Bringing Regularized Optimal Transport to Lightspeed

In this seminar, Mikael Johansson will introduce an efficient algorithm for regularized optimal transport, a crucial technique in various domains. What sets this method apart from previous approaches is its utilization of the Douglas-Rachford splitting technique, enabling the development of an efficient solver capable of handling a broad class of regularizers. This innovative algorithm boasts strong global convergence guarantees, a low per-iteration cost, and the remarkable ability to harness GPU parallelization. As a result, it significantly outpaces the state-of-the-art in solving many complex problems. The seminar will also feature practical demonstrations of its competitiveness in applications such as domain adaptation and learning of generative models. 

Joint Work

The research seminar will be based on collaborative research conducted with Jacob Lindbäck, Zesen Wang, and Vien Van Mai, further enhancing its depth and relevance.