Confidence-Controlled Hebbian Learning Efficiently Extracts Category Membership From Stimuli Encoded in View of a...

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Perceptual decision-making involves learning categorization tasks via trial-and-error.
• Attractor neural networks are models for decision-making dynamics and behavioral confidence.
• Reward-modulated Hebbian learning (RMHL) is a potential mechanism for learning in neural networks.

Purpose of the Study:

• To investigate if reward-modulated Hebbian learning (RMHL) can efficiently learn categorization tasks.
• To propose and evaluate a novel confidence-controlled, reward-based Hebbian learning (CC-RHL) rule for improved learning.
• To determine if CC-RHL approximates gradient descent on a reward maximization cost function.

Main Methods:

• Simulated a decision-making attractor network with a stimulus-encoding layer.
• Implemented and tested RMHL with optimized coding layers.
• Developed and applied CC-RHL, using network confidence to modulate learning rate.
• Analyzed the learning rule's locality and performance compared to RMHL.

Main Results:

• RMHL failed to efficiently extract category membership from an optimized coding layer.
• CC-RHL demonstrated efficient extraction of categorical information.
• CC-RHL is a local learning rule and does not require storing past rewards.
• CC-RHL achieved near-optimal performance and approximated gradient descent.

Conclusions:

• Confidence-based modulation is crucial for efficient Hebbian learning in optimized neural networks.
• CC-RHL offers an efficient, local, and near-optimal learning mechanism for categorization tasks.
• The findings suggest a biologically plausible learning rule for decision-making systems.