Semi-supervised Learning Using Generative Adversarial Networks

Abstract

Machine learning is a powerful tool in many applications, but the most difficult process in machine learning is the collection of data and the labeling of data. Unsupervised and semi-supervised learning has thus become an important issue. In this paper, we introduce a semi-supervised learning approach which using generative adversarial networks to generate training samples. Those imitated samples were involved in training set to train the classifier, this can improve the stability and robustness of the classifier models. To demonstrate the performance of the proposed framework, four benchmarks including Iris, MNIST, CIFAR-10, and SVHN datasets were evaluated. The experimental results show that even in a small amount of training data, the proposed framework can predict more accurately than the existing methods.