Active Semi-supervised Learning Based on Self-expressive Correlation with Generative Adversarial Networks

Abstract

Typically in practical applications, the learning performance of a model is inclined to be jeopardized by the inadequacy of labeled instances and the unbalance within various classes. This paper aims to address the issues and develop a novel framework for effective and efficient model learning, which fully explores both labeled and unlabeled instances for robust training and meanwhile leverages reliable synthetic instances for further augmentation. We firstly present a self-expressive correlation estimation method to reveal the underlying inter-instance correlation. After that, a novel active semi-supervised learning with GANs (ASSL-GANs) is presented, which simultaneously maintains three component modules, i.e., a generator, a discriminator, and a classifier. The learners work with each other in either adversarial or cooperative manner to obtain a comprehensive perception of the entire data distribution. The whole architecture is trained end-to-end by jointly optimizing loss functions w.r.t. the corresponding component networks in an alternating update fashion. Experimental results validate the superiority of the proposed algorithm over state-of-the-art models.