DynGAN: Solving Mode Collapse in GANs With Dynamic Clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) are powerful tools for synthesizing realistic data.
• Mode collapse, a common GAN issue, significantly reduces the diversity of generated samples.
• This limitation hinders the broader application of GANs in complex data generation tasks.

Purpose of the Study:

• To theoretically analyze the root causes of mode collapse in GANs.
• To propose a novel framework, Dynamic GAN, for detecting and resolving mode collapse.
• To enhance the diversity and coverage of generated samples in GANs.

Main Methods:

• Theoretical analysis of the generator loss function's non-convexity.
• Development of a Dynamic GAN framework utilizing discriminator outputs for sample detection.
• Partitioning the training set and training dynamic conditional models on these partitions.

Main Results:

• Demonstrated that parameters leading to partial mode coverage are local minima of the generator loss.
• Dynamic GAN framework effectively detects collapsed samples.
• Experiments show Dynamic GAN achieves progressive mode coverage and outperforms other GAN variants.

Conclusions:

• Mode collapse in GANs stems from non-convex generator loss functions.
• Dynamic GAN provides a quantitative method to detect and resolve mode collapse.
• The proposed method significantly improves sample diversity and coverage in GANs.