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

Forward propagating reinforcement learning--biologically plausible learning method for multi-layer networks.

Masataka Watanabe1, Tomohiro Masuda, Kazuyuki Aihara

  • 1Department of Quantum Engineering and Systems Science, Graduate School of Engineering, The University of Tokyo, 7-3-1, Hongo Bunkyo-ku, Tokyo 113, Japan. watanabe@sk.q.t.u-tokyo.ac.jp

Bio Systems
|October 22, 2003
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

CauFinder: Steering Cell-State and Phenotype Transitions by Causal Disentanglement Learning.

Advanced science (Weinheim, Baden-Wurttemberg, Germany)·2026
Same author

Aberrant prefrontal activity and arousal level correlate with action initiation and response vigor in rats.

iScience·2026
Same author

Identifying the optimal rapid antigen test for screening and determining the end of isolation: A modeling study.

PLoS computational biology·2026
Same author

Force Learning in Balanced Cortical E-I Networks.

Neural computation·2026
Same author

Psychological distress among Japanese high school students during the COVID-19 pandemic: An energy landscape analysis.

PLoS medicine·2026
Same author

Stratification of viral shedding patterns in saliva of COVID-19 patients.

eLife·2026

This study presents a biologically plausible reinforcement learning method for neural networks. It localizes synaptic changes using inhibitory connections and bypass pathways for effective learning.

Area of Science:

  • Computational Neuroscience
  • Machine Learning
  • Artificial Intelligence

Background:

  • Reinforcement learning (RL) is crucial for decision-making.
  • Implementing RL in multi-layer neural networks faces challenges with biological plausibility.
  • Delayed reinforcement signals pose a significant hurdle for localized learning.

Purpose of the Study:

  • To propose a biologically plausible method for reinforcement learning in multi-layer neural networks.
  • To address the spatial localization of synaptic modulation by reinforcement signals.
  • To reconcile broadcast reinforcement signals with localized learning in neural systems.

Main Methods:

  • Introducing a novel method for synaptic modulation in neural networks.
  • Spatially localizing reinforcement learning signals downstream from initial to final layers.

Related Experiment Videos

  • Utilizing inhibitory backward connections and bypass pathways to output units.
  • Ensuring adherence to neurophysiological principles for localized delayed reinforcement.

    • Main Results:

    • Demonstrated a method for biologically plausible reinforcement learning.
    • Successfully localized synaptic modulation effects within multi-layer neural networks.
    • Showcased how inhibitory connections and bypass pathways enable localized learning from delayed reinforcement.

    Conclusions:

    • The proposed method offers a biologically plausible approach to reinforcement learning in neural networks.
    • This technique effectively addresses the challenge of localized synaptic plasticity with broadcast reinforcement signals.
    • The findings contribute to understanding neural computation and developing more sophisticated AI.