Evolutionary Generative Adversarial Networks

  • 2018-03-01 23:15:38
  • Chaoyue Wang, Chang Xu, Xin Yao, Dacheng Tao
  • 57


Generative adversarial networks (GAN) have been effective for learninggenerative models for real-world data. However, existing GANs (GAN and itsvariants) tend to suffer from training problems such as instability and modecollapse. In this paper, we propose a novel GAN framework called evolutionarygenerative adversarial networks (E-GAN) for stable GAN training and improvedgenerative performance. Unlike existing GANs, which employ a pre-definedadversarial objective function alternately training a generator and adiscriminator, we utilize different adversarial training objectives as mutationoperations and evolve a population of generators to adapt to the environment(i.e., the discriminator). We also utilize an evaluation mechanism to measurethe quality and diversity of generated samples, such that only well-performinggenerator(s) are preserved and used for further training. In this way, E-GANovercomes the limitations of an individual adversarial training objective andalways preserves the best offspring, contributing to progress in and thesuccess of GANs. Experiments on several datasets demonstrate that E-GANachieves convincing generative performance and reduces the training problemsinherent in existing GANs.


Introduction (beta)



Conclusion (beta)