InfoMax-GAN: Improved Adversarial Image Generation via Information Maximization and Contrastive Learning

While Generative Adversarial Networks (GANs) are fundamental to many\ngenerative modelling applications, they suffer from numerous issues. In this\nwork, we propose a principled framework to simultaneously mitigate two\nfundamental issues in GANs: catastrophic forgetting of the discriminator and\nmode collapse of the generator. We achieve this by employing for GANs a\ncontrastive learning and mutual information maximization approach, and perform\nextensive analyses to understand sources of improvements. Our approach\nsignificantly stabilizes GAN training and improves GAN performance for image\nsynthesis across five datasets under the same training and evaluation\nconditions against state-of-the-art works. In particular, compared to the\nstate-of-the-art SSGAN, our approach does not suffer from poorer performance on\nimage domains such as faces, and instead improves performance significantly.\nOur approach is simple to implement and practical: it involves only one\nauxiliary objective, has a low computational cost, and performs robustly across\na wide range of training settings and datasets without any hyperparameter\ntuning. For reproducibility, our code is available in Mimicry:\nhttps://github.com/kwotsin/mimicry.\n