Gated recurrence enables simple and accurate sequence prediction in stochastic, changing, and structured environments

  1. Cédric Foucault
  2. Florent Meyniel  Is a corresponding author
  1. INSERM, CEA, Université Paris-Saclay, France
  2. CEA, Sorbonne Université, France

Abstract

From decision making to perception to language, predicting what is coming next is crucial. It is also challenging in stochastic, changing, and structured environments; yet the brain makes accurate predictions in many situations. What computational architecture could enable this feat? Bayesian inference makes optimal predictions but is prohibitively difficult to compute. Here, we show that a specific recurrent neural network architecture enables simple and accurate solutions in several environments. This architecture relies on three mechanisms: gating, lateral connections, and recurrent weight training. Like the optimal solution and the human brain, such networks develop internal representations of their changing environment (including estimates of the environment's latent variables and the precision of these estimates), leverage multiple levels of latent structure, and adapt their effective learning rate to changes without changing their connection weights. Being ubiquitous in the brain, gated recurrence could therefore serve as a generic building block to predict in real-life environments.

Data availability

This paper presents no experimental data. All synthetic data are available in the code repository at https://github.com/cedricfoucault/networks_for_sequence_prediction and archived on Zenodo with DOI: 10.5281/zenodo.5707498.

The following data sets were generated
    1. Foucault C
    (2021) Networks for sequence prediction
    Publicly available at Zenodo (https://zenodo.org/).

Article and author information

Author details

  1. Cédric Foucault

    INSERM, CEA, Université Paris-Saclay, Gif sur Yvette, France
    Competing interests
    The authors declare that no competing interests exist.
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7247-6927
  2. Florent Meyniel

    NeuroSpin, CEA, Sorbonne Université, Gif sur Yvette, France
    For correspondence
    florent.meyniel@cea.fr
    Competing interests
    The authors declare that no competing interests exist.
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6992-678X

Funding

Ecole normale superieure Paris-Saclay (PhD fellowship Contrat doctoral spécifique normalien"")

  • Cédric Foucault

Agence Nationale de la Recherche (18-CE37-0010-01 CONFI LEARN"")

  • Florent Meyniel

H2020 European Research Council (ERC StG 947105 NEURAL PROB"")

  • Florent Meyniel

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

Reviewing Editor

  1. Srdjan Ostojic, Ecole Normale Superieure Paris, France

Version history

  1. Preprint posted: May 3, 2021 (view preprint)
  2. Received: June 30, 2021
  3. Accepted: December 1, 2021
  4. Accepted Manuscript published: December 2, 2021 (version 1)
  5. Version of Record published: January 6, 2022 (version 2)
  6. Version of Record updated: January 21, 2022 (version 3)
  7. Version of Record updated: February 3, 2022 (version 4)

Copyright

© 2021, Foucault & Meyniel

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

  • 1,207
    views
  • 183
    downloads
  • 5
    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. Cédric Foucault
  2. Florent Meyniel
(2021)
Gated recurrence enables simple and accurate sequence prediction in stochastic, changing, and structured environments
eLife 10:e71801.
https://doi.org/10.7554/eLife.71801

Share this article

https://doi.org/10.7554/eLife.71801

Further reading

    1. Neuroscience
    Daniel Hoops, Robert Kyne ... Cecilia Flores
    Short Report

    Dopamine axons are the only axons known to grow during adolescence. Here, using rodent models, we examined how two proteins, Netrin-1 and its receptor, UNC5C, guide dopamine axons toward the prefrontal cortex and shape behaviour. We demonstrate in mice (Mus musculus) that dopamine axons reach the cortex through a transient gradient of Netrin-1-expressing cells – disrupting this gradient reroutes axons away from their target. Using a seasonal model (Siberian hamsters; Phodopus sungorus) we find that mesocortical dopamine development can be regulated by a natural environmental cue (daylength) in a sexually dimorphic manner – delayed in males, but advanced in females. The timings of dopamine axon growth and UNC5C expression are always phase-locked. Adolescence is an ill-defined, transitional period; we pinpoint neurodevelopmental markers underlying this period.

    1. Neuroscience
    Baba Yogesh, Georg B Keller
    Research Article

    Acetylcholine is released in visual cortex by axonal projections from the basal forebrain. The signals conveyed by these projections and their computational significance are still unclear. Using two-photon calcium imaging in behaving mice, we show that basal forebrain cholinergic axons in the mouse visual cortex provide a binary locomotion state signal. In these axons, we found no evidence of responses to visual stimuli or visuomotor prediction errors. While optogenetic activation of cholinergic axons in visual cortex in isolation did not drive local neuronal activity, when paired with visuomotor stimuli, it resulted in layer-specific increases of neuronal activity. Responses in layer 5 neurons to both top-down and bottom-up inputs were increased in amplitude and decreased in latency, whereas those in layer 2/3 neurons remained unchanged. Using opto- and chemogenetic manipulations of cholinergic activity, we found acetylcholine to underlie the locomotion-associated decorrelation of activity between neurons in both layer 2/3 and layer 5. Our results suggest that acetylcholine augments the responsiveness of layer 5 neurons to inputs from outside of the local network, possibly enabling faster switching between internal representations during locomotion.