Title: Theoretical foundations of diffusion models

Speaker: Sitan Chen, Assistant Professor of Computer Science at Harvard's John A. Paulson School of Engineering and Applied Sciences

Abstract: Diffusion models are the de facto method for generative modeling across a range of data modalities including video, audio, molecules, and visuomotor policies. In recent years there has been a surge of interest in developing theory to understand and build upon this method’s empirical successes. In this talk I will survey what we know so far. At the core of this theory is the surprising finding that diffusion models can provably sample from any probability distribution provided one can solve a certain supervised learning problem called score estimation. After discussing the intuition behind this result, I will overview recent efforts to prove rigorous algorithmic guarantees for score estimation. I will then provide some vignettes illustrating what this theory can say about diffusion models in practice, including a new, training-free method for accelerating the generation process for these models.

Speaker Bio: Sitan Chen is an Assistant Professor of Computer Science at Harvard University, where he is a member of the Theory of Computation, the ML Foundations group, and the Harvard Quantum Initiative. Previously, he was an NSF math postdoc at UC Berkeley, after completing his PhD in EECS at MIT in 2021. His research centers around developing rigorous guarantees for fundamental algorithmic problems in machine learning and quantum information.