Reproduce “Functional 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
Gavin Kerrigan, Giosue Migliorini, Padhraic Smyth
Functional Flow Matching (FFM) introduces the first continuous-time normalizing flow model designed for infinite-dimensional function spaces, rigorously generalizing the Flow Matching framework. The model demonstrates superior performance in generating diverse functional data, including time series and PDE solutions, while requiring fewer function evaluations during sampling compared to existing function-space generative models.
Key insights
- Continuous-time normalizing flows can be rigorously extended to infinite-dimensional function spaces by focusing on measure-theoretic properties and carefully constructing paths of Gaussian measures.
- Analytically deriving conditional vector fields for specific Gaussian path choices (like OT and VP) allows for an efficient, simulation-free regression objective to train the function-space generative model.
- Fourier Neural Operators (FNOs) are effective architectures for learning the operator mapping between function spaces, enabling resolution-invariant generation and super-resolution capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.