From Score Estimation to Sampling

Abstract

Recent advances in the algorithmic generation of high-fidelity images, audio, and video have been largely driven by the success of score-based diffusion models. A key component in their implementation is score matching, which estimates the score function of the forward diffusion process from training data. In this work, we establish rate-optimal procedures for estimating the score function of smooth, compactly supported densities and explore their implications for density estimation and optimal transport.

Presenters

Harrison Zhou is the Henry Ford II Professor of Statistics and Data Science at Yale University. He earned his Ph.D. in Mathematics from Cornell University in 2004. A leading scholar in the statistical decision theory, Zhou is known for his work on the fundamental limits of statistical estimation. He currently serves as an Editor-in-Chief of The Annals of Statistics. During his tenure as department chair at Yale, he played a pivotal role in transforming the Department of Statistics into a full-fledged Data Science Department.