Spectral bounding: Strictly satisfying the 1-Lipschitz property for generative adversarial networks. (September 2020)

• Record Type: Journal Article
• Authors: Zhang, Zhihong; Zeng, Yangbin; Bai, Lu; Hu, Yiqun; Wu, Meihong; Wang, Shuai; Hancock, Edwin R.
• Is Part Of: Pattern recognition. Volume 105(2020:Sep.)
• Publication Date: 2020-09
• Subjects: Generative adversarial networks -- 1-Lipschitz constraint -- Spectral bounding -- Image generation; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.107179
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: The paper starts by highlighting the need for spectral bounding of weights in the discriminator for GAN training. The paper proposes to perform quick spectral bounding by using the 1 norm and infinity norms of the weight matrices to normalize the weights of the models. Extensive experimental results on CIFAR-10 and ImageNet dataset demonstrate that our approach can maintain more successfully the balance between generators and discriminators encountered prior to a Nash equilibrium having been reached. In so doing we can obtain a robust GAN model which accurately captures features of the statistical distribution for data samples used in training. Abstract: Imposing the 1-Lipschitz constraint is a problem of key importance in the training of Generative Adversarial Networks (GANs), which has been proved to productively improve stability of GAN training. Although some interesting alternative methods have been proposed to enforce the 1-Lipschitz property, these existing approaches (e.g., weight clipping, gradient penalty (GP), and spectral normalization (SN)) are only partially successful. In this paper, we propose a novel method, which we refer to as spectral bounding (SB) to strictly enforce the 1-Lipschitz constraint. Our method adopts very cost-effective terms of both 1-norm and ∞-norm, and yet allows us to efficiently approximate the upper bound of spectral norms. In this way, our method provide important insights to the relationship between an alternative of strictly satisfying the Lipschitz property and explainable training stability improvements of GAN. Our proposed method thus significantly enhances the stability of GAN training and the quality of generated images. Extensive experiments are conducted, showing that the proposed method outperforms GP and SN on both CIFAR-10 and ILSVRC2015 (ImagetNet) dataset in terms of the standard inception score.