Training Generative Adversarial Networks with Adaptive Composite Gradient
作者机构:School of Mathematical SciencesEast China Normal UniversityShanghai 200241China
出 版 物:《Data Intelligence》 (数据智能(英文))
年 卷 期:2024年第6卷第1期
页 面:120-157页
核心收录:
学科分类:08[工学] 0839[工学-网络空间安全] 081201[工学-计算机系统结构] 0812[工学-计算机科学与技术(可授工学、理学学位)] 10[医学]
基 金:This work is supported by the National Key Research and Development Program of China(No.2018AAA0101001) Science and Technology Commission of Shanghai Municipality(No.20511100200) supported in part by the Science and Technology Commission of Shanghai Municipality(No.18dz2271000)
主 题:Generative adversarial networks adaptive composite gradient semi-gradient free game theory bilinear game
摘 要:The wide applications of Generative adversarial networks benefit from the successful training methods,guaranteeing that an object function converges to the local ***,designing an efficient and competitive training method is still a challenging task due to the cyclic behaviors of some gradient-based ways and the expensive computational cost of acquiring the Hessian *** address this problem,we proposed the Adaptive Composite Gradients(ACG)method,linearly convergent in bilinear games under suitable *** analysis and toy-function experiments both suggest that our approach alleviates the cyclic behaviors and converges faster than recently proposed SOTA *** convergence speed of the ACG is improved by 33%than other *** ACG method is a novel Semi-Gradient-Free algorithm that can reduce the computational cost of gradient and Hessian by utilizing the predictive information in future *** mixture of Gaussians experiments and real-world digital image generative experiments show that our ACG method outperforms several existing technologies,illustrating the superiority and efficacy of our method.