A computational model for grid maps in neural populations

Fabio Anselmi, Benedetta Franceschiello, Micah M. Murray, Lorenzo Rosasco

TL;DR

This paper proposes a novel computational framework for how hexagonal grid cells in the brain might emerge from head direction cells. By leveraging modern Frame Theory and Hebbian learning, the model explains the observed hexagonal firing patterns and their remapping properties, offering a biologically plausible alternative to existing oscillatory and attractor models.

What Problem Does This Paper Address?

The brain's ability to represent an animal's position in space is a fundamental cognitive function. A key component of this spatial representation is found in the entorhinal cortex, specifically in "grid cells." These neurons exhibit a remarkable hexagonal firing pattern, forming a kind of internal coordinate system. While grid cells were discovered in 2005, a universally accepted model explaining how these precise hexagonal receptive fields emerge from neural activity has remained elusive. Existing models often rely on complex mechanisms like theta oscillations or specific synaptic connection symmetries, which have faced experimental challenges. This paper aims to bridge this gap by proposing a simpler, more biologically plausible model that explains the emergence of hexagonal grid maps from more basic neural inputs, like head direction cells.

Methods

The authors propose a two-phase learning process inspired by the Hubel and Wiesel model of visual cortex processing.

  • Phase 1: "Simple" HD Cells Learning: This phase models the emergence of head direction (HD) cells. Using a stationary input stimulus (analogous to natural images), the model learns receptive fields that are essentially Fourier components. This learning is driven by Oja's rule, a form of Hebbian learning that converges to principal components of the input. The key assumption here is that the underlying neuronal inputs have stationary second-order statistics.
  • Phase 2: "Complex" Grid Cell Learning: This phase explains how grid cells aggregate the activity of HD cells. The model posits that grid cells aim to encode the animal's position with maximal precision, which translates to minimizing the variance of their responses. This optimization problem is framed using modern Frame Theory, specifically focusing on equiangular frames. The core idea is that a population of HD cells, when aggregated by a "complex" grid cell, can form an optimal representation if their underlying frequency vectors form an equiangular frame. This aggregation is achieved through a gradient descent strategy with shrinkage to ensure sparsity in the selection of HD cells.

Datasets

Not discussed in the paper. The model is primarily theoretical and uses simulated inputs based on statistical properties of natural images and transformations.

Key Insights and Ideas

  • Frame Theory for Neural Coding: The paper ingeniously applies the mathematical framework of equiangular frames to explain the optimality of hexagonal grid cell receptive fields. This provides a novel perspective on why this specific geometric pattern is favored for spatial encoding.
  • Minimal Variance Encoding: The core principle driving the model is the brain's likely objective to encode spatial information with minimal variance (maximal precision). This principle, combined with Hebbian learning, naturally leads to the observed grid-like structures.
  • Emergence from Head Direction Cells: The model demonstrates a plausible pathway for grid cells to emerge from the activity of head direction cells, a more fundamental neural signal for orientation.
  • Biological Plausibility: By relying on Hebbian learning and avoiding the need for specific oscillatory dynamics or unrealistic synaptic symmetries, the model offers a more biologically grounded explanation for grid cell function.
  • Generalizability to Conceptual Spaces: The framework extends beyond purely spatial representations, suggesting that similar encoding principles could underlie "conceptual" spaces, as hinted at by experimental findings in [22].

Key Results

The model successfully predicts:

  • The emergence of hexagonal grid-like response patterns from the interference of planar waves, which are derived from the learned receptive fields of HD cells.
  • The alignment of axes in neighboring grid cells, a phenomenon explained by the shared orientation of their underlying frequency vectors.
  • The remapping of grid cells in response to environmental cue transformations (rotations, scaling), which is naturally explained by the corresponding transformations of the frequency vectors.
  • The optimality of the hexagonal grid for spatial encoding, as demonstrated through the minimization of the Fisher Information matrix, which is achieved when the underlying frequency vectors form an equiangular frame (specifically, a Mercedes-Benz frame for N=3 neurons).

Limitations and Future Work

  • Experimental Validation: While the model is theoretically sound and explains existing experimental findings, direct experimental validation of the proposed two-phase learning process and the role of equiangular frames in neural circuits would be beneficial.
  • Environmental Geometry Adaptation: The paper acknowledges that the model predicts a consistent hexagonal grid, but experimental evidence shows that grid receptive fields can distort to adapt to different environmental geometries. Extending the model to account for this dynamic adaptation is a key area for future work.
  • Higher Dimensions and Other Groups: The authors suggest that the framework can be generalized to higher dimensions and other group transformations (e.g., rotations), predicting grid-like geometries associated with Platonic Solids. Exploring these extensions and their experimental implications is a promising avenue.
  • Conceptual Encoding Details: While the paper outlines how the framework can apply to conceptual encoding, further detailed modeling and experimental testing are needed to fully elucidate this aspect.

Summary updated

Abstract

Grid cells in the entorhinal cortex, together with head direction, place, speed and border cells, are major contributors to the organization of spatial representations in the brain. In this work we introduce a novel theoretical and algorithmic framework able to explain the emergence of hexagonal grid-like response patterns from head direction cells' responses. We show that this pattern is a result of minimal variance encoding of neurons. The novelty lies into the formulation of the encoding problem through the modern Frame Theory language, specifically that of equiangular Frames, providing new insights about the optimality of hexagonal grid receptive fields. The model proposed overcomes some crucial limitations of the current attractor and oscillatory models. It is based on the well-accepted and tested hypothesis of Hebbian learning, providing a simplified cortical-based framework that does not require the presence of theta velocity-driven oscillations (oscillatory model) or translational symmetries in the synaptic connections (attractor model). We moreover demonstrate that the proposed encoding mechanism naturally explains axis alignment of neighbor grid cells and maps shifts, rotations and scaling of the stimuli onto the shape of grid cells' receptive fields, giving a straightforward explanation of the experimental evidence of grid cells remapping under transformations of environmental cues.
BibTeX
@article{anselmi2019computational,
  author = {Fabio Anselmi and Benedetta Franceschiello and Micah M. Murray and Lorenzo Rosasco},
  title = {{A computational model for grid maps in neural populations}},
  journal = {Journal of Computational Neuroscience},
  year = {2019},
  doi = {10.1007/s10827-020-00742-9},
}
Figure