MGGAN: Solving Mode Collapse Using Manifold-Guided Training
Duhyeon Bang, Hyunjung Shim
Abstract
Mode collapse is a critical problem in training generative adversarial networks. To alleviate mode collapse, several recent studies have introduced new objective functions, network architectures, or alternative training schemes. However, their achievement is often the result of sacrificing the image quality. In this paper, we propose a new algorithm, namely, the manifold-guided generative adversarial network (MGGAN), which leverages a guidance network on existing GAN architecture to induce the generator to learn the overall modes of a data distribution. The guidance network transforms an image into a learned manifold space, which is effective in representing the coverage of the overall modes. The characteristics of this guidance network helps penalize mode imbalance. Results of the experimental comparisons using various baseline GANs showed that MGGAN can be easily extended to existing GANs and resolve mode collapse without losing the image quality. Moreover, we extend the idea of manifold-guided GAN training to increase the original diversity of a data distribution. From the experiment, we confirmed that a GAN model guided by a joint manifold can sample data distribution with greater diversity. Results of the experimental analysis confirmed that MGGAN is an effective and efficient tool for improving the diversity of GANs.