Wasserstein Generative Adversarial Network

Recent advances in deep generative models give us new perspective on modeling highdimensional, nonlinear data distributions. Especially the GAN training can successfully produce sharp, realistic images. However, GAN sidesteps the use of traditional maximum likelihood learning and instead adopts an two-player game approach. This new training behaves very differently compared to ML learning. There are still many remaining problem of GAN training. In this thesis, we gives a comprehensive review of recently published methods or analysis on GAN training, especially the Wasserstein GAN and FlowGAN model. We also discuss the limitation of the later model and use this as the motivation to propose a novel generator architecture using mixture models. Furthermore, we also modify the discriminator architecture using similar ideas to allow ’personalized’ guidance. We refer the generator mixture model as Mixflow and mixture of discriminators as ’personalized GAN’ (PGAN). In experiment chapter, we demonstrate their performance advantages using toy examples compared to single flow model. In the end, we test their performance on MNIST dataset and the Mixflow model not only achieves the best log likelihood but also produce reasonable images compared to state-of-art DCGAN generation.