Unsupervised and Semi-supervised Learning with Categorical Generative\n Adversarial Networks

Abstract

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