Towards a better understanding of Wasserstein GANs
schedule le lundi 24 février 2020 de 17h30 à 18h30
Organisé par : F. Bechtold, W. Da Silva , A. Fermanian, S. Has, Y. Yu
Intervenant : Ugo Tanielian (LPSM et Criteo Research)
Lieu : Jussieu, Salle Paul Lévy (16-26-209)
Sujet : Towards a better understanding of Wasserstein GANs
In recent years, Generative Adversarial Networks (GANs), as proposed in Goodfellow et al. (2014), have seen an incredible success and shown state-of-the-art results in image, video and text generation. Arjovsky et al. (2017) have shown the benefits of the Wasserstein GANs architecture bringing stabilization in the training and solving the mode collapse phenomenon. Building on this work, we are thus interested in analyzing the theoretical properties of Wasserstein GANs. The goal of the paper is to make small theoretical steps in understanding the convergence of Wasserstein GANs regarding some optimization and asymptotic properties. Theoretical results will also be backed and explicated with empirical examples on synthetic and real-world datasets.