Learning Cognitive Map Representations for Navigation by Sensory-Motor Integration.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Central questions in navigation research involve transforming sensory-motor information into self-location memory and environmental map representations.
• Place cells in the rodent hippocampus are known to form dynamic cognitive maps of locations.

Purpose of the Study:

• To propose a biologically inspired neural network model, SeMINet, for learning cognitive maps.
• To integrate sensory and motor information for spatial representation in a virtual environment.

Main Methods:

• A deep neural network processes visual environmental features.
• A recurrent network of place units encodes spatial information via sensorimotor integration.
• A decoding network predicts agent location from spatial representations.

Main Results:

• The SeMINet model forms neural codes for location independent of head direction.
• Recurrent connections sustain network activity, mimicking agent motion.
• Competitive learning corrects path-integration errors and associates sensory input with place unit states.

Conclusions:

• SeMINet provides a brain-inspired model for cognitive map formation.
• The model updates spatial representations using both self-motion and visual cues.
• Reliable location decoding is achieved even with noisy movement data.