Supervised Contrastive Generative Adversarial Networks

Abstract

Generative Adversarial Networks (GANs) is becoming more and more popular, artists use them to find their own inspirations, computer scientists use it for data synthesis, workers use it for machine fault diagnosis and so on. However, GANs are flawed despite its popularity: they are unstable. GANs are based on game theory. In a typical GAN model, the generator and the discriminator are both improved by competing with each other. Therefore, in this highly competitive training process, GANs can easily run into trouble while they move towards the optimal solution. In most cases, the case of such instability arises from the loss function, or in other words, the gradient of the loss function. This research proposed a new set of GAN that replaces its objective function with supcon, or the supervised contrastive loss to solve gradient-related problems. We have also proved that under our model, the GANs are less likely to suffer from these two factors of instability. Finally, we have compared our model and the traditional generative adversarial nets.