Triple Generative Adversarial Networks

20 Dec 2019  ยท  Chongxuan Li, Kun Xu, Jiashuo Liu, Jun Zhu, Bo Zhang ยท

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.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Semi-Supervised Image Classification CIFAR-10, 1000 Labels Triple-GAN-V2 (ResNet-26) Accuracy 91.59 # 3
Semi-Supervised Image Classification CIFAR-10, 1000 Labels Triple-GAN-V2 (CNN-13, no aug) Accuracy 81.81 # 9
Semi-Supervised Image Classification CIFAR-10, 1000 Labels Triple-GAN-V2 (CNN-13) Accuracy 85.00 # 6
Semi-Supervised Image Classification CIFAR-10, 4000 Labels Triple-GAN-V2 (CNN-13) Percentage error 10.01 # 36
Semi-Supervised Image Classification CIFAR-10, 4000 Labels Triple-GAN-V2 (CNN-13, no aug) Percentage error 12.41 # 41
Semi-Supervised Image Classification CIFAR-10, 4000 Labels Triple-GAN-V2 (ResNet-26) Percentage error 6.54 # 30
Semi-Supervised Image Classification SVHN, 1000 labels Triple-GAN-V2 (CNN-13, no aug) Accuracy 96.04 # 15
Semi-Supervised Image Classification SVHN, 1000 labels Triple-GAN-V2 (CNN-13) Accuracy 96.55 # 8
Semi-Supervised Image Classification SVHN, 250 Labels Triple-GAN-V2 (CNN-13) Accuracy 96.52 # 6
Semi-Supervised Image Classification SVHN, 250 Labels Triple-GAN-V2 (CNN-13, no aug) Accuracy 95.81 # 9
Semi-Supervised Image Classification SVHN, 500 Labels Triple-GAN-V2 (CNN-13, no aug) Accuracy 96.16 # 3
Semi-Supervised Image Classification SVHN, 500 Labels Triple-GAN-V2 (CNN-13) Accuracy 96.39 # 1

Methods