Abstract
Despite excellent progress in recent years, mode collapse remains a majorunsolved problem in generative adversarial networks (GANs).In this paper, wepresent spectral regularization for GANs (SR-GANs), a new and robust method forcombating the mode collapse problem in GANs. Theoretical analysis shows thatthe optimal solution to the discriminator has a strong relationship to thespectral distributions of the weight matrix.Therefore, we monitor the spectraldistribution in the discriminator of spectral normalized GANs (SN-GANs), anddiscover a phenomenon which we refer to as spectral collapse, where a largenumber of singular values of the weight matrices drop dramatically when modecollapse occurs. We show that there are strong evidence linking mode collapseto spectral collapse; and based on this link, we set out to tackle spectralcollapse as a surrogate of mode collapse. We have developed a spectralregularization method where we compensate the spectral distributions of theweight matrices to prevent them from collapsing, which in turn successfullyprevents mode collapse in GANs. We provide theoretical explanations for whySR-GANs are more stable and can provide better performances than SN-GANs. Wealso present extensive experimental results and analysis to show that SR-GANsnot only always outperform SN-GANs but also always succeed in combating modecollapse where SN-GANs fail. The code is available athttps://github.com/max-liu-112/SRGANs-Spectral-Regularization-GANs-.