Introducing Adversarial Dropout in Generative Multi-Adversarial Networks

Gonçalo Mordido, Haojin Yang, Christoph Meinel

Abstract: We propose to extend the original generative adversarial networks (GANs) frame- work to multiple discriminators and omit, or dropout, the feedback of each dis- criminator with same probability at the end of each batch. Our approach forces the generator to not rely on a given discriminator to learn how to produce realistic looking samples, but, instead, on a dynamic ensemble of adversaries. This pro- motes variety of the generated samples, leading to a richer generator less prone to mode collapsing. We show preliminary results on MNIST and Fashion-MNIST that sustain our claims.

TL;DR: We propose to reduce mode collapse in GANs by training the generator against a dynamic ensemble of adversaries.

Keywords: GANs, multiple adversarial training, adversarial dropout