More Adversarial GANs
vendredi 29 novembre 2019, 14h00 - 15h00
The Generative Adversarial Networks (GAN) literature tackles the problem of imitating data from a dataset. Mathematically, GANs create data that are close to the dataset in terms of density distributions. One type of GANs are Wasserstein GANs because the criterion that is minimized is the Wasserstein Distance between generated and original data distributions. As a Wasserstein distance between distributions of points are built on a distance between these points (usually euclidean), this work focuses on the optimization of that distance on which we build our GAN. Surprisingly, new applications emerge from our new type of GANs and the optimized distances involved for data analysis.