ScoreMix: A Scalable Augmentation Strategy for Training GANs With Limited Data.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) often face overfitting challenges due to limited training data.
• Existing data augmentation techniques for GANs can be difficult to scale for practical applications.

Purpose of the Study:

• To introduce ScoreMix, a novel and scalable data augmentation approach for image synthesis tasks.
• To enhance GAN training by reducing overfitting and increasing data diversity.

Main Methods:

• ScoreMix generates augmented samples via convex combinations of real data.
• Augmented samples are optimized by minimizing data score norms, ensuring proximity to the data manifold.
• Scores are estimated using a deep network trained with multi-scale score matching.

Main Results:

• ScoreMix effectively mitigates the overfitting problem in GANs.
• The method enhances data diversity, leading to improved performance across various image synthesis tasks.
• GAN models integrated with ScoreMix show significant performance gains.

Conclusions:

• ScoreMix offers a scalable and effective data augmentation solution for GANs.
• The approach requires no hyperparameter tuning or architectural modifications, facilitating easy integration.
• ScoreMix demonstrates broad applicability and significant improvements in GAN-based image synthesis.