Dimensionality Reduction using Residual Encoding Net

How can we turn multiple factors into fewer features? We proposed a residual encoding net to convert high dimensional samples into low dimensional codings. In this paper, we introduced a network architecture to improve performance in reducing features of samples supported by residual networks and generative adversarial networks. First of all, the residual encoding net using discriminator to evaluate the quality of reconstruction data, at the same time, we use discriminator to minimize the discrepancy between samples and reconstruction data. In addition, the residual encoding net consists of many residual units, therefore, it can be very deep and turns out to work better than principle component analysis and auto-encoder in dimensionality reduction territory. We test our model by using MINIST dataset in the end.