Reproduce “Conditional Wasserstein Distances with Applications in Bayesian OT Flow Matching” 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
Jannis Chemseddine, Paul Hagemann, Gabriele Steidl, Christian Wald
This research introduces a conditional Wasserstein distance that precisely equals the expected conditional Wasserstein distance, addressing a theoretical gap in learning conditional generative models. The work uses this framework to develop Bayesian Optimal Transport Flow Matching, demonstrating improved posterior approximation in inverse problems and achieving a Frechet Inception Distance (FID) of 4.10 for class-conditional image generation on CIFAR10.
Key insights
- A conditional Wasserstein distance exists that is exactly equal to the expected Wasserstein distance of conditional measures, providing a theoretically sound metric for conditional generative modeling.
- Geodesics in this conditional Wasserstein space are driven by velocity fields with zero components corresponding to the conditioning variable (Y), implying optimal conditional transport involves no movement of conditioning information.
- The dual formulation of the conditional Wasserstein distance (for p=1) naturally aligns with the loss structure of conditional Wasserstein GANs, offering theoretical justification for their design.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.