Take a close look at mode collapse and vanishing gradient in GAN

Abstract

The Generative Adversarial Nets (GAN) has made great advances during these years. Many research branches of the GAN have appeared. However, the researchers have to face the difficulties of the several problems of the GAN. This paper presents a literature review of two problems in GAN over the years, i.e., mode collapse and vanishing gradient. We start with a comprehensive tutorial of GAN. Then, we review the solution for these two problems from the perspective of architecture and loss. Furthermore, some conclusions about these two main challenges are given. Finally, we look into the GAN's future by summarizing the various GAN applications.