Abstract
Connecting optimal transport and variational inference, we present aprincipled and systematic framework for sampling and generative modellingcentred around divergences on path space. Our work culminates in thedevelopment of the \emph{Controlled Monte Carlo Diffusion} sampler (CMCD) forBayesian computation, a score-based annealing technique that crucially adaptsboth forward and backward dynamics in a diffusion model. On the way, we clarifythe relationship between the EM-algorithm and iterative proportional fitting(IPF) for Schr{\"o}dinger bridges, deriving as well a regularised objectivethat bypasses the iterative bottleneck of standard IPF-updates. Finally, weshow that CMCD has a strong foundation in the Jarzinsky and Crooks identitiesfrom statistical physics, and that it convincingly outperforms competingapproaches across a wide array of experiments.