Unsupervised deep feature representation using adversarial auto-encoder

Abstract

Feature representation is typically applied to reducing the dimensions of high-dimensional data, so as to accelerate the data processing. However, it is difficult to label the data with high dimension for supervised learning, and make use of the information in high-dimensional form directly, which resulting in extensive discussion about unsupervised feature representation methods. 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. First, the architecture and training strategy of adversarial auto-encoder are presented, we attempt to learn a feature representation for high-dimensional data using adversarial auto-encoder. Then a number of comparative experiments are carried out. The comparisons contain six unsupervised feature representation methods performed on eight different publicly available image data sets. Finally, for evaluating their performance, we employ K-means clustering on the low-dimensional feature learned from each algorithm, and select three evaluation metrics including clustering accuracy, adjusted rand index and normalized mutual information, to provide comparisons. Extensive experiments indicate the effectiveness of the learned feature representation via adversarial auto-encoder on the tested data sets.