Marco Biroli (UChicago): Schmidt AI in Science Speaker Series

Organized by the University of Chicago’s Eric and Wendy Schmidt AI in Science Fellowship Program.

Agenda
4:00pm – 4:45pm:  Presentation
4:45pm – 5:00pm:  Q&A
5:00pm – 5:30pm: Reception

Meeting location
William Eckhardt Research Center. Room 401
5640 S Ellis Avenue, Chicago, IL 60637
Map It

Title: The assumptions of Variational AutoEncoders that nature disagrees with

Abstract: Variational AutoEncoders (VAEs) are often presented as flexible latent-variable models capable of discovering hidden structure in complex data. However, they have known and well documented limitations: blurry samples, posterior collapse, and poor performance on interacting systems. These limitations are usually explained heuristically, in this talk I argue that the core limitation of standard VAEs is a structural conditional-independence assumption. This perspective emerged unexpectedly from biological clock models, physical stochastic particle models and 1D Coulomb gas studies which made these structural flaws of VAE rise to the surface. I will show how VAEs can be used as a test to obtain structural information on the behavior of these natural systems, reframing VAEs not as universal generators, but as verifiable tests for conditional-independence, bridging modern representation learning with classical statistical physics and biological modeling.

Bio: Marco Biroli is an Eric and Wendy Schmidt AI in Science Fellow at UChicago under the supervision of Professor Vincenzo Vitelli. His research interests lie in non-equilibrium statistical physics, with particular focus on extreme value statistics of strongly correlated variables and interacting particle systems. He holds a Master’s degree in Theoretical Physics from École Normale Supérieure, Paris, and a Bachelor of Science with a double major in Mathematics and Physics from École Polytechnique, Palaiseau.

The goal of Marco Biroli’s research is to develop physically inspired machine learning models, drawing on advances at the interface of statistical physics, stochastic processes, and deep learning. The aim is to devise algorithms that respect underlying physical constraints (e.g., symmetries, conserved quantities, stochastic dynamics), thereby improving interpretability, generalization, and robustness.