Improving Generative Adversarial Networks With Local Coordinate Coding.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generative adversarial networks (GANs) excel at realistic data generation from prior distributions.
• Standard GANs often lose semantic data information due to independent prior distributions.
• Latent distributions learned from data can capture semantic information but pose sampling challenges.

Purpose of the Study:

• To propose a novel GAN model, LCCGAN, that improves image generation by incorporating local coordinate coding (LCC).
• To enhance LCCGAN with LCCGAN++ by introducing a higher-order term for improved generator approximation.
• To analyze the generalization capabilities of LCCGAN and LCCGAN++ and demonstrate their superiority.

Main Methods:

• Developed LCCGAN, a GAN model utilizing LCC sampling to extract meaningful points from the latent manifold.
• Introduced LCCGAN++, an advanced version with a higher-order term in the generator for better approximation.
• Derived generalization bounds for LCCGAN and LCCGAN++, proving sufficiency of low-dimensional input for good generalization.

Main Results:

• LCCGAN effectively exploits local information on the latent manifold, producing high-quality generated data.
• LCCGAN++ demonstrates further performance improvements due to enhanced generator approximation.
• Extensive experiments on benchmark datasets confirm the proposed methods outperform existing GAN techniques.

Conclusions:

• LCCGAN and LCCGAN++ offer significant advancements in GAN-based image generation by effectively utilizing latent manifold information.
• The proposed methods achieve superior data quality and generalization performance compared to conventional GAN approaches.
• Low-dimensional input is sufficient for achieving good generalization with LCCGAN and LCCGAN++.