Improved Generative Adversarial Networks for Intersection of Two Domains

Abstract

The goal of generative models is to capture domain distribution based on training samples. Generative Adversarial Networks (or GANs) are a successful framework for training a generative model. In this paper, we consider a process for training generative models using GAN when the target domain is an intersection of two target domains. When two target domains only share a small intersection domain, we have identified an issue referred to as canceling gradients, caused by unintended optimization of learning loss. We propose a simple method based on gradient scaling and perform experiments to verify our remedy.