2018/02/05 by Xu Chen, Jiang Wang, Chen, Xu +3
Computer Science · #Generative Adversarial Networks and Image Synthesis #Gaussian Processes and Bayesian Inference #Advanced Image Processing Techniques
paper · pdf · doi:10.48550/arxiv.1802.01765
We relate the minimax game of generative adversarial networks (GANs) to\nfinding the saddle points of the Lagrangian function for a convex optimization\nproblem, where the discriminator outputs and the distribution of generator\noutputs play the roles of primal variables and dual variables, respectively.\nThis formulation shows the connection between the standard GAN training process\nand the primal-dual subgradient methods for convex optimization. The inherent\nconnection does not only provide a theoretical convergence proof for training\nGANs in the function space, but also inspires a novel objective function for\ntraining. The modified objective function forces the distribution of generator\noutputs to be updated along the direction according to the primal-dual\nsubgradient methods. A toy example shows that the proposed method is able to\nresolve mode collapse, which in this case cannot be avoided by the standard GAN\nor Wasserstein GAN. Experiments on both Gaussian mixture synthetic data and\nreal-world image datasets demonstrate the performance of the proposed method on\ngenerating diverse samples.\n