Events Calendar

CS Colloquium: Stephen Tu (USC / ECE) - On the Effectiveness of Generative Modeling for Planning and Control

Speaker: Stephen Tu, USC / ECE

Talk Title: On the Effectiveness of Generative Modeling for Planning and Control

Abstract: Recent work has demonstrated that modern generative models—including diffusion models and flow matching methods—are a powerful tool for both representing control policies and also designing planning and control algorithms. However, despite strong empirical results, there is a lack of rigorous understanding for why these models work so well in very high-dimensional, autoregressive settings, and surprisingly do not seem to suffer from classic “curse of dimensionality” sample complexity barriers.   In this talk, we will shed some light on this phenomenon. First, we will show that shallow diffusion networks can be sample-efficiently learned in the presence of simple latent low-dimensional structures: the intrinsic dimension of the underlying distribution governs the sample complexity, rather than the ambient dimensionality of the problem. Second, we will show that diffusion/flow-matching models and losses are not necessary for learning performant policies in control tasks, and we can actually achieve similar performance using classic energy-based models trained with ranking noise-contrastive estimation—the latter which we prove is nearly asymptotically optimal. We will conclude with some exciting future directions for further investigation into the interplay between generative modeling, controls, and learning. 
 
This lecture satisfies requirements for CSCI 591: Research Colloquium

Biography: Stephen Tu is an assistant professor in the Department of Electrical and Computer Engineering at the University of Southern California, where he leads the Statistical Learning for Dynamics and Control group. His research interests span statistical learning theory, safe and optimal control, and robot learning. More specifically, his work has focused 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 Tu earned his Ph.D. in Electrical Engineering and Computer Sciences (EECS) from the University of California, Berkeley. Previous to joining USC, Stephen Tu was a research scientist at Google DeepMind Robotics where he focused on combining learning and control-theoretic approaches for robotics.

Host: CS Department

Location: Olin Hall of Engineering (OHE) - 132

Audiences: Everyone (USC) is invited

Contact: CS Faculty Affairs