Reproduce “Towards Foundation Inference Models that Learn ODEs In-Context” 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
Maximilian Mauel, Manuel Hinz, Patrick Seifner, David Berghaus, Ramses J. Sanchez
FIM-ODE estimates a system’s continuous-time vector field from observed trajectories, using them as context and allowing multiple trajectories from the same system. On a synthetic polynomial-vector-field test set, its held-out-trajectory R²-accuracy—the percentage of R² scores above 0.9—is 0.26, versus 0.19 for ODEFormer with one context trajectory. This regular-grid result does not establish real-world or sparse/noisy-data performance.
Key insights
- Neural operators offer superior flexibility and robustness in representing complex vector fields and handling sparse, noisy data compared to symbolic regression.
- Large-scale pretraining on diverse synthetic data allows Foundation Inference Models to generalize effectively to new ODE systems in a zero-shot manner.
- The ability to process multiple context trajectories significantly enhances the accuracy and global consistency of the inferred vector fields, improving generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.