CGE-GAN: Contrastive-guided evolutionary generative adversarial networks with dynamic adaptive weight sharing

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Generative adversarial networks (GANs) excel at image synthesis but struggle with mode collapse, training instability, and architecture search.
• Existing evolutionary GANs lack semantic alignment, effective weight reuse, and knowledge transfer between generations.

Purpose of the Study:

• To introduce contrastive-guided evolutionary GANs (CGE-GANs) for stable, high-fidelity image generation.
• To address limitations of current GANs, including semantic alignment and training efficiency.

Main Methods:

• Developed a hybrid Wasserstein-Contrastive loss function for semantic alignment and adversarial competitiveness.
• Incorporated dynamic adaptive weight sharing (DAWS) for efficient training.
• Utilized knowledge distillation-based crossover for feature preservation across generations.

Main Results:

• CGE-GAN achieved an Inception Score (IS) of 8.99 (CIFAR-10) and 10.46 (STL-10).
• Achieved a Fréchet Inception Distance (FID) of 9.74 (CIFAR-10) and 21.86 (STL-10).
• Reduced FID by up to 1.74 points compared to baselines with only 0.36 GPU days.

Conclusions:

• Contrastive-driven evolution is effective for stable, high-fidelity image generation.
• CGE-GAN demonstrates superior performance in semantic diversity and convergence efficiency.
• The proposed method offers a promising direction for advancing GANs.