Learning flexible sensori-motor mappings in a complex network.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• The brain's complex structure necessitates understanding how synaptic plasticity supports continuous learning and forgetting of associations.
• Visuomotor association tasks present a significant challenge for modeling learning due to dynamic environmental changes.

Purpose of the Study:

• To investigate how different reinforcement learning rules in a multilayer network can explain learning and forgetting in a dynamic visuomotor association task.
• To determine the specific conditions of synaptic plasticity required to accurately model monkey behavior.

Main Methods:

• Simulated a multilayer neural network using various reinforcement learning rules.
• Modeled monkey behavior in a visuomotor association task.
• Analyzed synaptic modification rules, focusing on pre- and postsynaptic activity dependence and intrinsic stochasticity.

Main Results:

• Model performance matched monkey learning only when synaptic modifications depended on pre- and postsynaptic activity and stochasticity was low.
• The favored learning rule, reward-modulated Hebbian synaptic plasticity, demonstrated robust performance.
• Network performance did not significantly degrade with increased network layers, even for complex tasks.

Conclusions:

• Reward-modulated Hebbian synaptic plasticity with low stochasticity is crucial for explaining learning and forgetting of associations in complex, dynamic environments.
• This plasticity mechanism offers a scalable solution for learning in deep neural networks, mirroring brain function.