Dynamic constraints on functional architecture during learning
Next we asked whether constraints on activity and functional connectivity could distinguish between different phases of value learning. We expected that brain dynamics would be more highly constrained across subjects during early learning than during late learning, as indicated by high values of ISC and ISFC. This hypothesis is driven by the notion that in the later stages of learning, which are characterized by high performance accuracy and increasing automaticity of responses, greater neural real estate might be available for contemporaneous, non-task-related processing as well as processing reflecting subject-specific learning strategies. In late learning, we expected that functional connectivity would remain more heavily constrained than activity, due to its temporally extended nature being driven by commonly reinforced patterns of stimulus response.
To assess the verity of these expectations, we first examined the dynamics of ISC during value learning (Fig. 3A). We found that the average ISC over all brain regions increased from the first day to the second day, and then subsequently decreased through the fourth day (one-way ANOVA: \(F(3,447)=7.4,p=7.59\ \times10^{-5}\)). We also found that the average ISC estimated during task performance over all four days of training was significantly greater than the ISC estimated from resting state data acquired over those same four days (\(t=11.49,p=1.28\times 10^{-20}\); Fig. S2A-B). The low ISC during rest is consistent with the existence of minimal constraints on activity in the absence of a common stimulus. These trends in task-related ISC dynamics suggest that ISC might support two phases of value learning: (i) increased ISC between day one and day two may be associated with increasing constraints on activity, perhaps as a result of common neurophysiological mechanisms across subjects that facilitate early stage learning of the task mechanics, and (ii) decreased ISC from day two through day four may be associated with less constrained dynamics, perhaps as subjects explore diverse cognitive strategies to further increase their performance on the task. In a complementary analysis, we also examined the dynamics of ISFC during value learning, and found that ISFC significantly increased from day one to day two, peaking at day two (Fig. 3B).
Next, we explicitly tested for potential interactions between ISC and ISFC over time. We found that ISC and ISFC were significantly positively correlated on each of the four days (day 1: \(r=0.4196,\ p=0.0430\), day 2: \(r=0.4291,\ p=0.0000\), day 3: \(r=0.3497,\ p=0.0000\), day 4: \(r=0.4555,\ p=0.0000\)), suggesting that increased group-level constraints on activity are related to increased group-level constraints on functional connectivity. Consistent with this inference, we also observed that ISC and ISFC increased during early stage learning on days one and two, suggesting that activity and functional connectivity were both constrained to common organizational rules across the group. Nevertheless, we also observed some evidence for a divergence in the constraints on ISC and ISFC during the later stages of learning; for these comparisons, both ISC and and ISFC were normalized by subtracting the minimum and dividing by the range. Specifically, on day four we observed that normalized ISFC values were significantly greater than normalized ISC values (Fig. 3C), suggesting that activity is more autonomous than functional connectivity. These dynamics point to a potential driver-follower mechanism of
constrained activity preceding constrained functional connectivity over the four days of value learning.
Regional variability of functional constraints during learning
If the dynamics of whole-brain inter-subject correlations map on to different phases of value learning, it is natural to ask which brain systems might be most complicit in these phases. To address this question, we first partitioned brain regions into objectively defined functional modules using the GenLouvain community detection algorithm \cite{de2011generalized} (see Materials and Methods and the SI). Briefly, community detection is applied to the functional network constructed from data of each task session (raw subject functional connectivity, no ISFC) and parses brain regions into functional modules such that brain regions within the same module exhibit strong functional connections, and brain regions between different modules exhibit weak functional connections. To obtain a single representative partitioning of brain regions across subjects and scans, we computed the module allegiance matrix \cite{Bassett2015}, which encodes the probability that any two regions belong to the same functional module (Fig. 4A). By applying a final round of community detection to the module allegiance matrix, we identified seven modules that were associated with different putative brain systems, including a fronto-temporal module (FT) which covered most of the limbic lobe, a sensorimotor module (SM), an auditory module (AUD) including hippocampus and amygdala, the common default mode system (DMN), a language module (LAN), a visual module (VIS), and a subcortical module composed of the putamen, caudate, and thalamus (PCT) (Fig. 4B).