机器学习与数据科学博士生系列论坛（第七十七期）—— Solving Inverse Problems Exactly with Diffusion Model: Sequential Monto Carlo
发文时间：2024-10-17
Speaker(s)：金开诚（北京大学）
Time：2024-10-17 16:00-17:00
Venue：腾讯会议 627-5441-1672
摘要：
Inverse problems are ubiquitous across several domains. The goal of solving an inverse problem is to recover samples from a Bayesian posterior or generate samples following guidance. A rich literature are dedicated to solve inverse problems efficiently. The tremendous success of diffusion-based generative models motivates new methods in tackling inverse problems. A plethora of methods have explored the potential of diffusion model in solving inverse problems, mainly through a backward diffusion process induced by carefully parametrized conditional score estimation.
While these methods circumvented the intractability of the posterior distribution and have led to success in applications, they don't guarantee exact recovery of posterior distribution theoretically. In this talk, we will first briefly introduce some of the basic algorithms solving inverse problem with diffusion model. Then, we will focus on methods based on Sequential Monto Carlo (SMC) that aims to sample from the true posterior distribution, approximated by a weighted interacting particle system. Additionally, we share some thoughts on extension of these methods and their connection with other fields.
论坛简介：该线上论坛是由张志华教授机器学习实验室组织，每两周主办一次（除了公共假期）。论坛每次邀请一位博士生就某个前沿课题做较为系统深入的介绍，主题包括但不限于机器学习、高维统计学、运筹优化和理论计算机科学。