The Wasserstein GAN (WGAN) framework, instead, uses the Wasserstein (Earth-Mover) distance between distributions, which in many cases does not suffer from loss explosion or vanishing gradient. In the WGAN framework, the loss functions of the generator and critic (which no longer emits a simple probability, but rather an approximation of the Wasserstein distance between the fake and real distributions) become the following:
In this, is the real distribution and is the distribution learned by the generator. The original WGAN ...
