Triple Generative Adversarial Nets

Generative Adversarial Nets (GANs) have shown promise in image generation and semi-supervised learning (SSL). However, existing GANs in SSL have two problems: (1) the generator and discriminator may compete in learning; and (2) the generator cannot generate images in a specific class. The problems essentially arise from the two-player formulation, where a single discriminator shares incompatible roles of identifying fake samples and predicting labels and it only estimates the data without considering labels. We address the problems by presenting triple generative adversarial net (Triple-GAN), a flexible game-theoretical framework for classification and class-conditional generation in SSL. Triple-GAN consists of three players---a generator, a discriminator and a classifier, where the generator and classifier characterize the conditional distributions between images and labels, and the discriminator solely focuses on identifying fake image-label pairs. We design compatible utilities to ensure that the distributions characterized by the classifier and generator both concentrate to the data distribution. Our results on various datasets demonstrate that Triple-GAN as a unified model can simultaneously (1) achieve state-of-the-art classification results among deep generative models, and (2) disentangle the classes and styles and transfer smoothly on the data level via interpolation in the latent space class-conditionally.