Joint BCAM-UPV/EHU Analysis and PDE seminar: From gradient flow force-balance to robust machine learning

In this talk, I will introduce a variational approach to robust learning algorithms that seek learning and optimization solutions under distribution shifts – discrepancy between the training and test distribution.
One of the recent theoretical progress for such learning tasks is the adoption of geometries over probability measures, such as using the Wasserstein distance and kernel maximum mean discrepancy. The heart of our approaches is the explicit parametrization of the generalized force, in the dual space, from the gradient flow perspective. I will demonstrate how this can be used for robust machine learning in a principled manner.