence Models\n\nAbstract: In reasoning about sequential events it is natural to pose probabilistic queries such as when will event A occur next or what is the probability of A occurring before B. However\, with ma chine learning shifting towards neural autoregressive models such as RNNs and transformers\, probabilistic querying has been largely restricted to s imple cases such as next-event prediction\, in part because future queryin g involves marginalization over large path spaces. In this talk\, we descr ibe a novel representation of querying for discrete sequential models\, al ong with approximation and search techniques to estimate these probabilist ic queries\, and touch on extensions to continuous-time events.\n\nhttps:/ /cml.ics.uci.edu/seminars/2022-10-31-alex-boyd X-ALT-DESC;FMTTYPE=text/html:Alex Boyd\, PhD Student\, D epartment of Statistics\, University of California\, Irvine

Titl e: Predictive Querying for Autoregressive Neural Sequence Models
Abstract: In reasoning about sequential events it is natural to p ose probabilistic queries such as when will event A occur next or what is the probability of A occurring before B. However\, with machine learning s hifting towards neural autoregressive models such as RNNs and transformers \, probabilistic querying has been largely restricted to simple cases such as next-event prediction\, in part because future querying involves margi nalization over large path spaces. In this talk\, we describe a novel repr esentation of querying for discrete sequential models\, along with approxi mation and search techniques to estimate these probabilistic queries\, and touch on extensions to continuous-time events.

https://cml.ics.uci.edu/semin ars/2022-10-31-alex-boyd URL:https://cml.ics.uci.edu/seminars/2022-10-31-alex-boyd END:VEVENT BEGIN:VEVENT UID:2022-11-07-yanning-shen@cml.ics.uci.edu DTSTAMP:20221107T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20221107T130000 DTEND;TZID=America/Los_Angeles:20221107T140000 SUMMARY:[CML Seminar] Yanning Shen: Adaptive Online Scalable Learning with Graph Feedback LOCATION:Donald Bren Hall 4011 DESCRIPTION:Yanning Shen\, Assistant Professor of Electrical Engineering an d Computer Science\, University of California\, Irvine\n\nTitle: Adaptive Online Scalable Learning with Graph Feedback\n\nAbstract: We live in an er a of data deluge\, where pervasive media collect massive amounts of data\, often in a streaming fashion. The sheer volume makes batch analytics impo ssible\, and data are noisy\, incomplete\, and prone to outliers\, so anal ytics must often be performed in real-time. This talk introduces an online scalable function approximation scheme that adaptively learns and tracks the sought nonlinear function on the fly with quantifiable performance gua rantees\, even in adversarial environments. Building on this framework\, a scalable online learning approach with graph feedback is outlined for onl ine learning with possibly related models\, showcased on several real-worl d datasets.\n\nhttps://cml.ics.uci.edu/seminars/2022-11-07-yanning-shen X-ALT-DESC;FMTTYPE=text/html:Yanning Shen\, Assistant Pr ofessor of Electrical Engineering and Computer Science\, University of Cal ifornia\, Irvine

Title: Adaptive Online Scalable Learning wi th Graph Feedback

Abstract: We live in an era of data deluge \, where pervasive media collect massive amounts of data\, often in a stre aming fashion. The sheer volume makes batch analytics impossible\, and dat a are noisy\, incomplete\, and prone to outliers\, so analytics must often be performed in real-time. This talk introduces an online scalable functi on approximation scheme that adaptively learns and tracks the sought nonli near function on the fly with quantifiable performance guarantees\, even i n adversarial environments. Building on this framework\, a scalable online learning approach with graph feedback is outlined for online learning wit h possibly related models\, showcased on several real-world datasets.
 http s://cml.ics.uci.edu/seminars/2022-11-07-yanning-shen URL:https://cml.ics.uci.edu/seminars/2022-11-07-yanning-shen END:VEVENT BEGIN:VEVENT UID:2022-11-14-muhao-chen@cml.ics.uci.edu DTSTAMP:20221114T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20221114T130000 DTEND;TZID=America/Los_Angeles:20221114T140000 SUMMARY:[CML Seminar] Muhao Chen: Robust and Indirectly Supervised Informat ion Extraction LOCATION:Donald Bren Hall 4011 DESCRIPTION:Muhao Chen\, Assistant Research Professor of Computer Science\, University of Southern California\n\nTitle: Robust and Indirectly Supervi sed Information Extraction\n\nAbstract: Information extraction (IE) is the process of automatically inducing structures of concepts and relations de scribed in natural language text. Despite its importance\, obtaining direc t supervision for IE tasks is difficult\, as it requires expert annotators to read through long documents and identify complex structures. This talk covers recent advances that grant robustness against noise and perturbati on\, prevent systematic errors caused by spurious correlations\, and provi de indirect supervision for label-efficient and logically consistent IE.\n \nhttps://cml.ics.uci.edu/seminars/2022-11-14-muhao-chen X-ALT-DESC;FMTTYPE=text/html:Muhao Chen\, Assistant Rese arch Professor of Computer Science\, University of Southern California

Title: Robust and Indirectly Supervised Information Extraction 
Abstract: Information extraction (IE) is the process of auto matically inducing structures of concepts and relations described in natur al language text. Despite its importance\, obtaining direct supervision fo r IE tasks is difficult\, as it requires expert annotators to read through long documents and identify complex structures. This talk covers recent a dvances that grant robustness against noise and perturbation\, prevent sys tematic errors caused by spurious correlations\, and provide indirect supe rvision for label-efficient and logically consistent IE.

https://cml.ics.uci .edu/seminars/2022-11-14-muhao-chen URL:https://cml.ics.uci.edu/seminars/2022-11-14-muhao-chen END:VEVENT BEGIN:VEVENT UID:2022-11-21-peter-orbanz@cml.ics.uci.edu DTSTAMP:20221121T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20221121T130000 DTEND;TZID=America/Los_Angeles:20221121T140000 SUMMARY:[CML Seminar] Peter Orbanz: Statistical implications of group invar iance of distributions LOCATION:Donald Bren Hall 4011 DESCRIPTION:Peter Orbanz\, Professor of Machine Learning\, Gatsby Computati onal Neuroscience Unit\, University College London\n\nTitle: Statistical i mplications of group invariance of distributions\n\nAbstract: Consider a l arge random structure — a random graph\, a stochastic process on the lin e\, a random field on the grid — and a function that depends only on a s mall part of the structure. Use a family of transformations to move the do main of the function over the structure\, collect each function value\, an d average. Under suitable conditions\, the law of large numbers generalize s to such averages. My recent work with Morgane Austern shows that central limit theorems and other higher-order properties also hold: if the i.i.d. assumption of classical statistics is substituted by suitable properties formulated in terms of groups\, the fundamental theorems of inference stil l hold.\n\nhttps://cml.ics.uci.edu/seminars/2022-11-21-peter-orbanz X-ALT-DESC;FMTTYPE=text/html:Peter Orbanz\, Professor of Machine Learning\, Gatsby Computational Neuroscience Unit\, University Co llege London

Title: Statistical implications of group invari ance of distributions

Abstract: Consider a large random stru cture — a random graph\, a stochastic process on the line\, a random fie ld on the grid — and a function that depends only on a small part of the structure. Use a family of transformations to move the domain of the func tion over the structure\, collect each function value\, and average. Under suitable conditions\, the law of large numbers generalizes to such averag es. My recent work with Morgane Austern shows that central limit theorems and other higher-order properties also hold: if the i.i.d. assumption of c lassical statistics is substituted by suitable properties formulated in te rms of groups\, the fundamental theorems of inference still hold.

< a href="https://cml.ics.uci.edu/seminars/2022-11-21-peter-orbanz">https:// cml.ics.uci.edu/seminars/2022-11-21-peter-orbanz URL:https://cml.ics.uci.edu/seminars/2022-11-21-peter-orbanz END:VEVENT BEGIN:VEVENT UID:2023-01-30-maarten-bos@cml.ics.uci.edu DTSTAMP:20230130T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230130T130000 DTEND;TZID=America/Los_Angeles:20230130T140000 SUMMARY:[CML Seminar] Maarten Bos: Behavioral science research at a corpora te research laboratory LOCATION:Donald Bren Hall 4011 DESCRIPTION:Maarten Bos\, Lead Research Scientist\, Snap Research\n\nTitle: Behavioral science research at a corporate research laboratory\n\nAbstrac t: Corporate research labs aim to push the scientific and technological fo refront of innovation outside traditional academia. Snap Inc. combines aca demia and industry by hiring academic researchers and doing application-dr iven research. In this talk I will give examples of research projects from my corporate research experience. My goal is to showcase the value of — and hurdles for — working both with and within corporate research labs\ , and how some of these values and hurdles are different from working in t raditional academia.\n\nhttps://cml.ics.uci.edu/seminars/2023-01-30-maarte n-bos X-ALT-DESC;FMTTYPE=text/html:Maarten Bos\, Lead Research Scientist\, Snap Research

Title: Behavioral science researc h at a corporate research laboratory

Abstract: Corporate res earch labs aim to push the scientific and technological forefront of innov ation outside traditional academia. Snap Inc. combines academia and indust ry by hiring academic researchers and doing application-driven research. I n this talk I will give examples of research projects from my corporate re search experience. My goal is to showcase the value of — and hurdles for — working both with and within corporate research labs\, and how some o f these values and hurdles are different from working in traditional acade mia.

https://cml.ics.uci.edu/seminars/2023-01-30-maarten-bos URL:https://cml.ics.uci.edu/seminars/2023-01-30-maarten-bos END:VEVENT BEGIN:VEVENT UID:2023-02-06-kolby-nottingham@cml.ics.uci.edu DTSTAMP:20230206T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230206T130000 DTEND;TZID=America/Los_Angeles:20230206T140000 SUMMARY:[CML Seminar] Kolby Nottingham: Large Language Models as External K nowledge Sources for Sequential Decision Making LOCATION:Donald Bren Hall 4011 DESCRIPTION:Kolby Nottingham\, PhD Student\, Department of Computer Science \, University of California\, Irvine\n\nTitle: Large Language Models as Ex ternal Knowledge Sources for Sequential Decision Making\n\nAbstract: While it's common for other machine learning modalities to benefit from model p retraining\, reinforcement learning (RL) agents still typically learn tabu la rasa. Large language models (LLMs)\, trained on internet text\, have be en used as external knowledge sources for RL\, but on their own they are n oisy and lack the grounding necessary to reason in interactive environment s. In this talk\, we will cover methods for grounding LLMs in environment dynamics and applying extracted knowledge to training RL agents. Finally\, we will demonstrate our newly proposed method for applying LLMs to improv ing RL sample efficiency through guided exploration.\n\nhttps://cml.ics.uc i.edu/seminars/2023-02-06-kolby-nottingham X-ALT-DESC;FMTTYPE=text/html:Kolby Nottingham\, PhD Stud ent\, Department of Computer Science\, University of California\, Irvine
Title: Large Language Models as External Knowledge Sources fo r Sequential Decision Making

Abstract: While it's common for other machine learning modalities to benefit from model pretraining\, rei nforcement learning (RL) agents still typically learn tabula rasa. Large l anguage models (LLMs)\, trained on internet text\, have been used as exter nal knowledge sources for RL\, but on their own they are noisy and lack th e grounding necessary to reason in interactive environments. In this talk\ , we will cover methods for grounding LLMs in environment dynamics and app lying extracted knowledge to training RL agents. Finally\, we will demonst rate our newly proposed method for applying LLMs to improving RL sample ef ficiency through guided exploration.

https://cml.ics.uci.edu/seminars/ 2023-02-06-kolby-nottingham URL:https://cml.ics.uci.edu/seminars/2023-02-06-kolby-nottingham END:VEVENT BEGIN:VEVENT UID:2023-02-13-noble-kennamer@cml.ics.uci.edu DTSTAMP:20230213T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230213T130000 DTEND;TZID=America/Los_Angeles:20230213T140000 SUMMARY:[CML Seminar] Noble Kennamer: Variational Methods for Bayesian Opti mal Experimental Design LOCATION:Donald Bren Hall 4011 DESCRIPTION:Noble Kennamer\, PhD Student\, Department of Computer Science\, University of California\, Irvine\n\nTitle: Variational Methods for Bayes ian Optimal Experimental Design\n\nAbstract: Bayesian optimal experimental design is a sub-field of statistics focused on developing methods to make efficient use of experimental resources. Any potential design is evaluate d in terms of a utility function\, such as the expected information gain ( EIG)\; unfortunately\, under most circumstances the EIG is intractable to evaluate. In this talk we build off successful variational approaches\, wh ich optimize a parameterized variational model with respect to bounds on t he EIG. We present a novel neural architecture that allows experimenters t o optimize a single variational model that can estimate the EIG for potent ially infinitely many designs. We demonstrate the effectiveness of our tec hnique on generalized linear models\, showing that our method greatly impr oves accuracy over existing approximation strategies with far better sampl e efficiency.\n\nhttps://cml.ics.uci.edu/seminars/2023-02-13-noble-kenname r X-ALT-DESC;FMTTYPE=text/html:Noble Kennamer\, PhD Studen t\, Department of Computer Science\, University of California\, Irvine

Title: Variational Methods for Bayesian Optimal Experimental De sign

Abstract: Bayesian optimal experimental design is a sub -field of statistics focused on developing methods to make efficient use o f experimental resources. Any potential design is evaluated in terms of a utility function\, such as the expected information gain (EIG)\; unfortuna tely\, under most circumstances the EIG is intractable to evaluate. In thi s talk we build off successful variational approaches\, which optimize a p arameterized variational model with respect to bounds on the EIG. We prese nt a novel neural architecture that allows experimenters to optimize a sin gle variational model that can estimate the EIG for potentially infinitely many designs. We demonstrate the effectiveness of our technique on genera lized linear models\, showing that our method greatly improves accuracy ov er existing approximation strategies with far better sample efficiency.

https://cml.ics.uci.edu/seminars/2023-02-13-noble-kennamer URL:https://cml.ics.uci.edu/seminars/2023-02-13-noble-kennamer END:VEVENT BEGIN:VEVENT UID:2023-03-06-shlomo-zilberstein@cml.ics.uci.edu DTSTAMP:20230306T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230306T130000 DTEND;TZID=America/Los_Angeles:20230306T140000 SUMMARY:[CML Seminar] Shlomo Zilberstein: Competence-Aware Systems LOCATION:Donald Bren Hall 4011 DESCRIPTION:Shlomo Zilberstein\, Professor of Computer Science\, University of Massachusetts\, Amherst\n\nTitle: Competence-Aware Systems\n\nAbstract : Competence is the ability to do something well. Competence awareness is the ability to represent and learn a model of self competence and use it t o decide how to best use the agent's own abilities as well as any availabl e human assistance. This capability is critical for the success and safety of autonomous systems that operate in the open world. In this talk\, I in troduce two types of competence-aware systems (CAS): Type I refers to a st and-alone system that can learn its own competence and fine-tune itself to the problem instance without human assistance\; Type II is a human-aware system that uses a self-competence model to optimize the utilization of co stly human assistive actions. I describe recent results that demonstrate t he benefits of the two types of competence awareness\, including autonomou s vehicle decision making.\n\nhttps://cml.ics.uci.edu/seminars/2023-03-06- shlomo-zilberstein X-ALT-DESC;FMTTYPE=text/html:Shlomo Zilberstein\, Profes sor of Computer Science\, University of Massachusetts\, Amherst

Title: Competence-Aware Systems

Abstract: Competence is the ability to do something well. Competence awareness is the ability to r epresent and learn a model of self competence and use it to decide how to best use the agent's own abilities as well as any available human assistan ce. This capability is critical for the success and safety of autonomous s ystems that operate in the open world. In this talk\, I introduce two type s of competence-aware systems (CAS): Type I refers to a stand-alone system that can learn its own competence and fine-tune itself to the problem ins tance without human assistance\; Type II is a human-aware system that uses a self-competence model to optimize the utilization of costly human assis tive actions. I describe recent results that demonstrate the benefits of t he two types of competence awareness\, including autonomous vehicle decisi on making.

https://cml.ics.uci.edu/seminars/2023-03-06-shlomo-zilber stein URL:https://cml.ics.uci.edu/seminars/2023-03-06-shlomo-zilberstein END:VEVENT BEGIN:VEVENT UID:2023-04-10-durk-kingma@cml.ics.uci.edu DTSTAMP:20230410T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230410T130000 DTEND;TZID=America/Los_Angeles:20230410T140000 SUMMARY:[CML Seminar] Durk Kingma: Understanding the Diffusion Objective LOCATION:Donald Bren Hall 4011 DESCRIPTION:Durk Kingma\, Research Scientist\, Google Research\n\nTitle: Un derstanding the Diffusion Objective\n\nAbstract: Some believe that maximum likelihood is incompatible with high-quality image generation. We provide counter-evidence: diffusion models with SOTA FIDs are actually optimized with the ELBO\, with very simple data augmentation (additive noise). We sh ow that diffusion models in the literature are optimized with various obje ctives that are special cases of a weighted loss\, where the weighting fun ction specifies the weight per noise level. Uniform weighting corresponds to maximizing the ELBO\, a principled approximation of maximum likelihood. We expose a direct relationship between the weighted loss (with any weigh ting) and the ELBO objective: the weighted loss can be written as a weight ed integral of ELBOs\, with one ELBO per noise level. If the weighting fun ction is monotonic\, then the weighted loss is a likelihood-based objectiv e. Our main contribution is a deeper theoretical understanding of the diff usion objective.\n\nhttps://cml.ics.uci.edu/seminars/2023-04-10-durk-kingm a X-ALT-DESC;FMTTYPE=text/html:Durk Kingma\, Research Scie ntist\, Google Research

Title: Understanding the Diffusion O bjective

Abstract: Some believe that maximum likelihood is i ncompatible with high-quality image generation. We provide counter-evidenc e: diffusion models with SOTA FIDs are actually optimized with the ELBO\, with very simple data augmentation (additive noise). We show that diffusio n models in the literature are optimized with various objectives that are special cases of a weighted loss\, where the weighting function specifies the weight per noise level. Uniform weighting corresponds to maximizing th e ELBO\, a principled approximation of maximum likelihood. We expose a dir ect relationship between the weighted loss (with any weighting) and the EL BO objective: the weighted loss can be written as a weighted integral of E LBOs\, with one ELBO per noise level. If the weighting function is monoton ic\, then the weighted loss is a likelihood-based objective. Our main cont ribution is a deeper theoretical understanding of the diffusion objective.

https://cml.ics.uci.edu/seminars/2023-04-10-durk-kingma URL:https://cml.ics.uci.edu/seminars/2023-04-10-durk-kingma END:VEVENT BEGIN:VEVENT UID:2023-04-17-danish-pruthi@cml.ics.uci.edu DTSTAMP:20230417T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230417T130000 DTEND;TZID=America/Los_Angeles:20230417T140000 SUMMARY:[CML Seminar] Danish Pruthi: Evaluating Explanations LOCATION:Donald Bren Hall 4011 DESCRIPTION:Danish Pruthi\, Assistant Professor\, Department of Computation al and Data Sciences\, Indian Institute of Science (IISc)\, Bangalore\n\nT itle: Evaluating Explanations\n\nAbstract: While large deep learning model s have become increasingly accurate\, concerns about their (lack of) inter pretability have taken center stage. In response\, a growing subfield on i nterpretability and analysis of these models has emerged. While hundreds o f techniques have been proposed to explain predictions of models\, what ai ms these explanations serve and how they ought to be evaluated are often u nstated. In this talk\, I will present a framework to quantify the value o f explanations\, along with specific applications in a variety of contexts . I would end with some of my thoughts on evaluating large language models and the rationales they generate.\n\nhttps://cml.ics.uci.edu/seminars/202 3-04-17-danish-pruthi X-ALT-DESC;FMTTYPE=text/html:Danish Pruthi\, Assistant P rofessor\, Department of Computational and Data Sciences\, Indian Institut e of Science (IISc)\, Bangalore

Title: Evaluating Explanatio ns

Abstract: While large deep learning models have become in creasingly accurate\, concerns about their (lack of) interpretability have taken center stage. In response\, a growing subfield on interpretability and analysis of these models has emerged. While hundreds of techniques hav e been proposed to explain predictions of models\, what aims these explana tions serve and how they ought to be evaluated are often unstated. In this talk\, I will present a framework to quantify the value of explanations\, along with specific applications in a variety of contexts. I would end wi th some of my thoughts on evaluating large language models and the rationa les they generate.

https://cml.ics.uci.edu/seminars/2023-04-17-danish-pru thi URL:https://cml.ics.uci.edu/seminars/2023-04-17-danish-pruthi END:VEVENT BEGIN:VEVENT UID:2023-04-24-anthony-chen@cml.ics.uci.edu DTSTAMP:20230424T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230424T130000 DTEND;TZID=America/Los_Angeles:20230424T140000 SUMMARY:[CML Seminar] Anthony Chen: Researching and Revising What Language Models Say\, Using Language Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Anthony Chen\, PhD Student\, Department of Computer Science\, U C Irvine\n\nTitle: Researching and Revising What Language Models Say\, Usi ng Language Models\n\nAbstract: As the strengths of large language models (LLMs) have become prominent\, so too have their weaknesses. A glaring wea kness of LLMs is their penchant for generating false\, biased\, or mislead ing claims in a phenomena broadly referred to as hallucinations. Most LLMs also do not ground their generations to any source. To enable attribution while still preserving all the powerful advantages of LLMs\, we propose R ARR (Retrofit Attribution using Research and Revision)\, a system that aut omatically retrieves evidence to support the output of any LLM and then po st-edits the output to fix any information that contradicts the retrieved evidence while preserving the original output as much as possible. When ap plied to several state-of-the-art LLMs\, RARR significantly improves attri bution.\n\nhttps://cml.ics.uci.edu/seminars/2023-04-24-anthony-chen X-ALT-DESC;FMTTYPE=text/html:Anthony Chen\, PhD Student\ , Department of Computer Science\, UC Irvine

Title: Research ing and Revising What Language Models Say\, Using Language Models

Abstract: As the strengths of large language models (LLMs) have beco me prominent\, so too have their weaknesses. A glaring weakness of LLMs is their penchant for generating false\, biased\, or misleading claims in a phenomena broadly referred to as hallucinations. Most LLMs also do not gro und their generations to any source. To enable attribution while still pre serving all the powerful advantages of LLMs\, we propose RARR (Retrofit At tribution using Research and Revision)\, a system that automatically retri eves evidence to support the output of any LLM and then post-edits the out put to fix any information that contradicts the retrieved evidence while p reserving the original output as much as possible. When applied to several state-of-the-art LLMs\, RARR significantly improves attribution.

< a href="https://cml.ics.uci.edu/seminars/2023-04-24-anthony-chen">https:// cml.ics.uci.edu/seminars/2023-04-24-anthony-chen URL:https://cml.ics.uci.edu/seminars/2023-04-24-anthony-chen END:VEVENT BEGIN:VEVENT UID:2023-05-01-hengrui-cai@cml.ics.uci.edu DTSTAMP:20230501T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230501T130000 DTEND;TZID=America/Los_Angeles:20230501T140000 SUMMARY:[CML Seminar] Hengrui Cai: Towards Causal Revolution: On Learning H eterogeneity and Non-Spuriousness in Causal Graphs LOCATION:Donald Bren Hall 4011 DESCRIPTION:Hengrui Cai\, Assistant Professor of Statistics\, University of California\, Irvine\n\nTitle: Towards Causal Revolution: On Learning Hete rogeneity and Non-Spuriousness in Causal Graphs\n\nAbstract: The causal re volution has spurred interest in understanding complex relationships in va rious fields. Under a general causal graph\, the exposure may have a direc t effect on the outcome and also an indirect effect regulated by a set of mediators. In this talk\, we introduce a new statistical framework to comp rehensively characterize causal effects with multiple mediators\, namely A Nalysis Of Causal Effects (ANOCE). Built upon such causal impact learning\ , we focus on two emerging challenges: heterogeneity and spuriousness. We conceptualize heterogeneous causal graphs (HCGs) and propose to learn a cl ass of necessary and sufficient causal graphs (NSCG) that only contain cau sally relevant variables by utilizing the probabilities of causation. Acro ss empirical studies\, we show that the proposed algorithms outperform exi sting ones and can reveal true heterogeneous and non-spurious causal graph s.\n\nhttps://cml.ics.uci.edu/seminars/2023-05-01-hengrui-cai X-ALT-DESC;FMTTYPE=text/html:Hengrui Cai\, Assistant Pro fessor of Statistics\, University of California\, Irvine

Title: Towards Causal Revolution: On Learning Heterogeneity and Non-Spuriousn ess in Causal Graphs

Abstract: The causal revolution has spu rred interest in understanding complex relationships in various fields. Un der a general causal graph\, the exposure may have a direct effect on the outcome and also an indirect effect regulated by a set of mediators. In th is talk\, we introduce a new statistical framework to comprehensively char acterize causal effects with multiple mediators\, namely ANalysis Of Causa l Effects (ANOCE). Built upon such causal impact learning\, we focus on tw o emerging challenges: heterogeneity and spuriousness. We conceptualize he terogeneous causal graphs (HCGs) and propose to learn a class of necessary and sufficient causal graphs (NSCG) that only contain causally relevant v ariables by utilizing the probabilities of causation. Across empirical stu dies\, we show that the proposed algorithms outperform existing ones and c an reveal true heterogeneous and non-spurious causal graphs.

https://cml.ic s.uci.edu/seminars/2023-05-01-hengrui-cai URL:https://cml.ics.uci.edu/seminars/2023-05-01-hengrui-cai END:VEVENT BEGIN:VEVENT UID:2023-05-08-pierre-baldi-and-alexander-shmakov@cml.ics.uci.edu DTSTAMP:20230508T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230508T130000 DTEND;TZID=America/Los_Angeles:20230508T140000 SUMMARY:[CML Seminar] Pierre Baldi and Alexander Shmakov: Deep Learning in Science LOCATION:Donald Bren Hall 4011 DESCRIPTION:Pierre Baldi and Alexander Shmakov\, Department of Computer Sci ence\, UC Irvine\n\nTitle: Deep Learning in Science\n\nAbstract: The Baldi group will present ongoing progress in the theory and applications of dee p learning. On the theory side\, we will discuss homogeneous activation fu nctions and their important connections to the concept of generalized neur al balance. On the application side\, we will present applications of neur al transformers to physics\, in particular for the assignment of observati on measurements to the leaves of partial Feynman diagrams in particle phys ics. In these applications\, the permutation invariance properties of tran sformers are used to capture fundamental symmetries (e.g. matter vs antima tter) in the laws of physics.\n\nhttps://cml.ics.uci.edu/seminars/2023-05- 08-pierre-baldi-and-alexander-shmakov X-ALT-DESC;FMTTYPE=text/html:Pierre Baldi and Alexander Shma kov\, Department of Computer Science\, UC Irvine

Title: Deep Learning in Science

Abstract: The Baldi group will pres ent ongoing progress in the theory and applications of deep learning. On t he theory side\, we will discuss homogeneous activation functions and thei r important connections to the concept of generalized neural balance. On t he application side\, we will present applications of neural transformers to physics\, in particular for the assignment of observation measurements to the leaves of partial Feynman diagrams in particle physics. In these ap plications\, the permutation invariance properties of transformers are use d to capture fundamental symmetries (e.g. matter vs antimatter) in the law s of physics.

https://cml.ics.uci.edu/seminars/2023- 05-08-pierre-baldi-and-alexander-shmakov URL:https://cml.ics.uci.edu/seminars/2023-05-08-pierre-baldi-and-alexander- shmakov END:VEVENT BEGIN:VEVENT UID:2023-05-15-guy-van-den-broeck@cml.ics.uci.edu DTSTAMP:20230515T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230515T130000 DTEND;TZID=America/Los_Angeles:20230515T140000 SUMMARY:[CML Seminar] Guy Van den Broeck: AI can Learn from Data. But can i t Learn to Reason? LOCATION:Donald Bren Hall 4011 DESCRIPTION:Guy Van den Broeck\, Associate Professor of Computer Science\, University of California\, Los Angeles\n\nTitle: AI can Learn from Data. B ut can it Learn to Reason?\n\nAbstract: Many expect that AI will go from p owering chatbots to providing mental health services\, from advertisement to deciding who is given bail. The expectation is that AI will solve socie ty's problems by simply being more intelligent than we are. Implicit in th is bullish perspective is the assumption that AI will naturally learn to r eason from data: that it can form trains of thought that make sense\, simi lar to how a mental health professional or judge might reason about a case \, or how a mathematician might prove a theorem. This talk will investigat e whether this behavior can be learned from data\, and how we can design t he next generation of AI techniques that can achieve such capabilities\, f ocusing on neuro-symbolic learning and tractable deep generative models.\n \nhttps://cml.ics.uci.edu/seminars/2023-05-15-guy-van-den-broeck X-ALT-DESC;FMTTYPE=text/html:Guy Van den Broeck\, Associ ate Professor of Computer Science\, University of California\, Los Angeles

Title: AI can Learn from Data. But can it Learn to Reason? 
Abstract: Many expect that AI will go from powering chatbots to providing mental health services\, from advertisement to deciding who is given bail. The expectation is that AI will solve society's problems by simply being more intelligent than we are. Implicit in this bullish persp ective is the assumption that AI will naturally learn to reason from data: that it can form trains of thought that make sense\, similar to how a men tal health professional or judge might reason about a case\, or how a math ematician might prove a theorem. This talk will investigate whether this b ehavior can be learned from data\, and how we can design the next generati on of AI techniques that can achieve such capabilities\, focusing on neuro -symbolic learning and tractable deep generative models.

https://cml .ics.uci.edu/seminars/2023-05-15-guy-van-den-broeck URL:https://cml.ics.uci.edu/seminars/2023-05-15-guy-van-den-broeck END:VEVENT BEGIN:VEVENT UID:2023-05-22-gabe-hope@cml.ics.uci.edu DTSTAMP:20230522T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230522T130000 DTEND;TZID=America/Los_Angeles:20230522T140000 SUMMARY:[CML Seminar] Gabe Hope: Semi-Supervised Learning with Prediction-C onstrained Variational Autoencoders LOCATION:Donald Bren Hall 4011 DESCRIPTION:Gabe Hope\, PhD Student\, Computer Science\, University of Cali fornia\, Irvine\n\nTitle: Semi-Supervised Learning with Prediction-Constra ined Variational Autoencoders\n\nAbstract: Variational autoencoders (VAEs) have proven to be an effective approach to modeling complex data distribu tions while providing compact representations useful for downstream predic tion tasks. In this work we train VAEs with the dual goals of good likelih ood-based generative modeling and good discriminative performance in super vised and semi-supervised prediction tasks. We show that the dominant appr oach to training semi-supervised VAEs has key weaknesses\, and propose a n ovel framework that maximizes generative likelihood subject to prediction quality constraints. To handle sparse labels\, we further enforce a consis tency constraint requiring predictions on reconstructed data to match thos e on the original data. Our experiments show that prediction and consisten cy constraints improve generative samples as well as image classification performance in semi-supervised settings.\n\nhttps://cml.ics.uci.edu/semina rs/2023-05-22-gabe-hope X-ALT-DESC;FMTTYPE=text/html:Gabe Hope\, PhD Student\, C omputer Science\, University of California\, Irvine

Title: S emi-Supervised Learning with Prediction-Constrained Variational Autoencode rs

Abstract: Variational autoencoders (VAEs) have proven to be an effective approach to modeling complex data distributions while prov iding compact representations useful for downstream prediction tasks. In t his work we train VAEs with the dual goals of good likelihood-based genera tive modeling and good discriminative performance in supervised and semi-s upervised prediction tasks. We show that the dominant approach to training semi-supervised VAEs has key weaknesses\, and propose a novel framework t hat maximizes generative likelihood subject to prediction quality constrai nts. To handle sparse labels\, we further enforce a consistency constraint requiring predictions on reconstructed data to match those on the origina l data. Our experiments show that prediction and consistency constraints i mprove generative samples as well as image classification performance in s emi-supervised settings.

https://cml.ics.uci.edu/seminars/2023-05-22-gabe-hop e URL:https://cml.ics.uci.edu/seminars/2023-05-22-gabe-hope END:VEVENT BEGIN:VEVENT UID:2023-06-05-sangeetha-abdu-jyothi@cml.ics.uci.edu DTSTAMP:20230605T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230605T130000 DTEND;TZID=America/Los_Angeles:20230605T140000 SUMMARY:[CML Seminar] Sangeetha Abdu Jyothi: CrystalBox: Future-Based Expla nations for Deep RL Network Controllers LOCATION:Donald Bren Hall 4011 DESCRIPTION:Sangeetha Abdu Jyothi\, Assistant Professor of Computer Science \, University of California\, Irvine\n\nTitle: CrystalBox: Future-Based Ex planations for Deep RL Network Controllers\n\nAbstract: Lack of explainabi lity is a key factor limiting the practical adoption of high-performant De ep Reinforcement Learning (DRL) controllers in systems environments. Expla inable RL for networking hitherto used salient input features to interpret a controller's behavior. However\, these feature-based solutions do not c ompletely explain the controller's decision-making process. In this talk\, I will present CrystalBox\, a framework that explains a controller's beha vior in terms of the future impact on key network performance metrics. Cry stalBox employs a novel learning-based approach to generate succinct and e xpressive explanations\, using reward components of the DRL controller as the basis. I will present three practical use cases: cross-state explainab ility\, guided reward design\, and network observability.\n\nhttps://cml.i cs.uci.edu/seminars/2023-06-05-sangeetha-abdu-jyothi X-ALT-DESC;FMTTYPE=text/html:Sangeetha Abdu Jyothi\, Ass istant Professor of Computer Science\, University of California\, Irvine
Title: CrystalBox: Future-Based Explanations for Deep RL Netw ork Controllers

Abstract: Lack of explainability is a key fa ctor limiting the practical adoption of high-performant Deep Reinforcement Learning (DRL) controllers in systems environments. Explainable RL for ne tworking hitherto used salient input features to interpret a controller's behavior. However\, these feature-based solutions do not completely explai n the controller's decision-making process. In this talk\, I will present CrystalBox\, a framework that explains a controller's behavior in terms of the future impact on key network performance metrics. CrystalBox employs a novel learning-based approach to generate succinct and expressive explan ations\, using reward components of the DRL controller as the basis. I wil l present three practical use cases: cross-state explainability\, guided r eward design\, and network observability.

https://cml.ics.uci.edu /seminars/2023-06-05-sangeetha-abdu-jyothi URL:https://cml.ics.uci.edu/seminars/2023-06-05-sangeetha-abdu-jyothi END:VEVENT BEGIN:VEVENT UID:2023-07-20-vincent-fortuin@cml.ics.uci.edu DTSTAMP:20230720T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20230720T110000 DTEND;TZID=America/Los_Angeles:20230720T120000 SUMMARY:[CML Seminar] Vincent Fortuin: Use Cases for Bayesian Deep Learning in the Age of ChatGPT LOCATION:Donald Bren Hall 3011 DESCRIPTION:Vincent Fortuin\, Research group leader in Machine Learning\, H elmholtz AI\n\nTitle: Use Cases for Bayesian Deep Learning in the Age of C hatGPT\n\nAbstract: Many researchers have pondered the same existential qu estions since the release of ChatGPT: Is scale really all you need? Will t he future of machine learning rely exclusively on foundation models? In th is talk\, I will try to make the case that the answer should be a convince d no and that now\, maybe more than ever\, should be the time to focus on fundamental questions in machine learning again. I will provide evidence b y presenting three modern use cases of Bayesian deep learning in the areas of self-supervised learning\, interpretable additive modeling\, and seque ntial decision making. Together\, these will show that the research field of Bayesian deep learning is very much alive and thriving.\n\nhttps://cml. ics.uci.edu/seminars/2023-07-20-vincent-fortuin X-ALT-DESC;FMTTYPE=text/html:Vincent Fortuin\, Research group leader in Machine Learning\, Helmholtz AI

Title: Use C ases for Bayesian Deep Learning in the Age of ChatGPT

Abstract: Many researchers have pondered the same existential questions since th e release of ChatGPT: Is scale really all you need? Will the future of mac hine learning rely exclusively on foundation models? In this talk\, I will try to make the case that the answer should be a convinced no and that no w\, maybe more than ever\, should be the time to focus on fundamental ques tions in machine learning again. I will provide evidence by presenting thr ee modern use cases of Bayesian deep learning in the areas of self-supervi sed learning\, interpretable additive modeling\, and sequential decision m aking. Together\, these will show that the research field of Bayesian deep learning is very much alive and thriving.

https://cml.ics.uci.edu/semi nars/2023-07-20-vincent-fortuin URL:https://cml.ics.uci.edu/seminars/2023-07-20-vincent-fortuin END:VEVENT BEGIN:VEVENT UID:2023-10-16-marius-kloft@cml.ics.uci.edu DTSTAMP:20231016T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20231016T130000 DTEND;TZID=America/Los_Angeles:20231016T140000 SUMMARY:[CML Seminar] Marius Kloft: Deep Anomaly Detection LOCATION:Donald Bren Hall 4011 DESCRIPTION:Marius Kloft\, Professor of Computer Science\, RPTU Kaiserslaut ern-Landau\, Germany\n\nTitle: Deep Anomaly Detection\n\nAbstract: Anomaly detection is one of the fundamental topics in machine learning and artifi cial intelligence. The aim is to find instances deviating from the norm — so-called anomalies. Anomalies can be observed in various scenarios\, from attacks on computer or energy networks to critical faults in a chemic al factory or rare tumors in cancer imaging data. In my talk\, I will firs t introduce the field of anomaly detection\, with an emphasis on deep anom aly detection. Then\, I will present recent algorithms and theory for deep anomaly detection\, with images as primary data type. I will demonstrate how these methods can be better understood using explainable AI methods. I will show new algorithms for deep anomaly detection on other data types\, such as time series\, graphs\, tabular data\, and contaminated data. Fina lly\, I will close my talk with an outlook on exciting future research dir ections in anomaly detection and beyond.\n\nhttps://cml.ics.uci.edu/semina rs/2023-10-16-marius-kloft X-ALT-DESC;FMTTYPE=text/html:Marius Kloft\, Professor of Computer Science\, RPTU Kaiserslautern-Landau\, Germany

Title: Deep Anomaly Detection

Abstract: Anomaly detection is on e of the fundamental topics in machine learning and artificial intelligenc e. The aim is to find instances deviating from the norm — so-called anom alies. Anomalies can be observed in various scenarios\, from attacks on co mputer or energy networks to critical faults in a chemical factory or rare tumors in cancer imaging data. In my talk\, I will first introduce the fi eld of anomaly detection\, with an emphasis on deep anomaly detection. The n\, I will present recent algorithms and theory for deep anomaly detection \, with images as primary data type. I will demonstrate how these methods can be better understood using explainable AI methods. I will show new alg orithms for deep anomaly detection on other data types\, such as time seri es\, graphs\, tabular data\, and contaminated data. Finally\, I will close my talk with an outlook on exciting future research directions in anomaly detection and beyond.

https://cml.ics.uci.edu/seminars/2023-10-16-marius- kloft URL:https://cml.ics.uci.edu/seminars/2023-10-16-marius-kloft END:VEVENT BEGIN:VEVENT UID:2023-10-23-sarah-wiegreffe@cml.ics.uci.edu DTSTAMP:20231023T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20231023T130000 DTEND;TZID=America/Los_Angeles:20231023T140000 SUMMARY:[CML Seminar] Sarah Wiegreffe: Towards Transparent Language Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Sarah Wiegreffe\, Postdoctoral Researcher\, Allen Institute for AI and University of Washington\n\nTitle: Towards Transparent Language Mo dels\n\nAbstract: Recently-released language models have attracted a lot o f attention for their major successes and (often more subtle\, but still p lentiful) failures. In this talk\, I will motivate why transparency into m odel operations is needed to rectify these failures and increase model uti lity in a reliable way. I will highlight how techniques must be developed in this changing NLP landscape for both open-source models and black-box m odels behind an API. I will provide an example of each from my recent work demonstrating how improved transparency can improve language model perfor mance on downstream tasks.\n\nhttps://cml.ics.uci.edu/seminars/2023-10-23- sarah-wiegreffe X-ALT-DESC;FMTTYPE=text/html:Sarah Wiegreffe\, Postdocto ral Researcher\, Allen Institute for AI and University of Washington
Title: Towards Transparent Language Models

Abstract: Recently-released language models have attracted a lot of attention for their major successes and (often more subtle\, but still plentiful) failur es. In this talk\, I will motivate why transparency into model operations is needed to rectify these failures and increase model utility in a reliab le way. I will highlight how techniques must be developed in this changing NLP landscape for both open-source models and black-box models behind an API. I will provide an example of each from my recent work demonstrating h ow improved transparency can improve language model performance on downstr eam tasks.

https://cml.ics.uci.edu/seminars/2023-10-23-sarah-wiegreffe< /a> URL:https://cml.ics.uci.edu/seminars/2023-10-23-sarah-wiegreffe END:VEVENT BEGIN:VEVENT UID:2023-10-30-noga-zaslavsky@cml.ics.uci.edu DTSTAMP:20231030T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20231030T130000 DTEND;TZID=America/Los_Angeles:20231030T140000 SUMMARY:[CML Seminar] Noga Zaslavsky: Losing bits and finding meaning: Effi cient compression underlies meaning in language LOCATION:Donald Bren Hall 4011 DESCRIPTION:Noga Zaslavsky\, Assistant Professor of Language Science\, Univ ersity of California\, Irvine\n\nTitle: Losing bits and finding meaning: E fficient compression underlies meaning in language\n\nAbstract: Our world is extremely complex\, and yet we are able to exchange our thoughts and be liefs about it using a relatively small number of words. What computationa l principles can explain this extraordinary ability? In this talk\, I argu e that in order to communicate and reason about meaning while operating un der limited resources\, both humans and machines must efficiently compress their representations of the world. In support of this claim\, I present a series of studies showing that: (i) human languages evolve under pressur e to efficiently compress meanings into words via the Information Bottlene ck (IB) principle\; (ii) the same principle can help ground meaning repres entations in artificial neural networks trained for vision\; and (iii) the se findings offer a new framework for emergent communication in artificial agents. Taken together\, these results suggest that efficient compression underlies meaning in language.\n\nhttps://cml.ics.uci.edu/seminars/2023-1 0-30-noga-zaslavsky X-ALT-DESC;FMTTYPE=text/html:Noga Zaslavsky\, Assistant Professor of Language Science\, University of California\, Irvine

Title: Losing bits and finding meaning: Efficient compression underl ies meaning in language

Abstract: Our world is extremely com plex\, and yet we are able to exchange our thoughts and beliefs about it u sing a relatively small number of words. What computational principles can explain this extraordinary ability? In this talk\, I argue that in order to communicate and reason about meaning while operating under limited reso urces\, both humans and machines must efficiently compress their represent ations of the world. In support of this claim\, I present a series of stud ies showing that: (i) human languages evolve under pressure to efficiently compress meanings into words via the Information Bottleneck (IB) principl e\; (ii) the same principle can help ground meaning representations in art ificial neural networks trained for vision\; and (iii) these findings offe r a new framework for emergent communication in artificial agents. Taken t ogether\, these results suggest that efficient compression underlies meani ng in language.

https://cml.ics.uci.edu/seminars/2023-10-30-noga-zaslavs ky URL:https://cml.ics.uci.edu/seminars/2023-10-30-noga-zaslavsky END:VEVENT BEGIN:VEVENT UID:2023-11-06-mariel-werner@cml.ics.uci.edu DTSTAMP:20231106T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20231106T130000 DTEND;TZID=America/Los_Angeles:20231106T140000 SUMMARY:[CML Seminar] Mariel Werner: Provably Personalized and Robust Feder ated Learning LOCATION:Donald Bren Hall 4011 DESCRIPTION:Mariel Werner\, PhD Student\, Department of Electrical Engineer ing and Computer Science\, UC Berkeley\n\nTitle: Provably Personalized and Robust Federated Learning\n\nAbstract: I will be discussing my recent wor k on personalization in federated learning. Federated learning is a powerf ul distributed optimization framework in which multiple clients collaborat ively train a global model without sharing their raw data. In this work\, we tackle the personalized version of the federated learning problem. In p articular\, we ask: throughout the training process\, can clients identify a subset of similar clients and collaboratively train with just those cli ents? In the affirmative\, we propose simple clustering-based methods whic h are provably optimal for a broad class of loss functions\, are robust to malicious attackers\, and perform well in practice.\n\nhttps://cml.ics.uc i.edu/seminars/2023-11-06-mariel-werner X-ALT-DESC;FMTTYPE=text/html:Mariel Werner\, PhD Student \, Department of Electrical Engineering and Computer Science\, UC Berkeley

Title: Provably Personalized and Robust Federated Learning 
Abstract: I will be discussing my recent work on personaliza tion in federated learning. Federated learning is a powerful distributed o ptimization framework in which multiple clients collaboratively train a gl obal model without sharing their raw data. In this work\, we tackle the pe rsonalized version of the federated learning problem. In particular\, we a sk: throughout the training process\, can clients identify a subset of sim ilar clients and collaboratively train with just those clients? In the aff irmative\, we propose simple clustering-based methods which are provably o ptimal for a broad class of loss functions\, are robust to malicious attac kers\, and perform well in practice.

https://cml.ics.uci.edu/seminars/202 3-11-06-mariel-werner URL:https://cml.ics.uci.edu/seminars/2023-11-06-mariel-werner END:VEVENT BEGIN:VEVENT UID:2023-11-13-yian-ma@cml.ics.uci.edu DTSTAMP:20231113T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20231113T130000 DTEND;TZID=America/Los_Angeles:20231113T140000 SUMMARY:[CML Seminar] Yian Ma: MCMC\, variational inference\, and reverse d iffusion Monte Carlo LOCATION:Donald Bren Hall 4011 DESCRIPTION:Yian Ma\, Assistant Professor\, Halıcıoğlu Data Science Inst itute\, University of California\, San Diego\n\nTitle: MCMC\, variational inference\, and reverse diffusion Monte Carlo\n\nAbstract: I will introduc e some recent progress towards understanding the scalability of Markov cha in Monte Carlo (MCMC) methods and their comparative advantage with respect to variational inference. I will fact-check the folklore that variational inference is fast but biased\, MCMC is unbiased but slow. I will then dis cuss a combination of the two via reverse diffusion\, which holds promise of solving some of the multi-modal problems. This talk will be motivated b y the need for Bayesian computation in reinforcement learning problems as well as the differential privacy requirements that we face.\n\nhttps://cml .ics.uci.edu/seminars/2023-11-13-yian-ma X-ALT-DESC;FMTTYPE=text/html:Yian Ma\, Assistant Profess or\, Halıcıoğlu Data Science Institute\, University of California\, San Diego

Title: MCMC\, variational inference\, and reverse dif fusion Monte Carlo

Abstract: I will introduce some recent pr ogress towards understanding the scalability of Markov chain Monte Carlo ( MCMC) methods and their comparative advantage with respect to variational inference. I will fact-check the folklore that variational inference is fa st but biased\, MCMC is unbiased but slow. I will then discuss a combinati on of the two via reverse diffusion\, which holds promise of solving some of the multi-modal problems. This talk will be motivated by the need for B ayesian computation in reinforcement learning problems as well as the diff erential privacy requirements that we face.

https://cml.ics.uci.edu/seminars/20 23-11-13-yian-ma URL:https://cml.ics.uci.edu/seminars/2023-11-13-yian-ma END:VEVENT BEGIN:VEVENT UID:2023-11-20-yuhua-zhu@cml.ics.uci.edu DTSTAMP:20231120T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20231120T130000 DTEND;TZID=America/Los_Angeles:20231120T140000 SUMMARY:[CML Seminar] Yuhua Zhu: Continuous-in-time Limit for Multi-armed B andit LOCATION:Donald Bren Hall 4011 DESCRIPTION:Yuhua Zhu\, Assistant Professor\, Halicioglu Data Science Insti tute and Department of Mathematics\, University of California\, San Diego\ n\nTitle: Continuous-in-time Limit for Multi-armed Bandit\n\nAbstract: In this talk\, I will build the connection between Hamilton-Jacobi-Bellman eq uations (HJB) and the multi-armed bandit (MAB) problems. HJB is an importa nt equation in solving stochastic optimal control problems. MAB is a widel y used paradigm for studying the exploration-exploitation trade-off in seq uential decision making under uncertainty. This is the first work that est ablishes this connection in a general setting. I will present an efficient algorithm for solving MAB problems based on this connection and demonstra te its practical applications.\n\nhttps://cml.ics.uci.edu/seminars/2023-11 -20-yuhua-zhu X-ALT-DESC;FMTTYPE=text/html:Yuhua Zhu\, Assistant Profe ssor\, Halicioglu Data Science Institute and Department of Mathematics\, U niversity of California\, San Diego

Title: Continuous-in-tim e Limit for Multi-armed Bandit

Abstract: In this talk\, I wi ll build the connection between Hamilton-Jacobi-Bellman equations (HJB) an d the multi-armed bandit (MAB) problems. HJB is an important equation in s olving stochastic optimal control problems. MAB is a widely used paradigm for studying the exploration-exploitation trade-off in sequential decision making under uncertainty. This is the first work that establishes this co nnection in a general setting. I will present an efficient algorithm for s olving MAB problems based on this connection and demonstrate its practical applications.

https://cml.ics.uci.edu/seminars/2023-11-20-yuhua-zhu URL:https://cml.ics.uci.edu/seminars/2023-11-20-yuhua-zhu END:VEVENT BEGIN:VEVENT UID:2023-11-21-yejin-choi@cml.ics.uci.edu DTSTAMP:20231121T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20231121T110000 DTEND;TZID=America/Los_Angeles:20231121T120000 SUMMARY:[CML Seminar] Yejin Choi: Possible Impossibilities and Impossible P ossibilities LOCATION:Donald Bren Hall 4011 DESCRIPTION:Yejin Choi\, Wissner-Slivka Professor of Computer Science and E ngineering\, University of Washington and Allen Institute for Artificial I ntelligence\n\nTitle: Possible Impossibilities and Impossible Possibilitie s\n\nAbstract: In this talk\, I will question if there can be possible imp ossibilities of large language models (i.e.\, the fundamental limits of tr ansformers\, if any) and the impossible possibilities of language models ( i.e.\, seemingly impossible alternative paths beyond scale\, if at all).\n \nhttps://cml.ics.uci.edu/seminars/2023-11-21-yejin-choi X-ALT-DESC;FMTTYPE=text/html:Yejin Choi\, Wissner-Slivka Professor of Computer Science and Engineering\, University of Washington and Allen Institute for Artificial Intelligence

Title: Possi ble Impossibilities and Impossible Possibilities

Abstract: I n this talk\, I will question if there can be possible impossibilities of large language models (i.e.\, the fundamental limits of transformers\, if any) and the impossible possibilities of language models (i.e.\, seemingly impossible alternative paths beyond scale\, if at all).

https://cml.ics.uci .edu/seminars/2023-11-21-yejin-choi URL:https://cml.ics.uci.edu/seminars/2023-11-21-yejin-choi END:VEVENT BEGIN:VEVENT UID:2023-11-27-tryphon-georgiou@cml.ics.uci.edu DTSTAMP:20231127T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20231127T130000 DTEND;TZID=America/Los_Angeles:20231127T140000 SUMMARY:[CML Seminar] Tryphon Georgiou: Stochastic thermodynamics: Diffusio n models for information and energy transfer LOCATION:Donald Bren Hall 4011 DESCRIPTION:Tryphon Georgiou\, Distinguished Professor of Mechanical and Ae rospace Engineering\, University of California\, Irvine\n\nTitle: Stochast ic thermodynamics: Diffusion models for information and energy transfer\n\ nAbstract: The energetic cost of information erasure and of energy transdu ction can be cast as the stochastic problem to minimize entropy production during thermodynamic transitions. This formalism of Stochastic Thermodyna mics allows quantitative assessment of work exchange and entropy productio n for systems that are far from equilibrium. In the talk we will highlight the cost of Landauer's bit-erasure in finite time and explain how to obta in bounds on the performance of Carnot-like thermodynamic engines and of p rocesses that are powered by thermal anisotropy.\n\nhttps://cml.ics.uci.ed u/seminars/2023-11-27-tryphon-georgiou X-ALT-DESC;FMTTYPE=text/html:Tryphon Georgiou\, Distingu ished Professor of Mechanical and Aerospace Engineering\, University of Ca lifornia\, Irvine

Title: Stochastic thermodynamics: Diffusio n models for information and energy transfer

Abstract: The e nergetic cost of information erasure and of energy transduction can be cas t as the stochastic problem to minimize entropy production during thermody namic transitions. This formalism of Stochastic Thermodynamics allows quan titative assessment of work exchange and entropy production for systems th at are far from equilibrium. In the talk we will highlight the cost of Lan dauer's bit-erasure in finite time and explain how to obtain bounds on the performance of Carnot-like thermodynamic engines and of processes that ar e powered by thermal anisotropy.

https://cml.ics.uci.edu/seminars/2023 -11-27-tryphon-georgiou URL:https://cml.ics.uci.edu/seminars/2023-11-27-tryphon-georgiou END:VEVENT BEGIN:VEVENT UID:2023-12-04-deying-kong@cml.ics.uci.edu DTSTAMP:20231204T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20231204T130000 DTEND;TZID=America/Los_Angeles:20231204T140000 SUMMARY:[CML Seminar] Deying Kong: Handformer2T: A Lightweight Regression-b ased model for Interacting Hands Pose Estimation from a single RGB Image LOCATION:Donald Bren Hall 4011 DESCRIPTION:Deying Kong\, Software Engineer\, Google\n\nTitle: Handformer2T : A Lightweight Regression-based model for Interacting Hands Pose Estimati on from a single RGB Image\n\nAbstract: Despite its extensive range of pot ential applications in virtual reality and augmented reality\, 3D interact ing hand pose estimation from an RGB image remains a very challenging prob lem\, due to appearance confusions between keypoints of the two hands\, an d severe hand-hand occlusion. Due to their ability to capture long range r elationships between keypoints\, transformer-based methods have gained pop ularity. However\, existing methods usually deploy tokens at keypoint leve l\, which results in high computational and memory complexity. In this tal k\, we propose a novel mechanism\, hand-level tokenization\, where we depl oy only one token for each hand. We also propose a pose query enhancer mod ule\, which refines the pose prediction iteratively. As a result\, our pro posed model\, Handformer2T\, can achieve high performance while remaining lightweight.\n\nhttps://cml.ics.uci.edu/seminars/2023-12-04-deying-kong X-ALT-DESC;FMTTYPE=text/html:Deying Kong\, Software Engi neer\, Google

Title: Handformer2T: A Lightweight Regression- based model for Interacting Hands Pose Estimation from a single RGB Image 
Abstract: Despite its extensive range of potential applicati ons in virtual reality and augmented reality\, 3D interacting hand pose es timation from an RGB image remains a very challenging problem\, due to app earance confusions between keypoints of the two hands\, and severe hand-ha nd occlusion. Due to their ability to capture long range relationships bet ween keypoints\, transformer-based methods have gained popularity. However \, existing methods usually deploy tokens at keypoint level\, which result s in high computational and memory complexity. In this talk\, we propose a novel mechanism\, hand-level tokenization\, where we deploy only one toke n for each hand. We also propose a pose query enhancer module\, which refi nes the pose prediction iteratively. As a result\, our proposed model\, Ha ndformer2T\, can achieve high performance while remaining lightweight.

http s://cml.ics.uci.edu/seminars/2023-12-04-deying-kong URL:https://cml.ics.uci.edu/seminars/2023-12-04-deying-kong END:VEVENT BEGIN:VEVENT UID:2024-01-08-fuxin-li@cml.ics.uci.edu DTSTAMP:20240108T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240108T130000 DTEND;TZID=America/Los_Angeles:20240108T140000 SUMMARY:[CML Seminar] Fuxin Li: From Heatmaps to Structural and Counterfact ual Explanations LOCATION:Donald Bren Hall 4011 DESCRIPTION:Fuxin Li\, Associate Professor of Electrical Engineering and Co mputer Science\, Oregon State University\n\nTitle: From Heatmaps to Struct ural and Counterfactual Explanations\n\nAbstract: This talk will focus on our endeavors in the past few years on explaining deep image models. Reali zing that an important missing piece for explaining neural networks is a r eliable heatmap visualization tool\, we developed I-GOS and iGOS++ which o ptimize with integrated gradients to avoid local optima in heatmap generat ions and improve performance in high-resolution heatmaps. During the devel opment of those visualizations\, we realize that for a significant number of images\, the classifier has multiple different paths to reach a confide nt prediction. This leads to our recent development of structural attentio n graphs\, an approach that utilizes beam search to locate multiple coarse heatmaps for a single image. Finally\, we present results traversing the latent space of variational autoencoders and generative adversarial networ ks (GANs)\, generating high-quality counterfactual explanations that visua lly show how to change one image so that CNNs predict them as another cate gory.\n\nhttps://cml.ics.uci.edu/seminars/2024-01-08-fuxin-li X-ALT-DESC;FMTTYPE=text/html:Fuxin Li\, Associate Profes sor of Electrical Engineering and Computer Science\, Oregon State Universi ty

Title: From Heatmaps to Structural and Counterfactual Exp lanations

Abstract: This talk will focus on our endeavors in the past few years on explaining deep image models. Realizing that an imp ortant missing piece for explaining neural networks is a reliable heatmap visualization tool\, we developed I-GOS and iGOS++ which optimize with int egrated gradients to avoid local optima in heatmap generations and improve performance in high-resolution heatmaps. During the development of those visualizations\, we realize that for a significant number of images\, the classifier has multiple different paths to reach a confident prediction. T his leads to our recent development of structural attention graphs\, an ap proach that utilizes beam search to locate multiple coarse heatmaps for a single image. Finally\, we present results traversing the latent space of variational autoencoders and generative adversarial networks (GANs)\, gene rating high-quality counterfactual explanations that visually show how to change one image so that CNNs predict them as another category.

https://cml.ic s.uci.edu/seminars/2024-01-08-fuxin-li URL:https://cml.ics.uci.edu/seminars/2024-01-08-fuxin-li END:VEVENT BEGIN:VEVENT UID:2024-01-29-gavin-kerrigan@cml.ics.uci.edu DTSTAMP:20240129T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240129T130000 DTEND;TZID=America/Los_Angeles:20240129T140000 SUMMARY:[CML Seminar] Gavin Kerrigan: Deep Generative Models in Infinite-Di mensional Spaces LOCATION:Donald Bren Hall 4011 DESCRIPTION:Gavin Kerrigan\, PhD Student\, Department of Computer Science\, UC Irvine\n\nTitle: Deep Generative Models in Infinite-Dimensional Spaces \n\nAbstract: Deep generative models have seen a meteoric rise in capabili ties across a wide array of domains\, ranging from natural language and vi sion to scientific applications such as precipitation forecasting and mole cular generation. However\, a number of important applications focus on da ta which is inherently infinite-dimensional\, such as time-series\, soluti ons to partial differential equations\, and audio signals. This relatively under-explored class of problems poses unique theoretical and practical c hallenges for generative modeling. In this talk\, we will explore recent d evelopments for infinite-dimensional generative models\, with a focus on d iffusion-based methodologies.\n\nhttps://cml.ics.uci.edu/seminars/2024-01- 29-gavin-kerrigan X-ALT-DESC;FMTTYPE=text/html:Gavin Kerrigan\, PhD Studen t\, Department of Computer Science\, UC Irvine

Title: Deep G enerative Models in Infinite-Dimensional Spaces

Abstract: De ep generative models have seen a meteoric rise in capabilities across a wi de array of domains\, ranging from natural language and vision to scientif ic applications such as precipitation forecasting and molecular generation . However\, a number of important applications focus on data which is inhe rently infinite-dimensional\, such as time-series\, solutions to partial d ifferential equations\, and audio signals. This relatively under-explored class of problems poses unique theoretical and practical challenges for ge nerative modeling. In this talk\, we will explore recent developments for infinite-dimensional generative models\, with a focus on diffusion-based m ethodologies.

https://cml.ics.uci.edu/seminars/2024-01-29-gavin-kerrigan URL:https://cml.ics.uci.edu/seminars/2024-01-29-gavin-kerrigan END:VEVENT BEGIN:VEVENT UID:2024-02-12-shivanshu-gupta@cml.ics.uci.edu DTSTAMP:20240212T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240212T130000 DTEND;TZID=America/Los_Angeles:20240212T140000 SUMMARY:[CML Seminar] Shivanshu Gupta: Informative Example Selection for In -Context Learning LOCATION:Donald Bren Hall 4011 DESCRIPTION:Shivanshu Gupta\, PhD Student\, Department of Computer Science\ , UC Irvine\n\nTitle: Informative Example Selection for In-Context Learnin g\n\nAbstract: In-context Learning (ICL) uses large language models (LLMs) for new tasks by conditioning them on prompts comprising a few task examp les. With the rise of LLMs that are intractable to train or hidden behind APIs\, the importance of such a training-free interface cannot be overstat ed. However\, ICL is known to be critically sensitive to the choice of in- context examples. Despite this\, the standard approach for selecting in-co ntext examples remains to use general-purpose retrievers due to the limite d effectiveness and training requirements of prior approaches. In this tal k\, I'll posit that good in-context examples demonstrate the salient infor mation necessary to solve a given test input. I'll present efficient appro aches for selecting such examples\, with a special focus on preserving the training-free ICL pipeline.\n\nhttps://cml.ics.uci.edu/seminars/2024-02-1 2-shivanshu-gupta X-ALT-DESC;FMTTYPE=text/html:Shivanshu Gupta\, PhD Stude nt\, Department of Computer Science\, UC Irvine

Title: Infor mative Example Selection for In-Context Learning

Abstract: I n-context Learning (ICL) uses large language models (LLMs) for new tasks b y conditioning them on prompts comprising a few task examples. With the ri se of LLMs that are intractable to train or hidden behind APIs\, the impor tance of such a training-free interface cannot be overstated. However\, IC L is known to be critically sensitive to the choice of in-context examples . Despite this\, the standard approach for selecting in-context examples r emains to use general-purpose retrievers due to the limited effectiveness and training requirements of prior approaches. In this talk\, I'll posit t hat good in-context examples demonstrate the salient information necessary to solve a given test input. I'll present efficient approaches for select ing such examples\, with a special focus on preserving the training-free I CL pipeline.

https://cml.ics.uci.edu/seminars/2024-02-12-shivanshu-gupt a URL:https://cml.ics.uci.edu/seminars/2024-02-12-shivanshu-gupta END:VEVENT BEGIN:VEVENT UID:2024-02-20-max-welling@cml.ics.uci.edu DTSTAMP:20240220T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240220T140000 DTEND;TZID=America/Los_Angeles:20240220T150000 SUMMARY:[CML Seminar] Max Welling: The Synergy between Machine Learning and the Natural Sciences LOCATION:Donald Bren Hall 6011 DESCRIPTION:Max Welling\, Professor and Research Chair in Machine Learning\ , University of Amsterdam\n\nTitle: The Synergy between Machine Learning a nd the Natural Sciences\n\nAbstract: Traditionally machine learning has be en heavily influenced by neuroscience (hence the name artificial neural ne tworks) and physics (e.g. MCMC\, Belief Propagation\, and Diffusion based Generative AI). We have recently witnessed that the flow of information ha s also reversed\, with new tools developed in the ML community impacting p hysics\, chemistry and biology. Examples include faster DFT\, Force-Field accelerated MD simulations\, PDE Neural Surrogate models\, generating drug like molecules\, and many more. In this talk I will review the exciting op portunities for further cross fertilization between these fields\, ranging from faster (classical) DFT calculations and enhanced transition path sam pling to traveling waves in artificial neural networks.\n\nhttps://cml.ics .uci.edu/seminars/2024-02-20-max-welling X-ALT-DESC;FMTTYPE=text/html:Max Welling\, Professor and Research Chair in Machine Learning\, University of Amsterdam

Ti tle: The Synergy between Machine Learning and the Natural Sciences

Abstract: Traditionally machine learning has been heavily influ enced by neuroscience (hence the name artificial neural networks) and phys ics (e.g. MCMC\, Belief Propagation\, and Diffusion based Generative AI). We have recently witnessed that the flow of information has also reversed\ , with new tools developed in the ML community impacting physics\, chemist ry and biology. Examples include faster DFT\, Force-Field accelerated MD s imulations\, PDE Neural Surrogate models\, generating druglike molecules\, and many more. In this talk I will review the exciting opportunities for further cross fertilization between these fields\, ranging from faster (cl assical) DFT calculations and enhanced transition path sampling to traveli ng waves in artificial neural networks.

https://cml.ics.uci.edu/seminars/20 24-02-20-max-welling URL:https://cml.ics.uci.edu/seminars/2024-02-20-max-welling END:VEVENT BEGIN:VEVENT UID:2024-03-04-bratin-saha@cml.ics.uci.edu DTSTAMP:20240304T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240304T110000 DTEND;TZID=America/Los_Angeles:20240304T120000 SUMMARY:[CML Seminar] Bratin Saha: Scaling Generative AI in the Enterprise LOCATION:Donald Bren Hall 4011 DESCRIPTION:Bratin Saha\, Vice President of Machine Learning and AI Service s\, Amazon Web Services\n\nTitle: Scaling Generative AI in the Enterprise\ n\nAbstract: Machine learning (ML) and generative artificial intelligence (AI) is one of the most transformational technologies that is opening up n ew opportunities for innovation in every domain across software\, finance\ , health care\, manufacturing\, media\, entertainment and others. This tal k will discuss the key trends that are driving AI/ML innovation\, how ente rprises are using AI/ML today to innovate how they run their businesses\, the key technology challenges in scaling out ML and generative AI across t he enterprise\, some of the key innovations from Amazon\, and how this fie ld is likely to evolve in the future.\n\nhttps://cml.ics.uci.edu/seminars/ 2024-03-04-bratin-saha X-ALT-DESC;FMTTYPE=text/html:Bratin Saha\, Vice Presiden t of Machine Learning and AI Services\, Amazon Web Services

Titl e: Scaling Generative AI in the Enterprise

Abstract: Mac hine learning (ML) and generative artificial intelligence (AI) is one of t he most transformational technologies that is opening up new opportunities for innovation in every domain across software\, finance\, health care\, manufacturing\, media\, entertainment and others. This talk will discuss t he key trends that are driving AI/ML innovation\, how enterprises are usin g AI/ML today to innovate how they run their businesses\, the key technolo gy challenges in scaling out ML and generative AI across the enterprise\, some of the key innovations from Amazon\, and how this field is likely to evolve in the future.

https://cml.ics.uci.edu/seminars/2024-03-04-bratin-sa ha URL:https://cml.ics.uci.edu/seminars/2024-03-04-bratin-saha END:VEVENT BEGIN:VEVENT UID:2024-03-07-terra-blevins@cml.ics.uci.edu DTSTAMP:20240307T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240307T110000 DTEND;TZID=America/Los_Angeles:20240307T120000 SUMMARY:[CML Seminar] Terra Blevins: Breaking the Curse of Multilinguality in Language Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Terra Blevins\, PhD Student\, School of Computer Science and En gineering\, University of Washington\n\nTitle: Breaking the Curse of Multi linguality in Language Models\n\nAbstract: While language models (or LMs\, à la ChatGPT) have become the predominant tool in natural language proce ssing\, their performance in non-English languages increasingly lags behin d. This gap is due to the curse of multilinguality\, which harms individua l language performance in multilingual models through inter-language compe tition for model capacity. In this talk\, I examine how current language m odels do and don't capture different languages and present new methods for fair modeling of all languages. First\, I demonstrate how LMs become mult ilingual through their data and training dynamics. I then characterize whe n multilingual models learn (and forget) languages during training to unco ver how the curse of multilinguality develops. These analyses provide key insights into developing more equitable multilingual models\, and I propos e a new language modeling approach for Cross-Lingual Expert Language Model s (X-ELM) that explicitly allocates model resources to reduce language com petition.\n\nhttps://cml.ics.uci.edu/seminars/2024-03-07-terra-blevins X-ALT-DESC;FMTTYPE=text/html:Terra Blevins\, PhD Student \, School of Computer Science and Engineering\, University of Washington
Title: Breaking the Curse of Multilinguality in Language Mode ls

Abstract: While language models (or LMs\, à la ChatGPT) have become the predominant tool in natural language processing\, their pe rformance in non-English languages increasingly lags behind. This gap is d ue to the curse of multilinguality\, which harms individual language perfo rmance in multilingual models through inter-language competition for model capacity. In this talk\, I examine how current language models do and don 't capture different languages and present new methods for fair modeling o f all languages. First\, I demonstrate how LMs become multilingual through their data and training dynamics. I then characterize when multilingual m odels learn (and forget) languages during training to uncover how the curs e of multilinguality develops. These analyses provide key insights into de veloping more equitable multilingual models\, and I propose a new language modeling approach for Cross-Lingual Expert Language Models (X-ELM) that e xplicitly allocates model resources to reduce language competition.

https ://cml.ics.uci.edu/seminars/2024-03-07-terra-blevins URL:https://cml.ics.uci.edu/seminars/2024-03-07-terra-blevins END:VEVENT BEGIN:VEVENT UID:2024-03-11-paola-cascante-bonilla@cml.ics.uci.edu DTSTAMP:20240311T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240311T110000 DTEND;TZID=America/Los_Angeles:20240311T120000 SUMMARY:[CML Seminar] Paola Cascante-Bonilla: More from Less: Learning with Limited Annotated Data in Vision and Language LOCATION:Donald Bren Hall 4011 DESCRIPTION:Paola Cascante-Bonilla\, Postdoctoral Associate\, University of Maryland Institute for Advanced Computer Studies (UMIACS)\n\nTitle: More from Less: Learning with Limited Annotated Data in Vision and Language\n\n Abstract: Despite the impressive results of deep learning models\, modern large-scale systems are required to be trained using massive amounts of ma nually annotated or freely available data on the Internet. But this data i n the wild is insufficient to learn specific structural patterns of the wo rld\, and existing large-scale models still fail on common sense tasks req uiring compositional inference. This talk will focus on answering three fu ndamental questions: (a) How can we create systems that can learn with lim ited annotated data and adapt to new tasks and novel criteria? (b) How can we create systems able to encode real-world concepts with granularity in a robust manner? (c) Is it possible to create such a system with alternati ve data\, complying with privacy protection principles and avoiding cultur al bias? I will conclude with my future plans to continue exploring hyper- realistic synthetic data generation techniques and the expressiveness of g enerative models to train multimodal systems able to perform well in real- world scenarios.\n\nhttps://cml.ics.uci.edu/seminars/2024-03-11-paola-casc ante-bonilla X-ALT-DESC;FMTTYPE=text/html:Paola Cascante-Bonilla\, Po stdoctoral Associate\, University of Maryland Institute for Advanced Compu ter Studies (UMIACS)

Title: More from Less: Learning with Li mited Annotated Data in Vision and Language

Abstract: Despit e the impressive results of deep learning models\, modern large-scale syst ems are required to be trained using massive amounts of manually annotated or freely available data on the Internet. But this data in the wild is in sufficient to learn specific structural patterns of the world\, and existi ng large-scale models still fail on common sense tasks requiring compositi onal inference. This talk will focus on answering three fundamental questi ons: (a) How can we create systems that can learn with limited annotated d ata and adapt to new tasks and novel criteria? (b) How can we create syste ms able to encode real-world concepts with granularity in a robust manner? (c) Is it possible to create such a system with alternative data\, comply ing with privacy protection principles and avoiding cultural bias? I will conclude with my future plans to continue exploring hyper-realistic synthe tic data generation techniques and the expressiveness of generative models to train multimodal systems able to perform well in real-world scenarios.

https://cml.ics.uci.edu/seminars/2024-03-11-paola-cascante-boni lla URL:https://cml.ics.uci.edu/seminars/2024-03-11-paola-cascante-bonilla END:VEVENT BEGIN:VEVENT UID:2024-03-14-xi-ye@cml.ics.uci.edu DTSTAMP:20240314T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240314T110000 DTEND;TZID=America/Los_Angeles:20240314T120000 SUMMARY:[CML Seminar] Xi Ye: Steering Textual Reasoning with Explanations LOCATION:Donald Bren Hall 4011 DESCRIPTION:Xi Ye\, PhD Student\, Department of Computer Science\, Universi ty of Texas at Austin\n\nTitle: Steering Textual Reasoning with Explanatio ns\n\nAbstract: Large language models (LLMs) have significantly extended t he boundaries of NLP's potential applications\, partially because of their increased ability to do complex reasoning. However\, LLMs have well-docum ented reasoning failures\, such as hallucinations and inability to systema tically generalize. In this talk\, I describe my work on enhancing LLMs in reliably performing textual reasoning\, with a particular focus on levera ging explanations. I will first introduce a framework for automatically as sessing the robustness of black-box models using explanations. I will then describe how to form effective explanations for better teaching LLMs to r eason. My work uses declarative formal specifications as explanations\, wh ich enables using an SMT solver to amend the limited planning capabilities of LLMs. Finally\, I will describe future directions for further enhancin g LLMs to better aid humans in challenging real-world applications demandi ng deep reasoning.\n\nhttps://cml.ics.uci.edu/seminars/2024-03-14-xi-ye X-ALT-DESC;FMTTYPE=text/html:Xi Ye\, PhD Student\, Depar tment of Computer Science\, University of Texas at Austin

Title: Steering Textual Reasoning with Explanations

Abstract: Large language models (LLMs) have significantly extended the boundaries of NLP's potential applications\, partially because of their increased abili ty to do complex reasoning. However\, LLMs have well-documented reasoning failures\, such as hallucinations and inability to systematically generali ze. In this talk\, I describe my work on enhancing LLMs in reliably perfor ming textual reasoning\, with a particular focus on leveraging explanation s. I will first introduce a framework for automatically assessing the robu stness of black-box models using explanations. I will then describe how to form effective explanations for better teaching LLMs to reason. My work u ses declarative formal specifications as explanations\, which enables usin g an SMT solver to amend the limited planning capabilities of LLMs. Finall y\, I will describe future directions for further enhancing LLMs to better aid humans in challenging real-world applications demanding deep reasonin g.

http s://cml.ics.uci.edu/seminars/2024-03-14-xi-ye URL:https://cml.ics.uci.edu/seminars/2024-03-14-xi-ye END:VEVENT BEGIN:VEVENT UID:2024-03-19-sai-praneeth-karimireddy@cml.ics.uci.edu DTSTAMP:20240319T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240319T110000 DTEND;TZID=America/Los_Angeles:20240319T120000 SUMMARY:[CML Seminar] Sai Praneeth Karimireddy: Building Planetary-Scale Co llaborative Intelligence LOCATION:Donald Bren Hall 4011 DESCRIPTION:Sai Praneeth Karimireddy\, Postdoctoral Researcher\, University of California\, Berkeley\n\nTitle: Building Planetary-Scale Collaborative Intelligence\n\nAbstract: Today\, access to high-quality data has become the key bottleneck to deploying machine learning. Often\, the data that is most valuable is locked away in inaccessible silos due to unfavorable inc entives and ethical or legal restrictions. This is starkly evident in heal th care\, where such barriers have led to highly biased and underperformin g tools. Using my collaborations with Doctors Without Borders and the Canc er Registry of Norway as case studies\, I will describe how collaborative learning systems\, such as federated learning\, provide a natural solution . Yet for these systems to truly succeed\, three fundamental challenges mu st be confronted: they need to 1) be efficient and scale to massive networ ks\, 2) manage the divergent goals of the participants\, and 3) provide re silient training and trustworthy predictions. I will discuss how tools fro m optimization\, statistics\, and economics can be leveraged to address th ese challenges.\n\nhttps://cml.ics.uci.edu/seminars/2024-03-19-sai-praneet h-karimireddy X-ALT-DESC;FMTTYPE=text/html:Sai Praneeth Karimireddy\, Postdoctoral Researcher\, University of California\, Berkeley

Ti tle: Building Planetary-Scale Collaborative Intelligence

Abs tract: Today\, access to high-quality data has become the key bottlene ck to deploying machine learning. Often\, the data that is most valuable i s locked away in inaccessible silos due to unfavorable incentives and ethi cal or legal restrictions. This is starkly evident in health care\, where such barriers have led to highly biased and underperforming tools. Using m y collaborations with Doctors Without Borders and the Cancer Registry of N orway as case studies\, I will describe how collaborative learning systems \, such as federated learning\, provide a natural solution. Yet for these systems to truly succeed\, three fundamental challenges must be confronted : they need to 1) be efficient and scale to massive networks\, 2) manage t he divergent goals of the participants\, and 3) provide resilient training and trustworthy predictions. I will discuss how tools from optimization\, statistics\, and economics can be leveraged to address these challenges. 
https://cml.ics.uci.edu/seminars/2024-03-19-sai-praneeth-karim ireddy URL:https://cml.ics.uci.edu/seminars/2024-03-19-sai-praneeth-karimireddy END:VEVENT BEGIN:VEVENT UID:2024-04-02-chen-wei@cml.ics.uci.edu DTSTAMP:20240402T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240402T110000 DTEND;TZID=America/Los_Angeles:20240402T120000 SUMMARY:[CML Seminar] Chen Wei: Learning Generalized Knowledge for AI with Limited Supervision LOCATION:Donald Bren Hall 4011 DESCRIPTION:Chen Wei\, PhD Student\, Department of Computer Science\, Johns Hopkins University\n\nTitle: Learning Generalized Knowledge for AI with L imited Supervision\n\nAbstract: As babies\, we begin to grasp the world th rough spontaneous observations\, gradually developing generalized knowledg e about the world. This foundational knowledge enables humans to effortles sly learn new skills without extensive teaching for each task. Can we deve lop a similar paradigm for AI? This talk describes how learning from limit ed supervision can address fundamental challenges in AI such as scalabilit y and generalization while embedding generalized knowledge. I will first t alk about our research in self-supervised learning\, utilizing natural ima ges and videos without human-annotated labels. The second part will descri be how to leverage non-curated image-text pairs\, through which we obtain textual representation of images. This representation comprehensively desc ribes semantic elements in an image and bridges various AI tools such as l arge language models (LLMs)\, enabling diverse vision-language application s.\n\nhttps://cml.ics.uci.edu/seminars/2024-04-02-chen-wei X-ALT-DESC;FMTTYPE=text/html:Chen Wei\, PhD Student\, De partment of Computer Science\, Johns Hopkins University

Title: Learning Generalized Knowledge for AI with Limited Supervision

Abstract: As babies\, we begin to grasp the world through spontaneou s observations\, gradually developing generalized knowledge about the worl d. This foundational knowledge enables humans to effortlessly learn new sk ills without extensive teaching for each task. Can we develop a similar pa radigm for AI? This talk describes how learning from limited supervision c an address fundamental challenges in AI such as scalability and generaliza tion while embedding generalized knowledge. I will first talk about our re search in self-supervised learning\, utilizing natural images and videos w ithout human-annotated labels. The second part will describe how to levera ge non-curated image-text pairs\, through which we obtain textual represen tation of images. This representation comprehensively describes semantic e lements in an image and bridges various AI tools such as large language mo dels (LLMs)\, enabling diverse vision-language applications.

https://cml.ics.u ci.edu/seminars/2024-04-02-chen-wei URL:https://cml.ics.uci.edu/seminars/2024-04-02-chen-wei END:VEVENT BEGIN:VEVENT UID:2024-04-05-kun-zhang@cml.ics.uci.edu DTSTAMP:20240405T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240405T150000 DTEND;TZID=America/Los_Angeles:20240405T160000 SUMMARY:[CML Seminar] Kun Zhang: Causal Representation Learning: Discovery of the Hidden World LOCATION:Donald Bren Hall 2011 DESCRIPTION:Kun Zhang\, Professor and Director\, Center for Integrative AI\ , Mohamed bin Zayed University of Artificial Intelligence\n\nTitle: Causal Representation Learning: Discovery of the Hidden World\n\nAbstract: Causa lity is a fundamental notion in science\, engineering\, and even in machin e learning. Uncovering the causal process behind observed data can natural ly help answer why and how questions\, inform optimal decisions\, and achi eve adaptive prediction. In many scenarios\, observed variables (such as i mage pixels and questionnaire results) are often reflections of the underl ying causal variables rather than being the causal variables themselves. C ausal representation learning aims to reveal the underlying high-level hid den causal variables and their relations. The modularity property of a cau sal system implies properties of minimal changes and independent changes o f causal representations\, and in this talk\, we show how such properties make it possible to recover the underlying causal representations from obs ervational data with identifiability guarantees. Various problem settings are considered\, involving i.i.d. data\, temporal data\, or data with dist ribution shift as input.\n\nhttps://cml.ics.uci.edu/seminars/2024-04-05-ku n-zhang X-ALT-DESC;FMTTYPE=text/html:Kun Zhang\, Professor and D irector\, Center for Integrative AI\, Mohamed bin Zayed University of Arti ficial Intelligence

Title: Causal Representation Learning: D iscovery of the Hidden World

Abstract: Causality is a fundam ental notion in science\, engineering\, and even in machine learning. Unco vering the causal process behind observed data can naturally help answer w hy and how questions\, inform optimal decisions\, and achieve adaptive pre diction. In many scenarios\, observed variables (such as image pixels and questionnaire results) are often reflections of the underlying causal vari ables rather than being the causal variables themselves. Causal representa tion learning aims to reveal the underlying high-level hidden causal varia bles and their relations. The modularity property of a causal system impli es properties of minimal changes and independent changes of causal represe ntations\, and in this talk\, we show how such properties make it possible to recover the underlying causal representations from observational data with identifiability guarantees. Various problem settings are considered\, involving i.i.d. data\, temporal data\, or data with distribution shift a s input.

https://cml.ics.uci.edu/seminars/2024-04-05-kun-zhang URL:https://cml.ics.uci.edu/seminars/2024-04-05-kun-zhang END:VEVENT BEGIN:VEVENT UID:2024-04-08-unnat-jain@cml.ics.uci.edu DTSTAMP:20240408T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240408T110000 DTEND;TZID=America/Los_Angeles:20240408T120000 SUMMARY:[CML Seminar] Unnat Jain: Jump-starting Embodied Intelligence LOCATION:Donald Bren Hall 4011 DESCRIPTION:Unnat Jain\, Postdoctoral Researcher\, Carnegie Mellon Universi ty\, Fundamental AI Research (FAIR) at Meta\n\nTitle: Jump-starting Embodi ed Intelligence\n\nAbstract: AI has revolutionized the way we interact onl ine. Despite this\, it hasn't quite made the leap when it comes to tasks l ike cooking dinner or cleaning our desks. Why has AI excelled in automatin g our digital interactions but not in assisting us with physical tasks? In my talk\, I will explore the challenges of applying AI to embodied tasks — those requiring physical interaction with the environment. To address these challenges\, I turn to the efficient pathways humans use to achieve embodied intelligence and propose three strategies to jump-start the learn ing process for embodied AI agents: (1) combining learning from both teach ers and own experience\, (2) leveraging external information or hints to s implify learning\, such as using maps to learn about physical spaces\, and (3) learning intelligent behaviors by simply observing others.\n\nhttps:/ /cml.ics.uci.edu/seminars/2024-04-08-unnat-jain X-ALT-DESC;FMTTYPE=text/html:Unnat Jain\, Postdoctoral R esearcher\, Carnegie Mellon University\, Fundamental AI Research (FAIR) at Meta

Title: Jump-starting Embodied Intelligence

A bstract: AI has revolutionized the way we interact online. Despite thi s\, it hasn't quite made the leap when it comes to tasks like cooking dinn er or cleaning our desks. Why has AI excelled in automating our digital in teractions but not in assisting us with physical tasks? In my talk\, I wil l explore the challenges of applying AI to embodied tasks — those requir ing physical interaction with the environment. To address these challenges \, I turn to the efficient pathways humans use to achieve embodied intelli gence and propose three strategies to jump-start the learning process for embodied AI agents: (1) combining learning from both teachers and own expe rience\, (2) leveraging external information or hints to simplify learning \, such as using maps to learn about physical spaces\, and (3) learning in telligent behaviors by simply observing others.

https://cml.ics.uci.edu/semi nars/2024-04-08-unnat-jain URL:https://cml.ics.uci.edu/seminars/2024-04-08-unnat-jain END:VEVENT BEGIN:VEVENT UID:2024-04-10-karen-ullrich@cml.ics.uci.edu DTSTAMP:20240410T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240410T110000 DTEND;TZID=America/Los_Angeles:20240410T120000 SUMMARY:[CML Seminar] Karen Ullrich: Challenges in Improving and Applying G enerative Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Karen Ullrich\, Research Scientist\, Fundamental AI Research (F AIR) at Meta\, New York\n\nTitle: Challenges in Improving and Applying Gen erative Models\n\nAbstract: The emergence of powerful\, ever more universa l models such as ChatGPT\, and stable Diffusion\, made generative modeling (GM) undoubtedly a focal point for modern AI research. In the talk\, we w ill discuss applications of GM and how GM fits into a vision of autonomous machine intelligence. We will critically examine the sustainability of sc aling AI models\, a prevalent approach driving remarkable advancements in GM. Despite significant successes\, I highlight the substantial physical\, economic\, and environmental limitations of continuous scaling\, question ing its long-term feasibility. Furthermore\, we will discuss inherent limi tations in current high performance models that lead to a lack of tractabi lity of statistical queries necessary to enable reasoning.\n\nhttps://cml. ics.uci.edu/seminars/2024-04-10-karen-ullrich X-ALT-DESC;FMTTYPE=text/html:Karen Ullrich\, Research Sc ientist\, Fundamental AI Research (FAIR) at Meta\, New York

Titl e: Challenges in Improving and Applying Generative Models

Ab stract: The emergence of powerful\, ever more universal models such as ChatGPT\, and stable Diffusion\, made generative modeling (GM) undoubtedl y a focal point for modern AI research. In the talk\, we will discuss appl ications of GM and how GM fits into a vision of autonomous machine intelli gence. We will critically examine the sustainability of scaling AI models\ , a prevalent approach driving remarkable advancements in GM. Despite sign ificant successes\, I highlight the substantial physical\, economic\, and environmental limitations of continuous scaling\, questioning its long-ter m feasibility. Furthermore\, we will discuss inherent limitations in curre nt high performance models that lead to a lack of tractability of statisti cal queries necessary to enable reasoning.

https://cml.ics.uci.edu/semina rs/2024-04-10-karen-ullrich URL:https://cml.ics.uci.edu/seminars/2024-04-10-karen-ullrich END:VEVENT BEGIN:VEVENT UID:2024-04-11-hila-gonen@cml.ics.uci.edu DTSTAMP:20240411T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240411T110000 DTEND;TZID=America/Los_Angeles:20240411T120000 SUMMARY:[CML Seminar] Hila Gonen: Unlocking Language Models: Controlling LM s to Enable NLP for All LOCATION:Donald Bren Hall 4011 DESCRIPTION:Hila Gonen\, Postdoctoral Researcher\, School of Computer Scien ce & Engineering\, University of Washington\n\nTitle: Unlocking Language M odels: Controlling LMs to Enable NLP for All\n\nAbstract: Large language m odels (LLMs) have soared in popularity in recent years\, thanks to their a bility to generate well-formed natural language answers for a myriad of to pics. Despite their astonishing capabilities\, they still suffer from vari ous limitations. This talk will focus on two of them: the limited control over LLMs\, and their failure to serve users from diverse backgrounds. I w ill start by presenting my research on controlling and enriching language models through the input (prompting). In the second part\, I will introduc e a novel algorithmic method to remove protected properties (such as gende r and race) from text representations\, which is crucial for preserving pr ivacy and promoting fairness. The third part of the talk will focus on my research efforts to develop models that support multiple languages\, and t he challenges faced when working with languages other than English.\n\nhtt ps://cml.ics.uci.edu/seminars/2024-04-11-hila-gonen X-ALT-DESC;FMTTYPE=text/html:Hila Gonen\, Postdoctoral R esearcher\, School of Computer Science &\; Engineering\, University of Washington

Title: Unlocking Language Models: Controlling LMs to Enable NLP for All

Abstract: Large language models (LLMs ) have soared in popularity in recent years\, thanks to their ability to g enerate well-formed natural language answers for a myriad of topics. Despi te their astonishing capabilities\, they still suffer from various limitat ions. This talk will focus on two of them: the limited control over LLMs\, and their failure to serve users from diverse backgrounds. I will start b y presenting my research on controlling and enriching language models thro ugh the input (prompting). In the second part\, I will introduce a novel a lgorithmic method to remove protected properties (such as gender and race) from text representations\, which is crucial for preserving privacy and p romoting fairness. The third part of the talk will focus on my research ef forts to develop models that support multiple languages\, and the challeng es faced when working with languages other than English.

https://cml.ics.uci .edu/seminars/2024-04-11-hila-gonen URL:https://cml.ics.uci.edu/seminars/2024-04-11-hila-gonen END:VEVENT BEGIN:VEVENT UID:2024-04-15-peter-west@cml.ics.uci.edu DTSTAMP:20240415T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20240415T110000 DTEND;TZID=America/Los_Angeles:20240415T120000 SUMMARY:[CML Seminar] Peter West: Hidden Capabilities and Counterintuitive Limits in Large Language Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Peter West\, PhD Student\, School of Computer Science & Enginee ring\, University of Washington\n\nTitle: Hidden Capabilities and Counteri ntuitive Limits in Large Language Models\n\nAbstract: Massive scale has be en a recent winning recipe in natural language processing and AI\, with ex treme-scale language models like GPT-4 receiving most attention. This is i n spite of staggering energy and monetary costs\, and further\, the contin uing struggle of even the largest models with concepts such as composition al problem solving and linguistic ambiguity. In this talk\, I will propose my vision for a research landscape where compact language models share th e forefront with extreme scale models\, working in concert with many piece s besides scale\, such as algorithms\, knowledge\, information theory\, an d more. I will cover alternative ingredients to scale\, and discuss counte rintuitive disparities in the capabilities of even extreme-scale models\, which can meet or exceed human performance in some complex tasks while tra iling behind humans in what seem to be much simpler tasks.\n\nhttps://cml. ics.uci.edu/seminars/2024-04-15-peter-west X-ALT-DESC;FMTTYPE=text/html:Peter West\, PhD Student\, School of Computer Science &\; Engineering\, University of Washington
Title: Hidden Capabilities and Counterintuitive Limits in Lar ge Language Models

Abstract: Massive scale has been a recent winning recipe in natural language processing and AI\, with extreme-scale language models like GPT-4 receiving most attention. This is in spite of staggering energy and monetary costs\, and further\, the continuing strugg le of even the largest models with concepts such as compositional problem solving and linguistic ambiguity. In this talk\, I will propose my vision for a research landscape where compact language models share the forefront with extreme scale models\, working in concert with many pieces besides s cale\, such as algorithms\, knowledge\, information theory\, and more. I w ill cover alternative ingredients to scale\, and discuss counterintuitive disparities in the capabilities of even extreme-scale models\, which can m eet or exceed human performance in some complex tasks while trailing behin d humans in what seem to be much simpler tasks.

https://cml.ics.uci.edu/semi nars/2024-04-15-peter-west URL:https://cml.ics.uci.edu/seminars/2024-04-15-peter-west END:VEVENT BEGIN:VEVENT UID:2024-10-07-harsh-trivedi@cml.ics.uci.edu DTSTAMP:20241007T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20241007T130000 DTEND;TZID=America/Los_Angeles:20241007T140000 SUMMARY:[CML Seminar] Harsh Trivedi: AppWorld: Reliable Evaluation of Inter active Agents in a World of Apps and People LOCATION:Donald Bren Hall 4011 DESCRIPTION:Harsh Trivedi\, PhD Student\, Department of Computer Science\, Stony Brook University\n\nTitle: AppWorld: Reliable Evaluation of Interact ive Agents in a World of Apps and People\n\nAbstract: We envision a world where AI agents (assistants) are widely used for complex tasks in our digi tal and physical worlds and are broadly integrated into our society. To mo ve towards such a future\, we need an environment for a robust evaluation of agents' capability\, reliability\, and trustworthiness. In this talk\, I'll introduce AppWorld\, which is a step towards this goal in the context of day-to-day digital tasks. AppWorld is a high-fidelity simulated world of people and their digital activities on nine apps like Amazon\, Gmail\, and Venmo. On top of this fully controllable world\, we build a benchmark of complex day-to-day tasks such as splitting Venmo bills with roommates\, which agents have to solve via interactive coding and API calls. Our benc hmarking evaluations show that even the best LLMs\, like GPT-4o\, can only solve ~30% of such tasks\, highlighting the challenging nature of the App World benchmark.\n\nhttps://cml.ics.uci.edu/seminars/2024-10-07-harsh-triv edi X-ALT-DESC;FMTTYPE=text/html:Harsh Trivedi\, PhD Student \, Department of Computer Science\, Stony Brook University

Title : AppWorld: Reliable Evaluation of Interactive Agents in a World of Ap ps and People

Abstract: We envision a world where AI agents (assistants) are widely used for complex tasks in our digital and physical worlds and are broadly integrated into our society. To move towards such a future\, we need an environment for a robust evaluation of agents' capab ility\, reliability\, and trustworthiness. In this talk\, I'll introduce A ppWorld\, which is a step towards this goal in the context of day-to-day d igital tasks. AppWorld is a high-fidelity simulated world of people and th eir digital activities on nine apps like Amazon\, Gmail\, and Venmo. On to p of this fully controllable world\, we build a benchmark of complex day-t o-day tasks such as splitting Venmo bills with roommates\, which agents ha ve to solve via interactive coding and API calls. Our benchmarking evaluat ions show that even the best LLMs\, like GPT-4o\, can only solve ~30% of s uch tasks\, highlighting the challenging nature of the AppWorld benchmark.

https://cml.ics.uci.edu/seminars/2024-10-07-harsh-trivedi URL:https://cml.ics.uci.edu/seminars/2024-10-07-harsh-trivedi END:VEVENT BEGIN:VEVENT UID:2024-10-14-kushagra-pandey@cml.ics.uci.edu DTSTAMP:20241014T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20241014T130000 DTEND;TZID=America/Los_Angeles:20241014T140000 SUMMARY:[CML Seminar] Kushagra Pandey: Conjugate Integrators for Fast Sampl ing in Diffusion Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Kushagra Pandey\, PhD Student\, Department of Computer Science\ , University of California\, Irvine\n\nTitle: Conjugate Integrators for Fa st Sampling in Diffusion Models\n\nAbstract: Diffusion models exhibit exce llent sample quality across multiple-generation tasks. However\, their inf erence process is iterative and often requires hundreds of function evalua tions. Moreover\, it is unclear if existing methods for accelerating diffu sion model sampling can generalize well across different types of diffusio n processes. In the first part of my talk\, I will introduce Conjugate Int egrators\, which project unconditional diffusion dynamics to an alternate space that is more amenable to faster sampling. In the second part of my t alk\, I will extend the idea of Conjugate Integrators from unconditional s ampling to conditional diffusion sampling in the context of solving invers e problems. Empirically\, on challenging inverse problems like 4x super-re solution on the ImageNet-256 dataset\, conditional Conjugate Integrators c an generate high-quality samples in as few as 5 conditional sampling steps .\n\nhttps://cml.ics.uci.edu/seminars/2024-10-14-kushagra-pandey X-ALT-DESC;FMTTYPE=text/html:Kushagra Pandey\, PhD Stude nt\, Department of Computer Science\, University of California\, Irvine

Title: Conjugate Integrators for Fast Sampling in Diffusion Mo dels

Abstract: Diffusion models exhibit excellent sample qua lity across multiple-generation tasks. However\, their inference process i s iterative and often requires hundreds of function evaluations. Moreover\ , it is unclear if existing methods for accelerating diffusion model sampl ing can generalize well across different types of diffusion processes. In the first part of my talk\, I will introduce Conjugate Integrators\, which project unconditional diffusion dynamics to an alternate space that is mo re amenable to faster sampling. In the second part of my talk\, I will ext end the idea of Conjugate Integrators from unconditional sampling to condi tional diffusion sampling in the context of solving inverse problems. Empi rically\, on challenging inverse problems like 4x super-resolution on the ImageNet-256 dataset\, conditional Conjugate Integrators can generate high -quality samples in as few as 5 conditional sampling steps.

https://cml .ics.uci.edu/seminars/2024-10-14-kushagra-pandey URL:https://cml.ics.uci.edu/seminars/2024-10-14-kushagra-pandey END:VEVENT BEGIN:VEVENT UID:2024-10-21-tianchen-qian@cml.ics.uci.edu DTSTAMP:20241021T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20241021T130000 DTEND;TZID=America/Los_Angeles:20241021T140000 SUMMARY:[CML Seminar] Tianchen Qian: Causal inference and machine learning in mobile health: Modeling time-varying effects using longitudinal functio nal data LOCATION:Donald Bren Hall 4011 DESCRIPTION:Tianchen Qian\, Assistant Professor of Statistics\, University of California\, Irvine\n\nTitle: Causal inference and machine learning in mobile health: Modeling time-varying effects using longitudinal functional data\n\nAbstract: To optimize mobile health interventions and advance dom ain knowledge on intervention design\, it is critical to understand how th e intervention effect varies over time and with contextual information. Th is study aims to assess how a push notification suggesting physical activi ty influences individuals' step counts using data from the HeartSteps micr o-randomized trial (MRT). We propose the first semiparametric causal excur sion effect model with varying coefficients to model the time-varying effe cts within a decision point and across decision points in an MRT. We propo se a two-stage causal effect estimator that uses machine learning and is r obust against a misspecified high-dimensional outcome regression nuisance model. Our analysis provides new insights into individuals' change in resp onse profiles due to the activity suggestions.\n\nhttps://cml.ics.uci.edu/ seminars/2024-10-21-tianchen-qian X-ALT-DESC;FMTTYPE=text/html:Tianchen Qian\, Assistant P rofessor of Statistics\, University of California\, Irvine

Title : Causal inference and machine learning in mobile health: Modeling tim e-varying effects using longitudinal functional data

Abstract: To optimize mobile health interventions and advance domain knowledge on intervention design\, it is critical to understand how the intervention e ffect varies over time and with contextual information. This study aims to assess how a push notification suggesting physical activity influences in dividuals' step counts using data from the HeartSteps micro-randomized tri al (MRT). We propose the first semiparametric causal excursion effect mode l with varying coefficients to model the time-varying effects within a dec ision point and across decision points in an MRT. We propose a two-stage c ausal effect estimator that uses machine learning and is robust against a misspecified high-dimensional outcome regression nuisance model. Our analy sis provides new insights into individuals' change in response profiles du e to the activity suggestions.

https://cml.ics.uci.edu/seminars/2024-10-2 1-tianchen-qian URL:https://cml.ics.uci.edu/seminars/2024-10-21-tianchen-qian END:VEVENT BEGIN:VEVENT UID:2024-10-28-jana-lipkova@cml.ics.uci.edu DTSTAMP:20241028T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20241028T130000 DTEND;TZID=America/Los_Angeles:20241028T140000 SUMMARY:[CML Seminar] Jana Lipkova: AI-based multimodal data fusion for out come prediction in oncology LOCATION:Donald Bren Hall 4011 DESCRIPTION:Jana Lipkova\, Assistant Professor\, Department of Pathology\, School of Medicine\, University of California\, Irvine\n\nTitle: AI-based multimodal data fusion for outcome prediction in oncology\n\nAbstract: In oncology\, the patient state is characterized by a spectrum of diverse med ical data\, each providing unique insights. The vast amount of data\, howe ver\, makes it difficult for experts to adequately assess patient prognosi s under the multimodal context. We present a deep learning-based multimoda l framework for integration of radiology\, histopathology\, and genomics d ata to improve patient outcome prediction. The framework does not require annotations\, tumor segmentation\, or hand-crafted features and can be eas ily applied to larger cohorts and diverse disease models. The feasibility of the model is tested on two external independent cohorts\, including gli oma and non-small cell lung cancer\, indicating benefits of multimodal dat a integration for patient risk stratification\, outcome prediction\, and p rognostic biomarker exploration.\n\nhttps://cml.ics.uci.edu/seminars/2024- 10-28-jana-lipkova X-ALT-DESC;FMTTYPE=text/html:Jana Lipkova\, Assistant Pr ofessor\, Department of Pathology\, School of Medicine\, University of Cal ifornia\, Irvine

Title: AI-based multimodal data fusion for outcome prediction in oncology

Abstract: In oncology\, the p atient state is characterized by a spectrum of diverse medical data\, each providing unique insights. The vast amount of data\, however\, makes it d ifficult for experts to adequately assess patient prognosis under the mult imodal context. We present a deep learning-based multimodal framework for integration of radiology\, histopathology\, and genomics data to improve p atient outcome prediction. The framework does not require annotations\, tu mor segmentation\, or hand-crafted features and can be easily applied to l arger cohorts and diverse disease models. The feasibility of the model is tested on two external independent cohorts\, including glioma and non-smal l cell lung cancer\, indicating benefits of multimodal data integration fo r patient risk stratification\, outcome prediction\, and prognostic biomar ker exploration.

https://cml.ics.uci.edu/seminars/2024-10-28-jana-lipkova< /a> URL:https://cml.ics.uci.edu/seminars/2024-10-28-jana-lipkova END:VEVENT BEGIN:VEVENT UID:2024-11-04-amir-rahmani@cml.ics.uci.edu DTSTAMP:20241104T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20241104T130000 DTEND;TZID=America/Los_Angeles:20241104T140000 SUMMARY:[CML Seminar] Amir Rahmani: Future Health: Harnessing Multimodal Da ta and GenAI for Health Promotion LOCATION:Donald Bren Hall 4011 DESCRIPTION:Amir Rahmani\, Professor of Nursing and Computer Science\, Univ ersity of California\, Irvine\n\nTitle: Future Health: Harnessing Multimod al Data and GenAI for Health Promotion\n\nAbstract: Future Health emphasiz es the importance of recognizing each individual's uniqueness\, which aris es from their specific omics\, lifestyle\, environmental\, and socioeconom ic conditions. Thanks to advancements in sensors\, mobile computing\, ubiq uitous computing\, and artificial intelligence (AI)\, we can now collect d etailed information about individuals. This data serves as the foundation for creating personal models\, offering predictive and preventive advice t ailored specifically to each person. In my presentation\, I will explore h ow AI\, including generative AI\, and wearable technology are revolutioniz ing the collection and analysis of big health data in everyday environment s. I will discuss the analytics used to evaluate physical and mental healt h and how smart recommendations can be made objectively. Moreover\, I will illustrate how leveraging Large Language Models (LLMs)-powered conversati onal health agents (CHAs) can integrate personal data\, models\, and knowl edge into healthcare chatbots. Additionally\, I will present our open-sour ce initiative on developing OpenCHA.\n\nhttps://cml.ics.uci.edu/seminars/2 024-11-04-amir-rahmani X-ALT-DESC;FMTTYPE=text/html:Amir Rahmani\, Professor of Nursing and Computer Science\, University of California\, Irvine

Title: Future Health: Harnessing Multimodal Data and GenAI for Healt h Promotion

Abstract: Future Health emphasizes the importanc e of recognizing each individual's uniqueness\, which arises from their sp ecific omics\, lifestyle\, environmental\, and socioeconomic conditions. T hanks to advancements in sensors\, mobile computing\, ubiquitous computing \, and artificial intelligence (AI)\, we can now collect detailed informat ion about individuals. This data serves as the foundation for creating per sonal models\, offering predictive and preventive advice tailored specific ally to each person. In my presentation\, I will explore how AI\, includin g generative AI\, and wearable technology are revolutionizing the collecti on and analysis of big health data in everyday environments. I will discus s the analytics used to evaluate physical and mental health and how smart recommendations can be made objectively. Moreover\, I will illustrate how leveraging Large Language Models (LLMs)-powered conversational health agen ts (CHAs) can integrate personal data\, models\, and knowledge into health care chatbots. Additionally\, I will present our open-source initiative on developing OpenCHA.

https://cml.ics.uci.edu/seminars/2024-11-04-amir-rahm ani URL:https://cml.ics.uci.edu/seminars/2024-11-04-amir-rahmani END:VEVENT BEGIN:VEVENT UID:2024-11-18-daniel-seita@cml.ics.uci.edu DTSTAMP:20241118T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20241118T130000 DTEND;TZID=America/Los_Angeles:20241118T140000 SUMMARY:[CML Seminar] Daniel Seita: In Pursuit of Dexterous and Generalizab le Robot Manipulation using Reinforcement Learning\, Imitation Learning\, and Foundation Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Daniel Seita\, Assistant Professor of Computer Science\, Univer sity of Southern California\n\nTitle: In Pursuit of Dexterous and Generali zable Robot Manipulation using Reinforcement Learning\, Imitation Learning \, and Foundation Models\n\nAbstract: The robotics community has seen sign ificant progress in applying machine learning for robot manipulation. Howe ver\, despite this progress\, developing a system capable of generalizable robot manipulation remains fundamentally difficult\, especially when mani pulating in clutter and adjusting deformable objects such as fabrics\, rop e\, and liquids. Some promising techniques for developing general robot ma nipulation systems include reinforcement learning\, imitation learning\, a nd more recently\, leveraging foundation models trained on internet-scale data\, such as GPT-4. In this talk\, I will discuss our recent work on (1) deep reinforcement learning for dexterous manipulation in clutter\, (2) f oundation models and imitation learning for bimanual manipulation\, and (3 ) our benchmarks and applications of foundation models for deformable obje ct manipulation.\n\nhttps://cml.ics.uci.edu/seminars/2024-11-18-daniel-sei ta X-ALT-DESC;FMTTYPE=text/html:Daniel Seita\, Assistant Pr ofessor of Computer Science\, University of Southern California

Title: In Pursuit of Dexterous and Generalizable Robot Manipulation us ing Reinforcement Learning\, Imitation Learning\, and Foundation Models

Abstract: The robotics community has seen significant progress in applying machine learning for robot manipulation. However\, despite th is progress\, developing a system capable of generalizable robot manipulat ion remains fundamentally difficult\, especially when manipulating in clut ter and adjusting deformable objects such as fabrics\, rope\, and liquids. Some promising techniques for developing general robot manipulation syste ms include reinforcement learning\, imitation learning\, and more recently \, leveraging foundation models trained on internet-scale data\, such as G PT-4. In this talk\, I will discuss our recent work on (1) deep reinforcem ent learning for dexterous manipulation in clutter\, (2) foundation models and imitation learning for bimanual manipulation\, and (3) our benchmarks and applications of foundation models for deformable object manipulation.

https://cml.ics.uci.edu/seminars/2024-11-18-daniel-seita URL:https://cml.ics.uci.edu/seminars/2024-11-18-daniel-seita END:VEVENT BEGIN:VEVENT UID:2024-11-25-yasaman-razeghi@cml.ics.uci.edu DTSTAMP:20241125T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20241125T130000 DTEND;TZID=America/Los_Angeles:20241125T140000 SUMMARY:[CML Seminar] Yasaman Razeghi: Evaluating Foundational Models Using Insights from Their Pretraining Data LOCATION:Donald Bren Hall 4011 DESCRIPTION:Yasaman Razeghi\, PhD Student\, Department of Computer Science\ , University of California\, Irvine\n\nTitle: Evaluating Foundational Mode ls Using Insights from Their Pretraining Data\n\nAbstract: Foundational mo dels have demonstrated exceptional performance on established academic ben chmarks\, often narrowing the gap between human reasoning and artificial i ntelligence. While the success of these models is widely attributed to the ir scale\, the critical role of pretraining data in shaping their capabili ties and limitations is often acknowledged but rarely studied. In this tal k\, I will argue that understanding the true performance of foundational m odels requires going beyond conventional benchmark testing. In particular\ , incorporating insights from their pretraining data is essential for comp rehensively evaluating and interpreting the models' capabilities and limit ations. I show that while models often excel in benchmark settings\, they can fail on basic\, trivial reasoning tasks\, raising concerns about their true robustness. This work cautions against overly optimistic interpretat ions of models' abilities based on canonical evaluation results.\n\nhttps: //cml.ics.uci.edu/seminars/2024-11-25-yasaman-razeghi X-ALT-DESC;FMTTYPE=text/html:Yasaman Razeghi\, PhD Stude nt\, Department of Computer Science\, University of California\, Irvine

Title: Evaluating Foundational Models Using Insights from Thei r Pretraining Data

Abstract: Foundational models have demons trated exceptional performance on established academic benchmarks\, often narrowing the gap between human reasoning and artificial intelligence. Whi le the success of these models is widely attributed to their scale\, the c ritical role of pretraining data in shaping their capabilities and limitat ions is often acknowledged but rarely studied. In this talk\, I will argue that understanding the true performance of foundational models requires g oing beyond conventional benchmark testing. In particular\, incorporating insights from their pretraining data is essential for comprehensively eval uating and interpreting the models' capabilities and limitations. I show t hat while models often excel in benchmark settings\, they can fail on basi c\, trivial reasoning tasks\, raising concerns about their true robustness . This work cautions against overly optimistic interpretations of models' abilities based on canonical evaluation results.

https://cml.ics.uci.ed u/seminars/2024-11-25-yasaman-razeghi URL:https://cml.ics.uci.edu/seminars/2024-11-25-yasaman-razeghi END:VEVENT BEGIN:VEVENT UID:2024-12-02-ali-younis@cml.ics.uci.edu DTSTAMP:20241202T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20241202T130000 DTEND;TZID=America/Los_Angeles:20241202T140000 SUMMARY:[CML Seminar] Ali Younis: End-to-end Learnable Particle Filters and Smoothers LOCATION:Donald Bren Hall 4011 DESCRIPTION:Ali Younis\, PhD Student\, Department of Computer Science\, Uni versity of California\, Irvine\n\nTitle: End-to-end Learnable Particle Fil ters and Smoothers\n\nAbstract: Estimating the temporal state of a system from image sequences is an important task for many vision and robotics app lications. A number of classical frameworks for state estimation have been proposed\, but often these methods require human experts to specify the s ystem dynamics and measurement model\, requiring simplifying assumptions t hat hurt performance. In this presentation\, I will develop end-to-end lea rnable particle filters and particle smoothers\, and show how to bring cla ssic state estimation methods into the age of deep learning. We first crea te an end-to-end learnable particle filter that uses flexible neural netwo rks to propagate multimodal\, particle-based representations of state unce rtainty. We then expand on our particle filtering method to create the fir st end-to-end learnable particle smoother\, which incorporates information from future as well as past observations\, and apply this particle smooth er to the real-world task of city-scale geo-localization using camera and planimetric map data.\n\nhttps://cml.ics.uci.edu/seminars/2024-12-02-ali-y ounis X-ALT-DESC;FMTTYPE=text/html:Ali Younis\, PhD Student\, Department of Computer Science\, University of California\, Irvine

Title: End-to-end Learnable Particle Filters and Smoothers

Abstract: Estimating the temporal state of a system from image seque nces is an important task for many vision and robotics applications. A num ber of classical frameworks for state estimation have been proposed\, but often these methods require human experts to specify the system dynamics a nd measurement model\, requiring simplifying assumptions that hurt perform ance. In this presentation\, I will develop end-to-end learnable particle filters and particle smoothers\, and show how to bring classic state estim ation methods into the age of deep learning. We first create an end-to-end learnable particle filter that uses flexible neural networks to propagate multimodal\, particle-based representations of state uncertainty. We then expand on our particle filtering method to create the first end-to-end le arnable particle smoother\, which incorporates information from future as well as past observations\, and apply this particle smoother to the real-w orld task of city-scale geo-localization using camera and planimetric map data.

https://cml.ics.uci.edu/seminars/2024-12-02-ali-younis URL:https://cml.ics.uci.edu/seminars/2024-12-02-ali-younis END:VEVENT BEGIN:VEVENT UID:2025-01-13-dongxia-wu@cml.ics.uci.edu DTSTAMP:20250113T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250113T130000 DTEND;TZID=America/Los_Angeles:20250113T140000 SUMMARY:[CML Seminar] Dongxia Wu: Uncertainty Quantification for Scientific Machine Learning LOCATION:Donald Bren Hall 4011 DESCRIPTION:Dongxia Wu\, PhD Student\, Dept. of Computer Science & Engineer ing\, University of California\, San Diego\n\nTitle: Uncertainty Quantific ation for Scientific Machine Learning\n\nAbstract: Scientific Machine Lear ning (SML) is an emerging interdisciplinary field with wide-ranging applic ations in domains such as public health\, climate science\, and drug disco very. The primary goal of SML is to develop data-driven surrogate models t hat can learn spatiotemporal dynamics or predict key system properties\, t hereby accelerating time-intensive simulations and reducing the need for r eal-world experiments. To make SML approaches truly reliable for domain ex perts\, Uncertainty Quantification (UQ) plays a critical role in enabling risk assessment and informed decision-making. In this presentation\, I wil l first introduce our recent advancements in UQ for spatiotemporal and mul ti-fidelity surrogate modeling with Bayesian deep learning\, focusing on a pplications in accelerating computational epidemiology simulations. Follow ing this\, I will demonstrate how quantified uncertainties can be leverage d to design sample-efficient algorithms for adaptive experimental design\, with a focus on Bayesian Active Learning and Bayesian Optimization.\n\nht tps://cml.ics.uci.edu/seminars/2025-01-13-dongxia-wu X-ALT-DESC;FMTTYPE=text/html:Dongxia Wu\, PhD Student\, Dept. of Computer Science &\; Engineering\, University of California\, San Diego

Title: Uncertainty Quantification for Scientific M achine Learning

Abstract: Scientific Machine Learning (SML) is an emerging interdisciplinary field with wide-ranging applications in d omains such as public health\, climate science\, and drug discovery. The p rimary goal of SML is to develop data-driven surrogate models that can lea rn spatiotemporal dynamics or predict key system properties\, thereby acce lerating time-intensive simulations and reducing the need for real-world e xperiments. To make SML approaches truly reliable for domain experts\, Unc ertainty Quantification (UQ) plays a critical role in enabling risk assess ment and informed decision-making. In this presentation\, I will first int roduce our recent advancements in UQ for spatiotemporal and multi-fidelity surrogate modeling with Bayesian deep learning\, focusing on applications in accelerating computational epidemiology simulations. Following this\, I will demonstrate how quantified uncertainties can be leveraged to design sample-efficient algorithms for adaptive experimental design\, with a foc us on Bayesian Active Learning and Bayesian Optimization.

https://cml.ics.uc i.edu/seminars/2025-01-13-dongxia-wu URL:https://cml.ics.uci.edu/seminars/2025-01-13-dongxia-wu END:VEVENT BEGIN:VEVENT UID:2025-01-27-satish-kumar-thittamaranahalli@cml.ics.uci.edu DTSTAMP:20250127T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250127T130000 DTEND;TZID=America/Los_Angeles:20250127T140000 SUMMARY:[CML Seminar] Satish Kumar Thittamaranahalli: Revisiting FastMap: N ew Applications LOCATION:Donald Bren Hall 4011 DESCRIPTION:Satish Kumar Thittamaranahalli\, Research Associate Professor\, Department of Computer Science\, University of Southern California\n\nTit le: Revisiting FastMap: New Applications\n\nAbstract: FastMap was first in troduced in the Data Mining community for generating Euclidean embeddings of complex objects. In this talk\, I will first generalize FastMap to gene rate Euclidean embeddings of graphs in near-linear time: The pairwise Eucl idean distances approximate a desired graph-based distance function on the vertices. I will then apply the graph version of FastMap to efficiently s olve various graph-theoretic problems of significant interest in AI: inclu ding shortest-path computations\, facility location\, top-K centrality com putations\, and community detection and block modeling. I will also presen t a novel learning framework\, called FastMapSVM\, by combining FastMap an d Support Vector Machines. I will then apply FastMapSVM to predict the sat isfiability of Constraint Satisfaction Problems and to classify seismogram s in Earthquake Science.\n\nhttps://cml.ics.uci.edu/seminars/2025-01-27-sa tish-kumar-thittamaranahalli X-ALT-DESC;FMTTYPE=text/html:Satish Kumar Thittamaranahalli\, Research Associate Professor\, Department of Computer Science\, Univ ersity of Southern California

Title: Revisiting FastMap: New Applications

Abstract: FastMap was first introduced in the Data Mining community for generating Euclidean embeddings of complex objec ts. In this talk\, I will first generalize FastMap to generate Euclidean e mbeddings of graphs in near-linear time: The pairwise Euclidean distances approximate a desired graph-based distance function on the vertices. I wil l then apply the graph version of FastMap to efficiently solve various gra ph-theoretic problems of significant interest in AI: including shortest-pa th computations\, facility location\, top-K centrality computations\, and community detection and block modeling. I will also present a novel learni ng framework\, called FastMapSVM\, by combining FastMap and Support Vector Machines. I will then apply FastMapSVM to predict the satisfiability of C onstraint Satisfaction Problems and to classify seismograms in Earthquake Science.

https://cml.ics.uci.edu/seminars/2025-01-27-sat ish-kumar-thittamaranahalli URL:https://cml.ics.uci.edu/seminars/2025-01-27-satish-kumar-thittamaranaha lli END:VEVENT BEGIN:VEVENT UID:2025-02-03-francesco-immorlano@cml.ics.uci.edu DTSTAMP:20250203T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250203T130000 DTEND;TZID=America/Los_Angeles:20250203T140000 SUMMARY:[CML Seminar] Francesco Immorlano: Transferring Climate Change Phys ical Knowledge LOCATION:Donald Bren Hall 4011 DESCRIPTION:Francesco Immorlano\, Postdoctoral Researcher\, Department of C omputer Science\, University of California\, Irvine\n\nTitle: Transferring Climate Change Physical Knowledge\n\nAbstract: Earth system models (ESMs) are the main tools currently used to project global mean temperature rise according to several future greenhouse gases emissions scenarios. Accurat e and precise climate projections are required for climate adaptation and mitigation\, but these models still exhibit great uncertainties that are a major roadblock for policy makers. Several approaches have been developed to reduce the spread of climate projections\, yet those methods cannot ca pture the non-linear complexity inherent in the climate system. Using a Tr ansfer Learning approach\, Machine Learning can leverage and combine the k nowledge gained from ESMs simulations and historical observations to more accurately project global surface air temperature fields in the 21st centu ry. This helps enhance the representation of future projections and their associated spatial patterns which are critical to climate sensitivity.\n\n https://cml.ics.uci.edu/seminars/2025-02-03-francesco-immorlano X-ALT-DESC;FMTTYPE=text/html:Francesco Immorlano\, Postd octoral Researcher\, Department of Computer Science\, University of Califo rnia\, Irvine

Title: Transferring Climate Change Physical Kn owledge

Abstract: Earth system models (ESMs) are the main to ols currently used to project global mean temperature rise according to se veral future greenhouse gases emissions scenarios. Accurate and precise cl imate projections are required for climate adaptation and mitigation\, but these models still exhibit great uncertainties that are a major roadblock for policy makers. Several approaches have been developed to reduce the s pread of climate projections\, yet those methods cannot capture the non-li near complexity inherent in the climate system. Using a Transfer Learning approach\, Machine Learning can leverage and combine the knowledge gained from ESMs simulations and historical observations to more accurately proje ct global surface air temperature fields in the 21st century. This helps e nhance the representation of future projections and their associated spati al patterns which are critical to climate sensitivity.

https://cml. ics.uci.edu/seminars/2025-02-03-francesco-immorlano URL:https://cml.ics.uci.edu/seminars/2025-02-03-francesco-immorlano END:VEVENT BEGIN:VEVENT UID:2025-02-10-kolby-nottingham@cml.ics.uci.edu DTSTAMP:20250210T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250210T130000 DTEND;TZID=America/Los_Angeles:20250210T140000 SUMMARY:[CML Seminar] Kolby Nottingham: Aligning Language Model Agents to E nvironment Dynamics LOCATION:Donald Bren Hall 4011 DESCRIPTION:Kolby Nottingham\, PhD Student\, Department of Computer Science \, University of California\, Irvine\n\nTitle: Aligning Language Model Age nts to Environment Dynamics\n\nAbstract: Language model agents are tacklin g challenging tasks from embodied planning to web navigation to programmin g. These models are a powerful artifact of natural language processing res earch that are being applied to interactive environments traditionally res erved for reinforcement learning. However\, many environments are not nati vely expressed in language\, resulting in poor alignment between language representations and true states and actions. Additionally\, while language models are generally capable\, their biases from pretraining can be unali gned with specific environment dynamics. In this talk\, I cover our resear ch into rectifying these issues through methods such as: (1) mapping high- level language model plans to low-level actions\, (2) optimizing language model agent inputs using reinforcement learning\, and (3) in-context polic y improvement for continual task adaptation.\n\nhttps://cml.ics.uci.edu/se minars/2025-02-10-kolby-nottingham X-ALT-DESC;FMTTYPE=text/html:Kolby Nottingham\, PhD Stud ent\, Department of Computer Science\, University of California\, Irvine
Title: Aligning Language Model Agents to Environment Dynamics

Abstract: Language model agents are tackling challenging ta sks from embodied planning to web navigation to programming. These models are a powerful artifact of natural language processing research that are b eing applied to interactive environments traditionally reserved for reinfo rcement learning. However\, many environments are not natively expressed i n language\, resulting in poor alignment between language representations and true states and actions. Additionally\, while language models are gene rally capable\, their biases from pretraining can be unaligned with specif ic environment dynamics. In this talk\, I cover our research into rectifyi ng these issues through methods such as: (1) mapping high-level language m odel plans to low-level actions\, (2) optimizing language model agent inpu ts using reinforcement learning\, and (3) in-context policy improvement fo r continual task adaptation.

https://cml.ics.uci.edu/seminars/2025-02- 10-kolby-nottingham URL:https://cml.ics.uci.edu/seminars/2025-02-10-kolby-nottingham END:VEVENT BEGIN:VEVENT UID:2025-02-24-valentina-pyatkin@cml.ics.uci.edu DTSTAMP:20250224T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250224T110000 DTEND;TZID=America/Los_Angeles:20250224T120000 SUMMARY:[CML Seminar] Valentina Pyatkin: Training Precise Language Models f or Imprecise Humans LOCATION:Donald Bren Hall 4011 DESCRIPTION:Valentina Pyatkin\, Postdoctoral Researcher\, Allen Institute f or AI and University of Washington\n\nTitle: Training Precise Language Mod els for Imprecise Humans\n\nAbstract: This talk examines methods for enhan cing language model capabilities through post-training. While large langua ge models have led to major breakthroughs in natural language processing\, significant challenges persist due to the inherent ambiguity and underspe cification in language. I will present a spectrum ranging from underspecif ication (preference modeling) to full specification (precise instruction f ollowing with verifiable constraints)\, and propose modeling approaches to increase language models' contextual robustness and precision. I will dem onstrate how models can become more precise instruction followers through synthetic data\, preference tuning\, and reinforcement learning from verif iable rewards. On the preference data side\, I will illustrate patterns of divergence in annotations\, showing how disagreements stem from underspec ification\, and propose alternatives to the Bradley-Terry reward model for capturing pluralistic preferences. The talk concludes by connecting under specification and reinforcement learning through a novel method: reinforce d clarification question generation\, which helps models obtain missing co ntextual information that is consequential for making predictions.\n\nhttp s://cml.ics.uci.edu/seminars/2025-02-24-valentina-pyatkin X-ALT-DESC;FMTTYPE=text/html:Valentina Pyatkin\, Postdoc toral Researcher\, Allen Institute for AI and University of Washington

Title: Training Precise Language Models for Imprecise Humans

Abstract: This talk examines methods for enhancing language mo del capabilities through post-training. While large language models have l ed to major breakthroughs in natural language processing\, significant cha llenges persist due to the inherent ambiguity and underspecification in la nguage. I will present a spectrum ranging from underspecification (prefere nce modeling) to full specification (precise instruction following with ve rifiable constraints)\, and propose modeling approaches to increase langua ge models' contextual robustness and precision. I will demonstrate how mod els can become more precise instruction followers through synthetic data\, preference tuning\, and reinforcement learning from verifiable rewards. O n the preference data side\, I will illustrate patterns of divergence in a nnotations\, showing how disagreements stem from underspecification\, and propose alternatives to the Bradley-Terry reward model for capturing plura listic preferences. The talk concludes by connecting underspecification an d reinforcement learning through a novel method: reinforced clarification question generation\, which helps models obtain missing contextual informa tion that is consequential for making predictions.

https://cml.ics.uc i.edu/seminars/2025-02-24-valentina-pyatkin URL:https://cml.ics.uci.edu/seminars/2025-02-24-valentina-pyatkin END:VEVENT BEGIN:VEVENT UID:2025-02-25-yanai-elazar@cml.ics.uci.edu DTSTAMP:20250225T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250225T110000 DTEND;TZID=America/Los_Angeles:20250225T120000 SUMMARY:[CML Seminar] Yanai Elazar: Understanding Generative Models Inside Out: From Representation to Data LOCATION:Donald Bren Hall 4011 DESCRIPTION:Yanai Elazar\, Postdoctoral Researcher\, Allen Institute for AI and University of Washington\n\nTitle: Understanding Generative Models In side Out: From Representation to Data\n\nAbstract: Generative models\, suc h as ChatGPT and DALL-E\, are used by millions of people daily for tasks r anging from programming and content creation to resume filtering. These mo dels often create the impression of being intelligent\, which can incentiv ize careless use in critical applications. While generative models are emp owering\, they appear to be black boxes\, and their misuse can result in h armful or unlawful outcomes. In this talk\, I will present algorithms and tools for dissecting and analyzing generative models using holistic\, caus al\, and data-centric approaches. By applying these methods to state-of-th e-art models\, we can foster trust in these technologies by uncovering hum an-interpretable concepts that underpin their behavior\, scrutinizing thei r extensive training data\, and evaluating their learning processes. Final ly\, I will reflect on how generative models have transformed the field of AI and discuss the challenges that remain in ensuring their responsible d evelopment and use.\n\nhttps://cml.ics.uci.edu/seminars/2025-02-25-yanai-e lazar X-ALT-DESC;FMTTYPE=text/html:Yanai Elazar\, Postdoctoral Researcher\, Allen Institute for AI and University of Washington

Title: Understanding Generative Models Inside Out: From Representati on to Data

Abstract: Generative models\, such as ChatGPT and DALL-E\, are used by millions of people daily for tasks ranging from prog ramming and content creation to resume filtering. These models often creat e the impression of being intelligent\, which can incentivize careless use in critical applications. While generative models are empowering\, they a ppear to be black boxes\, and their misuse can result in harmful or unlawf ul outcomes. In this talk\, I will present algorithms and tools for dissec ting and analyzing generative models using holistic\, causal\, and data-ce ntric approaches. By applying these methods to state-of-the-art models\, w e can foster trust in these technologies by uncovering human-interpretable concepts that underpin their behavior\, scrutinizing their extensive trai ning data\, and evaluating their learning processes. Finally\, I will refl ect on how generative models have transformed the field of AI and discuss the challenges that remain in ensuring their responsible development and u se.

https://cml.ics.uci.edu/seminars/2025-02-25-yanai-elazar URL:https://cml.ics.uci.edu/seminars/2025-02-25-yanai-elazar END:VEVENT BEGIN:VEVENT UID:2025-02-27-guandao-yang@cml.ics.uci.edu DTSTAMP:20250227T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250227T110000 DTEND;TZID=America/Los_Angeles:20250227T120000 SUMMARY:[CML Seminar] Guandao Yang: Toward Spatial Intelligence with Limite d Data LOCATION:Donald Bren Hall 4011 DESCRIPTION:Guandao Yang\, Postdoctoral Scholar\, Stanford University\n\nTi tle: Toward Spatial Intelligence with Limited Data\n\nAbstract: Current su ccess in artificial intelligence relies heavily on internet-scale data wit h unified representations. However\, such large-scale homogeneous data is not readily available for spatial computing applications involving 3D geom etry. In this talk\, I will present approaches to building spatial intelli gence systems with limited 3D data by combining existing mathematical mode ls into existing machine learning pipelines. I will share my work applying these approaches to develop data-driven methods that can synthesize and a nalyze 3D geometry. Finally\, I will discuss future opportunities and chal lenges of data-efficient spatial intelligence.\n\nhttps://cml.ics.uci.edu/ seminars/2025-02-27-guandao-yang X-ALT-DESC;FMTTYPE=text/html:Guandao Yang\, Postdoctoral Scholar\, Stanford University

Title: Toward Spatial Intelli gence with Limited Data

Abstract: Current success in artific ial intelligence relies heavily on internet-scale data with unified repres entations. However\, such large-scale homogeneous data is not readily avai lable for spatial computing applications involving 3D geometry. In this ta lk\, I will present approaches to building spatial intelligence systems wi th limited 3D data by combining existing mathematical models into existing machine learning pipelines. I will share my work applying these approache s to develop data-driven methods that can synthesize and analyze 3D geomet ry. Finally\, I will discuss future opportunities and challenges of data-e fficient spatial intelligence.

https://cml.ics.uci.edu/seminars/2025-02-27 -guandao-yang URL:https://cml.ics.uci.edu/seminars/2025-02-27-guandao-yang END:VEVENT BEGIN:VEVENT UID:2025-03-11-sadhika-malladi@cml.ics.uci.edu DTSTAMP:20250311T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250311T110000 DTEND;TZID=America/Los_Angeles:20250311T120000 SUMMARY:[CML Seminar] Sadhika Malladi: Deep Learning Theory in the Age of G enerative AI LOCATION:Donald Bren Hall 4011 DESCRIPTION:Sadhika Malladi\, PhD Student\, Department of Computer Science\ , Princeton University\n\nTitle: Deep Learning Theory in the Age of Genera tive AI\n\nAbstract: Modern deep learning has achieved remarkable results\ , but the design of training methodologies largely relies on guess-and-che ck approaches. Thorough empirical studies of recent massive language model s (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 theoretic al 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 empi rical settings published after the theory.\n\nhttps://cml.ics.uci.edu/semi nars/2025-03-11-sadhika-malladi X-ALT-DESC;FMTTYPE=text/html:Sadhika Malladi\, PhD Stude nt\, Department of Computer Science\, Princeton University

Title : Deep Learning Theory in the Age of Generative AI

Abstract: Modern deep learning has achieved remarkable results\, but the design of training methodologies largely relies on guess-and-check approaches. T horough empirical studies of recent massive language models (LMs) is prohi bitively expensive\, underscoring the need for theoretical insights\, but classical ML theory struggles to describe modern training paradigms. I pre sent a novel approach to developing prescriptive theoretical results that can directly translate to improved training methodologies for LMs. My rese arch has yielded actionable improvements in model training across the LM d evelopment pipeline — for example\, my theory motivates the design of Me ZO\, a fine-tuning algorithm that reduces memory usage by up to 12x and ha lves 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 p ublished after the theory.

https://cml.ics.uci.edu/seminars/2025-03-11- sadhika-malladi URL:https://cml.ics.uci.edu/seminars/2025-03-11-sadhika-malladi END:VEVENT BEGIN:VEVENT UID:2025-03-13-tianyu-gao@cml.ics.uci.edu DTSTAMP:20250313T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250313T110000 DTEND;TZID=America/Los_Angeles:20250313T120000 SUMMARY:[CML Seminar] Tianyu Gao: Enabling Language Models to Process Infor mation at Scale LOCATION:Donald Bren Hall 4011 DESCRIPTION:Tianyu Gao\, PhD Student\, Department of Computer Science\, Pri nceton University\n\nTitle: Enabling Language Models to Process Informatio n at Scale\n\nAbstract: Language models (LMs) can effectively internalize knowledge from vast amounts of pre-training data\, enabling them to achiev e remarkable performance on exam-style benchmarks. Expanding their ability to compile\, synthesize\, and reason over large volumes of information on the fly will further unlock transformative applications\, ranging from AI literature assistants to generative search engines. In this talk\, I will present my research on advancing LMs for processing information at scale. (1) I will present my evaluation framework for LM-based information-seeki ng systems\, emphasizing the importance of providing citations for verifyi ng the model-generated answers. (2) I will then introduce my foundational work on using contrastive learning to produce high-performing text embeddi ngs\, which form the cornerstone of effective and scalable search. (3) In addition to building systems that can process large-scale information\, I will discuss my contributions to creating efficient pre-training and custo mization methods for LMs. Finally\, I will share my vision for the next ge neration of autonomous information processing systems.\n\nhttps://cml.ics. uci.edu/seminars/2025-03-13-tianyu-gao X-ALT-DESC;FMTTYPE=text/html:Tianyu Gao\, PhD Student\, Department of Computer Science\, Princeton University

Title: Enabling Language Models to Process Information at Scale

Abstra ct: Language models (LMs) can effectively internalize knowledge from v ast amounts of pre-training data\, enabling them to achieve remarkable per formance on exam-style benchmarks. Expanding their ability to compile\, sy nthesize\, and reason over large volumes of information on the fly will fu rther unlock transformative applications\, ranging from AI literature assi stants to generative search engines. In this talk\, I will present my rese arch on advancing LMs for processing information at scale. (1) I will pres ent my evaluation framework for LM-based information-seeking systems\, emp hasizing the importance of providing citations for verifying the model-gen erated answers. (2) I will then introduce my foundational work on using co ntrastive learning to produce high-performing text embeddings\, which form the cornerstone of effective and scalable search. (3) In addition to buil ding systems that can process large-scale information\, I will discuss my contributions to creating efficient pre-training and customization methods for LMs. Finally\, I will share my vision for the next generation of auto nomous information processing systems.

https://cml.ics.uci.edu/seminars/2025 -03-13-tianyu-gao URL:https://cml.ics.uci.edu/seminars/2025-03-13-tianyu-gao END:VEVENT BEGIN:VEVENT UID:2025-03-17-anand-bhattad@cml.ics.uci.edu DTSTAMP:20250317T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250317T110000 DTEND;TZID=America/Los_Angeles:20250317T120000 SUMMARY:[CML Seminar] Anand Bhattad: What Generative Visual Models Understa nd (and Don't) About the Physical World LOCATION:Donald Bren Hall 4011 DESCRIPTION:Anand Bhattad\, Research Assistant Professor\, Toyota Technolog ical Institute at Chicago\n\nTitle: What Generative Visual Models Understa nd (and Don't) About the Physical World\n\nAbstract: Generative visual mod els like Stable Diffusion and Sora generate photorealistic images and vide os that are nearly indistinguishable from real ones to a naive observer. H owever\, their grasp of the physical world remains an open question: Do th ey understand 3D geometry\, light\, and object interactions\, or are they mere pixel parrots of their training data? Through systematic probing\, I will demonstrate that these models surprisingly learn fundamental scene pr operties — intrinsic images such as surface normals\, depth\, albedo\, a nd shading — without explicit supervision\, which enables applications l ike image relighting. But I will also show that this knowledge is insuffic ient. Careful analysis reveals unexpected failures: inconsistent shadows\, multiple vanishing points\, and scenes that defy basic physics. All these findings suggest these models excel at local texture synthesis but strugg le with global reasoning: a crucial gap between imitation and true underst anding. I will then conclude by outlining a path toward generative world m odels that emulate global and counterfactual reasoning\, causality\, and p hysics.\n\nhttps://cml.ics.uci.edu/seminars/2025-03-17-anand-bhattad X-ALT-DESC;FMTTYPE=text/html:Anand Bhattad\, Research As sistant Professor\, Toyota Technological Institute at Chicago

Ti tle: What Generative Visual Models Understand (and Don't) About the Ph ysical World

Abstract: Generative visual models like Stable Diffusion and Sora generate photorealistic images and videos that are near ly indistinguishable from real ones to a naive observer. However\, their g rasp of the physical world remains an open question: Do they understand 3D geometry\, light\, and object interactions\, or are they mere pixel parro ts of their training data? Through systematic probing\, I will demonstrate that these models surprisingly learn fundamental scene properties — int rinsic images such as surface normals\, depth\, albedo\, and shading — w ithout explicit supervision\, which enables applications like image religh ting. But I will also show that this knowledge is insufficient. Careful an alysis reveals unexpected failures: inconsistent shadows\, multiple vanish ing points\, and scenes that defy basic physics. All these findings sugges t these models excel at local texture synthesis but struggle with global r easoning: a crucial gap between imitation and true understanding. I will t hen conclude by outlining a path toward generative world models that emula te global and counterfactual reasoning\, causality\, and physics.

< a href="https://cml.ics.uci.edu/seminars/2025-03-17-anand-bhattad">https:/ /cml.ics.uci.edu/seminars/2025-03-17-anand-bhattad URL:https://cml.ics.uci.edu/seminars/2025-03-17-anand-bhattad END:VEVENT BEGIN:VEVENT UID:2025-03-18-wenting-zhao@cml.ics.uci.edu DTSTAMP:20250318T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250318T110000 DTEND;TZID=America/Los_Angeles:20250318T120000 SUMMARY:[CML Seminar] Wenting Zhao: Reasoning in the Wild LOCATION:Donald Bren Hall 4011 DESCRIPTION:Wenting Zhao\, PhD Student\, Department of Computer Science\, C ornell University\n\nTitle: Reasoning in the Wild\n\nAbstract: In this tal k\, I will discuss how to build natural language processing (NLP) systems that solve real-world problems requiring complex reasoning. I will address three key challenges. First\, because real-world reasoning tasks often di ffer from the data used in pretraining\, I will introduce WildChat\, a dat aset of reasoning questions collected from users\, and demonstrate how tra ining on it enhances language models' reasoning abilities. Second\, becaus e supervision is often limited in practice\, I will describe my approach t o enabling models to perform multi-hop reasoning without direct supervisio n. Finally\, since many real-world applications demand reasoning beyond na tural language\, I will introduce a language agent capable of acting on ex ternal feedback. I will conclude by outlining a vision for training the ne xt generation of AI reasoning models.\n\nhttps://cml.ics.uci.edu/seminars/ 2025-03-18-wenting-zhao X-ALT-DESC;FMTTYPE=text/html:Wenting Zhao\, PhD Student\ , Department of Computer Science\, Cornell University

Title: Reasoning in the Wild

Abstract: In this talk\, I will discu ss how to build natural language processing (NLP) systems that solve real- world problems requiring complex reasoning. I will address three key chall enges. First\, because real-world reasoning tasks often differ from the da ta used in pretraining\, I will introduce WildChat\, a dataset of reasonin g questions collected from users\, and demonstrate how training on it enha nces language models' reasoning abilities. Second\, because supervision is often limited in practice\, I will describe my approach to enabling model s to perform multi-hop reasoning without direct supervision. Finally\, sin ce many real-world applications demand reasoning beyond natural language\, I will introduce a language agent capable of acting on external feedback. I will conclude by outlining a vision for training the next generation of AI reasoning models.

https://cml.ics.uci.edu/seminars/2025-03-18-wenting- zhao URL:https://cml.ics.uci.edu/seminars/2025-03-18-wenting-zhao END:VEVENT BEGIN:VEVENT UID:2025-03-19-zongyi-li@cml.ics.uci.edu DTSTAMP:20250319T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250319T110000 DTEND;TZID=America/Los_Angeles:20250319T120000 SUMMARY:[CML Seminar] Zongyi Li: Neural Operator for Scientific Computing LOCATION:Donald Bren Hall 4011 DESCRIPTION:Zongyi Li\, PhD Student\, Computing and Mathematical Sciences\, Caltech\n\nTitle: Neural Operator for Scientific Computing\n\nAbstract: S cientific computing\, which aims to accurately simulate complex physical p henomena\, often requires substantial computational resources. By viewing data as continuous functions\, we leverage the smoothness structures of fu nction spaces to enable efficient large-scale simulations. We introduce th e neural operator\, a machine learning framework designed to approximate s olution operators in infinite-dimensional spaces\, achieving scalable phys ical simulations across diverse resolutions and geometries. Beginning with the Fourier Neural Operator\, we explore recent advancements including sc ale-consistent learning techniques and adaptive mesh methods. We demonstra te the real-world impact of our framework through applications in weather prediction\, carbon capture\, and plasma dynamics\, achieving speedups of several orders of magnitude.\n\nhttps://cml.ics.uci.edu/seminars/2025-03-1 9-zongyi-li X-ALT-DESC;FMTTYPE=text/html:Zongyi Li\, PhD Student\, C omputing and Mathematical Sciences\, Caltech

Title: Neural O perator for Scientific Computing

Abstract: Scientific comput ing\, which aims to accurately simulate complex physical phenomena\, often requires substantial computational resources. By viewing data as continuo us functions\, we leverage the smoothness structures of function spaces to enable efficient large-scale simulations. We introduce the neural operato r\, a machine learning framework designed to approximate solution operator s in infinite-dimensional spaces\, achieving scalable physical simulations across diverse resolutions and geometries. Beginning with the Fourier Neu ral Operator\, we explore recent advancements including scale-consistent l earning techniques and adaptive mesh methods. We demonstrate the real-worl d impact of our framework through applications in weather prediction\, car bon capture\, and plasma dynamics\, achieving speedups of several orders o f magnitude.

https://cml.ics.uci.edu/seminars/2025-03-19-zongyi-li URL:https://cml.ics.uci.edu/seminars/2025-03-19-zongyi-li END:VEVENT BEGIN:VEVENT UID:2025-04-01-sarah-wiegreffe@cml.ics.uci.edu DTSTAMP:20250401T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250401T110000 DTEND;TZID=America/Los_Angeles:20250401T120000 SUMMARY:[CML Seminar] Sarah Wiegreffe: Demystifying the Inner Workings of L anguage Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Sarah Wiegreffe\, Postdoctoral Researcher\, Allen Institute for AI and University of Washington\n\nTitle: Demystifying the Inner Workings of Language Models\n\nAbstract: Large language models (LLMs) power a rapi dly-growing and increasingly impactful suite of AI technologies. However\, due to their scale and complexity\, we lack a fundamental scientific unde rstanding of much of LLMs' behavior\, even when they are open source. The black-box nature of LMs not only complicates model debugging and evaluatio n\, but also limits trust and usability. In this talk\, I will describe ho w my research on interpretability (i.e.\, understanding models' inner work ings) has answered key scientific questions about how models operate. I wi ll then demonstrate how deeper insights into LLMs' behavior enable both 1) targeted performance improvements and 2) the production of transparent\, trustworthy explanations for human users.\n\nhttps://cml.ics.uci.edu/semin ars/2025-04-01-sarah-wiegreffe X-ALT-DESC;FMTTYPE=text/html:Sarah Wiegreffe\, Postdocto ral Researcher\, Allen Institute for AI and University of Washington
Title: Demystifying the Inner Workings of Language Models

Abstract: Large language models (LLMs) power a rapidly-growing and increasingly impactful suite of AI technologies. However\, due to their sc ale and complexity\, we lack a fundamental scientific understanding of muc h of LLMs' behavior\, even when they are open source. The black-box nature of LMs not only complicates model debugging and evaluation\, but also lim its trust and usability. In this talk\, I will describe how my research on interpretability (i.e.\, understanding models' inner workings) has answer ed key scientific questions about how models operate. I will then demonstr ate how deeper insights into LLMs' behavior enable both 1) targeted perfor mance improvements and 2) the production of transparent\, trustworthy expl anations for human users.

https://cml.ics.uci.edu/seminars/2025-04-01-s arah-wiegreffe URL:https://cml.ics.uci.edu/seminars/2025-04-01-sarah-wiegreffe END:VEVENT BEGIN:VEVENT UID:2025-04-04-xudong-wang@cml.ics.uci.edu DTSTAMP:20250404T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250404T140000 DTEND;TZID=America/Los_Angeles:20250404T150000 SUMMARY:[CML Seminar] XuDong Wang: Advancing Multimodal Models Beyond Human Supervision LOCATION:Donald Bren Hall 4011 DESCRIPTION:XuDong Wang\, PhD Student\, Berkeley AI Research Lab\, Universi ty of California\, Berkeley\n\nTitle: Advancing Multimodal Models Beyond H uman Supervision\n\nAbstract: To advance AI toward true artificial general intelligence\, it is crucial to incorporate a wider range of sensory inpu ts\, including physical interaction\, spatial navigation\, and social dyna mics. However\, achieving the successes of self-supervised Large Language Models (LLMs) across other modalities in our physical and digital environm ents remains a significant challenge. In this talk\, I will discuss how se lf-supervised learning methods can be harnessed to advance multimodal mode ls beyond the need for human supervision. Firstly\, I will highlight a ser ies of research efforts on self-supervised visual scene understanding that leverage the capabilities of self-supervised models to segment anything w ithout the need for 1.1 billion labeled segmentation masks. Secondly\, I w ill demonstrate how generative and understanding models can work together synergistically. Lastly\, I will explore the increasingly important techni ques for learning from unlabeled or imperfect data within the context of d ata-centric representation learning.\n\nhttps://cml.ics.uci.edu/seminars/2 025-04-04-xudong-wang X-ALT-DESC;FMTTYPE=text/html:XuDong Wang\, PhD Student\, Berkeley AI Research Lab\, University of California\, Berkeley

Title: Advancing Multimodal Models Beyond Human Supervision

Abstract: To advance AI toward true artificial general intelligence\, it is crucial to incorporate a wider range of sensory inputs\, including p hysical interaction\, spatial navigation\, and social dynamics. However\, achieving the successes of self-supervised Large Language Models (LLMs) ac ross other modalities in our physical and digital environments remains a s ignificant challenge. In this talk\, I will discuss how self-supervised le arning methods can be harnessed to advance multimodal models beyond the ne ed for human supervision. Firstly\, I will highlight a series of research efforts on self-supervised visual scene understanding that leverage the ca pabilities of self-supervised models to segment anything without the need for 1.1 billion labeled segmentation masks. Secondly\, I will demonstrate how generative and understanding models can work together synergistically. Lastly\, I will explore the increasingly important techniques for learnin g from unlabeled or imperfect data within the context of data-centric repr esentation learning.

https://cml.ics.uci.edu/seminars/2025-04-04-xudong-wan g URL:https://cml.ics.uci.edu/seminars/2025-04-04-xudong-wang END:VEVENT BEGIN:VEVENT UID:2025-04-21-felix-draxler@cml.ics.uci.edu DTSTAMP:20250421T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250421T130000 DTEND;TZID=America/Los_Angeles:20250421T140000 SUMMARY:[CML Seminar] Felix Draxler: Fast and Flexible Generative Modeling with Free-Form Flows LOCATION:Donald Bren Hall 4011 DESCRIPTION:Felix Draxler\, Postdoctoral Researcher\, Department of Compute r Science\, University of California\, Irvine\n\nTitle: Fast and Flexible Generative Modeling with Free-Form Flows\n\nAbstract: Generative models ha ve achieved remarkable quality and success in a variety of machine learnin g applications\, promising to become the standard paradigm for regression. However\, each predominant approach comes with drawbacks in terms of infe rence speed\, sample quality\, training stability\, or flexibility. In thi s talk\, I will propose Free-Form Flows\, a new generative model that offe rs fast data generation at high quality and flexibility. I will guide you through the fundamentals and showcase a variety of scientific applications .\n\nhttps://cml.ics.uci.edu/seminars/2025-04-21-felix-draxler X-ALT-DESC;FMTTYPE=text/html:Felix Draxler\, Postdoctora l Researcher\, Department of Computer Science\, University of California\, Irvine

Title: Fast and Flexible Generative Modeling with Fr ee-Form Flows

Abstract: Generative models have achieved rema rkable quality and success in a variety of machine learning applications\, promising to become the standard paradigm for regression. However\, each predominant approach comes with drawbacks in terms of inference speed\, sa mple quality\, training stability\, or flexibility. In this talk\, I will propose Free-Form Flows\, a new generative model that offers fast data gen eration at high quality and flexibility. I will guide you through the fund amentals and showcase a variety of scientific applications.

https://cml.i cs.uci.edu/seminars/2025-04-21-felix-draxler URL:https://cml.ics.uci.edu/seminars/2025-04-21-felix-draxler END:VEVENT BEGIN:VEVENT UID:2025-04-28-matu-s-dopiriak@cml.ics.uci.edu DTSTAMP:20250428T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250428T130000 DTEND;TZID=America/Los_Angeles:20250428T140000 SUMMARY:[CML Seminar] Matúš Dopiriak: Radiance Fields Advancing 3D Scene Understanding for Robotics and Autonomous Driving LOCATION:Donald Bren Hall 4011 DESCRIPTION:Matúš Dopiriak\, PhD Student\, Department of Computers and In formatics\, Technical University in Košice\n\nTitle: Radiance Fields Adva ncing 3D Scene Understanding for Robotics and Autonomous Driving\n\nAbstra ct: Since emerging in 2020\, neural radiance fields (NeRFs) have marked a transformative breakthrough in representing photorealistic 3D scenes. In t he years that followed\, numerous variants have evolved\, enhancing perfor mance\, enabling the capture of dynamic changes over time\, and tackling c hallenges in large-scale environments. Among these\, NVIDIA's Instant-NGP stood out\, earning recognition as one of TIME Magazine's Best Inventions of 2022. Radiance fields now facilitate advanced 3D scene understanding\, leveraging large language models (LLMs) and diffusion models to enable sop histicated scene editing and manipulation. Their applications span robotic s\, where they support planning\, navigation\, and manipulation. In autono mous driving\, they serve as immersive simulation systems or can be used a s digital twins for video compression integrated in edge computing archite ctures. This lecture explores the evolution\, capabilities\, and practical impact of radiance fields in these cutting-edge domains.\n\nhttps://cml.i cs.uci.edu/seminars/2025-04-28-matu-s-dopiriak X-ALT-DESC;FMTTYPE=text/html:Matúš Dopiriak\, PhD Stud ent\, Department of Computers and Informatics\, Technical University in Ko šice

Title: Radiance Fields Advancing 3D Scene Understandin g for Robotics and Autonomous Driving

Abstract: Since emergi ng in 2020\, neural radiance fields (NeRFs) have marked a transformative b reakthrough in representing photorealistic 3D scenes. In the years that fo llowed\, numerous variants have evolved\, enhancing performance\, enabling the capture of dynamic changes over time\, and tackling challenges in lar ge-scale environments. Among these\, NVIDIA's Instant-NGP stood out\, earn ing recognition as one of TIME Magazine's Best Inventions of 2022. Radianc e fields now facilitate advanced 3D scene understanding\, leveraging large language models (LLMs) and diffusion models to enable sophisticated scene editing and manipulation. Their applications span robotics\, where they s upport planning\, navigation\, and manipulation. In autonomous driving\, t hey serve as immersive simulation systems or can be used as digital twins for video compression integrated in edge computing architectures. This lec ture explores the evolution\, capabilities\, and practical impact of radia nce fields in these cutting-edge domains.

https://cml.ics.uci.edu/semin ars/2025-04-28-matu-s-dopiriak URL:https://cml.ics.uci.edu/seminars/2025-04-28-matu-s-dopiriak END:VEVENT BEGIN:VEVENT UID:2025-05-05-davide-corsi@cml.ics.uci.edu DTSTAMP:20250505T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250505T160000 DTEND;TZID=America/Los_Angeles:20250505T170000 SUMMARY:[CML Seminar] Davide Corsi: Safe Reinforcement Learning: Building A gents You Can Trust LOCATION:Interdisciplinary Science & Engineering Building 1010 DESCRIPTION:Davide Corsi\, Postdoctoral Researcher\, Department of Computer Science\, University of California\, Irvine\n\nTitle: Safe Reinforcement Learning: Building Agents You Can Trust\n\nAbstract: Reinforcement learnin g is increasingly used to train robots for tasks where safety is critical\ , such as autonomous surgery and navigation. However\, when combined with deep neural networks\, these systems can become unpredictable and difficul t to trust in contexts where even a single error is often unacceptable. Th is talk explores two complementary paths toward safer reinforcement learni ng: making agents more reliable through constrained training\, and adding formal guarantees through techniques such as verification and shielding. I n the second part of the talk\, we will look at the growing role of world modeling in robotics and how this\, together with the rise of large founda tion models\, opens up new challenges for ensuring safety in complex\, rea l-world environments.\n\nhttps://cml.ics.uci.edu/seminars/2025-05-05-david e-corsi X-ALT-DESC;FMTTYPE=text/html:Davide Corsi\, Postdoctoral Researcher\, Department of Computer Science\, University of California\, Irvine

Title: Safe Reinforcement Learning: Building Agents Y ou Can Trust

Abstract: Reinforcement learning is increasingl y used to train robots for tasks where safety is critical\, such as autono mous surgery and navigation. However\, when combined with deep neural netw orks\, these systems can become unpredictable and difficult to trust in co ntexts where even a single error is often unacceptable. This talk explores two complementary paths toward safer reinforcement learning: making agent s more reliable through constrained training\, and adding formal guarantee s through techniques such as verification and shielding. In the second par t of the talk\, we will look at the growing role of world modeling in robo tics and how this\, together with the rise of large foundation models\, op ens up new challenges for ensuring safety in complex\, real-world environm ents.

https://cml.ics.uci.edu/seminars/2025-05-05-davide-corsi< /html> URL:https://cml.ics.uci.edu/seminars/2025-05-05-davide-corsi END:VEVENT BEGIN:VEVENT UID:2025-08-07-julia-vogt@cml.ics.uci.edu DTSTAMP:20250807T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250807T130000 DTEND;TZID=America/Los_Angeles:20250807T140000 SUMMARY:[CML Seminar] Julia Vogt: From Models to Medicine: Theoretical Foun dations and Practical Impact of Machine Learning in Healthcare LOCATION:Donald Bren Hall 4011 DESCRIPTION:Julia Vogt\, Assistant Professor\, Department of Computer Scien ce\, ETH Zurich\n\nTitle: From Models to Medicine: Theoretical Foundations and Practical Impact of Machine Learning in Healthcare\n\nAbstract: The f usion of artificial intelligence and medicine is driving a healthcare tran sformation\, enabling personalized treatment tailored to each patient's un ique needs. By leveraging advances in machine learning\, we can transform challenging and diverse patient data into actionable insights that have th e potential to reshape clinical practice. In this presentation\, I will co ver both the theoretical foundations and practical applications of machine learning in healthcare. I will highlight innovative approaches designed t o enhance diagnostics\, improve patient care\, and make medical decision-m aking more accessible and reliable. I will touch on five fundamental\, int erconnected areas: multimodal data integration\, unsupervised structure de tection\, longitudinal data analysis\, transparent model development\, and development of decision support tools. Through real-world medical example s\, I aim to illuminate the collaborative future of machine learning and h ealthcare.\n\nhttps://cml.ics.uci.edu/seminars/2025-08-07-julia-vogt X-ALT-DESC;FMTTYPE=text/html:Julia Vogt\, Assistant Prof essor\, Department of Computer Science\, ETH Zurich

Title: F rom Models to Medicine: Theoretical Foundations and Practical Impact of Ma chine Learning in Healthcare

Abstract: The fusion of artific ial intelligence and medicine is driving a healthcare transformation\, ena bling personalized treatment tailored to each patient's unique needs. By l everaging advances in machine learning\, we can transform challenging and diverse patient data into actionable insights that have the potential to r eshape clinical practice. In this presentation\, I will cover both the the oretical foundations and practical applications of machine learning in hea lthcare. I will highlight innovative approaches designed to enhance diagno stics\, improve patient care\, and make medical decision-making more acces sible and reliable. I will touch on five fundamental\, interconnected area s: multimodal data integration\, unsupervised structure detection\, longit udinal data analysis\, transparent model development\, and development of decision support tools. Through real-world medical examples\, I aim to ill uminate the collaborative future of machine learning and healthcare.
https:/ /cml.ics.uci.edu/seminars/2025-08-07-julia-vogt URL:https://cml.ics.uci.edu/seminars/2025-08-07-julia-vogt END:VEVENT BEGIN:VEVENT UID:2025-08-28-daniel-neider@cml.ics.uci.edu DTSTAMP:20250828T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250828T130000 DTEND;TZID=America/Los_Angeles:20250828T140000 SUMMARY:[CML Seminar] Daniel Neider: A Gentle Introduction to Neural Networ k Verification LOCATION:Donald Bren Hall 4011 DESCRIPTION:Daniel Neider\, Professor of Verification and Formal Guarantees of Machine Learning\, Department of Computer Science\, TU Dortmund\n\nTit le: A Gentle Introduction to Neural Network Verification\n\nAbstract: Arti ficial Intelligence has become ubiquitous in modern life. This Cambrian ex plosion of intelligent systems has been made possible by extraordinary adv ances in machine learning\, especially in training deep neural networks an d their ingenious architectures. However\, like traditional hardware and s oftware\, neural networks often have defects\, which are notoriously diffi cult to detect and correct. Consequently\, deploying them in safety-critic al settings remains a substantial challenge. Motivated by the success of f ormal methods in establishing the reliability of safety-critical hardware and software\, numerous formal verification techniques for deep neural net works have emerged recently. As a guide through the vibrant and rapidly ev olving field of neural network verification\, this talk will give an overv iew of the fundamentals and core concepts of the field\, discuss prototypi cal examples of various existing verification approaches\, and showcase ho w generative AI can improve verification results.\n\nhttps://cml.ics.uci.e du/seminars/2025-08-28-daniel-neider X-ALT-DESC;FMTTYPE=text/html:Daniel Neider\, Professor o f Verification and Formal Guarantees of Machine Learning\, Department of C omputer Science\, TU Dortmund

Title: A Gentle Introduction t o Neural Network Verification

Abstract: Artificial Intellige nce has become ubiquitous in modern life. This Cambrian explosion of intel ligent systems has been made possible by extraordinary advances in machine learning\, especially in training deep neural networks and their ingeniou s architectures. However\, like traditional hardware and software\, neural networks often have defects\, which are notoriously difficult to detect a nd correct. Consequently\, deploying them in safety-critical settings rema ins a substantial challenge. Motivated by the success of formal methods in establishing the reliability of safety-critical hardware and software\, n umerous formal verification techniques for deep neural networks have emerg ed recently. As a guide through the vibrant and rapidly evolving field of neural network verification\, this talk will give an overview of the funda mentals and core concepts of the field\, discuss prototypical examples of various existing verification approaches\, and showcase how generative AI can improve verification results.

https://cml.ics.uci.edu/seminars/2025-0 8-28-daniel-neider URL:https://cml.ics.uci.edu/seminars/2025-08-28-daniel-neider END:VEVENT BEGIN:VEVENT UID:2026-05-21-dr-aahlad-puli@cml.ics.uci.edu DTSTAMP:20260521T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20260521T130000 DTEND;TZID=America/Los_Angeles:20260521T140000 SUMMARY:[CML Seminar] Dr. Aahlad Puli: Making the most of your Healthcare D ata for Reliable AI LOCATION:Donald Bren Hall 4011 DESCRIPTION:Dr. Aahlad Puli\, Faculty Fellow\, Center for Data Science at N YU\n\nTitle: Making the most of your Healthcare Data for Reliable AI\n\nAb stract: Scaling has become the dominant recipe for building AI systems tha t perform reliably across settings. In healthcare\, however\, data cannot be scaled in the same way as on the internet: it is expensive to collect\, hard to share\, and constrained by the availability of human expertise. T he central challenge\, then\, is how to build reliable healthcare AI by ma king the most of the data we already have. In this talk\, I present a new CHD risk score learned from EHR data\, then turn to the problem of interpr etation and trustworthiness. I introduce encoding as a fundamental obstacl e to interpreting black-box predictions with feature attributions\, and pr esent the first general formalization of this phenomenon. I then show that the CHD risk score exhibits poor transportability in certain settings\, a nd describe an approach to improve transportability without additional sup ervision. I conclude by highlighting opportunities for AI to improve evalu ation and decision-making in healthcare\, with a particular focus on adapt ing LLMs and world models to clinical settings.\n\nhttps://cml.ics.uci.edu /seminars/2026-05-21-dr-aahlad-puli X-ALT-DESC;FMTTYPE=text/html:Dr. Aahlad Puli\, Faculty F ellow\, Center for Data Science at NYU

Title: Making the mos t of your Healthcare Data for Reliable AI

Abstract: Scaling has become the dominant recipe for building AI systems that perform reliab ly across settings. In healthcare\, however\, data cannot be scaled in the same way as on the internet: it is expensive to collect\, hard to share\, and constrained by the availability of human expertise. The central chall enge\, then\, is how to build reliable healthcare AI by making the most of the data we already have. In this talk\, I present a new CHD risk score l earned from EHR data\, then turn to the problem of interpretation and trus tworthiness. I introduce encoding as a fundamental obstacle to interpretin g black-box predictions with feature attributions\, and present the first general formalization of this phenomenon. I then show that the CHD risk sc ore exhibits poor transportability in certain settings\, and describe an a pproach to improve transportability without additional supervision. I conc lude by highlighting opportunities for AI to improve evaluation and decisi on-making in healthcare\, with a particular focus on adapting LLMs and wor ld models to clinical settings.

https://cml.ics.uci.edu/seminars/2026-05 -21-dr-aahlad-puli URL:https://cml.ics.uci.edu/seminars/2026-05-21-dr-aahlad-puli END:VEVENT BEGIN:VEVENT UID:2026-06-10-riccardo-de-santi@cml.ics.uci.edu DTSTAMP:20260610T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20260610T130000 DTEND;TZID=America/Los_Angeles:20260610T140000 SUMMARY:[CML Seminar] Riccardo De Santi: Foundations of Generative Discover y Beyond the Data with Flow and Diffusion Models LOCATION:Donald Bren Hall 4011 DESCRIPTION:Riccardo De Santi\, PhD Student\, ETH AI Center\n\nTitle: Found ations of Generative Discovery Beyond the Data with Flow and Diffusion Mod els\n\nAbstract: Recent progress in flow and diffusion models has made gen erative models powerful priors over complex scientific design spaces\, whi le reward-guided adaptation offers a practical way to steer them toward de sired properties. This talk asks what is needed to turn such steering into discovery. I will first discuss tail-aware reward guidance: rather than m aximizing average reward\, one may deliberately sacrifice expected reward to concentrate probability on rare\, high-value samples in the top tail of the reward distribution. I will then argue that discovery also requires d ebiasing the generative model itself. In the natural sciences\, available data are local\, limited\, and biased by prior discoveries and measurement processes\, so distribution matching can hide valid low-probability modes . I will present Flow Density Control (FDC) as a framework for distributio nal fine-tuning of flow and diffusion models\, allowing to perform tasks i ncluding entropy-driven mode discovery\, risk-sensitive adaptation\, and e xperimental design. Finally\, I will introduce mathematical foundations fo r out-of-distribution flow modeling through generable-set expansion rather than standard distribution matching. And present Active Flow Expansion (A ctFlow)\, a synthetic pre-training method that uses verifier feedback and active exploration in learned flow representations to expand coverage over valid molecular\, peptide\, and protein space\, while enjoying first-of-t heir-kind statistical guarantees for out-of-distribution generative modeli ng.\n\nhttps://cml.ics.uci.edu/seminars/2026-06-10-riccardo-de-santi X-ALT-DESC;FMTTYPE=text/html:Riccardo De Santi\, PhD Stu dent\, ETH AI Center

Title: Foundations of Generative Discov ery Beyond the Data with Flow and Diffusion Models

Abstract: Recent progress in flow and diffusion models has made generative models p owerful priors over complex scientific design spaces\, while reward-guided adaptation offers a practical way to steer them toward desired properties . This talk asks what is needed to turn such steering into discovery. I wi ll first discuss tail-aware reward guidance: rather than maximizing averag e reward\, one may deliberately sacrifice expected reward to concentrate p robability on rare\, high-value samples in the top tail of the reward dist ribution. I will then argue that discovery also requires debiasing the gen erative model itself. In the natural sciences\, available data are local\, limited\, and biased by prior discoveries and measurement processes\, so distribution matching can hide valid low-probability modes. I will present Flow Density Control (FDC) as a framework for distributional fine-tuning of flow and diffusion models\, allowing to perform tasks including entropy -driven mode discovery\, risk-sensitive adaptation\, and experimental desi gn. Finally\, I will introduce mathematical foundations for out-of-distrib ution flow modeling through generable-set expansion rather than standard d istribution matching. And present Active Flow Expansion (ActFlow)\, a synt hetic pre-training method that uses verifier feedback and active explorati on in learned flow representations to expand coverage over valid molecular \, peptide\, and protein space\, while enjoying first-of-their-kind statis tical guarantees for out-of-distribution generative modeling.

https:/ /cml.ics.uci.edu/seminars/2026-06-10-riccardo-de-santi URL:https://cml.ics.uci.edu/seminars/2026-06-10-riccardo-de-santi END:VEVENT END:VCALENDAR