Adaptive Weighted Discriminator for Training Generative Adversarial Networks.

Area of Science:

• Machine Learning
• Computer Vision
• Deep Learning

Background:

• Generative Adversarial Networks (GANs) are key unsupervised learning models.
• Existing discriminator loss functions, summing real and fake losses equally, can cause training instability and mode collapse.
• A need exists for improved GAN training methodologies to enhance stability and generation quality.

Purpose of the Study:

• Introduce a novel family of adaptive weighted loss functions (aw-loss) for GAN discriminators.
• Address the instability and mode collapse issues inherent in traditional GAN training.
• Improve the performance and stability of GANs in image generation tasks.

Main Methods:

• Developed adaptive weighted loss functions (aw-loss) by introducing adaptive weights to the sum of real and fake discriminator losses.
• Utilized the gradients of the real and fake loss components to dynamically adjust weights, guiding the discriminator training.
• Applied the aw-loss functions to established GAN architectures including SN-GAN, AutoGAN, and BigGAN.

Main Results:

• Demonstrated significant improvements in GAN stability and a reduction in mode collapse.
• Achieved superior performance on unconditional and conditional image generation tasks.
• Showcased enhanced results on CIFAR-10, STL-10, and CIFAR-100 datasets, validated by Inception Scores (IS) and Fréchet Inception Distance (FID) metrics.

Conclusions:

• Adaptive weighted loss functions offer a promising approach to stabilize GAN training and mitigate mode collapse.
• The proposed aw-loss method enhances image generation quality and performance across various datasets and GAN architectures.
• This technique provides a broadly applicable strategy for improving GAN training dynamics in unsupervised machine learning.