On Positive-Unlabeled Classification From Corrupted Data in GANs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) traditionally use a fixed positive-negative classification for discriminator training.
• This approach struggles with the evolving quality of generated data and potential real-world data corruption.

Purpose of the Study:

• To propose a novel approach for stabilizing GAN discriminator training.
• To address the challenge of corrupted data in real-world GAN applications.
• To improve the overall performance and robustness of GANs.

Main Methods:

• Introduced a positive and unlabeled (PU) classification framework for GANs, leading to Positive-Unlabeled GAN (PUGAN).
• Developed PUGAN-C to handle corrupted real data by treating real data as unlabeled and generated data as positive.
• Theoretically analyzed the global optimality and optimization goals of the proposed models.

Main Results:

• PUGAN achieves comparable or superior performance to existing discriminator stabilization methods.
• PUGAN-C effectively handles corrupted datasets in image generation tasks.
• Experimental results demonstrate the effectiveness and generalization capabilities of PUGAN-C.

Conclusions:

• The PU classification approach offers a more robust and adaptable method for GAN training.
• PUGAN and PUGAN-C provide effective solutions for GAN stability and corrupted data challenges.
• The proposed methods show significant promise for real-world GAN applications.