Learning Dynamical Transport without Data

In this seminar talk from the New Technologies in Mathematics series, Harvard researcher Michael Albergo explores innovative approaches to dynamical transport without relying on training data. Discover how score-based diffusion models and dynamical transport of measure have revolutionized generative modeling, and learn about a novel algorithm that generates samples from target distributions using only the unnormalized log-likelihood or energy function. The presentation covers applications in statistical physics, chemistry, and Bayesian inference, introducing an approach that enhances annealed importance sampling and sequential Monte Carlo methods. Albergo also discusses extending these concepts to discrete distribution dynamics, sharing insights from collaborative work with Eric Vanden-Eijnden, Peter Holderrieth, and Tommi Jaakkola.