Luc Rey-Bellet – UMASS
Date/Time/Location
Thursday, April 30, 2026, 12:10 pm; Hill Center 705
Proximal optimal transport divergences for generative modeling and sampling.
The motivation for this work comes from generative modeling and sampling problems where the target distributions are either supported on low-dimensional structure and/or known only through empirical samples.
Through proximal regularization we construct new information theoretic divergences which combine in a flexible manner the desirable properties of relative entropy and optimal transport (Wasserstein distances). Using these divergences, associated Wasserstein gradient flows, and a mixture of implicit and explicit schemes we build efficient sampling and generative modeling algorithms.