Triple Generative Adversarial Nets

Generative adversarial nets (GANs) are good at generating realistic images and have been extended for semi-supervised classification. However, under a two-player formulation, existing work shares competing roles of identifying fake samples and predicting labels via a single discriminator network, which can lead to undesirable incompatibility. We present triple generative adversarial net (Triple-GAN), a flexible game-theoretical framework for classification and class-conditional generation in semi-supervised learning. 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. With designed utilities, the distributions characterized by the classifier and generator both concentrate to the data distribution under nonparametric assumptions. We further propose unbiased regularization terms to make the classifier and generator strongly coupled and some biased techniques to boost the performance of Triple-GAN in practice. Our results on several datasets demonstrate the promise in semi-supervised learning, where Triple-GAN achieves comparable or superior performance than state-of-the-art classification results among DGMs; it is also able to disentangle the classes and styles and transfer smoothly on the data level via interpolation on the latent space class-conditionally.