Towards a Better Global Loss Landscape of GANs

• DOI: 10.48550/arxiv.2011.04926
• Publication Date: 2020-01-01
• Authors: Ruoyu Sun; Tiantian Fang; Alexander G. Schwing
• Keywords: Maxima and minima; Code (set theory)

Understanding of GAN training is still very limited. One major challenge its non-convex-non-concave min-max objective, which may lead to sub-optimal local minima. In this work, we perform a global landscape analysis the empirical loss GANs. We prove that class separable-GAN, including original JS-GAN, has exponentially many bad basins are perceived as mode-collapse. also study relativistic pairing (RpGAN) couples generated samples and true samples. RpGAN no basins. Experiments on synthetic data show predicted basin can indeed appear in training. experiments support our theory better than separable-GAN. For instance, empirically performs separable-GAN with relatively narrow neural nets. The code available at https://github.com/AilsaF/RS-GAN.