ML/AI Seminar: Stephen Tu, University of Southern California

**REGISTRATION DEADLINE: **The Columbia Morningside campus is open to the Columbia community. If you do not have an active CUID, the deadline to register is at 12:00 PM the day before the event. External guests will receive a QR code by email to enter campus.

**Speaker**

Stephen Tu, is an assistant professor in the Department of Electrical and Computer Engineering at the University of Southern California. His research interests span statistical learning theory, safe and optimal control, and generative modeling. Specifically, his work focuses on non-asymptotic guarantees for learning dynamical systems, rigorous analysis of distribution shift in feedback settings, safe control synthesis, and more recently foundations of generative modeling. Stephen earned his Ph.D. in Electrical Engineering and Computer Sciences (EECS) from the University of California, Berkeley. Previous to joining USC, Stephen was a research scientist at Google DeepMind Robotics where he focused on combining learning and control-theoretic approaches for robotics.

**Abstract**

**Multistep Backwards Latent Losses Learn Unstable Linear Dynamics**

Model-based control synthesis relies on an accurate description of the system dynamics and the environment. When such models are unavailable or difficult to derive from first principles, they must be learned from system trajectories. Latent world models have recently emerged as a promising general-purpose approach, learning low-dimensional representations together with dynamics that evolve these representations to predict future behavior. However, it remains unclear which dynamical properties these models preserve, and whether they support control synthesis with guarantees of closed-loop stability and safety. Resolving these questions is essential to establishing latent world modeling as a reliable tool for control design. In this talk, we initiate a study of the dynamical properties captured by latent world model designs. For trajectory data generated by a linear dynamical system, we show that multi-step backward latent reconstruction losses recover the system’s unstable subspace, provided the latent dimension is large enough to represent the unstable modes. We then show that an H-infinity controller synthesized using the learned latent dynamics stabilizes the original system. Crucially, both multi-step prediction and reconstruction in the original space (i.e., backward reconstruction) are necessary for this guarantee: removing either can lead to an unstable closed-loop system. Although our analysis focuses on linear systems, experiments on high-dimensional nonlinear systems reveal behavior consistent with our theoretical findings.