Learning produces an orthogonalized state machine in the hippocampus.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Cognitive maps in the hippocampus enable flexible intelligence but their formation mechanisms are unclear.
• Understanding these mechanisms is crucial for advancing both biological and artificial intelligence.

Purpose of the Study:

• To investigate the algorithmic form and learning mechanisms of cognitive maps in the hippocampus.
• To elucidate how neural activity in the CA1 region represents and learns spatial relationships.

Main Methods:

• Large-scale, longitudinal two-photon calcium imaging in mice performing a virtual reality reward collection task.
• Recording activity from thousands of CA1 neurons across multiple learning stages.
• Utilizing computational modeling, including clone-structured causal graphs, to analyze neural dynamics.

Main Results:

• Hippocampal neural activity and animal behavior showed staged improvements mirroring task learning.
• Progressive decorrelation of neural activity led to orthogonalized representations resembling a state machine.
• Individual neurons developed task-state-specific responses ('state cells') driving this decorrelation.
• A clone-structured causal graph uniquely reproduced the observed learning trajectory and final neural states.

Conclusions:

• Cognitive map formation involves progressive decorrelation and orthogonalization of hippocampal neural representations.
• The emergence of 'state cells' and state-machine-like dynamics suggests hidden state inference as a core computational principle.
• Findings provide insights into hippocampal function and inform the development of more sophisticated artificial intelligence systems.