Triple Generative Adversarial Networks
This addresses the challenge of leveraging limited labeled data for both classification and generation tasks in machine learning, offering a flexible framework that integrates semi-supervised classifiers and GAN architectures, though it is incremental as it builds upon existing GAN and semi-supervised learning methods.
The paper tackles the problem of classification and conditional image generation with limited supervision by proposing Triple Generative Adversarial Networks (Triple-GAN), a three-player minimax game involving a generator, classifier, and discriminator, achieving excellent classification results and meaningful sample generation in semi-supervised and low-data settings, outperforming extensive semi-supervised learning methods on over 10 benchmarks.
We propose a unified game-theoretical framework to perform classification and conditional image generation given limited supervision. It is formulated as a three-player minimax game consisting of a generator, a classifier and a discriminator, and therefore is referred to as Triple Generative Adversarial Network (Triple-GAN). The generator and the classifier characterize the conditional distributions between images and labels to perform conditional generation and classification, respectively. The discriminator solely focuses on identifying fake image-label pairs. Under a nonparametric assumption, we prove the unique equilibrium of the game is that the distributions characterized by the generator and the classifier converge to the data distribution. As a byproduct of the three-player mechanism, Triple-GAN is flexible to incorporate different semi-supervised classifiers and GAN architectures. We evaluate Triple-GAN in two challenging settings, namely, semi-supervised learning and the extreme low data regime. In both settings, Triple-GAN can achieve excellent classification results and generate meaningful samples in a specific class simultaneously. In particular, using a commonly adopted 13-layer CNN classifier, Triple-GAN outperforms extensive semi-supervised learning methods substantially on more than 10 benchmarks no matter data augmentation is applied or not.