Predictive reward-prediction errors of climbing fiber inputs integrate modular reinforcement learning with supervised...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• The cerebellum is traditionally linked to supervised learning.
• Emerging evidence suggests its involvement in reward processing and reinforcement learning (RL).

Purpose of the Study:

• To investigate the cerebellum's role in executing RL algorithms.
• To understand the function of reward prediction errors within cerebellar circuits.

Main Methods:

• Utilized a Q-learning model to analyze mouse licking responses in a Go/No-go auditory-discrimination task.
• Performed tensor component analysis on calcium imaging data from over 6,000 Purkinje cells.
• Developed a spiking neural network model of the cerebellum.

Main Results:

• Identified distinct climbing fiber inputs activated by Go and No-go cues, inversely related to predictive reward errors.
• The neural network model replicated behavioral and neural changes during discrimination learning.
• Demonstrated that Purkinje cells develop context-specific motor commands guided by reward prediction errors.

Conclusions:

• The cerebellum contributes to modular reinforcement learning through context-specific computations.
• Findings suggest an integration of reinforcement and supervised learning mechanisms within the cerebellum.