BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Talks.cam//talks.cam.ac.uk// X-WR-CALNAME:Talks.cam BEGIN:VEVENT SUMMARY:Control and Adaptation Under Latent Risk and Nonstationary Interac tions - Yorie Nakahira\, Carnegie Mellon University DTSTART:20251127T140000Z DTEND:20251127T150000Z UID:TALK240547@talks.cam.ac.uk CONTACT:Fulvio Forni DESCRIPTION:Autonomous systems must operate safely in uncertain\, interact ive\, and nonstationary environments alongside humans. To enhance their ca pabilities\, we are developing techniques for stochastic safe control\, un certainty and risk quantification\, adaptation\, and language-guided contr ol. \n\nIn this talk\, we begin with quantifying and assuring safety from data in the presence of latent risks. Many systems contain unobservable va riables that render system dynamics partially unidentifiable or induce dis tribution shifts between offline and online statistics\, even when the und erlying dynamics remain unchanged. Such “spurious” distribution shifts often break standard approaches for risk quantification and stochastic sa fe control. To overcome this\, we propose a framework for designing data-d riven safety certificates for systems with latent risks. On the inference side\, the framework employs physics-informed learning/RL to estimate long -term risk from data without sufficient risk events\, while exploiting str uctural properties such as low-dimensional representations of risk or grap h decompositions of multi-agent systems. On the control side\, it builds o n a new notion of invariance\, termed probabilistic invariance\, which all ows safety conditions to be constructed from data\, despite spurious distr ibution shifts and partially unidentifiable dynamics.\n\nNext\, we introdu ce our work toward achieving lifelong safety in systems with self-seeking humans or adaptive opponents. Through modeling and experiments with human subjects\, we find that worst-case control can inadvertently induce advers arial opponent adaptation that increases risk in future interactions. This observation mirrors the empirical literature on social dilemmas and human risk compensation\, yet its implications for lifelong risk in nonstationa ry interactions have rarely been investigated. This result also suggests a n underexplored potential to proactively shape desirable opponent adaptati ons for enhanced performance and safety. \n\nFinally\, we will present our ongoing work on uncertainty quantification in neural networks\, sequentia l fine-tuning of Bayesian transformers\, and language-guided control. At t he core of our approach are analytic solutions for the moments of random v ariables passed through nonlinear activation functions. These solutions en able a moment propagation method that tracks mean vectors and covariance m atrices across networks\, providing sample-free tools for uncertainty quan tification and robustness analysis. Building on this\, we introduce a doub le-Bayesian framework for sequential transformer fine-tuning. This framewo rk reformulates fine-tuning as posterior inference across layers and time (data)\, which is reducible to one-pass propagation of analytic formulas\, eliminating the need for iterative gradient computation. These techniques are particularly useful for language-guided control when uncertainty esti mates are needed for robust decision-making.\n\nBio: Yorie Nakahira is an Assistant Professor in the Department of Electrical and Computer Engineeri ng at Carnegie Mellon University. She received B.E. in Control and Systems Engineering from Tokyo Institute of Technology and Ph.D. in Control and D ynamical Systems from California Institute of Technology. Her research goa l is to develop control and learning techniques that enhance the capabilit ies of autonomous systems. On the algorithm side\, her group studies robus t and safe control\, uncertainty and risk quantification\, adaptation algo rithms\, and language-guided control. On the application side\, her group explores diverse topics\, ranging from autonomous control systems to human sensorimotor control to poverty alleviation policy design. She has receiv ed four prestigious young investigator awards\, including the NSF CAREER\, and holds a part-time position at the Research and Development Center for Large Language Models\, where she applies control theory to LLM-jp. Her g roup will be recruiting Ph.D. students or postdoctoral researchers in 2026 .\n LOCATION:LR10\, Department of Engineering END:VEVENT END:VCALENDAR