2/3/26 (in-person) Prof. Michael Albergo, Harvard University
PSC 3150
Title: New frontiers in dynamical control of generative models
Abstract: Dynamical measure transport based generative models have seen great progress in the past 5 years. These methods work by learning a dynamical flow which connects samples from a simple distribution to samples under a distribution known through sample data. An open question is now how to best adapt these flows such that they satisfy certain control constraints, e.g. in AI for science and reasoning models. I will discuss new directions in this front as they pertain to overcoming efficiency challenges in these algorithms and demonstrate their efficacy at scale.Â
Further details https://go.umd.edu/statphys