Unsupervised discriminative feature representation via adversarial auto-encoder

Abstract

Feature representation is generally applied to reducing the dimensions of high-dimensional data to accelerate the process of data handling and enhance the performance of pattern recognition. However, the dimensionality of data nowadays appears to be a rapidly increasing trend. Existing unsupervised feature representation methods are susceptible to the rapidly increasing dimensionality of data, which may result in learning a meaningless feature that in turn affect their performance in other applications. In this paper, an unsupervised adversarial auto-encoder network is studied. This network is a probability model that combines generative adversarial networks and variational auto-encoder to perform variational inference and aims to generate reconstructed data similar to original data as much as possible. Due to its adversarial training, this model is relatively robust in feature learning compared with other methods. First, the architecture and training strategy of adversarial auto-encoder are presented. We attempt to learn a discriminative feature representation for high-dimensional image data via adversarial auto-encoder and take its advantage into image clustering, which has become a difficult computer vision task recently. Then amounts of comparative experiments are carried out. The comparison contains eight feature representation methods and two recently proposed deep clustering methods performed on eight different publicly available image data sets. Finally, to evaluate their performance, we utilize a K-means clustering on the low-dimensional feature learned from each feature representation algorithm, and select three evaluation metrics including clustering accuracy, adjusted rand index and normalized mutual information, to provide a comparison. Comprehensive experiments prove the usefulness of the learned discriminative feature via adversarial auto-encoder in the tested data sets.