Difference-Seeking Generative Adversarial Network

Abstract

We propose a novel algorithm, Difference-Seeking Generative Adversarial Network (DSGAN), developed from traditional GAN. DSGAN considers the scenario that the training samples of target distribution, p_t, are difficult to collect. Suppose there are two distributions p_d and p_d such that the density of the target distribution can be the differences between the densities of p_d and p_d. We show how to learn the target distribution p_t only via samples from p_d and p_d (relatively easy to obtain). DSGAN has the flexibility to produce samples from various target distributions (e.g. the out-of-distribution). Two key applications, semi-supervised learning and adversarial training, are taken as examples to validate the effectiveness of DSGAN. We also provide theoretical analyses about the convergence of DSGAN.

Tasks

Reproductions