SILO: Few-Step Likelihoods of Flow and Diffusion Models: Applications to Images and Molecules

Abstract

Flow and diffusion models define processes that generate samples from progressively less noisy versions of the data distribution. I will present a line of work on estimating the marginal probability densities along these processes, culminating in the density of the data distribution itself. Our methods are both sample-efficient to train and computationally efficient to evaluate. I will showcase their application to natural images and molecular configurations. For natural images, the estimated density correctly recovers likely and unlikely frequency patterns. For molecular configurations, it enables the computation of importance weights that correct model-generated samples using physical knowledge encoded by an energy function.

Bio

Omar Chehab is joining UW–Madison as a RISE-AI Assistant Professor in the Department of Electrical and Computer Engineering. He completed his graduate training in France, earning a PhD in Mathematical Computer Science at Inria under the supervision of Aapo Hyvärinen and Alexandre Gramfort. He subsequently held postdoctoral positions in the Department of Statistics at ENSAE/CREST, working with Anna Korba, and in the Machine Learning Department at Carnegie Mellon University, working with Pradeep Ravikumar.

His research focuses on principled methods for efficient inference from complex probability distributions. This includes estimating likelihoods from data, generating samples from unnormalized densities, as well as learning representations and discovering causal structure from brain imaging data. His work draws on a range of modern methods, including diffusion models, annealed MCMC, score matching, multi-view independent component analysis, and noise-contrastive estimation. More broadly, he studies these algorithms through the lens of computational and statistical efficiency, aiming to understand their fundamental limits and guide their design.

He regularly publishes at leading machine learning conferences such as NeurIPS, ICML, and ICLR.