On Data Augmentation for GAN Training.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) benefit from large datasets, but data collection is often costly, especially in medical imaging.
• Classical Data Augmentation (DA) techniques can inadvertently shift the data distribution, potentially hindering GAN performance.

Purpose of the Study:

• To propose a novel framework, Data Augmentation Optimized for GAN (DAG), to effectively utilize augmented data in GAN training.
• To ensure augmented data aids in learning the original data distribution rather than an altered one.

Main Methods:

• Developed a principled framework (DAG) to integrate augmented data into GAN training.
• Provided theoretical analysis demonstrating DAG's alignment with GAN objectives (minimizing Jensen-Shannon divergence).
• Applied DAG to various GAN architectures (unconditional, conditional, self-supervised, CycleGAN) on natural and medical image datasets.

Main Results:

• DAG consistently improved performance across different GAN models and datasets.
• The framework effectively leveraged augmented data to enhance both discriminator and generator learning.
• State-of-the-art Fréchet Inception Distance (FID) scores were achieved with DAG in certain GAN models.

Conclusions:

• DAG offers a robust method for incorporating augmented data into GANs, overcoming limitations of classical DA.
• The proposed framework leads to significant improvements in GAN training and generative model performance.
• DAG is a versatile tool applicable to diverse GAN architectures and data types, including sensitive medical imaging data.