Reproduce “From GAN to WGAN” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lilian Weng
The paper details how Wasserstein GANs (WGANs) address the training instability and gradient issues of traditional GANs by replacing the Jensen-Shannon divergence with the Wasserstein distance. This modification enables the model to provide meaningful gradients even when real and generated data distributions are disjoint, leading to more stable training and often higher quality generated samples.
Key insights
- The Wasserstein distance provides meaningful gradients even when the real and generated distributions have non-overlapping supports, a common scenario in high-dimensional spaces.
- Unlike Jensen-Shannon divergence, the Wasserstein distance does not saturate to a constant value when distributions are disjoint, offering continuous learning signals to the generator.
- The Earth Mover's interpretation of Wasserstein distance offers a robust measure of dissimilarity, providing a more stable and informative metric for distribution similarity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.