CP-GAN: Towards a Better Global Landscape of GANs

GANs have been very popular in data generation and unsupervised learning, but our understanding of GAN training is still very limited. One major reason is that GANs are often formulated as non-convex-concave min-max optimization. As a result, most recent studies focused on the analysis in the local region around the equilibrium. In this work, we perform a global analysis of GANs from two perspectives: the global landscape of the outer-optimization problem and the global behavior of the gradient descent dynamics. We find that the original GAN has exponentially many bad strict local minima which are perceived as mode-collapse, and the training dynamics (with linear discriminators) cannot escape mode collapse. To address these issues, we propose a simple modification to the original GAN, by coupling the generated samples and the true samples. We prove that the new formulation has no bad basins, and its training dynamics (with linear discriminators) has a Lyapunov function that leads to global convergence. Our experiments on standard datasets show that this simple loss outperforms the original GAN and WGAN-GP.