Information-Based Boundary Equilibrium Generative Adversarial Networks with Interpretable Representation Learning.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) are powerful for image generation but often lack interpretability.
• Existing GANs can suffer from convergence issues and unstable training dynamics.
• Learning interpretable representations from images is crucial for controllable generation.

Purpose of the Study:

• To develop a novel image generation algorithm based on GANs with enhanced interpretability.
• To address the convergence problems commonly faced by GANs.
• To enable manipulation of generated images through control of latent codes.

Main Methods:

• An information-theoretic extension to an autoencoder-based discriminator was employed.
• The model minimizes Wasserstein distance-based losses for generator and discriminator.
• Mutual information between latent variables and observations is maximized.
• Proportional control theory was used to balance generator and discriminator training.

Main Results:

• The proposed method successfully learns interpretable representations from input images.
• The GAN mitigates convergence problems and achieves stable training.
• Generated images can be manipulated by controlling latent codes.
• Visual quality of generated images is effectively maintained.

Conclusions:

• The developed GAN provides a stable and controllable approach to interpretable image generation.
• The method demonstrates effective control over generated image attributes via latent codes.
• Future work will focus on learning disentangled factors for improved interpretability.