Reproduce “Evolutionary Generative Adversarial Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chaoyue Wang, Chang Xu, Xin Yao, Dacheng Tao
Evolutionary Generative Adversarial Networks (E-GAN), developed by researchers at University of Technology Sydney, The University of Sydney, Southern University of Science and Technology, and University of Birmingham, integrates evolutionary algorithms to enhance training stability and mitigate mode collapse in GANs. It achieves superior mode coverage and robustness across diverse architectural variations, demonstrating state-of-the-art generative performance on various image datasets.
Key insights
- Dynamically combining diverse adversarial objectives within an evolutionary framework allows E-GAN to leverage their individual strengths and adaptively apply the most effective training strategy at different learning stages.
- The system's adaptive nature is evident as non-saturating objectives (heuristic, least-squares) are favored in early training, while the minimax objective gains preference as distributions overlap closer to convergence.
- Explicitly incorporating a diversity fitness score, based on the smoothness of discriminator gradients, effectively counters mode collapse by encouraging generators to explore the full data distribution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.