Inferring learning rules from animal decision-making.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Understanding animal learning mechanisms is a key challenge in neuroscience.
• Reinforcement learning models often focus on artificial agents, not empirical animal behavior.
• Existing models struggle to fully explain the nuances of how animals acquire new skills.

Purpose of the Study:

• To develop a novel modeling framework for inferring empirical learning rules in animals.
• To decompose trial-to-trial policy changes into learning and noise components.
• To compare different learning rules and identify parameters influencing policy updates.

Main Methods:

• Developed a computational framework to analyze animal policy updates.
• Inferred learning rules and learning rates from behavioral data.
• Validated the model on simulated data and applied it to rodent perceptual decision-making tasks.

Main Results:

• The framework successfully identified key learning rules governing animal behavior.
• A modified reinforcement learning rule, including baseline parameters, explained 92% of policy updates in mice, significantly outperforming the conventional REINFORCE rule (30%).
• Inferred learning rates and baseline values suggest animal policy updates do not solely aim to maximize expected reward.

Conclusions:

• The developed framework provides a powerful tool for dissecting animal learning processes.
• Findings challenge the assumption that animal learning always optimizes for immediate reward.
• This research offers valuable insights into biological learning algorithms for both neuroscience and machine learning communities.