GL2019 Seminar Talk: (q,p)-Wasserstein GANs

Anton Mallasto
University of Copenhagen
Department of Computer Science

Vanilla WGANs minimize the 1-Wasserstein distance with l^2-metric as the ground distance. I will explain, how general p-Wasserstein metrics over l^q-metrics can be incorporated in the GAN setting through approximating the c-transforms of Kantorovich potentials (that are called discriminators in the GAN context).

Back to Geometry and Learning from Data in 3D and Beyond