Towards Better Global Landscape of GAN: How Two Lines of Code Change Makes a Difference
Prof. Ruoyu Sun University of Illinois at Urbana-Champaign (UIUC) Our understanding of GAN (generative adversarial net) training is still very limited since it is a non-convex-non-concave min-max optimization. As a result, most recent studies focused on local analysis. In this talk, we discuss how to perform a global analysis of GANs. We prove that the original JS-GAN has exponentially many bad strict local minima which are perceived as mode-collapse. We show that a 2-line modification to JS-GAN called relativistic standard GAN (RS-GAN) eliminates all bad basins. We extend the two results to a large class of losses as well: for separable GANs (including JS-GAN, hinge-GAN, LS-GAN) exponentially many bad basins exist, while for R-GANs (relativistic GANs) no bad basins exist. The effectiveness of R-GANs were verified by a few empirical works before (e.g. ESR-GAN in super resolution). Based on theory, we predict that R-GANs has a bigger advantage for narrower neural nets, and our experiments verify that R-GANs (e.g. RS-GAN) can beat their separable counter-parts (e.g. JS-GAN) by 5-10 FID scores in narrower nets. We demonstrate that for high-resolution images like LSUN, while JS-GAN often only generates noise, RS-GAN can generate quite good images.
Prof. Ruoyu Sun University of Illinois at Urbana-Champaign (UIUC) Our understanding of GAN (generative adversarial net) training is still very limited since it is a non-convex-non-concave min-max optimization. As a result, most recent studies focused on local analysis. In this talk, we discuss how to perform a global analysis of GANs. We prove that the original JS-GAN has exponentially many bad strict local minima which are perceived as mode-collapse. We show that a 2-line modification to JS-GAN called relativistic standard GAN (RS-GAN) eliminates all bad basins. We extend the two results to a large class of losses as well: for separable GANs (including JS-GAN, hinge-GAN, LS-GAN) exponentially many bad basins exist, while for R-GANs (relativistic GANs) no bad basins exist. The effectiveness of R-GANs were verified by a few empirical works before (e.g. ESR-GAN in super resolution). Based on theory, we predict that R-GANs has a bigger advantage for narrower neural nets, and our experiments verify that R-GANs (e.g. RS-GAN) can beat their separable counter-parts (e.g. JS-GAN) by 5-10 FID scores in narrower nets. We demonstrate that for high-resolution images like LSUN, while JS-GAN often only generates noise, RS-GAN can generate quite good images.