On the convergence and mode collapse of GAN

Generative adversarial network (GAN) is a powerful generative model. However, it suffers from several problems, such as convergence instability and mode collapse. To overcome these drawbacks, this paper presents a novel architecture of GAN, which consists of one generator and two different discriminators. With the fact that GAN is the analogy of a minimax game, the proposed architecture is as follows. The generator (G) aims to produce realistic-looking samples to fool both of two discriminators. The first discriminator (D₁) rewards high scores for samples from the data distribution, while the second one (D₂) favors samples from the generator conversely. Specifically, the ResBlock and minibatch discrimination (MD) architectures are adopted in D₁ to improve the diversity of the samples. The leaky rectified linear unit (Leaky ReLU) and batch normalization (BN) are replaced by the scaled exponential linear unit (SELU) in D₂ to alleviate the convergence problem. A new loss function that minimizes the KL divergence is designed to better optimize the model. Extensive experiments on CIFAR-10/100 datasets demonstrate that the proposed method can effectively solve the problems of convergence and mode collapse.