Semi-Supervised Learning with Flexible Discriminator Objective in Generative Adversarial Networks Framework

Semi-supervised learning aims to learn a discriminative classifier with partially labeled data. This could leverage the large amount of unlabeled data which is relatively unlimited to enhance the model's generalization performance. The capability of prevalent deep models could be further improved with good semi-supervised mechanisms. Generative adversarial networks (GAN) has been adopted in this field and obtained strong empirical results, which mainly results from the adversarially generated samples that could help to build the decision boundaries in more reasonable places. In this paper, a flexible discriminator objective in GAN framework is proposed, which involves a supervised loss and a conditional entropy loss both added to the discriminator objective conditionally. A competitive semi-supervised classification accuracy could be obtained with our method.