Abstract:Generative adversarial networks (GANs) have shown extraordinary performance in generating high quality samples in domains including image, video, and text. GANs therefore have great potential in learning complex probability distributions in high dimensional spaces. However, current methods often miss capturing some of the modes in the examples, known as the mode collapse problem. The reason for this issue can be traced to that the initial generated manifold fails to cover the whole data manifold, while the training process is hard to recover from this failure. In this paper, we propose GANs with supervision signal (SSGAN), which introduces a supervision signal to alleviate this issue. The supervision signal tells the generator an approximate output corresponding to the input noise, which ensures the generated manifold to be close to the data manifold. Therefore, the generator could be able to better capture the whole data distribution. We have conducted experiments on MNIST, CIFAR 10 and CelebA datasets. The results show that our method outperforms several SoTA approaches measured by the inception score, mode score, and the newly proposed matching score.

On Mode Collapse in Generative Adversarial Networks