Contracting model of the basal ganglia

Benoît Girard, Nicolas Tabareau, Jean-Jacques Slotine, Alain Berthoz2005

TL;DR

This paper introduces a novel computational model of the basal ganglia that improves upon existing action selection models by incorporating a previously overlooked neural projection and leveraging contraction analysis for guaranteed stability. The resulting model demonstrates enhanced "perfect selection" capabilities, meaning it can more reliably and precisely choose one action from many competing options, even under challenging input conditions.

What Problem Does This Paper Address?

The basal ganglia are a group of subcortical nuclei in the brain crucial for selecting appropriate actions from a multitude of possibilities. Existing computational models, like the GPR model, have successfully captured some aspects of this selection process. However, these models often lack formal guarantees of stability and can exhibit undesirable behaviors, such as failing to select a single action when multiple competing inputs have similar strengths. This paper aims to address these limitations by developing a more robust and theoretically sound model that can reliably perform "perfect selection" – where one action is clearly chosen and others are suppressed.

Methods

The core of this paper's contribution lies in two key methodological advancements:

  • Incorporation of a Neglected Neural Projection: The authors include a projection from the external globus pallidus (GPe) to the striatum, a connection that was omitted in previous influential models. This addition is motivated by recent neuroscientific findings and is hypothesized to enhance the model's efficiency and selectivity.
  • Contraction Analysis for Stability Guarantees: To formally prove the stability and reliable convergence of their model, the researchers employ "contraction analysis." This theoretical framework allows them to analyze the dynamic behavior of nonlinear systems without relying on linearization. By demonstrating that their model is "contracting," they can guarantee that its trajectories will exponentially converge to a stable equilibrium point, regardless of initial conditions or minor perturbations. This is a significant improvement over models that only analyze equilibrium points without guaranteeing their reachability.

Datasets

Not discussed in the paper. The authors use simulated data generated by their model and compare it to simulation results from the GPR model.

Key Insights and Ideas

  • The Power of Underappreciated Connections: The inclusion of the GPe-to-striatum projection, though seemingly a minor detail, significantly impacts the model's performance, leading to sharper selection and reduced variability in the inhibition of unselected actions. This highlights how even subtle anatomical details can have profound functional consequences in neural circuits.
  • Contraction Analysis as a Rigorous Tool: The application of contraction analysis provides a strong theoretical foundation for the model's stability. This approach moves beyond simply observing behavior to formally proving its predictable and reliable convergence, a crucial step for building trustworthy computational neuroscience models.
  • "Perfect Selection" as a Measurable Goal: The paper defines and rigorously tests for "perfect selection," where the output inhibition of the selected channel is zero and others are fully inhibited. This precise definition allows for clear quantitative evaluation of the model's performance.
  • Biomimetic Enhancements: The model's ability to handle situations with similar input saliences more effectively than previous models suggests a more accurate representation of how the brain might arbitrate between competing actions.

Key Results

  • The proposed model demonstrates "perfect selection" in the standard GPR test scenario, achieving a more consistent and sharper selection than the GPR model.
  • In a test involving 1000 random salience vectors, the model achieved "perfect selection" in all but a few rare cases where input saliences were too low to overcome neural thresholds. This is a significant improvement over the GPR model, which failed "perfect selection" in a much higher percentage of cases.
  • The model exhibits enhanced contrast enhancement, effectively sorting out the winning channel even when input saliences are close. The limit of discrimination between two saliences is estimated to be very low, potentially a few percent.
  • The contraction analysis proved that the model is stable and always converges to an equilibrium point, regardless of the number of competing channels.

Limitations and Future Work

  • Simplified Neuron Models: The model uses leaky-integrator neuron models, which may not fully capture the complex dynamics of biological neurons, particularly the "up/down state" behavior observed in striatal neurons.
  • Omitted Connections: The authors acknowledge omitting certain documented connections, such as STN projections to the striatum and D1 striatal projections to GPe, which could further refine the model.
  • GPe-Striatum Projection Nuances: The model assumes a channel-to-channel selectivity for GPe-striatum projections, while some evidence suggests these projections might primarily target GABAergic interneurons, implying a more diffuse regulatory effect. This warrants further investigation.
  • Future Work: The authors plan to explore the impact of the omitted connections, investigate alternative models for GPe-striatum projections, and extend the model to include thalamo-cortical components to further enhance its selection capabilities, aiming to preserve its contraction properties.

Summary updated

Abstract

It is thought that one role of the basal ganglia is to constitute the neural substrate of action selection. We propose here a modification of the action selection model of the basal ganglia of (Gurney et al., 2001a,b) so as to improve its dynamical features. The dynamic behaviour of this new model is assessed by using the theoretical tool of contraction analysis. We simulate the model in the standard test defined in (Gurney et al., 2001b) and also show that it performs perfect selection when presented a thousand successive random entries. From a biomimetical point of view, our model takes into account a usually neglected projection from GPe to the striatum, which enhances its efficiency.
BibTeX
@misc{girard2005contracting,
  author = {Benoît Girard and Nicolas Tabareau and Jean-Jacques Slotine and Alain Berthoz},
  title = {{Contracting model of the basal ganglia}},
  year = {2005},
}
Figure