Modeling Interactions of the Rat's Place and Head Direction Systems

A. David Redish, David S. TouretzkyNeurIPS 1995

Introduction

When a rat enters a familiar room, it must align its internal spatial representations with the physical world. It relies on two primary sources of information: path integration (an internal "dead reckoning" system based on self-motion) and visual landmarks (external cues like a white card on a wall).

The challenge arises when these sources conflict. For example, if a rat is placed in an arena where the landmarks have been rotated or duplicated, how does its brain decide where it is? Previous experiments have shown complex, sometimes counter-intuitive behaviors—sometimes the rat’s internal map rotates with the cues, and sometimes it stays fixed. This paper introduces the CRAWL model, a computational framework designed to explain how rodents reconcile internal "maps" (place cells) with directional "compasses" (head direction cells) through a hierarchy of reset priorities.

The CRAWL Model Architecture

The authors propose the CRAWL model, which integrates four distinct spatial modules to simulate navigation and spatial awareness:

CRAWL Model Architecture
  1. Local View: Processes egocentric visual information, such as the distance and bearing to landmarks.
  2. Head Direction (HD) System: Acts as an internal compass representing the animal's current heading.
  3. Path Integrator (PI): Tracks the animal’s coordinates $(x, y)$ based on self-motion.
  4. Place Code: The "map" where neurons (place cells) fire based on a combination of visual input and path integration.

The core of the model relies on a consistency metric ($\kappa$). When a rat enters an environment, the model evaluates four possible outcomes to resolve conflicts between internal estimates and visual cues:

  1. Retain both initial HD and PI estimates.
  2. Reset HD (trust the landmarks for direction).
  3. Reset PI (trust the landmarks for location).
  4. Reset both.

The model uses a prioritized logic: it prefers to keep its initial estimates unless the inconsistency with visual cues exceeds a specific threshold ($\kappa^*$). If a reset is necessary, it prioritizes resetting the head direction over the path integrator.

Key Insights and Results

The model was benchmarked against classical experimental results involving rats navigating a cylindrical arena with white cue cards.

Simulated Place Field
  • Replication of Biological Data: The model accurately simulated biological observations, including cases where place fields remained stable despite doubled cue cards, and cases where they rotated 180° when the entry point was changed.
  • The Power of Entry Points: The model highlights that the entry point acts as a powerful anchor for the path integrator. Rats develop a strong expectation of their starting coordinates based on their entry location.
  • Priority Hierarchy: The brain does not simply average conflicting signals; it follows a "reset hierarchy." Resetting the internal compass (HD) takes priority over resetting the internal map (PI) when landmarks shift.
  • Quantitative Threshold: The authors empirically derived a consistency constant ($\kappa^* = 0.04$) that allows the model to match the behavioral switching points seen in real rats.
  • Predictive Power: The model successfully predicts scenarios not previously tested, such as how place fields would react to two rotated cue cards when entering from a specific direction.

Limitations and Future Work

While the CRAWL model provides a robust framework for understanding spatial alignment, it has limitations:

  • Static Tuning: The current model assumes the underlying associations between landmarks and locations are fixed. It does not account for long-term learning or "remapping," where place cell tuning changes over multiple sessions.
  • Mental State: Variations in an animal's attention or "mental state"—which might cause it to ignore certain cues—are not yet captured.
  • Future Experiments: The authors suggest future work should involve simultaneous recording from both place cells and head direction cells to verify the model's prediction that the HD system remains stable even when the place code shifts in specific single-cue scenarios.

Summary updated

BibTeX
@inproceedings{redish1995modeling,
  author = {A. David Redish and David S. Touretzky},
  title = {{Modeling Interactions of the Rat's Place and Head Direction Systems}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1995},
}
Figure