RDCGAN: Unsupervised Representation Learning With Regularized Deep Convolutional Generative Adversarial Networks

Abstract

In Recent years, Representation learning as one of the information extraction and data mapping methods in machine learning systems has received huge attention. Artificial deep neural networks are considered as one of the basic structures capable of representation learning. However, a large number of standard representation learning methods are supervised and requires a lot of labeled data. In this paper, we introduce an unsupervised representation learning by designing and implementing deep neural networks (DNNs) in combination with Generative Adversarial Networks (GANs). The main idea behind the proposed method, which causes the superiority of this method over others is representation learning via the generative models and encoder networks altogether. In this research, encoders are utilized in addition to the generative models to help the more features to be extracted. It is shown that the proposed method not only help feature extraction but accelerate and improve the performance of the learning in GANs which lead to better feature extraction. The results confirm the superiority of the proposed approach regarding classification accuracy by 2% to 6% improvement over other unsupervised feature learning methods.