Inductive biases of neural network modularity in spatial navigation.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• The brain's modular architecture, with functionally specialized circuits, may optimize learning and generalization.
• Understanding this modularity could inform the design of more effective artificial systems.

Purpose of the Study:

• To test the hypothesis that modular neural architectures enhance learning and generalization compared to less specialized ones.
• To investigate the computational mechanisms underlying modular processing in a naturalistic task.

Main Methods:

• Trained reinforcement learning agents with diverse neural architectures on a naturalistic navigation task.
• Analyzed the agents' learning efficiency, generalization capabilities, and internal representations.

Main Results:

• A modular agent, segregating state representation, value, and action computations, demonstrated superior learning and generalization.
• The agent's state representation integrated predictive and observational information based on uncertainty, resembling Bayesian inference.
• The agent's behavior showed similarities to macaque behavior.

Conclusions:

• Modular neural architectures provide a potential advantage for learning and generalization in complex tasks.
• Insights from brain modularity can guide the development of more capable artificial intelligence systems.
• The study offers a computational perspective on the evolutionary rationale for brain modularity.