Evolutionary Generative Adversarial Networks

被引:219
作者
Wang, Chaoyue [1 ,2 ]
Xu, Chang [1 ,2 ]
Yao, Xin [3 ,4 ]
Tao, Dacheng [1 ,2 ]
机构
[1] Univ Sydney, UBTECH Sydney Artificial Intelligence Ctr, Fac Engn & Informat Technol, Darlington, NSW 2008, Australia
[2] Univ Sydney, Sch Comp Sci, Darlington, NSW 2008, Australia
[3] Southern Univ Sci & Technol, Dept Comp Sci & Engn, Shenzhen Key Lab Computat Intelligence, Univ Key Lab Evolving Intelligent Syst Guangdong, Shenzhen 518055, Guangdong, Peoples R China
[4] Univ Birmingham, Sch Comp Sci, CERCIA, Birmingham B15 2TT, W Midlands, England
基金
英国工程与自然科学研究理事会; 澳大利亚研究理事会;
关键词
Gallium nitride; Training; Generators; Generative adversarial networks; Evolutionary computation; Task analysis; Computational modeling; Deep generative models; evolutionary computation; generative adversarial networks (GANs); IMAGE SYNTHESIS; TEXT;
D O I
10.1109/TEVC.2019.2895748
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Generative adversarial networks (GANs) have been effective for learning generative models for real-world data. However, accompanied with the generative tasks becoming more and more challenging, existing GANs (GAN and its variants) tend to suffer from different training problems such as instability and mode collapse. In this paper, we propose a novel GAN framework called evolutionary GANs (E-GANs) for stable GAN training and improved generative performance. Unlike existing GANs, which employ a predefined adversarial objective function alternately training a generator and a discriminator, we evolve a population of generators to play the adversarial game with the discriminator. Different adversarial training objectives are employed as mutation operations and each individual (i.e., generator candidature) are updated based on these mutations. Then, we devise an evaluation mechanism to measure the quality and diversity of generated samples, such that only well-performing generator(s) are preserved and used for further training. In this way, E-GAN overcomes the limitations of an individual adversarial training objective and always preserves the well-performing offspring, contributing to progress in, and the success of GANs. Experiments on several datasets demonstrate that E-GAN achieves convincing generative performance and reduces the training problems inherent in existing GANs.
引用
收藏
页码:921 / 934
页数:14
相关论文
共 76 条
[1]  
Alec R., 2016, 4 INT C LEARN REPR I, P1
[2]  
[Anonymous], P 34 INT C MACH LEAR
[3]  
[Anonymous], 2009, TR2009 U TOR DEP COM
[4]  
[Anonymous], 2017, PROC CVPR IEEE, DOI DOI 10.1109/CVPR.2017.632
[5]  
[Anonymous], 2015, COMPUT SCI
[6]  
[Anonymous], EXPRESSIVE POWER PAR
[7]  
[Anonymous], 2018, SELF ATTENTION GENER
[8]  
[Anonymous], 2014, ARXIV PREPRINT ARXIV
[9]  
[Anonymous], 2016, ADV NEURAL INFORM PR
[10]  
Arjovsky M., 2017, P 34 INT C MACH LEAR, V70, P214