KDE-GAN

Generative Adversarial Networks (GAN) is a powerful algorithm to reconstruct artificial data that is similar to a set of given data. Evolutionary GAN (E-GAN) is a state-of-the-art variant of GAN. E-GAN combines evolutionary elements such as mutation, fitness evaluation and selection to address vanishing gradient and mode collapse problems of classic GAN. This paper further improves E-GAN by proposing Knowledge Distillation and Knowledge Transfer. The new method, known as Knowledge Distillation E-GAN (KDE-GAN), incorporates the student-teacher architecture and fine-tuning of transfer learning to help the evolutionary GAN training process. Our experiments on benchmark data sets show that KDE-GAN can improve performance and efficiency. Knowledge distillation and knowledge transfer can indeed accelerate the learning process yet reach better FIDs (Frechet Inception Distance).