Mode Collapse in Generative Adversarial Networks: An Overview
Youssef Kossale, Mohammed Airaj, Aziz Darouichi
Abstract
With the rise of a new framework known as Generative Adversarial Networks (GANs), generative models have gained considerable amount of attention in the area of unsupervised learning. GANs have been thoroughly studied since their emergence in 2014, leading to an enormous amount of new models and applications built on this said framework. Although despite their success, GANs suffer from some notorious problems during training, hindering further advances in the field. This paper seeks to highlight one of the most encountered problems in GAN training, namely the “Helvetica scenario” as stated by its authors or “mode collapse” as widely known. We will try to provide an overview of this said challenge, what is it, why it occurs, and some suggested workarounds to reduce its impact on training.