David Banks obtained a Ph.D. in Statistics in 1984. He won an NSF Postdoctoral Research Fellowship in the Mathematical Sciences, which he took at Berkeley, working with David Blackwell. In 1986 he was a visiting assistant lecturer at the University of Cambridge, and then joined the Department of Statistics at Carnegie Mellon in 1987. In 1997 he went to the National Institute of Standards and Technology, then served as Chief Statistician of the U.S. Department of Transportation, and finally joined the U.S. Food and Drug Administration in 2002. In 2003, he returned to academics at Duke University.

He was the coordinating editor of the Journal of the American Statistical Association. He co-founded the journal Statistics and Public Policy and served as its editor. He co-founded the American Statistical Association's Sections on National Defense and Homeland Security and Text Analysis, and has chaired those sections, as well as the sections on Risk Analysis and Statistical Learning and Data Mining. David Banks is past-president of the Classification Society and the International Society for Business and Industrial Statistics. He has twice served on the Board of Directors of the American Statistical Association. He is a fellow of the American Statistical Association, the Institute of Mathematical Statistics, and the American Association for the Advancement of Science. He won the American Statistical Association's Founders Award, the De Groot Award, and gave the William Sealy Gosset and Deming Lectures. From January 2018 to Sept., 2021, he was the director of Statistical and Applied Mathematical Sciences Institute. His research areas include computational advertising, dynamic text networks, adversarial risk analysis (i.e., Bayesian behavioral game theory), human rights statistics, agent-based models, forensics, and certain topics in high-dimensional data analysis.

Abstract: The ground is changing under our feet. Our publication processes are antique. AI will affect our work in many ways, some of which are more predictable than others. Our educational curricula should be re-examined to better align our students with an emerging economy that favors data engineering over statistical theory. And although most of our funding mechanisms will never change, there would be huge benefits if they did.

Dean P. Foster studied AI at Rutgers and Statistics at the University of Maryland—before those fields merged into what we now call Machine Learning. He spent time as a professor at the University of Chicago and later at the University of Pennsylvania. In 2015, he made the leap from academia to industry, joining Amazon in New York City, where he's been ever since. His current research focuses on machine learning, reinforcement learning, and large language models (LLMs).

Dean helped pioneer two major areas in game theory: stochastic evolutionary game dynamics and calibrated learning. In both, he developed the theoretical tools needed to prove convergence to equilibrium. The calibrated learning strategies he introduced stemmed from his early work on individual sequences—work that has since become foundational in theoretical machine learning. His calibration and no-internal-regret algorithms were among the first learning methods proven to converge to a correlated equilibrium.

In statistics, he's best known for his work on large-scale regression problems. His early research on risk inflation was one of the first to seriously consider models with thousands—or even millions—of potential variables. More recently, his work on alpha-investing offers both a theoretical foundation for variable selection and a practical algorithm that's fast enough to keep pace with streaming data.

At Amazon, Dean founded a reinforcement learning team in New York City. The group is responsible for figuring out how much of each of Amazon's 30 million products to purchase each year—a $300 billion decision-making problem. Once those purchases are made, the team's work continues: routing inventory across hundreds of warehouses so that products can reach customers quickly. This entire pipeline is guided by a multi-agent reinforcement learning system.

These days, Dean is especially interested in how LLMs can be used not just to write code, but to prove theorems about code—a step toward systems that can reason about software as well as generate it.

Abstract: TBA

Bin Yu is Chancellor’s Distinguished Professor in the Departments of Statistics and Electrical Engineering and Computer Sciences at University of California, Berkeley. A member of both the National Academy of Sciences and the American Academy of Arts and Sciences, she is widely recognized for pioneering the Predictability–Computability–Stability (PCS) framework for veridical data science, which has become an influential approach for transparent, reproducible, and trustworthy data science and AI. Her group develops efficient and interpretable ML/AI methods and theory, including influential contributions such as iterative random forests (iRF), contextual decomposition for transformers, adaptive wavelet distillation, and LoRA+ methods for deep learning fine-tuning. Her research spans statistical theory, interpretable machine learning, causal inference, and scientific discovery, with applications in neuroscience, genomics, remote sensing, and precision medicine through close interdisciplinary collaborations with domain experts.

Professor Yu received her B.S. in Mathematics from Peking University and her M.S. and Ph.D. in Statistics from University of California, Berkeley. Her major honors include the COPSS Distinguished Achievement Award and Lectureship, the COPSS E. L. Scott Award, and a Guggenheim Fellowship. She has delivered many distinguished lectures, including the Breiman Lecture at NeurIPS, the Wald Memorial Lectures of the Institute of Mathematical Statistics, and the COPSS Distinguished Achievement Award and Lectureship at JSM. Professor Yu has also served in numerous leadership roles in the statistics and data science communities, including President of the IMS, Chair of the Berkeley Statistics Department, and advisory and editorial positions for leading scientific institutes and journals.

Abstract: Deep learning achieves remarkable predictive performance, yet trustworthiness demands more—reliable evaluation and interpretable structure. This talk presents two pillars of veridical deep learning grounded in the PCS (Predictability, Computability, Stability) framework. First, we address rigorous evaluation of LLMs beyond held-out accuracy by introducing Green Shielding, a user-centric research agenda for building evidence-backed deployment guidance by characterizing how benign input variation shifts LLM behavior. Second, we explore compositional interpretability: how deep models encode interactions among input features, illustrated through SPEX and related methods for extracting sparse, human-readable explanations.