Predicting non-linear dynamics by stable local learning in a recurrent spiking neural network

  1. Aditya Gilra  Is a corresponding author
  2. Wulfram Gerstner
  1. Ecole Polytechnique Federale de Lausanne, Switzerland

Abstract

The brain needs to predict how the body reacts to motor commands, but how a network of spiking neurons can learn non-linear body dynamics using local, online and stable learning rules is unclear. Here, we present a supervised learning scheme for the feedforward and recurrent connections in a network of heterogeneous spiking neurons. The error in the output is fed back through fixed random connections with a negative gain, causing the network to follow the desired dynamics. The rule for Feedback-based Online Local Learning Of Weights (FOLLOW) is local in the sense that weight changes depend on the presynaptic activity and the error signal projected onto the postsynaptic neuron. We provide examples of learning linear, non-linear and chaotic dynamics, as well as the dynamics of a two-link arm. Under reasonable approximations, we show, using the Lyapunov method, that FOLLOW learning is uniformly stable, with the error going to zero asymptotically.

Article and author information

Author details

  1. Aditya Gilra

    School of Computer and Communication Sciences, Ecole Polytechnique Federale de Lausanne, Lausanne, Switzerland
    For correspondence
    aditya.gilra@epfl.ch
    Competing interests
    The authors declare that no competing interests exist.
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-8628-1864
  2. Wulfram Gerstner

    School of Computer and Communication Sciences, Ecole Polytechnique Federale de Lausanne, Lausanne, Switzerland
    Competing interests
    The authors declare that no competing interests exist.

Funding

European Research Council (Multirules 268 689)

  • Aditya Gilra
  • Wulfram Gerstner

Swiss National Science Foundation (CRSII2_147636)

  • Aditya Gilra
  • Wulfram Gerstner

European Commission Horizon 2020 Framework Program (Human Brain Project 720270)

  • Wulfram Gerstner

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Reviewing Editor

  1. Peter Latham, University College London, United Kingdom

Version history

  1. Received: May 3, 2017
  2. Accepted: November 22, 2017
  3. Accepted Manuscript published: November 27, 2017 (version 1)
  4. Version of Record published: December 14, 2017 (version 2)
  5. Version of Record updated: February 23, 2018 (version 3)

Copyright

© 2017, Gilra & Gerstner

This article is distributed under the terms of the Creative Commons Attribution License permitting unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 4,850
    views
  • 908
    downloads
  • 59
    citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Aditya Gilra
  2. Wulfram Gerstner
(2017)
Predicting non-linear dynamics by stable local learning in a recurrent spiking neural network
eLife 6:e28295.
https://doi.org/10.7554/eLife.28295

Share this article

https://doi.org/10.7554/eLife.28295

Further reading

    1. Immunology and Inflammation
    2. Neuroscience
    Irini Papazian, Maria Kourouvani ... Lesley Probert
    Research Article

    Autoimmune diseases of the central nervous system (CNS) such as multiple sclerosis (MS) are only partially represented in current experimental models and the development of humanized immune mice is crucial for better understanding of immunopathogenesis and testing of therapeutics. We describe a humanized mouse model with several key features of MS. Severely immunodeficient B2m-NOG mice were transplanted with peripheral blood mononuclear cells (PBMCs) from HLA-DRB1-typed MS and healthy (HI) donors and showed rapid engraftment by human T and B lymphocytes. Mice receiving cells from MS patients with recent/ongoing Epstein–Barr virus reactivation showed high B cell engraftment capacity. Both HLA-DRB1*15 (DR15) MS and DR15 HI mice, not HLA-DRB1*13 MS mice, developed human T cell infiltration of CNS borders and parenchyma. DR15 MS mice uniquely developed inflammatory lesions in brain and spinal cord gray matter, with spontaneous, hCD8 T cell lesions, and mixed hCD8/hCD4 T cell lesions in EAE immunized mice, with variation in localization and severity between different patient donors. Main limitations of this model for further development are poor monocyte engraftment and lack of demyelination, lymph node organization, and IgG responses. These results show that PBMC humanized mice represent promising research tools for investigating MS immunopathology in a patient-specific approach.

    1. Neuroscience
    Ju-Young Lee, Dahee Jung, Sebastien Royer
    Research Article

    Animals can use a repertoire of strategies to navigate in an environment, and it remains an intriguing question how these strategies are selected based on the nature and familiarity of environments. To investigate this question, we developed a fully automated variant of the Barnes maze, characterized by 24 vestibules distributed along the periphery of a circular arena, and monitored the trajectories of mice over 15 days as they learned to navigate towards a goal vestibule from a random start vestibule. We show that the patterns of vestibule visits can be reproduced by the combination of three stochastic processes reminiscent of random, serial, and spatial strategies. The processes randomly selected vestibules based on either uniform (random) or biased (serial and spatial) probability distributions. They closely matched experimental data across a range of statistical distributions characterizing the length, distribution, step size, direction, and stereotypy of vestibule sequences, revealing a shift from random to spatial and serial strategies over time, with a strategy switch occurring approximately every six vestibule visits. Our study provides a novel apparatus and analysis toolset for tracking the repertoire of navigation strategies and demonstrates that a set of stochastic processes can largely account for exploration patterns in the Barnes maze.