Place-cell capacity and volatility with grid-like inputs

Man Yi Yim, Lorenzo Sadun, Ila Fiete, Thibaud TaillefumiereLife 2021 · DOI

TL;DR

This paper investigates how the brain might generate stable spatial maps using grid cells and place cells. They model place cells as simple decision-making units (perceptrons) that receive input from grid cells. Their findings suggest that while grid cells provide a rich "scaffold" for spatial representation, the specific arrangement of place cell fields is highly constrained by this scaffold, leading to both robustness and potential volatility in spatial memory.

What Problem Does This Paper Address?

The brain needs to create stable and flexible representations of space, especially in environments where external cues might be sparse or unreliable. Hippocampal place cells are known to fire in specific locations (place fields), and these fields can change or multiply across different environments or within large spaces. A key question is what determines the possible arrangements of these place fields. Grid cells, with their periodic firing patterns, are a major input to place cells, and it's hypothesized that they provide an internal "scaffold" for spatial navigation. This paper asks: how much freedom do place cells have in arranging their fields given these structured grid cell inputs, and what are the consequences of this structure for spatial memory?

Methods

The authors model place cells as perceptrons. In machine learning, a perceptron is a simple model of a neuron that takes weighted inputs and outputs a binary decision (e.g., fire or not fire) based on a threshold. They use the known properties of grid cell inputs – their periodic nature and modular structure – as the input patterns for these perceptrons.

Their approach involves:

  • Analytical modeling: They use mathematical tools to derive theoretical predictions about the number of possible place field arrangements (repertoire) and the spatial range over which any arrangement can be realized (capacity).
  • Geometric interpretation: They leverage the structured nature of grid cell inputs to represent them as points in a geometric space (a polytope). This allows them to apply concepts from geometry and combinatorics to understand which arrangements are "linearly separable" by the perceptron, which corresponds to realizable place field configurations.
  • Comparison with alternative codes: They compare the properties of grid-like inputs to other types of input patterns (random, one-hot, binary) to highlight the unique contributions of grid cell structure.
  • Robustness analysis: They investigate how well these place field arrangements withstand noise in the inputs or weights.
  • Volatility analysis: They simulate how changes to the grid-place cell connections (plasticity) affect existing and new place fields.

Datasets

Not discussed in the paper. The study is theoretical and relies on simulated input patterns based on the known properties of grid cells.

Key Insights and Ideas

  • Grid cells provide a rich but constrained scaffold: Grid cells offer a vast number of unique input combinations, allowing for a large repertoire of potential place field arrangements. However, this repertoire is a tiny fraction of all possible arrangements, meaning the structure of grid inputs strongly dictates where and how place fields can form.
  • Capacity is limited by grid periods: The spatial range over which a place cell can achieve any field arrangement (its capacity) is surprisingly small, scaling with the sum of the grid cell periods, not the much larger range of unique grid input states. This implies that while the overall grid code is rich, the local "realizable" space for place cells is more restricted.
  • Robustness to noise: Place fields generated from grid-like inputs are significantly more robust to random noise in neural activity or weights compared to those generated from random or shuffled inputs. This suggests the structured nature of grid inputs provides stable representations.
  • Volatility due to plasticity: If the connections between grid cells and place cells are plastic (changeable), attempting to create a new place field can drastically alter existing fields or create new ones in uncontrolled locations. This "volatility" could explain observed changes in place cell representations over time.
  • Fixed scaffold vs. plastic connections: The findings suggest a trade-off: either the grid-to-place cell connections (the scaffold) are relatively fixed and non-plastic to maintain stability, or plasticity in these connections leads to volatile but potentially adaptable spatial maps.

Key Results

  • The number of realizable place field arrangements (repertoire) is very large, growing exponentially with the number of grid modules.
  • The spatial range over which any field arrangement can be realized (separating capacity) scales with the sum of grid periods, which is much smaller than the range of unique grid input states.
  • Grid-driven place fields exhibit significantly larger margins (a measure of robustness) compared to those driven by shuffled or random inputs.
  • Simulations show that modifying grid-place weights to insert a new field can lead to the appearance or disappearance of other fields, demonstrating high volatility.

Limitations and Future Work

  • Simplistic perceptron model: The model uses a simplified perceptron for place cells, not accounting for more complex dendritic integration or network interactions within the hippocampus.
  • Idealized grid cells: The model assumes ideal grid cell inputs without considering the biological variability or noise in grid cell firing.
  • Sparse vs. dense arrangements: While the paper analyzes ultra-sparse (few fields) and dense (all possible arrangements) cases, the intermediate regime of sparse but not ultra-sparse arrangements remains an open question.
  • Future work: The authors suggest exploring the dependence of robustness on graded tuning curves and phase resolution, investigating the role of dendritic nonlinearities, and analyzing how return projections from place cells to grid cells might affect the system's capacity. They also highlight the need to experimentally distinguish between plastic and fixed grid-to-place cell weights.

Summary updated

Abstract

What factors constrain the arrangement of the multiple fields of a place cell? By modeling place cells as perceptrons that act on multiscale periodic grid-cell inputs, we analytically enumerate a place cell's <i>repertoire</i> - how many field arrangements it can realize without external cues while its grid inputs are unique - and derive its <i>capacity</i> - the spatial range over which it can achieve any field arrangement. We show that the repertoire is very large and relatively noise-robust. However, the repertoire is a vanishing fraction of all arrangements, while capacity scales only as the sum of the grid periods so field arrangements are constrained over larger distances. Thus, grid-driven place field arrangements define a large response scaffold that is strongly constrained by its structured inputs. Finally, we show that altering grid-place weights to generate an arbitrary new place field strongly affects existing arrangements, which could explain the volatility of the place code.
BibTeX
@inproceedings{yim2021place,
  author = {Man Yi Yim and Lorenzo Sadun and Ila Fiete and Thibaud Taillefumier},
  title = {{Place-cell capacity and volatility with grid-like inputs}},
  booktitle = {eLife},
  year = {2021},
  doi = {10.7554/elife.62702},
}
Figure