Theory for diffusion models

Data Science Seminar

Sitan Chen (Harvard)

Abstract

In this talk I will survey our recent efforts to develop rigorous theory for understanding diffusion generative modeling. The first part will cover discretization analyses that prove that diffusion models can approximately sample from arbitrary probability distributions provided one can has a sufficiently accurate estimate for the score function. The second part will cover new algorithms for score estimation that, in conjunction with the results in the first part, imply state-of-the-art bounds for learning Gaussian mixture models. Time permitting, the third part will then use the lens of mixture models to shed light on two intriguing empirical phenomena of diffusion models: the behavior of diffusion models under guidance, and the emergence of features in narrow windows of time during the generation process.