MGGAN: Improving sample generations of Generative Adversarial Networks

Generative adversarial networks (GANs) are powerful generative models that are widely used to produce synthetic data. This paper proposes a Multi-Group Generative Adversarial Network (MGGAN), a framework that consists of multiple generative groups for addressing the mode collapse problem and creating high-quality samples with less time cost. The idea is intuitive yet effective. The distinguishing characteristic of MGGAN is that a generative group includes a fixed generator but a dynamic discriminator. All the generators need to combine with a random discriminator from other generative groups after a certain number of training iterations, which is called regrouping. The multiple generative groups are trained simultaneously and independently without sharing the parameters. The learning rate and the regrouping interval are adjusted dynamically in the training process. We conduct extensive experiments on the synthetic and real-world datasets. The experimental results show the superior performance of our MGGAN in generating high quality and diverse samples with less training time.