DG-GAN: the GAN with the duality gap

Cheng Peng, Hao Wang, Xiao Wang, Zhouwang Yang

Abstract: Generative Adversarial Networks (GANs) are powerful, but difficult to understand and train because GANs is a min-max problem. This paper understand GANs with duality gap that comes from game theorem and show that duality gap can be a kind of metric to evolution the difference between the true data distribution and the distribution generated by generator with given condition. And train the networks using duality gap can get some better results. Furthermore, the paper calculates the generalization bound of duality gap to estimate the help design the neural networks and select the sample size.

Keywords: GAN, duality gap, metric, saddle point, game