Score Kalman Filter accepted as NeurIPS spotlight, scales Bayesian filtering to n=20

GhaffariMaani · x · 2026-09-25

Maani Ghaffari's team announces SKF (The Score Kalman Filter) is accepted as a NeurIPS spotlight. The method, led by student Kaito Iwasaki, combines moment representation with score matching and Stein's identity, replacing exponentially costly partition-function integrals in MaxEnt moment filtering with a single linear solve. It reduces to information-form Kalman filtering as a special case, runs to n=20 on coupled-oscillator networks, and beats EKF/UKF/EnKF and particle-filter baselines on RMSE.

Original post →

More from Research

Research channel →