Fast adversarial training for semi-supervised learning
In semi-supervised learning, Bad GAN approach is one of the most attractive method due to the intuitional simplicity and powerful performances. Bad GAN learns a classifier with bad samples distributed on complement of the support of the input data. But Bad GAN needs additional architectures, a generator and a density estimation model, which involves huge computation and memory consumption cost. VAT is another good semi-supervised learning algorithm, which utilizes unlabeled data to improve the invariance of the classifier with respect to perturbation of inputs. In this study, we propose a new method by combining the ideas of Bad GAN and VAT. The proposed method generates bad samples of high-quality by use of the adversarial training used in VAT. We give theoretical explanations why the adversarial training is good at both generating bad samples and semi-supervised learning. An advantage of the proposed method is to achieve the competitive performances with much fewer computations. We demonstrate advantages our method by various experiments with well known benchmark image datasets.
PDF Abstract
CIFAR-10
CIFAR-100
MNIST
