Improving Generative Adversarial Networks using Game Theory and Statistics

Generative Adversarial Networks (aka GANs) are a recently proposed approach for learning samplers of high-dimensional distributions with intricate structure, such as distributions over natural images, given samples from these distributions. They are trained by setting up a two-player zero-sum game between two neural networks, which learn statistics of a target distribution by adapting their strategies in the game using gradient descent. Despite their intriguing performance in practice, GANs pose great challenges to both Optimization and Statistics. Their training suffers from oscillations, and they are difficult to scale to high-dimensional settings. We study how game-theoretic and statistical techniques can be brought to bare on these important challenges. We use Game Theory towards improving GAN training, and Statistics towards scaling up the dimensionality of the generated distributions.