Self-supervised learning with consistency loss for improving GANs

Abstract

After much research and advancements, GANs have achieved great success but still face many challenges. In this paper, we adopt self-supervised learning based on rotation angles to overcome the catastrophic forgetting of the discriminator. Self-supervision encourages the discriminator to learn meaningful feature representations that are not forgotten during training. Meanwhile, this paper adopts consistent adversarial training to alleviate the mode collapse of the generator. The consistency constraint condition encourages the discriminator to explore more features, which helps the generator achieve more significant improvement space. This deep generative model improves unsupervised image generation tasks by simultaneously alleviating two critical issues in GANs. Experimental results demonstrate that our model achieves competitive scores.