Dynamic network perspective represents major departure from localist models; instead of cognitive functions mapping to discrete neural regions/connects, mental operations are suggested to be supported by unique conjunctions of distributed brain regions

Neuroimaging evidence for a network sampling theory of individual differences in human intelligence test performance. Eyal Soreq, Ines R. Violante, Richard E. Daws & Adam Hampshire . Nature Communications volume 12, Article number: 2072. Apr 6 2021. https://www.nature.com/articles/s41467-021-22199-9

Abstract: Despite a century of research, it remains unclear whether human intelligence should be studied as one dominant, several major, or many distinct abilities, and how such abilities relate to the functional organisation of the brain. Here, we combine psychometric and machine learning methods to examine in a data-driven manner how factor structure and individual variability in cognitive-task performance relate to dynamic-network connectomics. We report that 12 sub-tasks from an established intelligence test can be accurately multi-way classified (74%, chance 8.3%) based on the network states that they evoke. The proximities of the tasks in behavioural-psychometric space correlate with the similarities of their network states. Furthermore, the network states were more accurately classified for higher relative to lower performing individuals. These results suggest that the human brain uses a high-dimensional network-sampling mechanism to flexibly code for diverse cognitive tasks. Population variability in intelligence test performance relates to the fidelity of expression of these task-optimised network states.

Introduction

The question of whether human intelligence is dominated by a single general ability, ‘g’1, or by a mixture of psychological processes2,3,4,5,6, has been the focus of debate for over a century. While performance across cognitive tests does tend to positively correlate, population-level studies of intelligence have clearly demonstrated that tasks which involve similar mental operations form distinct clusters within a positive correlation manifold. These task clusters exhibit distinct relationships with various sociodemographic factors that are not observable when using aggregate measures of intelligence, such as ‘g’2,7.

Recent advances in network science offer the potential to resolve these contrasting views. It has been proposed that transient coalitions of brain regions form to meet the computational needs of the current task8,9,10. These dynamic functional networks are thought to be heavily overlapping, such that any given brain region can express flexible relationships with many networks, depending on the cognitive context8,9,11,12,13. This dynamic network perspective represents a major departure from localist models of brain functional organisation. Instead of cognitive functions mapping to discrete neural regions or specific connections, mental operations are suggested to be supported by unique conjunctions of distributed brain regions, en masse. The set of possible conjunctions can be considered as the repertoire of dynamic network states and the expression of these states may differ across individuals and relate to cognitive performance.

This conceptual shift motivates us to propose a network sampling theory of intelligence, which is conceptually framed by Thomson’s classic sampling theory14. Thomson originally proposed that ‘every mental test randomly taps a number of ‘bonds’ from a shared pool of neural resources, and the correlation between any two tests is the direct function of the extent of overlap between the bonds, or processes, sampled by different tests’. Extending this hypothesis, network sampling theory views the set of connections in the brain that constitute a task-evoked dynamic network state to be equivalent to Thomson’s ‘bonds’; therefore, the set of available brain regions is equivalent to the ‘shared pool of neural resources’. The distinctive clusters within the positive manifold reflect the tendency of operationally similar tasks to rely on similar dynamic networks2,15,16,17. From this perspective, the general intelligence factor ‘g’ is proposed to be a composite measure of the brain’s capacity to switch away from the steady state, as measured in resting-state analyses, in order to adopt information processing configurations that are optimal for each specific task. When recast in this framework, classic models of unitary and multiple-factorial intelligence1,14 are reconciled as different levels of summary description of the same high-dimensional dynamic network mechanism. The notion of domain-general systems such as ‘task active’ or ‘multiple-demand’ cortex is also reconciled within this framework. Specifically, each brain region can be characterised by the diversity of network states they are active members of. Brain regions that classic mapping studies define as ‘domain-general’ place at one extreme of the membership continuum, whereas areas ascribed specific functions, e.g., sensory or motor, place at the other extreme. The aim of this study was to test key predictions of network sampling theory using 12 cognitive tasks and machine learning techniques applied to functional MRI (fMRI) and psychometric data. First, we test the hypothesis that cognitive tasks evoke distinct configurations of activity and connectivity in the brain. We predicted that these configurations would be sufficient to reliably classify individual tasks, and that this would be the case even when focusing on brain regions at the domain-general extreme of the network membership continuum. We then tested Thomson’s hypothesis that similarity between cognitive tasks maps to the ‘overlap’ of the neural resources being tapped. Subsequently, it was predicted that the ability to classify pairs of tasks would negatively correlate with their behavioural-psychometric similarity, with tasks that are less similar being classified more reliably. Next, we hypothesised that individual functional dynamic repertoires would positively correlate with task performance, with the top performers expressing task configurations that would be more reliably classified. We also tested the prediction that classification success rates should have a basis in a combination of the distinct visual (VS), motor and cognitive sub-processes of the tasks. Finally, we hypothesised that task performance would be associated with optimal perturbation of the network architecture from the steady state, and that certain features within the network would have more general and more prominent roles in intelligence test performance.

Discussion

The results presented here are highly compatible with a network science interpretation of Thomson’s sampling theory14. Indeed, as has been noted by others, the relationship between the classic notion of a flexible pool of bonds and the analysis of the brain’s dynamic networks as applied a century later is striking17. Thomson proposed that mental tests tap bonds from a shared pool of neural resources, which is confirmed by our observation that different cognitive tasks tend to recruit unique but heavily overlapping networks of brain regions. Furthermore, when testing Thomson’s proposal that the correlation between any two tasks is a function of the extent of overlap between their bonds, we confirmed that the similarities of tasks in multi-factor behavioural psychometric space correlated strongly with the similarities in the dynamic network states that they evoked. These findings corroborate the key tenets of network sampling theory, further predictions of which were tested utilising a combination of machine learning techniques applied to the fMRI and psychometric data.

From a network science perspective, our results showing that the tasks were 12-way classifiable with high accuracy based on their dynamic network states is highly relevant. Indeed, the 74% accuracy achieved by the CRTX stack model was surprising, given that chance was 8.3% and we used just 1 min, comprising 30 images, of task performance data per classified sample. Although activity and connectivity provided complementary information when combined in the stack models, classification accuracy was consistently higher for connectivity when the measures were analysed independently. These results strongly support the hypothesis that the human brain is able to support diverse cognitive tasks because it can rapidly reconfigure its connectivity state in a manner that is optimal for processing their unique computational demands8,9,12,17. A key finding was that the task-evoked dynamic network states were consistent across individuals; i.e., our trained 12-way classification models operated with high accuracy when applied in a robust CV pipeline to data from individuals to whom they were completely naive. This was with an out of the box classifier with no CV optimisation, which is important, because it means that the features that drove accurate classification must reflect on a fundamental level how networks in the human brain are prewired to flexibly support diverse tasks.

At a finer grain, these task-optimised network states are most accurately described as a perturbation away from the RSN architecture of the brain12,29. More specifically, it was not simply the case that the relative levels of activity or connectivity within each RSN change, i.e., reflecting different mixtures dependent on task demands; instead, the features that were most specific to a given task-evoked state were predominantly the inter-RSN connections. Put another way, task-evoked states are not a simple blending of RSNs, but a dissolution of the RSN structure. This extends the findings of another recent study, where we used a similar analysis pipeline to examine how different aspects of working memory affected brain activity and connectivity12. Mirroring the current findings, we found that behaviourally distinct aspects of working memory mapped to distinct but densely overlapping patterns of activity and connectivity within the brain. Taken together, these results do not accord well with the hypothesis that the human brain is organised into discrete static networks. Instead, it would appear that the dynamic network coding mechanism is very high-dimensional, relating to the greater number of possible combinations of nodes8,9. There are dependencies whereby some nodes operate together more often than others, but these canonical network states, which are consistently evident in data-driven analyses of the resting state brain, are statistical rather than absolute. Our more holistic interpretation of the relationship between network states and cognitive processes is further supported by the analysis of the classifiability of task clusters when grouped according to their behavioural dimensions. Specifically, when grouped by psychometric, motor or VS characteristics, the clusters were more classifiable than random task groupings in all cases. It was notable though that psychometric and motor characteristics provided a stronger basis for classification. This is interesting, because it pertains to how the most prominent factors of human intelligence differ operationally. For example, it accords well with process overlap theory17, which proposes that general intelligence relates most closely to processes that are common across many different cognitive tasks.

More generally, the fact that inter-individual differences in the classifiability of the tasks predicted variability in a general measure of behavioural task performance provides further evidence that cognitive faculties relate to the way in which the brain expresses these task-optimal network states. Previous research into the neural basis of human intelligence has typically emphasised the role of flexible FP brain regions2,30,31,32. In this context, our focused analysis of the INTR ROI set warrants further discussion. Brain regions within the INTR ROI set belong to the classical MD cortical volume, which has been closely associated with general intelligence. MD includes the FP brain regions that have the broadest involvement in cognitively demanding tasks19,20,30; this includes executive functions, which enable us to perform complex mental operations33,34 and that have been proposed to relate closely to the ‘g’ factor17. From a graph theoretic perspective, MD ROIs have been reported to have amongst the broadest membership of dynamic networks of any brain regions35 and it has been shown that inter-individual variability in the flexibility of MD nodes, as measured by the degree of their involvement in different functional networks, correlates positively with individuals’ abilities to perform specific tasks, e.g., motor skill learning36 and working memory37. Collectively, these findings highlight a strong relationship between the flexibility of nodes within MD cortex and cognitive ability.

Here, we reconfirmed that MD ROIs were amongst the most consistently active across the 12 tasks. However, we also demonstrated that these ROIs were highly heterogeneous with respect to their activation profiles across those tasks. Furthermore, in many cases they were significantly active for most but not all tasks. This variability in the activation profiles even amongst the most commonly recruited areas of the brain aligns with the idea that MD cortex flexibly codes for diverse tasks in a high-dimensional manner. More critically, the internal activity and connectivity of the INTR ROI set was not strongly predictive of behavioural task performance. Nor did it provide the most accurate basis for classification overall, or correspondence to psychometric structure. Extending to the MDDM set provides an improvement, but it was inclusion of the whole cortex ROI set that provided the best predictor of task and behavioural performance. Furthermore, connections between the core set of INTR regions and the rest of the brain featured prominently in all of the above cases. This finding accords with bonds theory, insofar as that theory pertains to the wide variety of bonds that contribute to diverse behavioural abilities. It also accords particularly well with the core tenet of network science that cognitive processes are emergent properties of interactions that occur across large-scale distributed networks in the brain10,12.

An intriguing aside pertains to the phenomena of ‘factor differentiation’. It was originally noted by Spearman38 that ‘g’ explains a greater proportion of variance individuals who perform lower on intelligence tests. This finding has been robustly replicated over the subsequent century5. Our results provide a simple explanation for factor differentiation. When individuals of higher intelligence perform different cognitive tasks, the dynamic network states that they evoke are more specific. Therefore, there is less overlap in the neural resources that they recruit to perform the tasks. Given the relationship observed here between network similarity and behavioural-psychometric distance, this would be expected to reduce bivariate correlations in task performances and produce a corresponding reduction in the proportion of variance explained by ‘g’.

The boosted ensemble of regression trees provided a simple way to extend the individual differences analysis in order to capture not just mixtures but also interactions between network features when predicting behavioural performance. We observed that increased connectivity between DA and VS systems strongly associated with better performance, whilst increased connectivity within the DMN combined with decreased connectivity between either DA to VS or DM to FP associated with lower performance. This accords well with previous studies that have shown that these networks update their connectivity patterns according to the task context35,39,40,41,42,43. However, it was particularly notable that inter-RSN connections again played the most prominent role insofar as they formed the roots of all of the trees, meaning they had the broadest relevance across individuals. This further accords with the view that task-evoked network states are best described as a perturbation from the RSN architecture12,29.

In summary, we validated multiple key predictions of network sampling theory. This theory can potentially explain key findings from behavioural psychometrics, experimental psychology and functional neuroimaging research within the same overarching network-neuroscience framework, and bridges the classic divide between unitary and multi-factorial models of intelligence. Given that our machine learning analysis pipeline aligns naturally with multivariate network coding whereas more commonly applied univariate methods do not, we believe that the analysis of multivariate network states as applied here has untapped potential in clinical research; e.g., providing functional markers for quantifying the impact of pathologies and interventions on the brain’s capacity to flexibly express task optimised network states11,29.