Self-Supervised Learning of Representations for Space Generates Multi-Modular Grid Cells

Rylan Schaeffer, Mikail Khona, Tzuhsuan Ma, Cristobal Eyzaguirre, Sanmi Koyejo, Ila Rani Fiete

Advances in Neural Information Processing Systems Accepted

December 2023

Neuroscience Grid Cells Self-Supervised Learning Representation Learning

PDF

Abstract

We show that self-supervised learning objectives applied to spatial representation learning generate multi-modular grid cell-like representations, providing a normative account of grid cell emergence.

Summary

Self-supervised learning on spatial tasks generates multi-modular grid cell-like representations.

Summary

This work investigates whether self-supervised learning objectives can explain the emergence of grid cells in the entorhinal cortex. We show that networks trained with self-supervised objectives on spatial tasks develop multi-modular grid cell-like representations.

Key Contributions:

1. Demonstration that self-supervised learning generates grid-like representations
2. The emergent representations exhibit multi-modularity, a key property of biological grid cells
3. Provides a normative account for why the brain might use grid-like representations