Re-GAN: Data-Efficient GANs Training via Architectural Reconfiguration.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) typically require large datasets for high-fidelity image generation.
• Research on GANs 'lottery tickets' indicates sparse sub-networks can achieve superior performance with less data.
• Discovering these sub-networks conventionally involves a costly train-prune-retrain process.

Purpose of the Study:

• Introduce Re-GAN, a novel, data-efficient training approach for GANs.
• Address the limitations of resource-intensive methods for uncovering efficient GAN sub-networks.
• Expand the study of Re-GAN to new applications, datasets, and compatibility with data augmentation.

Main Methods:

• Re-GAN dynamically reconfigures GAN architecture during training.
• Employs iterative pruning of non-important connections and subsequent regrowth.
• Maintains model representational strength by preventing premature loss of critical features.

Main Results:

• Re-GAN demonstrates stable and efficient GAN training with limited data.
• The method serves as an alternative to progressive growing and GANs lottery ticket approaches.
• Expanded study validates Re-GAN's effectiveness in Image-to-Image translation and diverse scenarios.

Conclusions:

• Re-GAN establishes itself as a generic and adaptable GAN training methodology.
• The approach significantly improves GAN performance and efficiency, especially with data constraints.
• Code availability facilitates further research and application of Re-GAN.