Related Experiment Videos
Combining Hebbian and reinforcement learning in a minibrain model.
R J C Bosman1, W A van Leeuwen, B Wemmenhove
1Institute for Theoretical Physics, University of Amsterdam, Valckenierstraat 65, 1018 XE Amsterdam, The Netherlands.
Summary
This study introduces a toy neural network model combining Hebbian and reinforcement learning. By balancing Hebbian and reinforcement terms, the model significantly reduces learning steps and mitigates memory loss, improving neural network efficiency.
Area of Science:
- Computational Neuroscience
- Artificial Intelligence
Background:
- Neural networks can suffer from 'path interference', leading to rapid forgetting of learned information.
- Combining Hebbian learning and reinforcement learning presents a potential solution to enhance memory retention.
Purpose of the Study:
- To investigate a toy model of a neural network incorporating both Hebbian and reinforcement learning mechanisms.
- To address the 'path interference' problem in neural networks.
Main Methods:
- A toy neural network model was developed.
- A modified learning rule was implemented, integrating a Hebbian term (proportional to learning rate nu) with a reinforcement term (proportional to delta).
Main Results:
- The addition of the Hebbian term effectively combats the 'path interference' issue.
- Optimal performance, indicated by a considerable reduction in learning steps, was observed when the ratio of the Hebbian term to the reinforcement term (nu/delta) was between 1/4 and 1/2.
Conclusions:
- Balancing Hebbian and reinforcement learning components is crucial for efficient neural network learning.
- The proposed model demonstrates a method to improve the stability and efficiency of learning in neural networks by carefully tuning learning rule parameters.