Simplified Fréchet Distance for Generative Adversarial Nets.

Area of Science:

• Machine Learning
• Deep Learning
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) produce realistic data but often suffer from unstable training.
• The Fréchet distance (FD) offers a theoretical solution but is computationally infeasible due to its covariance term.
• Existing GAN models like Boundary Equilibrium GAN (BEGAN) can still exhibit instability and mode collapse.

Purpose of the Study:

• To introduce a novel, computationally feasible distance metric, the Simplified Fréchet distance (SFD).
• To propose a new GAN model, the Simplified Fréchet GAN (SFGAN), utilizing SFD for enhanced training stability.
• To evaluate SFGAN's performance against established GAN architectures and distance metrics.

Main Methods:

• Developed the Simplified Fréchet distance (SFD) by removing the covariance term from the Fréchet distance (FD).
• Integrated SFD into the loss function of a GAN architecture, creating the Simplified Fréchet GAN (SFGAN), based on BEGAN.
• Conducted experiments on datasets like CelebA and CIFAR-10, comparing SFGAN with BEGAN using various metrics.

Main Results:

• SFGAN exhibited significantly improved training stability compared to BEGAN.
• Mode collapse and mode drop were observed much later in SFGAN (after 3000k steps) compared to BEGAN (457k-968k steps).
• SFD proved more effective in stabilizing GAN training than other distance metrics, enhancing BEGAN's network structure.

Conclusions:

• The Simplified Fréchet distance (SFD) is a suitable and effective metric for improving GAN stability.
• SFGAN offers a robust alternative to existing GAN models, mitigating issues like mode collapse.
• This research highlights the potential of SFD for advancing generative modeling techniques.