Stabilized training of generative adversarial networks by a genetic algorithm

Abstract

Generative adversarial networks (GAN) facilitate the learning of probability distributions of complex data in the real world, and allow neural networks to generate the distribution. GANs (GAN and its variants) exhibit excellent performance in applications like image generation and video generation. However, GANs sometimes experience problems during training with regard to the distribution of real data. We applied a genetic algorithm to improve and optimize the GAN's training performance. As a result, the convergence speed and stability during the training process improved compared to the conventional GAN.