Alleviating limit cycling in training GANs with an optimization technique  

作　　者：Keke Li Liping Tang Xinmin Yang 

机构地区：[1]National Center for Applied Mathematics in Chongqing,Chongqing Normal University,Chongqing,401331,China [2]School of Mathematical Sciences,University of Electronic Science and Technology of China,Chengdu,611731,China

出　　处：《Science China Mathematics》2024年第6期1287-1316,共30页中国科学（数学）（英文版）

基　　金：supported by the Major Program of National Natural Science Foundation of China(Grant Nos.11991020 and 11991024);the Team Project of Innovation Leading Talent in Chongqing(Grant No.CQYC20210309536);NSFC-RGC(Hong Kong)Joint Research Program(Grant No.12261160365);the Scientific and Technological Research Program of Chongqing Municipal Education Commission(Grant No.KJQN202300528)。

摘　　要：In this paper,we undertake further investigation to alleviate the issue of limit cycling behavior in training generative adversarial networks(GANs)through the proposed predictive centripetal acceleration algorithm(PCAA).Specifically,we first derive the upper and lower complexity bounds of PCAA for a general bilinear game,with the last-iterate convergence rate notably improving upon previous results.Then,we combine PCAA with the adaptive moment estimation algorithm(Adam)to propose PCAA-Adam,for practical training of GANs to enhance their generalization capability.Finally,we validate the effectiveness of the proposed algorithm through experiments conducted on bilinear games,multivariate Gaussian distributions,and the CelebA dataset,respectively.

关 键 词：GANs general bilinear game predictive centripetal acceleration algorithm lower and upper complexity bounds PCAA-Adam 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]