Unsupervised and Semi-supervised Learning with Categorical Generative Adversarial Networks

In this paper we present a method for learning a discriminative classifier unlabeled or partially labeled data. Our approach is based on an objective that trades-off mutual information between observed examples and their categorical class distribution, against robustness of the classifier an adversarial generative model. The resulting algorithm can either be as a natural generalization of the generative adversarial networks (GAN) framework or as an extension of the regularized information maximization (RIM) framework to robust classification against an optimal adversary. We evaluate our method - which we dub categorical generative networks (or CatGAN) - on synthetic data as well as on challenging classification tasks, demonstrating the robustness of the learned. We further qualitatively assess the fidelity of samples generated the adversarial generator that is learned alongside the discriminative, and identify links between the CatGAN objective and discriminative algorithms (such as RIM).