Beyond pseudo-time: GP-SDE Matching learns stochastic dynamics from irregular observations
canaesseth · x · 2026-10-13
The author rounds up recent methods moving diffusion models, flow matching, and stochastic interpolants beyond pseudo-time toward real physical time. The linked arXiv paper proposes GP-SDE Matching, a simulation-free variational framework for learning Gaussian-process SDEs from noisy, irregularly sampled observations. It analytically marginalizes the sparse GP posterior into a drift-matching objective capturing both mean and uncertainty, plus an irregular-time-aware variational state posterior. On stochastic Lorenz-63 it substantially improves drift recovery and state reconstruction, and it stays robust under sparse observations across five system-identification benchmarks, beating latent-SDE and state-space baselines.
More from Research
- AI Math Breakthroughs Are Mostly Brute-Force Counterexamples, Not Insight — jonippolito · 2026-10-13
- Proposing a 'Millennium Problem of Robotics': ten long-horizon research challenges listed — DJiafei · 2026-10-13
- Fields Medalist Timothy Gowers gives talk: "What AI did during my summer holidays" — tak3sh8 · 2026-10-13
- Why mathematicians should race for early access to advanced math models — haltakov · 2026-10-13
- LightOn's first finetunes land: medieval Latin and cultural heritage models at 0.8B and 4B — IgorCarron · 2026-10-13
- From tens of millions to 100 billion: the next leap for cell atlas efforts — EricTopol · 2026-10-13