机器学习与数据科学博士生系列论坛（第一百零九期）——High-Accuracy Sampling Using Only Scores: From Diffusion Models to Log-Concave Distributions
报告人：葛琦晖（北京大学）
时间：2026-11-05 16:00-17:00
地点：腾讯会议：441-3293-3791
摘要：
Diffusion models rely on learned score functions for sampling, but discretization error often causes the number of sampling steps to grow polynomially with inverse target accuracy. Can this dependence be reduced to polylogarithmic using only score evaluations?
In this talk, we introduce the core idea of first-order rejection sampling (FORS): using randomized estimates of log-density differences to implement rejection sampling with gradient information alone. We explain its application to the reverse transitions of diffusion models and the resulting high-accuracy guarantees in the bounded Lipschitz metric, under a finite second-moment assumption and suitable control of score estimation error in L^2. We conclude by discussing the impact of score estimation error and briefly outlining an application to log-concave sampling.
论坛简介：该线上论坛是由张志华教授机器学习实验室组织，每两周主办一次（除了公共假期）。论坛每次邀请一位博士生就某个前沿课题做较为系统深入的介绍，主题包括但不限于机器学习、高维统计学、运筹优化和理论计算机科学。
