The Department of Mathematical Sciences will host Giacomo Borghi of Heriot-Watt University in Edinburgh, Scotland, on Thursday, October 31, from 12:45 p.m. – 1:45 p.m., as part of its math seminar series. Mr. Borghi will present, “Optimization in Bures-Wasserstein space via interacting particle systems.” This free event will take place in Armitage Hall, Room 121. Contact sm1380@camden.rutgers.edu for more information.
Abstract
Motivated by the development of computational algorithms dealing with probabilities, we present in this talk a novel optimization method based on Gaussian particles. The particles interact via a stochastic consensus-type dynamics exploiting the Riemannian structure of the Bures-Wasserstein space. We will also present an extension to the more general case of 2-Wasserstein space and compare the method with gradient flow dynamics in benchmark problems. This is joint work with M. Herty and A. Stavitskiy.