Source-linked AI summary
Accurate path integration in continuous attractor network models of grid cells
Yoram Burak, Ila R. Fiete
TL;DR
Previous grid-cell models accumulated velocity-integration errors rapidly, raising questions about their dead-reckoning capability. This paper analyzes continuous attractor networks using velocity and heading inputs, finding accurate path integration across periodic and aperiodic networks over ranges dependent on network size, organization, and noise.
Problem
Previous models accumulated velocity-integration errors rapidly, while accurate grid-cell activity requires precise velocity integration over behavioral trajectories.
Method
The paper analyzes continuous attractor networks using velocity and heading inputs while comparing boundary conditions, network size, and stochasticity.
Results
1–10 minutes and 10–100 meters are the maximum attainable ranges of accurate path integration across biologically plausible parameters.
Takeaways & Limitations
Continuous attractor dynamics may underlie velocity integration in dMEC, and the simulations provide upper bounds for attainable integration accuracy.
Takeaways & Limitations
The paper notes that previous theta-oscillation models fail to produce accurate spatial grids over behavioral dead-reckoning timescales when oscillations deviate from pure sine waves.
Abstract
from arXiv · showhide
Grid cells in the rodent entorhinal cortex display strikingly regular firing responses to the animal's position in 2-D space, and have been hypothesized to form the neural substrate for dead-reckoning. However, in previous models suggested for grid cell activity, errors accumulate rapidly in the integration of velocity inputs. To produce grid-cell like responses, these models would require frequent resets triggered by external sensory cues, casting doubt on the dead-reckoning potential of the grid cell system. Here we focus on the accuracy of path integration in continuous attractor models of grid cell activity. We show, in contrast to previous models, that continuous attractor models can generate regular triangular grid responses, based on inputs that encode only the rat's velocity and heading. We consider the role of the network boundary in integration performance, and show that both periodic and aperiodic networks are capable of accurate path integration, despite important differences in their attractor manifolds. We show that the rate at which errors accumulate in the integration of velocity depends on the network's organization and size, and on the intrinsic noise within the network. With a plausible range of parameters and the inclusion of spike variability, our model can accurately integrate velocity inputs over a maximum of ~10-100 m and ~1-10 minutes. These findings form a proof-of-concept that continuous attractor dynamics may underlie velocity integration in dMEC. The simulations also generate pertinent upper bounds on the accuracy of integration that may be achieved by continuous attractor dynamics in the grid cell network. We suggest experiments to test the continuous attractor model and differentiate it from models in which single cells establish their responses independently of each other.
Author Summary
The paper proposes a continuous-attractor network that integrates velocity and heading inputs to produce position-dependent, grid-cell-like responses. It argues that this neural computation can support accurate path integration and motivates experiments to test the mechanism in rats.
- Grid cells fire when the rat occupies vertices of a triangular spatial grid, linking their activity to dead reckoning.
- The model receives information about the rat’s velocity and heading, then integrates these inputs to produce responses dependent on absolute position.This provides a network-level account of path integration rather than independent position responses.
- The proposed neural network can integrate position accurately while reproducing grid-cell-like responses similar to experimental observations.
- The authors suggest experiments to determine whether this mechanism contributes to grid-cell emergence and path integration in the rat brain.
Introduction
The introduction frames a gap in path-integration models: previous mechanisms accumulate errors or rely on unrealistic assumptions, motivating analysis of continuous attractor networks. The paper examines how topology, size, and stochasticity constrain integration accuracy and reports accurate integration over biologically plausible behavioral ranges.
- Behavioral benchmark: Grid-cell tuning remains relatively accurate over ∼100 meters and ∼10 minutes in darkness, although olfactory and tactile cues were not eliminated.
- Research question: The paper asks whether continuous attractor networks can accurately estimate position from velocity cues without depending on nearly continuous external corrections.
- Motivation and prior models: Previous grid-cell models either integrate velocity inaccurately or require assumptions that limit their ability to maintain stable spatial patterns.Reported problems include sensitivity to oscillation phase, poor integration, absent grid patterns, or continuously modulated recurrent weights.
- Approach: The analysis compares periodic and aperiodic recurrent networks while varying boundary conditions, network size, and stochasticity in neural firing.
- Main result: 1–10 minutes and 10–100 meters are the maximum attainable ranges of accurate path integration across biologically plausible parameters.Larger, less noisy networks occupy the high end of the range, whereas smaller, more stochastic networks occupy the low end.
- Implications: The paper proposes experiments to quantify integration accuracy, test the continuous-attractor hypothesis, and distinguish recurrent-network mechanisms from single-cell computations.
Results
Continuous attractor networks generated coherent triangular grid responses and accurately integrated velocity inputs in both periodic and aperiodic configurations. Accuracy depended on boundary design, network size, and spiking variability, while noise caused diffusive drift and limited integration duration.
- The population response formed discrete activity blobs arranged on a regular triangular lattice, and pattern flow produced periodic intervals in single-neuron responses.
- < 15 cm total error accumulated over ∼260 m and 20 minutes in the periodic network, versus a grid period of about 48 cm.The corresponding average errors were less than 0.1 cm per meter and less than 0.01 cm per second over input speeds of 0-1 m/s.
- Coherent single-neuron grids were equivalent to accurate full-trajectory path integration, with linear network-flow velocity across input directions.These equivalent conditions applied to both periodic and aperiodic networks.
- Aperiodic networks: Aperiodic networks achieved periodic-case accuracy when boundary activity faded smoothly, whereas sharper fading caused pinning, rotation, and loss of single-neuron grids.Tapered input profiles dramatically improved performance by reducing boundary influence on network dynamics.
Predictions of the attractor model
The attractor model yields experimentally testable differences between periodic and aperiodic networks and between continuous-attractor and independent-neuron models. These predictions concern phase stability, boundary firing, defects, perturbation recovery, and responses to environmental changes.
- Phase stability: Periodic networks should preserve phase relationships over days, whereas aperiodic networks may rotate and preserve them for only 1-10 minutes.Periodic population-pattern rotations are forbidden, while aperiodic networks can rotate on minute-to-tens-of-minutes timescales.
- Comparison with independent-neuron models: Independent-neuron models predict phase relationships that drift or random-walk on the same short timescale as individual-cell phase drift.Continuous-attractor network interactions tether the phases of different neurons.
- Perturbation tests: Perturbations without an appreciable attractor-manifold component should be restored by network interactions, leaving absolute phases and phase relationships unchanged.Incoherent perturbations can instead produce a coherent translation of the entire population pattern along the attractor manifold.
- Defects: A stable defect appearing in every single-neuron spatial response would provide strong evidence for a continuous attractor network.Periodic-network defects are expected to recur across cells up to a global phase shift.
- Environmental changes: After spatial stretching, unchanged phase relationships and velocity-modulation decreases matching the percentage stretch support the continuous-attractor model.Changes in phase relationships or no corresponding velocity-modulation change would count as evidence against it.
- Causal manipulation: Blocking dMEC spiking without blocking its inputs should abolish periodic spatial responsiveness, unlike models that generate grids independently in single cells.
- Boundary signatures: Aperiodic networks should show a wide distribution of maximal firing rates because boundary neurons receive fading input and are less active than bulk neurons.Roughly equal maximal rates across same-type cells would be inconsistent with an aperiodic network, although a wide distribution would not prove it.
Discussion
The paper argues that continuous attractor networks can support accurate grid-cell path integration under reasonable conditions, while identifying network size, topology, noise, and input assumptions that constrain performance.
- Main contributions: The work’s three contributions are modeling accurate grid-cell integration, estimating upper bounds for dMEC, and proposing falsifiable experiments.The predictions are intended to distinguish continuous-attractor networks from independent cell computations.
- Main contributions: Continuous attractor networks can generate grid-cell-like responses and accurately integrate velocity inputs.The model assumes accurate velocity inputs and focuses on how well the network integrates them.
- Accuracy limits: With neural noise, estimated integration accuracy is approximately 1–10 minutes; larger, less stochastic networks perform better.Noise-free large networks also have finite accuracy, while smaller, more stochastic networks integrate shorter paths consistent with behavioral scales.
- Assumptions and limitations: The model’s accuracy estimates exclude noise and biases in velocity inputs, which can further reduce the behavioral range of dead reckoning.Uncertainty about variability in dMEC connectivity also leaves the degree of attractor pinning unresolved.
- Network size: Accurate path integration requires many neurons per grid, implying fewer distinct grids than representational-capacity considerations would favor.The proposed network size of 10^3–10^4 neurons is broadly consistent with estimates for the entorhinal cortex.
- Network topology: Both periodic and aperiodic networks can integrate accurately, but aperiodic networks require fine-tuning and remain more sensitive to parameters.Periodic networks generally perform better because their population pattern cannot rotate.
Methods
The model uses rate-based neurons on a 2-D sheet with recurrent center-surround connectivity and velocity-modulated inputs to drive a triangular lattice pattern.
- Network dynamics: Rate-based neuron dynamics combine recurrent synaptic activation with feedforward inputs encoding the rat’s velocity.The neural response has a 10 ms time constant and uses a rectification nonlinearity.
- Network organization: Neurons are arranged on a 2-D sheet with uniformly tiled preferred directions, restricted to N, S, E, and W for convenience.The model notes that rat preferences might instead span the continuum [0, 2π].
- Recurrent connectivity: The recurrent weight matrix has a center-surround shape centered at a location shifted according to each neuron’s preferred direction.This connectivity assumes neurons are topographically arranged.
- Velocity-driven flow: With zero velocity coupling, the network generates a static triangular lattice pattern, whereas nonzero coupling drives pattern flow.The shift and velocity-gain parameters determine how strongly and how quickly velocity inputs move the pattern while preserving lattice stability for small shifts.
- Boundary and input modulation: The envelope function spatially scales inputs while preserving the lattice, and velocity inputs are envelope-modulated to maintain the same flow rate in faded regions.Local flow rate is determined by the velocity-modulated feedforward component divided by total feedforward input.
- Simulation protocol: The simulations compare periodic and aperiodic networks using velocity steps, real rat velocity trajectories, and spike trains with controlled variability.Pattern flow and lattice orientation are tracked to quantify integration and network-state changes.