How Accurate and Creative is Diffusion Model? A Quantification of its Generation Accuracy and Generalization Capability

Abstract

Diffusion model is a prevailing generative AI paradigm, and this talk will focus on quantifying its performance, for the sake of certifying its results, designing its hyperparameters, and assessing its applicability to downstream tasks.

The first part of the talk will focus on bounding diffusion model’s generation accuracy. The importance of this problem already led to a rich and substantial literature; however, prior theoretical investigations assumed that an approximate score function was already oracle-given with at most epsilon error, and focused on just analyzing diffusion model’s inference process. I will instead describe a first, end-to-end quantitative understanding of the entire generative pipeline, including both score training (optimization) and inference (sampling). The resulting error analysis will lead to insights on how to design both the training and inference processes for efficacious generation.

Then, diffusion model’s generalization capability will be discussed - when it is not memorizing the training data, what new samples will it generate? This question is not only pertinent to privacy and copyright considerations, but also important for understanding whether/how diffusion model creates new knowledge. The inductive bias of diffusion model’s generation will be made explicit, leading to a quantification of how diffusion model generalizes. This quantification is purely based on the empirical distribution without considering any population limit.

Bio

Molei Tao is a Professor and Richard Duke Fellow at Georgia Tech, a founding director of GT AI4Science Center, and currently a visiting professor at the Simons Institute at UC Berkeley. He received B.S. from Tsinghua Univ. and Ph.D. from Caltech, and worked as a Courant Instructor before moving to Georgia Tech. He is a recipient of W.P. Carey Ph.D. Prize in Applied Mathematics (2011), American Control Conference Best Student Paper Finalist (2013), NSF CAREER Award (2019), AISTATS best paper award (2020), IEEE EFTF-IFCS Best Student Paper Finalist (2021), Cullen-Peck Scholar Award (2022), GT-Emory AI.Humanity Award (2023), SONY Faculty Innovation Award (2024), Best Poster Award at the Recent Advances and Future Directions for Sampling conference (2024), and Richard Duke Fellowship (2025). Molei works on diffusion generative models and sampling, deep learning theory and optimization, and AI4Science.