Michael W. Cole, Jeremy R. Reynolds, Jonathan D. Power, Grega Repovs, Alan Anticevic, & Todd S. Braver
Nature Neuroscience (2013)
See the Washington University press release
Full Supplementary Results
What are flexible hubs?
Flexible hubs are brain regions that coordinate activity throughout the brain to implement tasks – like the conductor of a large orchestra. However, unlike a typical conductor who has months or years of practice playing a piece with an orchestra, flexible hubs must coordinate brain activity for tasks that have never been performed previously.
The Flexible Hub Theory suggests this is possible because flexible hubs build up a repertoire of task component connectivity patterns that are highly practiced and can be reused in novel combinations in situations requiring high adaptivity. It's as if a conductor practiced short sound sequences with each section of the orchestra separately, then on the day of the performance began gesturing to some sections to play back what they learned, creating a new song that has never been played or heard before.
What does this study show that is novel?
This study tested a novel hypothesis for how humans are able to implement highly adaptive behavior, such as rapidly learning new tasks from instructions. This hypothesis suggests select regions of the brain are 'flexible hubs' that coordinate brain networks to perform new tasks (e.g., orchestrating visual and motor information in a novel visuo-motor task).
How does the Flexible Hub Theory differ from other theories?
As outlined in the manuscript, the Flexible Hub Theory is built upon previous theories of attention and task conflict resolution. Specifically, the Flexible Hub Theory is based on the Guided Activation Theory of Miller & Cohen (2001), which was in turn based on the Biased Competition Theory of Desimone & Duncan (1995). These theories postulate that the lateral prefrontal cortex maintains task context and biases activity throughout the brain (directly or indirectly) in order to overcome conflict from strong preexisting associations. In contrast to these theories, the Flexible Hub Theory postulates: 1) This mechanism is highly adaptive, allowing for rapid updating of top-down influences to quickly learn novel tasks in humans; 2) Functional connectivity patterns are reused compositionally by flexible hubs to facilitate rapid updating of top-down influences in a coherent, adaptive manner; 3) These mechanisms are generalized to other regions of the fronto-parietal control network (i.e., it's not restricted to lateral prefrontal cortex).
Do flexible hubs act as hubs in all task contexts?
This is unlikely, though it is possible in some task contexts. The Flexible Hub Theory characterizes flexible hubs as having widespread intrinsic functional connectivity (measurable imperfectly using, e.g., resting state fMRI), but that a flexible hub's functional connectivity will flexibly shift to whatever configuration is necessary to coordinate the functioning of task-relevant regions – perhaps reducing to a highly selective set of functional connections in any given task context. Thus, it is likely the case that a flexible hub – despite having adequate direct + indirect structural connectivity to be considered a hub – will nonetheless not act as a global hub during any given task state.
Is a flexible hub a truly explanatory construct, or does it simply describe a homunculus (little man) that implements control?
Flexible hubs are a systems-level mechanism that is distinct from properties typical of whole persons in several respects. First, there are many flexible hubs working together, with the implementation of adaptive task control likely possible only with all (or most) of the repertoire of flexible hubs available. Second, the identified flexible hubs appear to be specialized for adaptive task control, such that they are not able to do much that an entire person can do (e.g., process emotions, perceive colors). Finally, unlike whole persons, flexible hubs likely require other supporting mechanisms to function. It will be left to future research to identify the supporting sub-systems-level mechanisms (e.g., local within-region neural interactions) and supporting systems-level mechanisms (e.g., interactions with thalamus and/or basal ganglia) required for flexible hubs to function. See Cole, Laurent, et al. (2013) for a theoretical account of these supporting mechanisms.
What is the fronto-parietal control network (FPN)?
The FPN is a neocortical set of regions – primarily in lateral frontal and posterior parietal cortices – that are highly inter-connected both structurally and functionally. We've operationally defined the network using the regions and network partition from Power et al. (2011). The partition consists of 25 regions in neocortex, with most in lateral prefrontal cortex, some in posterior parietal cortex, and one in medial prefrontal cortex.
What is global variable connectivity?
This is a theoretical concept that is predicted by the Flexible Hub Theory. It describes brain regions that shift their functional connectivity (strength of temporal correlation) with many other regions across many task contexts.
What is compositional coding?
This is also a theoretical concept predicted by the Flexible Hub Theory. It describes the reuse of functional connectivity patterns (which likely represent task information) across multiple contexts, allowing these patterns to act as components that retain their function when combined with other pattern components. This allows for rapidly novel functionality from new mixtures of pattern components.
What is the global variability coefficient (GVC)?
GVC is a new graph theoretical measure for identifying global variable connectivity that consists of: 1) Estimating the standard deviation of a given node's (i.e., brain region's) connectivity strengths with all other nodes across many contexts, 2) Averaging these standard deviations across all nodes, summarizing the source node's tendency to shift connectivity across many nodes across many contexts.
What is the participation coefficient (of variable connectivity)?
The participation coefficient estimates the uniformity of a node's connectivity across the available networks/modules. A value of 0 means the source node's connectivity remains completely within that node's network, while a value of 1 means the source node's connectivity is uniformly distributed among all networks. We applied this measure to variable connectivity (the standard deviations across the 64 task contexts) in order to determine the uniformity of flexible hub connectivity shifts across networks.
What is gPPI?
Psycho-physiological interaction (PPI) estimates changes in 'functional connectivity' (functional connectivity is the similarity of activity time series across parts of the brain) across task contexts. Generalized psycho-physiological interaction (gPPI) was developed by McLaren et al. (2012) as an alternative to the classic PPI approach (Friston et al., 1997). The gPPI approach was developed, in part, to allow for estimating more than two task regressors per statistical model. This ability was essential for the present study due to the need to model 64 task states.
How was gPPI applied differently in this study?
In contrast to the gPPI implementation described by McLaren et al. (2012), we removed the deconvolution step. We did this for two reasons: 1) The deconvolution does not take into account known differences in hemodynamic response shapes across regions and subjects, which can result in distortion of data; and 2) It has not been empirically shown that the deconvolution step improves estimates relative to any sort of gold standard. In addition to differences from the gPPI approach, there are also differences from an approach similar to gPPI used by the FSL group (O'Reilly et al., 2012). The main difference is the use of binary (i.e., 'boxcar') task timing as the "psychological" parameter in the PPI, which (in combination with not de-meaning the regressors) results in rest-period fluctuations being removed from the PPI regressors. This likely reduces noise in the estimation of PPI effects.
What is rapid instructed task learning (RITL)?
RITL (pronounced "rittle") is "the ability to rapidly learn task procedures from instructions" (Cole, Laurent, et al., 2013). This cognitive control ability is most highly developed in humans, with only minimal competency in non-human primates. The Flexible Hub Theory suggests that flexible hubs implement key mechanisms allowing for the implementation of RITL by the human brain.
Are there likely to be other kinds of flexible hubs?
We have identified 'adaptive task control' flexible hubs by analyzing the distribution of functional connections across performance of many novel tasks. As noted in the manuscript (see the Discussion section), there are likely flexible hubs for other sorts of task control, such as stable task control over long time periods of task performance (and/or when task performance is highly practiced). There may also be flexible hubs for coordinating shifts of attention (perhaps in the dorsal attention network). Further, flexible connectivity may be a general property of brain dynamics, with incoming activity patterns to a brain region determining where output signals go (i.e., based on the context). Therefore it may be the case that local/provincial flexible hubs exist throughout the brain (e.g., perhaps V4 in the visual system). There may also be global/connector flexible hubs for other functions requiring whole-brain coordination, such as arousal based on emotional context.
Are there likely to be adaptive task control flexible hubs in networks other than the fronto-parietal network?
In exploratory analyses of the reported data we found nodes in other networks (e.g., the dorsal attention network) that had high GVC and flexible connectivity participation. However, the fronto-parietal network had by far the highest density of flexible hubs.
What other sorts of hubs are there than 'flexible' hubs?
We hypothesize that the major alternative to flexible hubs is 'broadcast hubs'. In contrast to flexible hubs, broadcast hubs send signals to all (or most) of its connections when active, rather than sending signal to a select set of connections depending on context. It will be important for future work to test this hypothesis.
How was the flexible connectivity participation coefficient analysis performed?
The analysis procedure involved calculating the variable connectivity matrix (the same as used for GVC), thresholding that matrix by density, and calculating the participation coefficient on the resulting (thresholded) matrix. All of this was done in MATLAB, with the density thresholding accomplished with the BCT function threshold_proportional and the final step accomplished using the BCT function participation_coef.
Global variability coefficient (MATLAB code): gvc.m
Miller, E., & Cohen, J. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24, 167–202.
Desimone, R., & Duncan, J. (1995). Neural mechanisms of selective visual attention. Annual Review of Neuroscience, 18, 193–222.
Cole, M. W., Laurent, P., & Stocco, A. (2013). Rapid instructed task learning: A new window into the human brain's unique capacity for flexible cognitive control. Cognitive, affective & behavioral neuroscience, 13(1), 1–22. doi:10.3758/s13415-012-0125-7
Power, J. D., Cohen, A. L., Nelson, S. M., Wig, G. S., Barnes, K. A., Church, J. A., et al. (2011). Functional Network Organization of the Human Brain. Neuron, 72(4), 665–678. doi:10.1016/j.neuron.2011.09.006
McLaren, D. G., Ries, M. L., Xu, G., & Johnson, S. C. (2012). A generalized form of context-dependent psychophysiological interactions (gPPI): a comparison to standard approaches. NeuroImage, 61(4), 1277–1286. doi:10.1016/j.neuroimage.2012.03.068
Friston, K. J., Buechel, C., Fink, G. R., Morris, J., Rolls, E., & Dolan, R. J. (1997). Psychophysiological and modulatory interactions in neuroimaging. NeuroImage, 6(3), 218–229. doi:10.1006/nimg.1997.0291
O'reilly, J. X., Woolrich, M. W., Behrens, T. E. J., Smith, S. M., & Johansen-Berg, H. (2012). Tools of the trade: psychophysiological interactions and functional connectivity. Social Cognitive and Affective Neuroscience, 7(5), 604–609. doi:10.1093/scan/nss055
Cole, M. W., Bagic, A., Kass, R., & Schneider, W. (2010). Prefrontal dynamics underlying rapid instructed task learning reverse with practice. The Journal of neuroscience : the official journal of the Society for Neuroscience, 30(42), 14245–14254. doi:10.1523/JNEUROSCI.1662-10.2010