Emergent elasticity in the neural code for space

Samuel A. Ocko, Kiah Hardcastle, Lisa M. Giocomo, Surya Ganguli

TL;DR

This paper presents a novel computational model that explains how animals can build a consistent internal map of their environment by combining self-motion cues with landmark information. The model reveals an emergent "elasticity" in neural circuits, where learned landmark associations act like springs, pulling the internal map into alignment with the external world. This framework unifies disparate experimental observations about grid cell behavior and makes testable predictions about how these spatial maps are formed and maintained.

What Problem Does This Paper Address?

Navigating a novel environment requires animals to simultaneously build a map of that environment and determine their own location within it. This is a classic "chicken-and-egg" problem: you need a map to localize yourself, but you need to localize yourself to build a map. Neural circuits must integrate two key sources of information: self-motion cues (path integration) and sensory cues from landmarks. While we know that certain neurons (like grid cells and border cells) are involved in spatial representation, the precise mechanisms by which their activity patterns self-organize into a coherent, consistent spatial map, especially through learning and exploration, remain unclear. This paper aims to bridge this gap by proposing a biologically plausible model that explains how known neural dynamics and plasticity rules can lead to the emergence of such a consistent spatial map.

Methods

The authors develop an analytical theory based on a neural attractor network model. This model incorporates:

  • Attractor Network Dynamics: A network of neurons with short-range excitation and long-range inhibition, capable of generating stable, localized "bump" activity patterns. These patterns are parameterized by a phase variable, representing a manifold of possible states.
  • Path Integration: Velocity-tuned cells provide input to the attractor network, causing its phase to advance proportionally to the animal's velocity. This allows the network to track the animal's movement through space.
  • Landmark Cells: Sensory neurons that fire when the animal is in a specific location, providing input to the attractor network.
  • Hebbian Plasticity: Synaptic weights between landmark cells and the attractor network are modified based on the Hebbian learning rule. This means synapses are strengthened when the landmark cell and the attractor network are active simultaneously.
  • Model Reduction: The complex dynamics of the full neural network are reduced to a simpler, low-dimensional system. This allows for analytical insights into the emergent properties of the system. The key insight here is that the attractor network's phase can be treated as a single variable, and landmark cell influences can be represented by learned "pinning phases."

This framework is then generalized from 1D to 2D, and the learning dynamics are interpreted as an "elastic relaxation process" between landmark representations.

Datasets

Not discussed in the paper. The study is primarily theoretical and uses simulations to test its predictions.

Key Insights and Ideas

  • Emergent Elasticity: The core idea is that learned associations between landmarks and internal states act like springs. When the animal explores, these "springs" pull the internal spatial representation into alignment with the external environment, ensuring consistency between path integration and landmark cues.
  • Self-Organization of Spatial Maps: The model demonstrates how a consistent spatial map can emerge purely from the interaction of basic neural dynamics (attractor networks, path integration) and a simple learning rule (Hebbian plasticity) without requiring an external teacher signal or pre-existing map.
  • Unified Explanation for Grid Cell Behavior: The model provides a single framework that can explain several previously disparate experimental findings:
  • Path-dependent shifts: Grid cell firing fields can shift slightly depending on the animal's recent trajectory, even in a learned environment.
  • Deformations in irregular environments: Grid cell patterns can bend and deform to conform to the shape of the environment, akin to elastic solids.
  • Topological defects: In specific environments, the model predicts the possibility of "topological defects" in grid patterns, where traversing the same path in different directions can lead to different phase shifts.
  • Landmark Cells as "Particles": In a simplified view, landmark cells' learned associations can be thought of as particles in physical space connected by damped springs. The learning process is then an elastic relaxation that settles these particles into positions consistent with the environment's geometry.

Key Results

  • The model analytically demonstrates how path integration and Hebbian learning of landmark associations can lead to a self-consistent neural map of space.
  • The model predicts systematic path-dependent shifts in grid cell firing fields, which are experimentally validated (shifts of ~1.5-3 cm observed).
  • The model reproduces observed deformations of grid cell firing fields in irregular environments.
  • The model predicts the possibility of topological defects in grid cell firing patterns under specific environmental conditions.
  • The framework provides a mechanistic explanation for how neuronal dynamics and synaptic plasticity sculpt spatial representations.

Limitations and Future Work

  • Simplifications: The model uses linearized dynamics and simplified representations (e.g., single variables for landmark states) to gain analytical tractability. The authors acknowledge that the full, non-linear dynamics might exhibit richer behaviors.
  • Experimental Validation: While the model makes several testable predictions that are supported by existing data, further experimental validation of specific predictions (like topological defects) is needed.
  • Heterogeneity: The authors suggest exploring how neural heterogeneity (observed in real brain regions like MEC) might influence the model's behavior.
  • Reliability of Cues: Future work could investigate how changes in the reliability of sensory and self-motion cues might be differentially regulated by higher-order mechanisms.
  • Generalization: The framework is a starting point for understanding how environmental geometry and exploration statistics sculpt neural representations, opening avenues for more complex scenarios.

Summary updated

Abstract

Upon encountering a novel environment, an animal must construct a consistent environmental map, as well as an internal estimate of its position within that map, by combining information from two distinct sources: self-motion cues and sensory landmark cues. How do known aspects of neural circuit dynamics and synaptic plasticity conspire to accomplish this feat? Here we show analytically how a neural attractor model that combines path integration of self-motion cues with Hebbian plasticity in synaptic weights from landmark cells can self-organize a consistent map of space as the animal explores an environment. Intriguingly, the emergence of this map can be understood as an elastic relaxation process between landmark cells mediated by the attractor network. Moreover, our model makes several experimentally testable predictions, including (<i>i</i>) systematic path-dependent shifts in the firing fields of grid cells toward the most recently encountered landmark, even in a fully learned environment; (<i>ii</i>) systematic deformations in the firing fields of grid cells in irregular environments, akin to elastic deformations of solids forced into irregular containers; and (<i>iii</i>) the creation of topological defects in grid cell firing patterns through specific environmental manipulations. Taken together, our results conceptually link known aspects of neurons and synapses to an emergent solution of a fundamental computational problem in navigation, while providing a unified account of disparate experimental observations.
BibTeX
@inproceedings{ocko2018emergent,
  author = {Samuel A. Ocko and Kiah Hardcastle and Lisa M. Giocomo and Surya Ganguli},
  title = {{Emergent elasticity in the neural code for space}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2018},
  doi = {10.1073/pnas.1805959115},
}

Loading related content...

Figure