We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Statistics > Machine Learning

Title: Online Natural Gradient as a Kalman Filter

Authors: Yann Ollivier
Abstract: We cast Amari's natural gradient in statistical learning as a specific case of Kalman filtering. Namely, applying an extended Kalman filter to estimate a fixed unknown parameter of a probabilistic model from a series of observations, is rigorously equivalent to estimating this parameter via an online stochastic natural gradient descent on the log-likelihood of the observations.
In the i.i.d. case, this relation is a consequence of the "information filter" phrasing of the extended Kalman filter. In the recurrent (state space, non-i.i.d.) case, we prove that the joint Kalman filter over states and parameters is a natural gradient on top of real-time recurrent learning (RTRL), a classical algorithm to train recurrent models.
This exact algebraic correspondence provides relevant interpretations for natural gradient hyperparameters such as learning rates or initialization and regularization of the Fisher information matrix.
Comments: 3rd version: expanded intro
Subjects: Machine Learning (stat.ML); Optimization and Control (math.OC)
Cite as: arXiv:1703.00209 [stat.ML]
  (or arXiv:1703.00209v3 [stat.ML] for this version)

Submission history

From: Yann Ollivier [view email]
[v1] Wed, 1 Mar 2017 10:13:52 GMT (26kb)
[v2] Thu, 27 Apr 2017 16:45:48 GMT (27kb)
[v3] Mon, 27 Aug 2018 18:45:10 GMT (33kb)

Link back to: arXiv, form interface, contact.