Incorporating Least-Effort Loss to Stabilize Training of Wasserstein GAN

In order to further improve the convergence properties of generative adversarial networks, in this paper, we analyze how the stability can be affected by the so-called best-effort manner of the discriminator in the minimax game. We point out that this manner can cause the multistate problem and the optimization entangling problem. To alleviate these, we proposed an alternative least-effort loss to regularize the training behaviors of the discriminator. With this loss, the discriminator only updates when it is unable to distinguish distributions. To evaluate the effectiveness of the least-effort loss, we introduce it into Wasserstein GAN. Experiments on Dirac delta distribution and image datasets demonstrate that the least-effort loss can effectively improve the convergence properties and generation quality of WGAN. Furthermore, the behaviors of the discriminator and generator during the training show that, with the least-effort loss, the state space of the discriminator shrinks, and the optimization of the discriminator and the generator disentangles in some way.