Title: Self-Correcting Discriminator Optimization for Image and Speech Enhancement GANs
Abstract: Generative adversarial network (GAN) has become one of the most important neural network models for classical unsupervised machine learning. Various discriminator loss functions have been developed to train GAN's discriminators, most of which have a common structure: a sum of real and fake losses that depend only on the actual and generated data, respectively. One challenge associated with an equally weighted sum of two losses is that the training may benefit one loss but harm the other. We present self-correcting optimization for training a GAN discriminator, which helps avoid "harmful" training directions for parts of the discriminator loss function. Experiments validated the effectiveness of our loss functions on conditional and unconditional image generation tasks as well as speech enhancement tasks.