BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning probabilistic filters for data assimilation - Eviatar Bac
 h (University of Reading)
DTSTART:20250627T091500Z
DTEND:20250627T101500Z
UID:TALK232345@talks.cam.ac.uk
DESCRIPTION:Co-authors: Ricardo Baptista\, Edoardo Calvello\, Bohan Chen\,
  Enoch Luk\, Andrew Stuart\nFiltering &ndash\; the task of estimating the 
 conditional distribution for states of a dynamical system given partial an
 d noisy observations &ndash\; is important in many areas of science and en
 gineering\, including weather and climate prediction. However\, the filter
 ing distribution is generally intractable to obtain for high-dimensional\,
  nonlinear systems. Filters used in practice\, such as the ensemble Kalman
  filter (EnKF)\, provide biased probabilistic estimates for nonlinear syst
 ems and have numerous tuning parameters.\nI will present a framework for l
 earning a parameterized analysis map &ndash\; the transformation that take
 s samples from a forecast distribution\, and combines with an observation\
 , to update the approximate filtering distribution &ndash\; using variatio
 nal inference. In principle this can lead to a better approximation of the
  filtering distribution\, and hence smaller bias. We show that this method
 ology can be used to learn the gain matrix\, in an affine analysis map\, f
 or filtering linear and nonlinear dynamical systems\; we also study the le
 arning of inflation and localization parameters for an EnKF. The framework
  developed here can also be used to learn new filtering algorithms with mo
 re general forms for the analysis map.\nI will also present some recent wo
 rk on learning corrections to the EnKF using permutation-invariant neural 
 architectures\, leading to superior performance compared to leading method
 s in filtering chaotic systems. Lastly\, I will present some ideas for lea
 rning filters using other probabilistic cost functions.
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR