Tue, 26 May 2026
12:00 -
13:00
C5
Understanding and mitigating the bias of Diffusion Posterior Sampling algorithm
Dr. Matias Delgadino
(University of Texas at Austin)
Abstract
We identify the bias in the Diffusion Posterior Sampling algorithm by the use of the classical Feynman-Kac formula. This analysis, the first of its kind, allows us to understand correction/improvements to the algorithm from first principles. We show how STSL, a better performing variant of DPS, can be derived from first principles using this analysis.