Wasserstein Proximal of GANs

We introduce a new method for training generative adversarial networks by applying the Wasserstein-2 metric proximal on generators. The approach is based Wasserstein information geometry. It defines parametrization invariant natural gradient pulling back optimal transport structures from probability space to parameter space. obtain easy-to-implement iterative regularizers updates of implicit deep models. Our experiments demonstrate that this improves speed and stability in terms wall-clock time Fréchet Inception Distance.