DAG

DAG optimizes the integration of augmented data into Generative Adversarial Network (GAN) training to align the learned model distribution with the original data distribution and improve generator and discriminator learning.

Key Features:

• Principled Framework: Integrates augmented data into GAN training while maintaining alignment with the original data distribution.
• Theoretical Foundation: Provides theoretical analysis demonstrating reduction of Jensen-Shannon (JS) divergence between original and model distributions.
• Enhanced Learning: Leverages augmented data to improve learning of both the discriminator and the generator within GANs.
• Architecture Versatility: Applies across multiple GAN variants without altering the core alignment objective.
• Empirical Performance: Demonstrates improved generation quality and diversity, including state-of-the-art Fréchet Inception Distance (FID) scores on certain architectures.

Scientific Applications:

• GAN Architectures: Applied to unconditional GAN, conditional GAN, self-supervised GAN, and CycleGAN.
• Datasets and Domains: Evaluated on natural images and medical images.

Methodology:

Integrates augmented data into the GAN training process in a manner that aligns with the original GAN objectives.