Source-linked AI summary
Learning-Induced Autonomy of Sensorimotor Systems
Danielle S. Bassett, Muzhi Yang, Nicholas F. Wymbs, Scott T. Grafton
TL;DR
The paper examines how distributed brain circuits dynamically reorganize as people learn a motor skill. Using time-resolved network analysis of fMRI data, it finds that training increases sensorimotor autonomy and that frontal and anterior cingulate disengagement predicts individual learning differences.
Problem
The dynamic integration of distributed neural circuits during the transformation of a motor skill from challenging to automatic has been difficult to describe with existing analysis frameworks.
Method
The study uses time-resolved network analysis of fMRI data to quantify module allegiance and examine recruitment and integration across motor-skill learning.
Results
Learning increases the autonomy of sensorimotor systems, while disengagement of distributed frontal and anterior cingulate regions predicts individual differences in learning.
Takeaways & Limitations
The results provide a statistically principled account of changing distributed neural-circuit dynamics during human skill learning.
Takeaways & Limitations
The task-based network architecture likely differs significantly from architecture measured during task-free conditions.
Abstract
from arXiv · showhide
Distributed networks of brain areas interact with one another in a time-varying fashion to enable complex cognitive and sensorimotor functions. Here we use novel network analysis algorithms to test the recruitment and integration of large-scale functional neural circuitry during learning. Using functional magnetic resonance imaging (fMRI) data acquired from healthy human participants, from initial training through mastery of a simple motor skill, we investigate changes in the architecture of functional connectivity patterns that promote learning. Our results reveal that learning induces an autonomy of sensorimotor systems and that the release of cognitive control hubs in frontal and cingulate cortices predicts individual differences in the rate of learning on other days of practice. Our general statistical approach is applicable across other cognitive domains and provides a key to understanding time-resolved interactions between distributed neural circuits that enable task performance.
Results
Across learning, the brain retained stable motor and visual modules while reducing their interaction and the recruitment of non-motor, non-visual regions. These changes indicate increasing sensorimotor autonomy and reduced cognitive-system involvement with practice.
- Summary Architecture of Learning: The study quantified module allegiance as the probability that two brain areas were assigned to the same functional community across time-resolved network partitions.The allegiance matrix summarized the architecture across subjects, sessions, sequence types, and trial blocks.
- Summary Architecture of Learning: The summary architecture identified stable motor and visual modules, with motor and secondary sensorimotor areas grouped separately from primary visual cortex.Their dissociation reflected significantly different BOLD activation time courses.
- Summary Architecture of Learning: The non-motor, non-visual set showed 29% same-community probability, compared with 80% for visual and 65% for motor modules.These regions include areas supporting attention, executive function, and cognitive control.
- Dynamic Architecture of Learning: Motor and visual modules remained present across naive, early, middle, and late learning, while their interaction strength decreased with task practice.The non-motor, non-visual set also showed reduced module allegiance during practice.
- Dynamic Architecture of Learning: Recruitment measures captured each group’s task involvement, whereas normalized integration measured interaction strength between different groups while accounting for group-strength differences.These recruitment and integration measures were termed brain network diagnostics.
- Dynamic Architecture of Learning: Motor and visual recruitment stayed steady, but motor-visual integration decreased with training intensity and duration, indicating increasing autonomy between these systems.As autonomy increased, recruitment of non-motor, non-visual areas decreased.
System-Level Correlates of Performance and Learning
Training-dependent changes in non-motor, non-visual recruitment, rather than motor-visual integration, tracked individual learning differences. A predictive network of frontal and cingulate connections characterized this relationship.
- Individual differences: Training-dependent modulation of motor-visual integration and non-motor, non-visual recruitment varied across individuals.Greater sensorimotor autonomy was hypothesized to distinguish better learners from participants retaining stronger cross-system recruitment.
- System-level correlates: Motor-visual integration showed no significant relationship with learning rate (r = 0.42, p = 0.0650).The separation between motor and visual modules was associated with task practice across participants rather than individual learning rate.
- System-level correlates: Non-motor, non-visual recruitment modulation strongly correlated with learning rate (r = 0.59, p = 0.0062).Participants who disbanded extraneous areas during training learned better than those who maintained them later in learning.
- Predictive network: The predictive-network analysis correlated each edge’s training-induced modulation with individual learning rate using Pearson’s r.Edges meeting p < 0.05 uncorrected were collectively defined as the predictive network.
- Predictive network: The predictive network comprised 180 functional connections spanning approximately 96% of the non-motor, non-visual set.Its regional strength distribution was highly skewed (s = 1.89; non-parametric test p = 0.00009), with high-strength areas predominantly in frontal and cingulate cortices.
Learning-Induced Changes in Sensorimotor Systems
Across learning, motor and visual systems became more autonomous while broader non-motor, non-visual recruitment and integration declined. Frontal and anterior cingulate disengagement predicted individual learning differences, although the measurements came from a limited scanning schedule.
- Learning-induced changes: Motor and visual systems formed separate cohesive modules whose recruitment remained stable while their integration decreased as performance became automatic.The study examined network adaptations across a continuum of learning rather than only early or late learning.
- Sensorimotor autonomy: Motor and visual systems transitioned from heavy early integration to later autonomy, supporting independent computations with distinct BOLD temporal profiles.Early practice required coordinated visual decoding, attention, and precise movement; later practice reduced dependence on visual cues as sequences became learned.
- Network recruitment: As sensorimotor autonomy increased, the non-motor, non-visual network decreased in recruitment.Participants who disbanded this network learned better, whereas motor-visual recruitment or integration did not correlate with learning.
- Cognitive-control systems: Training-dependent release of frontal and anterior cingulate connections predicted individual differences in learning.These regions are hubs of frontal-parietal and cingulo-opercular cognitive-control systems that became disengaged through training.
- Methodological considerations: Neurophysiological measures came from 4 scans over 6 weeks, while behavioral learning estimates came from home sessions between scans.This temporal separation was used to characterize the findings as predictive rather than simply correlative.
Conclusions
Using dynamic network neuroscience on motor-sequence learning, the study identified learning-related autonomy in sensorimotor systems and predictive disengagement of frontal and cingulate networks. The experiment followed healthy participants across repeated fMRI and training sessions.
- Conclusions: Dynamic network neuroscience exposed learning-induced autonomy of sensorimotor systems and a frontal-anterior cingulate network whose disengagement predicted learning differences.The conclusion addresses time-resolved interactions among distributed neural circuits during skill acquisition.
- Training design: Participants practiced sequences during naive, early, middle, and late scanning sessions, with intervening home practice and differing exposure levels.The design included extensively, moderately, and minimally practiced sequences.
Experiment Setup and Procedure
Participants learned six visually presented finger-movement sequences across repeated scanning and home-training sessions. The experiment varied sequence exposure and measured behavior during a discrete sequence-production task using matched response hardware.
- Participants and sessions: Twenty participants completed at least 30 behavioral sessions, three fMRI test sessions, and a pre-training fMRI session.Two participants were excluded, one for noncompletion and one for excessive head motion.
- Task: The task required right-hand responses to visually presented 10-element sequences using a five-position stimulus array.Stimulus locations mapped from left to right onto the thumb through smallest finger.
- Trial structure: Each trial began with a 2-second sequence-identity cue, followed by an imperative stimulus initiating sequence production.Each of the six trained sequences had a unique identity cue.
- Training exposure: Training sessions included extensively, moderately, and minimally practiced sequences with different exposure levels.The session structure used two sequences in each exposure category, with extensive sequences receiving 64 trials and moderate sequences 10 trials.
- Experiment timeline: Scanning sessions occurred after approximately 10 home-training sessions and included repeated practice of the same sequences.The experiment timeline interleaved scanner training with home practice across the learning period.
- Scanning procedure: Each fMRI test session used the same task and sequences, with unlimited trial time and instructions to respond quickly while maintaining accuracy.Responses were collected with a fiber-optic button box configured similarly to the laptop training setup.
- Implementation: Stimulus presentation and response collection used laptop-based task software and a custom fiber-optic button box.Training and testing used separate Octave/PsychtoolBox and MATLAB configurations.
Behavioral Estimates of Learning
The study estimated motor learning from movement-time data collected across home training and related these behavioral estimates to independently measured brain-network structure. Learning rate was modeled with a robust double-exponential fit, while fMRI data were organized using a 112-region atlas.
- Behavioral Estimates of Learning: Learning estimates were derived from movement-time data collected during six weeks of home training, independently from four fMRI scanning sessions.This separation was used to ensure independence between behavioral learning and brain-network measures.
- Behavioral Estimates of Learning: Movement time was defined as the interval between the first and last button presses in a sequence.
- Behavioral Estimates of Learning: A robust double-exponential function estimated fast and slow improvement rates for each sequence type.The model used outlier correction and was fit separately for sequences 1–6.
- Behavioral Estimates of Learning: The learning parameter κ quantified the fast rate of improvement, with larger κ values indicating steeper movement-time declines and quicker learning.The fitting approach estimates learning rate independently of initial performance and performance ceiling.
- Behavioral Estimates of Learning: fMRI acquisition and preprocessing used BOLD imaging, spatial normalization, smoothing, and a 112-region cortical and subcortical atlas.Regional mean BOLD time series were obtained by averaging voxels within each atlas-defined region after gray-matter restriction.
Wavelet Decomposition
The analysis isolated frequency-specific fMRI dynamics with wavelet decomposition and represented functional connectivity in short, block-specific temporal windows. These windowed connectivity matrices were then assembled into multilayer networks.
- Wavelet Decomposition: Wavelet decomposition was used to detect small signal changes in non-stationary fMRI time series with noisy backgrounds.
- Wavelet Decomposition: With a 2-second sampling interval, wavelet scale one covered 0.125–0.25 Hz and scale two covered 0.06–0.125 Hz.
- Wavelet Decomposition: Wavelet coefficients were extracted for all 112 regions within approximately 60-TR windows aligned to experimental trial blocks.The analysis included EXT, MOD, and MIN sequences across multiple blocks and participants.
- Wavelet Decomposition: Each block-specific dataset produced a 112 × 112 adjacency matrix containing pairwise functional connections for that temporal window.
- Wavelet Decomposition: Consecutive block-level adjacency matrices were combined into rank-3 adjacency tensors representing time-dependent networks for each sequence type, participant, and scan.
Dynamic Community Detection
Dynamic community detection identified groups of brain regions whose connectivity patterns evolved together across network layers. The multilayer model used a modularity-based objective with explicit within-layer and between-layer structure.
- Dynamic Community Detection: Community detection identifies putative functional modules as groups of regions with similar trajectories through time.
- Dynamic Community Detection: The modularity framework groups nodes with stronger within-group connections than connections to other groups.
- Dynamic Community Detection: The multilayer formulation incorporates adjacency matrices, a null model, structural resolution, community assignments, and inter-layer coupling.
- Dynamic Community Detection: Inter-layer coupling was restricted to neighboring layers, with ω = 1 and γ = 1.The model therefore linked adjacent time layers while using constant structural resolution.
Supplementary Results
Supplementary analyses tested whether module-allegiance patterns and training-related diagnostics depended on statistical thresholding. The reported recruitment, integration, and architectural features were preserved under these robustness checks.
- Supplementary Results: Module allegiance P_ij represents the probability that regions i and j occupy the same functional community across partitions.
- Supplementary Results: The null model randomly reassigned nodes to communities within each partition to estimate chance co-assignment.
- Supplementary Results: 38.89% was the average chance co-assignment rate across 100 partitions, and values below the maximum random-association entry were removed.
- Supplementary Results: After thresholding, motor recruitment, visual recruitment, motor–visual integration, and non-motor, non-visual recruitment remained preserved.
- Supplementary Results: Training-related recruitment and integration results were robust to whether statistically thresholded or unthresholded module-allegiance matrices were used.
Training-Dependent Modulation of Intra-Module Integration for Motor and Visual Systems
Training alters integration within the motor module but not the visual module. Several motor regions disengage from their module as practice increases, whereas visual-module integration shows no significant training-related change.
- Motor system: Motor-region disengagement was quantified through training-dependent modulation of intra-module integration, based on correlations with trials practiced.Intra-module integration sums functional connections between a region and the other regions in its module.
- Motor system: Seven of 12 motor-module regions significantly disengaged from the motor module with training after Bonferroni correction.Significant regions included bilateral SMA and precentral areas, left postcentral, left superior parietal, and left parietal operculum.
- Visual system: No visual-module regions showed significant enhanced integration or disengagement with training.
- Predictive network: The predictive network’s node-strength distribution was significantly more skewed than expected under a random graph null model.The observed skewness was s = 1.89, with p = 0.00009, using 100,000 Erdős-Rényi graphs with 90 nodes and 180 edges.
Methodological Considerations
The analyses compare dynamic module-allegiance measures with functional-connectivity alternatives while addressing block-design and region-size influences. Dynamic analyses operate at slower time scales than task block effects, and region size was not identified as the driver of module organization.
- Block-design effects: Task-related on-off cycles fall within the 0.06–0.12 Hz wavelet band and likely influence single-window correlation patterns.Trials last 4–6 TRs, producing one task-related on-off cycle per trial.
- Block-design effects: Dynamic module-allegiance analyses probe 0.0083–0.012 Hz activity, about an order of magnitude below the frequencies associated with block-design effects.This slower time scale permits early- and extended-learning effects to be examined independently of block-design effects.
- Region size: The analysis tested whether region size explained the motor and visual module organization and concluded that it did not.Region size was estimated from voxel counts averaged over participants and compared across motor, visual, and non-motor, non-visual regions.
- Method comparison: Module allegiance provides greater sensitivity to learning-related network changes than functional-connectivity matrices alone.The comparison covers summary and dynamic learning architecture as well as training-modulated recruitment and integration.
- Network measures: Functional connectivity W uses wavelet coherence between pairs of regional BOLD time series, whereas module allegiance P captures shared community assignment.Recruitment and integration can be computed from either representation, but their learning sensitivity differs.
- Method comparison: Alternative diagnostics based on averaged functional-connectivity matrices showed no training-dependent modulation consistent with trials practiced.These diagnostics instead showed effects associated independently with training intensity, training duration, or neither.
Summary Architecture of Learning
The paper presents summary and dynamic network architectures through module-allegiance and functional-connectivity matrices. The figures compare these representations and organize brain regions using a standard cortical and subcortical parcellation scheme.
- Summary and dynamic architecture: Figure 6 contrasts module-allegiance matrices with functional-connectivity matrices for summary and dynamic learning architectures.The top panel shows summary architecture and the bottom panel shows dynamic architecture; functional-connectivity matrices are normalized by their mean.
- Parcellation: The brain regions are represented using the Harvard-Oxford Cortical and Subcortical Parcellation Scheme provided by FSL.