Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

"FORCE" learning in recurrent neural networks as data assimilation.

Gregory S Duane1

  • 1Geophysical Institute, University of Bergen, Postboks 7803, 5020 Bergen, Norway and Department of Atmospheric and Oceanic Sciences, University of Colorado, UCB 311, Boulder, Colorado 80309, USA.

Chaos (Woodbury, N.Y.)
|January 1, 2018
PubMed
Summary
This summary is machine-generated.

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Simulating climate with a synchronization-based supermodel.

Chaos (Woodbury, N.Y.)·2018
Same author

Role of atmosphere-ocean interactions in supermodeling the tropical Pacific climate.

Chaos (Woodbury, N.Y.)·2018
Same author

Introduction to focus issue: Synchronization in large networks and continuous media-data, models, and supermodels.

Chaos (Woodbury, N.Y.)·2018
Same author

A pulse sequence optimization method for assessment of nucleus size in q-space analysis of idealized cells.

Journal of magnetic resonance (San Diego, Calif. : 1997)·2013
Same author

A "cellular neuronal" approach to optimization problems.

Chaos (Woodbury, N.Y.)·2009
Same author

Synchronization of extended systems from internal coherence.

Physical review. E, Statistical, nonlinear, and soft matter physics·2009

The FORCE algorithm for neuronal network learning is a Kalman Filter, offering new insights for initialization and extending to complex weight interactions and multiple outputs.

Area of Science:

  • Computational Neuroscience
  • Machine Learning
  • Signal Processing

Background:

  • The study addresses the "FORCE" algorithm, a method for training recurrent neural networks.
  • Existing methods for neuronal network learning often lack a unified theoretical framework.

Purpose of the Study:

  • To re-interpret the FORCE algorithm within the established framework of Kalman Filtering.
  • To explore the implications of this new interpretation for the algorithm's practical application and theoretical extensions.

Main Methods:

  • The FORCE algorithm's update rules were analyzed and reformulated.
  • Mathematical connections were established between the algorithm's parameters and Kalman Filter components, specifically state-dependent background error covariances.

Related Experiment Videos

Main Results:

  • The FORCE algorithm was successfully cast as a specific type of Kalman Filter.
  • This interpretation provides a principled approach to initializing the learning algorithm.
  • The framework allows for extensions to model interactions between weight updates and to handle multiple output signals.

Conclusions:

  • Viewing the FORCE algorithm as a Kalman Filter offers a powerful new perspective.
  • This unification facilitates algorithm initialization, extension to complex network dynamics, and representation of multi-output systems.