Generative Adversarial Networks Based on a General Parameterized Family of Generator Loss Functions

Abstract

This thesis introduces a unifying parameterized generator loss function for generative adversarial networks (GANs). We establish an equilibrium theorem for our resulting GAN system under a canonical discriminator in terms of the so-called Jensen-$f$-divergence, a natural generalization of the Jensen-Shannon divergence to the $f$-divergence. We also show that our result recovers as special cases several GANs from the literature, including the original GAN, least square GAN (LSGAN), $\alpha$-GAN and others. Finally, we systematically conduct experiments on three image datasets for different manifestations of our GAN system to illustrate their performance and stability.