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 the discriminator (i.e. the classifier) may not be optimal at the same time; and (2) the generator cannot control the semantics of the generated samples. 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 the labels. To address the problems, we present triple generative adversarial net (Triple-GAN), which consists of three players---a generator, a discriminator and a classifier. The generator and the 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 the generator both converge to the data distribution. Our results on various datasets demonstrate that Triple-GAN as a unified model can simultaneously (1) achieve the state-of-the-art classification results among deep generative models, and (2) disentangle the classes and styles of the input and transfer smoothly in the data space via interpolation in the latent space class-conditionally.

In this paper, the authors propose a new formulation of adversarial networks for image generation, that incorporates three networks instead of the usual generator G and discriminator D. In addition, they include a classifier C, which cooperates with G to learn a compatible joint distribution (X,Y) over images and labels. The authors show how this formulation overcomes pitfalls of previous class-conditional GANs; namely that class-conditional generator and discriminator networks have competing objectives that may prevent them from learning the true distribution and preventing G from accurately generating class-conditional samples. The authors identify the following deficiency in class-conditional GAN setups: "The competition between G and D essentially arises from their two-player formulation, where a single discriminator network has to play two incompatible roles--identifying fake samples and predicting labels". The argument goes that if G were perfect, then a class-conditional D has an equal incentive to output 0 since the sample comes from G, and to output 1 since the image matches the label. This might force D to systematically underperform as a classifier, and therefore prevent G from learning to produce accurate class-conditional samples.