A Multi-Region Brain Model to Elucidate the Role of Hippocampus in Spatially Embedded Decision-Making

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Brains demonstrate remarkable data-efficient learning and robust decision-making.
• Understanding neural architectures can guide the development of more effective deep learning models.
• Structured memory circuits play a key role in spatial decision-making tasks.

Purpose of the Study:

• To explore the normative role of structured memory circuits in spatial decision-making.
• To compare learning performance and neural representations of reinforcement learning (RL) agents with different brain model architectures.
• To identify brain-inspired architectures that enhance RL efficiency.

Main Methods:

• Development of a multi-region brain model incorporating grid cells, place cells, and a recurrent neural network for action selection.
• Counterfactual comparison of RL agents with various interaction architectures between entorhinal cortex and hippocampal components.
• Analysis of learning efficiency and neural representations under different architectural configurations.

Main Results:

• A specific architecture, where grid cells jointly encode self-movement velocity and decision evidence, significantly optimizes learning efficiency.
• This optimized architecture better reproduces experimental observations compared to alternative models.
• The study suggests that structured, brain-inspired architectures can lead to more efficient RL.

Conclusions:

• Brain-inspired structured architectures offer a promising direction for developing efficient reinforcement learning systems.
• The findings predict that grid cells must conjunctively encode position and evidence for effective spatial decision-making.
• This research motivates new neurophysiological experiments to investigate information flow in the entorhinal-hippocampal-neocortical circuit.