- Keywords: GANs, Mode collapse, Gradient exploding, Stability
- TL;DR: We propose a novel GAN training method by considering certain fake samples as real to alleviate mode collapse and stabilize training process.
- Abstract: In order to alleviate the notorious mode collapse phenomenon in generative adversarial networks (GANs), we propose a novel training method of GANs in which certain fake samples can be reconsidered as real ones during the training process. This strategy can reduce the gradient value that generator receives in the region where gradient exploding happens. We show that the theoretical equilibrium between the generators and discriminations actually can be seldom realized in practice. And this results in an unbalanced generated distribution that deviates from the target one, when fake datepoints overfit to real ones, which explains the non-stability of GANs. We also prove that, by penalizing the difference between discriminator outputs and considering certain fake datapoints as real for adjacent real and fake sample pairs, gradient exploding can be alleviated. Accordingly, a modified GAN training method is proposed with a more stable training process and a better generalization. Experiments on different datasets verify our theoretical analysis.