Generalization analysis of conditional generative adversarial networks

Abstract

Analyzing the generalization performance of conditional Generative Adversarial Networks (cGANs) is known to be a very challenging problem. This difficulty arises from the issue inherent in Conditional Density Estimation (CDE), where information about the conditional distribution at a specific value of the conditioning variable is difficult to use for predicting the conditional distribution at other values of the conditioning variable. In this thesis, we introduce correlations between conditional distributions generated by two main methods. Firstly, we impose Lipschitz continuity with respect to the conditioning variable on the generator network of the cGAN. Secondly, we transform the problem into approximating a vicinal estimate (VE) that combines conditional distributions at adjacent values of the conditioning variable instead of the original distribution. Using these imposed correlations, we analyze the generalization performance of the corresponding model in cGANs for the first time. Furthermore, from the analysis of this generalization performance, we propose new regularization methods and loss functions, and experimentally confirm that they result in improved generalization performance.