A Unifying Generator Loss Function for Generative Adversarial Networks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) are powerful deep learning models for generating data.
• Existing GAN variants use diverse objective functions, leading to fragmentation in research and application.
• A unified framework is needed to generalize and simplify GAN training.

Purpose of the Study:

• To introduce a unifying α-parametrized generator loss function for dual-objective GANs.
• To develop the Lα-GAN system, a novel framework that generalizes existing GAN loss functions.
• To demonstrate the theoretical underpinnings and practical applicability of the Lα-GAN.

Main Methods:

• Introduced a unifying α-parametrized generator loss function (Lα) for GANs.
• Developed the Lα-GAN system, integrating Lα with canonical discriminator losses.
• Analyzed the generator's optimization problem under an optimal discriminator, linking it to Jensen-fα-divergence.
• Showcased Lα-GAN as a generalization of VanillaGAN, LSGAN, LkGAN, and (αD,αG)-GAN.

Main Results:

• The Lα-GAN framework unifies several existing GAN loss functions.
• The generator's optimization problem is shown to minimize a Jensen-fα-divergence.
• Experimental validation on MNIST, CIFAR-10, and Stacked MNIST datasets demonstrates Lα-GAN's effectiveness.
• The proposed method offers a flexible and generalized approach to GAN training.

Conclusions:

• The Lα-GAN provides a unified and flexible framework for generative adversarial networks.
• This generalization simplifies the understanding and application of various GAN models.
• The Lα-GAN system shows promising performance across diverse image generation tasks.