An integrated deep learning-based model of spatial cells that combines self-motion with sensory information

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Spatial cells, including place cells, grid cells, and head direction (HD) cells, are crucial for building the brain's spatial map.
• Understanding the computational mechanisms underlying spatial cell function is a key challenge in neuroscience.

Purpose of the Study:

• To present a general deep learning-based modeling framework for understanding spatial cell responses.
• To explain the emergence of spatial cell activity, integrating path integration and visual cues.

Main Methods:

• Developed a deep learning model with layers for head direction (HD) cells and path integration (PI) using oscillatory neurons.
• Applied Principal Component Analysis (PCA) to analyze PI cell responses and identified grid-like periodicity.
• Utilized Bessel functions to describe the response properties of emergent grid-like cells.
• Employed a stack of autoencoders trained on PI layer outputs to model complex spatial representations.

Main Results:

• The model successfully generated neurons exhibiting responses characteristic of both grid cells and place cells.
• Emergent cells in the PI layer showed spatial periodicity consistent with grid cell firing patterns.
• The framework demonstrated the ability to model responses combining path integration and visual information.

Conclusions:

• The proposed deep learning framework provides a unified model for understanding the emergence of diverse spatial cell types.
• The model's architecture offers a potential explanation for how the brain constructs spatial representations.
• The framework has broader applicability for modeling neural systems beyond the simulated studies.