Rapid path planning and preplay in maze{like environments using attractor networks

Dane Corneil, Wulfram GerstnerInfoscience 2015

Abstract

Biological organisms exhibit remarkable "one-shot learning," navigating to reward locations after a single visit. This capability suggests a flexible "cognitive map" in the hippocampus. We present a computational model using a Continuous Attractor Neural Network (CANN) that achieves rapid path planning and replicates "preplay" sequences. By incorporating multi-scale, geodesic place fields, the network inherently respects environmental geometry, allowing neural activity to flow around obstacles toward a goal.

Introduction

The hippocampal formation is widely regarded as the brain's internal GPS. A key challenge in neuroscience is understanding how this system supports arbitrary goal encoding and planning in complex, maze-like environments. Experimental observations of "preplay"—the sequential firing of place cells representing a future trajectory—suggest that the brain pre-calculates paths before movement. However, standard neural models often fail to account for physical barriers. This work addresses how these sequences emerge and navigate around obstacles.

Methodology: Multi-scale Geodesic Attractors

The proposed model utilizes a Continuous Attractor Neural Network (CANN), or "bump attractor," to represent spatial position. To enable long-range planning, the authors incorporate several key features:

  • Multi-scale Place Fields: The network integrates large place fields (typical of the ventral hippocampus) with smaller, more precise fields (dorsal hippocampus). Large fields provide the necessary connectivity for distant goal signals to influence local activity.
  • Geodesic Constraints: Place fields are defined using geodesic distances rather than Euclidean ones. This ensures that neural connectivity respects the maze's topology, preventing activity from "bleeding" through walls.
  • Pseudo-Inverse Weighting: Recurrent weights are determined via the pseudo-inverse method, allowing for an accurate representation of the spatial manifold.
  • Goal-Directed Stimulation: Path planning is initiated by stimulating neurons at a target location. This creates a gradient that pulls the activity "bump" from the current position toward the goal.

Results and Insights

The model demonstrates robust pathfinding capabilities across various maze geometries.

Path planning and activity propagation. The vector fields (left and right) indicate the directional bias toward the goal (red dot). The central panels show the temporal evolution of the activity bump as it flows around obstacles.

The Role of Scale and Geometry

A critical insight is that large place fields act as a bridge, allowing top-down goal signals to guide local sequences. Because these fields are geodesic, the resulting activity bump follows the shortest valid path around obstacles rather than a straight line. Furthermore, the model demonstrates that a single network can store and navigate multiple distinct environments (a "multichart" attractor), mirroring the hippocampus's ability to represent different contexts.

Replication of Preplay Sequences

The network successfully replicates the sequential firing patterns observed in rodent hippocampal recordings. These sequences represent a "mental rehearsal" of the path to be taken, occurring before the animal physically moves.

Raster plots showing sequential firing of neural units over time, representing the 'preplay' of a future trajectory.

Discussion

This work provides a biologically plausible mechanism for rapid path planning. A significant prediction of the model is that inactivating regions with large place fields (intermediate/ventral hippocampus) should impair long-distance pathfinding and abolish long-range preplay in the dorsal hippocampus.

While the pseudo-inverse method for setting weights is mathematically efficient, future research should explore biologically grounded learning rules, such as Spike-Timing-Dependent Plasticity (STDP), to explain how these networks are formed through experience. Additionally, extending the model to dynamic environments where obstacles change position remains an open challenge.

Summary updated

BibTeX
@inproceedings{corneil2015rapid,
  author = {Dane Corneil and Wulfram Gerstner},
  title = {{Rapid path planning and preplay in maze\{like environments using attractor networks}},
  booktitle = {Infoscience},
  year = {2015},
}
Figure