Reward-modulated Hebbian learning of decision making.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Decision Theory

Background:

• Traditional decision-making models often lack biological plausibility.
• Efficient learning mechanisms in neural systems are not fully understood.
• Hebbian learning offers a simple, biologically inspired learning rule.

Purpose of the Study:

• To introduce a novel framework for decision making based on biologically plausible learning.
• To reduce decision-making learning to Hebbian learning on a linear neuron.
• To investigate a reward-modulated Bayesian-Hebbian learning rule.

Main Methods:

• Formulating a Bayesian-Hebbian learning rule as a reinforcement learning problem.
• Proving exponential convergence of synaptic weights to log-odds of reward.
• Implementing a winner-take-all operation for action selection.
• Modulating synaptic updates with a global reward signal.

Main Results:

• Synaptic weights converge exponentially fast to the log-odds of reward.
• The learning rule is a pure Hebb update, dependent on coactivation.
• The framework enables fast learning of decisions.
• The approach bypasses complex calculations of weighted inputs.

Conclusions:

• The proposed Bayesian-Hebbian learning framework offers a simple and biologically plausible model for decision making.
• This model provides new hypotheses for neural computation in cortical areas involved in decision making.
• Fast learning of decisions is achievable through reward-modulated Hebbian plasticity.