Go to AABI 2025 Workshop Track homepageDiffusion-based Sampling via Amortized Posterior Inference
Keywords: diffusion model; sampling; posterior inference
TL;DR: train a diffusion model to sample from unnormalized distributions, where the training target is constructed by posterior inference results
Abstract: Recent years have seen a surge of interest in utilizing diffusion models for sampling from an unnormalized target density without access to data. A common approach is to construct a reverse diffusion SDE whose terminal distribution matches the desired target. In this work, we propose a novel method that trains a diffusion model to sample from unnormalized distributions by reformulating the sampling problem as a sequence of posterior inference tasks. To effectively solve these inference problems, we leverage existing Monte Carlo methods and design a training objective based on the inference results. Empirically, our method demonstrates competitive performance on Gaussian mixture densities and achieves better performance on the DW-4 particle system, compared to other diffusion-based sampling methods. These results highlight the potential of integrating diffusion models with advanced posterior inference techniques for improved sampling in complex distributions.
Submission Number: 29
Loading