Reproduce “Improving Neural Optimal Transport via Displacement Interpolation” 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
Jaemoo Choi, Yongxin Chen, Jaewoong Choi
Researchers from Georgia Institute of Technology and Sungkyunkwan University developed the Displacement Interpolation Optimal Transport Model (DIOTM), which leverages the theoretical properties of displacement interpolation and a Hamilton-Jacobi-Bellman (HJB) regularizer. This approach provides enhanced training stability and accuracy for neural optimal transport map learning, achieving competitive performance on image-to-image translation tasks.
Key insights
- McCann's Displacement Interpolation, which describes the optimal path between distributions, can be reformulated into a dual max-min problem involving intermediate potential functions and transport maps.
- The optimal value function within this dynamic optimal transport context satisfies a specific Hamilton-Jacobi-Bellman (HJB) equation, which serves as a powerful theoretical condition for regularization.
- Leveraging the entire continuous trajectory of displacement interpolation, rather than just the static endpoints, provides richer information that dramatically enhances the stability and accuracy of neural optimal transport learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.