Flexible inference for animal learning rules using neural networks

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Understanding animal learning is crucial for neuroscience and artificial intelligence (AI) development.
• Existing methods often use fixed learning rules (e.g., Q-learning) that may not reflect complex animal behavior.
• A flexible, data-driven approach is needed to infer learning rules from observed behavior.

Purpose of the Study:

• To develop a framework for inferring animal learning rules directly from behavioral data during novel task learning.
• To model decision policies using generalized linear models (GLMs) and learning rules using deep neural networks (DNNs).
• To capture complex, multi-trial learning dynamics using recurrent neural networks (RNNs).

Main Methods:

• Developed a framework to infer learning rules from behavioral data during de novo task learning.
• Used a generalized linear model (GLM) for decision policies and a deep neural network (DNN) for learning rules.
• Introduced a recurrent neural network (RNN) variant to model multi-trial learning dependencies.

Main Results:

• Simulations successfully recovered ground-truth learning rules.
• DNN and RNN methods outperformed traditional reinforcement learning (RL) rules in predicting mouse learning trajectories on a sensory decision-making task.
• Inferred learning rules showed reward-history-dependent dynamics, with larger updates after rewarded trial sequences.

Conclusions:

• The proposed framework offers a flexible, data-driven method for inferring learning rules from behavioral data in novel learning tasks.
• These methods can enhance animal training protocols and contribute to the development of behavioral digital twins.
• The findings advance our understanding of animal learning mechanisms and their application in AI.