Talk Announcement: 1W-MINDS Seminar — Learning Enhanced Ensemble Filters
Table of Contents
Overview
I’m pleased to share that I’ll be speaking in the 1W-MINDS seminar series. The talk will present new results on learning-augmented ensemble filtering for nonlinear, non-Gaussian state estimation.
Event Details
- Date: Thursday, November 6, 2025
- Time: 2:30 PM – 3:30 PM Eastern (11:30 AM – 12:30 PM Pacific)
- Format: Online (Zoom)
- Zoom: Join the webinar
- Password: 101 (likely not required when using the full link)
Talk
Title
Learning Enhanced Ensemble Filters
Abstract
The filtering distribution in hidden Markov models evolves according to a mean-field law on the joint state–observation space. While the ensemble Kalman filter (EnKF) provides a robust particle approximation, its Gaussian ansatz limits accuracy for non-Gaussian dynamics. We address this with a measure neural mapping (MNM)—a neural operator defined on probability measures—which yields the MNM-enhanced ensemble filter (MNMEF) in both the mean-field limit and as an interacting-particle algorithm. The ensemble instantiation uses a permutation-invariant set transformer, enabling a single parameterization to operate across varying ensemble sizes, with light size-specific fine-tuning further improving accuracy. Empirically, MNMEF achieves superior RMSE to leading methods on Lorenz-96 and Kuramoto–Sivashinsky benchmarks.
A central theoretical contribution establishes a continuum limit for attention on measures: attention layers fed empirical measures are consistent with their continuous-measure counterparts, with outputs converging to those from the underlying measure in Wasserstein distance as the sample size goes to infinity. This result justifies cross-size parameter sharing, explains the observed stability of MNMEF when changing ensemble sizes.
Presenter
Bohan Chen, California Institute of Technology, U.S.
Add to Calendar
- Eastern Time: Nov 6, 2025, 2:30–3:30 PM ET
- Pacific Time: Nov 6, 2025, 11:30 AM–12:30 PM PT
