Semi-Supervised Generative Adversarial Networks Based On Scalable Support Vector Machines And Manifold Regularization

Generative adversarial networks (GANs) are potential models in semi-supervised learning because of the excellent performance of GANs. However, most GAN-based semi-supervised models are sensitive to local perturbation, which means those models are not stable enough. Besides, Softmax classifier is the first choice of those models. In this paper, a novel method is proposed by introducing a discriminator using scalable SVM classifier with manifold regularization. Scalable SVM classifiers typically perform better in small sample data sets compared with other classifiers, which is consistent with the feature that semi-supervised learning consists of a few labeled data and a large number of unlabeled data. Manifold regularization forces discriminator to keep invariable to local perturbations. By incorporating into feature-matching GAN architecture, the proposed GANs-based semi-supervised learning algorithm has advantages over other methods on the Cifar-10, SVHN and Cifar-100 datasets. The results show that the proposed model SSVM-GAN has good robustness and strong generalization ability.