Closing the Gap Between Theory and Practice During Alternating Optimization for GANs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) aim to synthesize high-quality, diverse samples.
• Mode collapse remains a significant challenge, hindering GANs' ability to capture target data distributions.
• Current GAN training practices, particularly alternating optimization, are not fully aligned with theoretical objectives.

Purpose of the Study:

• To re-evaluate alternating optimization in GANs and its theoretical underpinnings.
• To propose a novel generator loss function that mitigates mode collapse.
• To enhance the generator's capacity to learn the true data distribution.

Main Methods:

• Rethinking the theoretical basis of alternating optimization in GANs.
• Introducing a novel generator loss function designed for alternating optimization.
• Theoretically optimizing the reverse Kullback-Leibler divergence between model and target distributions.

Main Results:

• The proposed loss function provides an appropriate objective for the generator under alternating optimization.
• The novel approach theoretically optimizes reverse Kullback-Leibler divergence, encouraging learning of the target distribution.
• Extensive experiments show consistent performance improvements across diverse datasets and network architectures.

Conclusions:

• The novel generator loss function effectively addresses mode collapse in GANs.
• The method enhances the generator's ability to capture the target data distribution.
• This approach offers a practical solution for improving GAN performance in real-world applications.