Source-linked AI summary
A computational model of inhibitory control in frontal cortex and basal ganglia
Thomas V. Wiecki, Michael J. Frank
TL;DR
The paper addresses the lack of a coherent framework integrating findings on frontal-basal-ganglia response inhibition and component effects on behavior. It presents a dynamic neural network model of selective and global response inhibition, capturing control over prepotent responses and multiple levels of analysis while acknowledging limitations.
Problem
There is no coherent framework integrating findings on response inhibition and the effects of component parts on behavior.
Method
The paper presents a dynamic neural network model of selective and global response inhibition grounded in established neuroanatomical and neurotransmitter-related complexity.
Results
The model captures control over prepotent responses and a wealth of data across multiple levels of analysis, while providing multiple mechanisms.
Takeaways & Limitations
The model suggests that observed deficits in inhibitory-control paradigms do not necessarily reflect dysfunctional inhibitory control.
Takeaways & Limitations
The authors acknowledge that the model contains many errors and discuss salient limitations.
Abstract
from arXiv · showhide
Planning and executing volitional actions in the face of conflicting habitual responses is a critical aspect of human behavior. At the core of the interplay between these two control systems lies an override mechanism that can suppress the habitual action selection process and allow executive control to take over. Here, we construct a neural circuit model informed by behavioral and electrophysiological data collected on various response inhibition paradigms. This model extends a well established model of action selection in the basal ganglia by including a frontal executive control network which integrates information about sensory input and task rules to facilitate well-informed decision making via the oculomotor system. Our simulations of the antisaccade, Simon and saccade-override task ensue in conflict between a prepotent and controlled response which causes the network to pause action selection via projections to the subthalamic nucleus. Our model reproduces key behavioral and electrophysiological patterns and their sensitivity to lesions and pharmacological manipulations. Finally, we show how this network can be extended to include the inferior frontal cortex to simulate key qualitative patterns of global response inhibition demands as required in the stop-signal task.
1 Introduction
Response inhibition requires overriding prepotent actions through interacting frontal and basal-ganglia processes, but no coherent model previously integrated these findings. The paper introduces a neural network account spanning selective and global inhibition, behavioral and electrophysiological patterns, and perturbations.
- Background: Selective inhibition tasks require replacing a prepotent response with a controlled alternative, whereas stop-signal inhibition requires only stopping the planned response.The antisaccade, Simon, and saccade-override tasks require subsequent response initiation; the stop-signal task does not.
- Background: Frontal and subcortical frontostriatal regions are implicated in response inhibition, and disrupting either level can cause inhibition deficits.The implicated regions include DLPFC, SEF, pre-SMA, FEF, striatum, STN, and superior colliculus.
- Problem: No coherent framework previously integrated electrophysiological findings with selective disruptions of component parts and their behavioral effects.The authors identify this integration as the central modeling gap.
- Contribution: The paper extends an established basal-ganglia model with frontal executive control to formalize separable, interacting neural processes underlying inhibition and volitional action.The model addresses action selection, inhibitory control, conflict-induced slowing, and volitional action generation.
- Results: The model reproduces behavioral reaction-time patterns, electrophysiological activity across frontal and subcortical regions, and effects of psychiatric, developmental, lesion, and pharmacological manipulations.The authors also report that extending the model with rIFG recovers key stop-signal task patterns.
- Results: Selective response inhibition combines conflict-induced global slowing through the hyperdirect pathway with selective NoGo inhibition of the prepotent response.The model predicts that tonic dopamine affects selective but not global response inhibition.
2 Neural Network Model
The model extends basal-ganglia action selection into a dynamic oculomotor circuit linking sensory input, frontal planning, executive rules, and response gating. Its mechanisms combine direct and indirect striatal pathways with dopamine modulation and conflict-sensitive STN threshold control.
- Implementation and scope: The model uses one core parameter set across simulations, tests each reported simulation on 8 randomly initialized networks, and omits rule-learning dynamics.The paper describes the rule representations as the post-learning state of the network.
- Core architecture: The network adapts an established basal-ganglia model to rapid eye movements, using FEF for action planning and SC threshold crossing for saccade generation.Response time is defined as the time from trial onset until an SC unit crosses threshold.
- Basal-ganglia gating: Direct striatal Go activity inhibits SNr and disinhibits SC, whereas indirect NoGo activity increases SNr inhibition to prevent response gating.These pathways implement opposing Go and NoGo functions for saccade selection.
- Neuromodulation: Dopamine amplifies active Go units and inhibits NoGo units, while the present simulations omit learning and focus on already learned associations.The model retains dopamine-dependent pathway modulation but does not simulate acquisition of the stimulus-response mappings.
- Conflict control: STN provides a diffuse global NoGo signal that raises the gating threshold when frontal conflict increases during early response selection.Unlike a static threshold, STN activity is driven by response conflict through frontal input.
- Executive control: An executive control layer summarizing DLPFC, SEF, and pre-SMA integrates task rules with sensory information to select FEF responses and bias basal-ganglia gating.This layer enables controlled responding in tasks involving prepotent response competition.
- Mechanistic account: Conflict produces global threshold adjustment while executive control selectively inhibits the prepotent response, implementing two distinct suppression mechanisms.The authors characterize this as a biologically plausible implementation of cognitive control.
2.1 Selective Response Inhibition
The model represents selective response inhibition as executive control overriding prepotent responses in incongruent trials. It reproduces behavioral, electrophysiological, and manipulation-related patterns across this control process.
- Model implementation: The model encodes prepotent stimulus-response biases in FEF and uses DLPFC sensory-rule integration to activate the task-appropriate response.Strong input-to-FEF weights facilitate congruent responses but bias incongruent responding; DLPFC-to-FEF projections are stronger and support controlled selection.
- Behavior: Incongruent error trials had faster RTs than correct incongruent trials because prepotent response capture preceded controlled selection.This reproduces the reported pattern that erroneous responses are initiated before the volitional response and are associated with shorter reaction times.
- Manipulations: Increased tonic dopamine and disrupted STN function increased incongruent errors, whereas STN dysfunction produced fast but inaccurate responding.The model attributes STN-related errors to failure to raise the threshold required to prevent prepotent response gating.
- Manipulations: Increasing FEF→striatum connectivity sped responses and reduced accuracy, whereas increasing STN→SNr connectivity slowed responses and improved accuracy.These routes provide distinct mechanisms for modulating the speed-accuracy tradeoff and decision threshold.
- Neurophysiology: Conflict-related dACC and STN activity occurred before correct incongruent responses but after erroneous responses, matching monkey and human electrophysiological patterns.The model also reproduced FEF/SC dynamics and striatal Go/NoGo activity associated with suppressing prepotent responses and selecting controlled actions.
2.2 Global Response Inhibition
The model extends its selective inhibition architecture to global stopping by adding an rIFG–STN pathway. It reproduces stop-signal patterns while distinguishing fast global inhibition from slower selective control.
- Task and method: The stop-signal task requires outright response inhibition after a variable stop-signal delay, assessed by successful stopping across delays and SSRT estimation.The model uses a race-model interpretation and a dynamic one-up/one-down staircase to adjust stop-signal delays.
- Model extension: Adding rIFG with direct STN projections enables the model to simulate the stop-signal task.Stop signals activate rIFG, which transiently excites STN and prevents striatal response gating when the response has not already crossed threshold.
- Model extension: The architecture combines fast, global but transient inhibition with slower, selective but lasting inhibition.rIFG–STN activity globally pauses gating, while DLPFC-driven striatal NoGo activity selectively inhibits the associated response after STN activity returns to baseline.
- Behavioral results: The probability of correctly stopping decreased monotonically as stop-signal delay increased, and stop-signal inhibition and error distributions matched up to SSD+SSRT.These patterns correspond to qualitative stop-signal benchmark results reproduced by the model.
- Manipulations: Dopamine manipulations sped GoRT while leaving SSRT largely unaffected, whereas reduced frontal gain and STN or rIFG lesions increased SSRT.Accuracy emphasis slowed GoRT but produced faster SSRT, consistent with more effective inhibition.
- Neurophysiology: DLPFC, SEF, and pre-SMA activation followed SSRT, supporting fast rIFG–STN global stopping before delayed executive control contributes to selective inhibition.Executive control may participate after the global response pause, including selective inhibition and correct-response activation in stop-change tasks.
3 Discussion
The model links executive control with basal-ganglia action selection to explain selective and global response inhibition across multiple tasks. It reproduces behavioral, electrophysiological, lesion, and pharmacological patterns while identifying mechanisms regulating response gating.
- Conflict between prepotent and controlled responses pauses action selection through subthalamic-nucleus projections and raises the decision threshold for gating.The threshold subsequently collapses dynamically over time as conflict resolves.
- The model extends a basal-ganglia action-selection framework with frontal executive-control regions to simulate selective and global response inhibition.Its network includes distributed computations across frontal cortex and basal ganglia, with one intact parameterization across simulations.
- Incongruent trials produced more errors and slower responses, with errors especially likely when networks responded quickly.Reduced DLPFC connectivity also degraded incongruent-trial accuracy.
- The model reproduced dACC activity before correct incongruent responses, after incorrect incongruent responses, and at baseline during congruent responses.This supports a framework in which dACC activity reflects conflict and the value of an alternative action.
- Adding an rIFG layer generalized the model to stop-signal tasks by detecting salient events and engaging a global response-pause mechanism.STN activity surged similarly across successful and failed stops, whereas SNr activity differentiated the outcomes; errors arose from variable Go-process timing.
- In the stop-signal task, SSRT improved with increased tonic rIFG activity but was unaffected by mechanisms that merely slowed overall responding.This distinguishes active stop-process engagement from general response slowing as a source of inhibitory control.
4 Limitations
The authors identify omissions, possible unsupported commissions, uncertain frontal-region roles, and hard-coded task rules as limitations of the model.
- The model contains acknowledged omissions and possible errors of commission, although its assumptions and simulations are largely orthogonal and individually falsifiable.
- Hard-coded task rules and input-output weights bypass computational problems such as rule retrieval, sensory integration, and motor-sequence computation.The authors chose this approach partly because specific learning phenomena for the antisaccade and stop-signal tasks are not well documented.
- How the necessary executive computations can be implemented dynamically remains unresolved despite existing models with more detailed prefrontal representations.
- The individual contributions and interactions of frontal regions remain uncertain and may require revision as more data become available.DLPFC, SEF, pre-SMA, FEF, and dACC are less firmly established than the basal-ganglia component.
5 Conclusions
The paper presents a biologically plausible model of global and selective response inhibition that integrates basal-ganglia action selection with frontal executive control. Augmenting the basal-ganglia model with FEF, DLPFC, and rIFG captures data across multiple analytical levels and offers mechanisms relevant to inhibitory-control disruptions.
- The model is presented as a comprehensive account of global and selective response inhibition.
- The model links neuronal mechanisms with findings from cognitive science, electrophysiology, imaging, and pharmacological experiments.
- Augmenting the basal-ganglia model with FEF, DLPFC, and rIFG simulates control over prepotent responses across multiple levels of analysis.
- The framework provides multiple mechanisms that can disrupt inhibitory-control processes and inform interpretation of data from patients with SZ and ADHD.
- Observed deficits in inhibitory-control paradigms do not necessarily reflect dysfunctional inhibitory control itself.
7 Appendix
The appendix describes the model’s conductance-based neuron implementation, activation dynamics, time conversion, and inhibitory competition mechanisms. It also explains kWTA inhibition as a computational approximation with adjustable rigidity and biological plausibility limitations.
- Implementation details: Emergent simulator cycles are converted to milliseconds by multiplying by 4 to approximately match behavioral and electrophysiological data.The model and scripts are available through the cited project website.
- Implementation details: The simulator uses point neurons whose excitatory, inhibitory, and leak conductances determine an integrated membrane potential and rate-coded output.Discrete spiking is also available but produces noisier results.
- Implementation details: Overall conductance combines a dynamic component dependent on model state with a constant component controlling each conductance’s relative influence.The model distinguishes excitatory, leak, and inhibitory input channels.
- Implementation details: The equilibrium potential expresses a balance between excitation and opposing leak and inhibitory forces, with excitatory, leak, and inhibitory driving potentials simplified to 1, 0, and 0.This equilibrium form is interpreted within a Bayesian decision-making framework.
- Implementation details: Excitatory net input is computed from sending-unit activations weighted by connection strengths, while inhibition is supplied by kWTA competition or modeled inhibitory interneurons and leak remains constant.Activation communicated to other cells is a thresholded sigmoidal function of membrane potential with gain γ, followed by projection scaling factors.
- Inhibition within and between layers: kWTA assigns uniform inhibitory current so the kth most excited unit is generally above threshold while the k + 1th is generally below threshold.The method approximates inhibitory interneuron dynamics, although its implementation requires global activation information and sorting mechanisms.
- Inhibition within and between layers: The kWTA implementation is computationally effective but biologically implausible because it requires global activation information and sorting mechanisms.This limitation concerns the implementation rather than the inhibitory competition behavior it approximates.
- Inhibition within and between layers: The basic kWTA version rigidly permits k units above threshold, typically uses q = .25, and is contrasted with a more flexible version typically using q = .6 for hidden layers.The flexible version allows the number of active neurons to vary with the activation distribution.