Learning place cells and remapping by decoding the cognitive map

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Hippocampal place cells encode location and form cognitive maps.
• Place cells exhibit remapping (rate changes) when environments change.
• Interactions between place, border, and grid cells remain unclear.

Purpose of the Study:

• To develop a normative computational model of place cell function and remapping.
• To investigate how neural networks can learn spatial representations and perform path integration.
• To explore potential mechanisms for interaction between different types of spatially tuned neurons.

Main Methods:

• A neural network model was developed for position reconstruction and path integration.
• A non-trainable decoding scheme was used to estimate position from network outputs.
• The model was trained in multiple simulated environments to observe remapping phenomena.

Main Results:

• Network output units developed place-like spatial representations.
• Upstream recurrent units became boundary-tuned.
• Place-like units demonstrated global, geometric, and rate remapping similar to biological cells.
• Place unit centers showed hexagonal lattice clustering, with preliminary evidence in mouse data.
• Remapping was supported by rate changes in upstream units.

Conclusions:

• The model provides a normative framework for understanding place cell field formation and remapping.
• The findings suggest a potential mechanism for interaction between place, border, and grid cells.
• This work offers new insights into the computational principles underlying spatial cognition.