Speaker: Prof. Michael (Misha) CHERTKOV - University of Arizona, Tucson, USA
Title: Path Integral Diffusion: From Analytic Bridges to Adaptive, Guided, and Mean-Field Sampling
Abstract
Diffusion models of generative AI are commonly formulated through learned score fields and reverse-time stochastic dynamics. I will present an alternative control-theoretic formulation—Path Integral Diffusion (PID)—in which generative sampling is posed as a stochastic bridge or optimal-transport problem with a potential. The optimal drift can then be expressed through forward and backward Green functions. For quadratic potentials, this construction is analytic and interpretable: the score, sampling drift, and predicted terminal state are available in closed form
