Optimal Transport-based Loss Functions for Machine Learning

Abstract

This short paper briefly reports the essential facets of the article (Kamsu-Foguem & al., 2022) presented and discussed as a Journal First paper. The article overviews generative neural networks whose loss functions are based on optimal transport with the Wasserstein distance. This tool of mathematical origin allows interesting automatic learning to be obtained in a reasoning time under Lipschitz constraints. As the proposed studies are based on Wasserstein Generative Adversarial Networks (WGAN), we conclude this report with a short discussion on how WGAN currently supports critical, intelligent applications in our society and nearly all industry sectors.