Monte Carlo Seminar| Luhan Wu| Reverse Diffusion Sequential Monte Carlo Samplers скачать в хорошем качестве

Monte Carlo Seminar| Luhan Wu| Reverse Diffusion Sequential Monte Carlo Samplers

Tuesday, March 10, 2026 [sites.google.com/view/monte-carlo-seminar] Speaker: Luhuan Wu (Flatiron, Johns Hopkins University) Title: Reverse Diffusion Sequential Monte Carlo Samplers Abstract: Diffusion models have emerged as a powerful paradigm for generative modeling. In this talk, we explore their use as annealing paths for sampling from unnormalized target distributions. Building on prior work, we first present a unifying framework that leverages Monte Carlo methods to estimate score functions and simulate diffusion-based sampling trajectories. However, such approaches can suffer from accumulated bias due to time discretization and imperfect score estimation. To address these challenges, we introduce a principled Sequential Monte Carlo (SMC) framework that formalizes diffusion-based samplers as proposal mechanisms while systematically correcting their biases. The key idea is to construct informative intermediate target distributions that progressively guide particles toward the final distribution of interest. Although the ideal targets are intractable, we derive exact approximations using quantities already available from the score-based proposal, requiring no extra inference overhead. The resulting method, Reverse Diffusion Sequential Monte Carlo, enables consistent sampling and unbiased estimation of the target normalization constant. We demonstrate our method on a range of synthetic targets and Bayesian regression tasks.