Discussion

In this study, we used intersubject analyses to determine the relative roles of activity and connectivity during the learning of object values over the course of four days of task practice. Importantly, our analyses were able to separate evidence for group constraints on large-scale neuronal processes supporting learning from evidence for idiosyncratic processes that may be relevant for task performance in single subjects. In general, our results suggest that connectivity and activity tend to be constrained to the group average in a similar manner across brain regions, but in a distinct temporal manner. Greater inter-subject correlations in both activity and connectivity exist in early versus late learning, suggesting that the early stages of learning recruit common neurophysiological processes necessary for encoding information. We observed greater inter-subject correlations in activity than functional connectivity on day one, likely because stimulus induced activity is very similar across individuals. However, we observed greater inter-subject correlations in functional connectivity than in activity late in learning, suggesting a role of connectivity in the consolidation of learned information and greater individual variation in activity patterns in late learning that might serve to optimize behavior in an idiosyncratic manner. Collectively, our findings demonstrate that subjects express common stimulus-induced activity and connectivity that evolves based on the stage of the learning process.

Intersubject brain network pattern during value learning

Are some cognitive systems more constrained in their activity and connectivity than others? We identified several data-derived network modules reminiscent of cognitive systems – visual, sensorimotor, and default mode modules – that demonstrated significantly greater intersubject correlation and intersubject functional connectivity than other modules. In particular, we observed significantly higher ISC and ISFC within lateral occipital cortex, lingual gyrus, fusiform gyrus, and pericalcarine cortex, as well as other areas of the occipital lobe, suggesting that value learning commonly recruits brain regions responsible for visual perception, visual memory, and decision making based on chosen value \cite{kuai2013learning}. We also observed significant ISC and ISFC in temporal and parietal areas involved in attention, mentalizing, and the attribution of self-belief  \cite{corbetta2002control,corbetta2011spatial}. It is intuitively plausible that the significant ISC and ISFC in visual and sensory motor systems is a direct result of shared sensory inputs. However, the significant ISC and ISFC in  the default mode system is unlikely to be adequately explained by the same mechanism, due to the tendency of this system to be active independent from stimulus input \cite{Margulies_2016}. An alternative explanation lies in the fact that default mode areas typically display high degree in network models, comprising a large part of the rich club \cite{Bertolero_2017}, supporting the notion that the default-mode forms a stable core of areas that exhibit similar activity and connectivity across subjects. 
Interestingly, we observed that several brain regions displayed ISFC that exceeded their ISC, including the amygdala, thalamus, precuneus, cuneus, cingulate gyrus, and lateral frontal cortex areas. Collectively, these regions are commonly associated with higher cognitive functions such as flexibility of thinking, problem solving, cognitive inhibition, and attentional control \cite{corbetta2002control}. One explanation for the high ISFC exhibited by these regions is that they represent important cognitive control areas that are likely to act as functional hubs that may be commonly integrated with other brain regions across subjects \cite{Hwang_2016,Bertolero_2017,Bertolero_2015}. Specifically, frontal pole frequently interacts with the anterior cingulate gyrus during reward-guided learning \cite{rushworth2011frontal}, and the precuneus – although widely known to serve as an important hub in the default mode \cite{raichle2001default} – is often activated during episodic memory retrieval and self-processing operations \cite{cavanna2006precuneus}. The higher ISFC at the precuneus might be related to the higher reaction times in correct value trials observed during value learning tasks \cite{oishi2005activation}, which is thought to be driven in part by a subject's memory of previous similar tasks. In summary, more constrained brain activity appears to occur within predominantly sensory-specific brain regions and more constrained functional connectivity appears to occur in integrative areas that are commonly associated with higher cognitive function.

Time-dependent variation in intersubject correlations during learning

An advantage of our experimental approach is the ability to examine dynamic changes in neuronal processes, and their variation across subjects, throughout different phases of value learning. During the early stages of learning, we observed high intersubject correlations in activity, suggesting that participants generally exhibited a similar time course of BOLD activation in response to the task demands. During later stages of learning, we observed a decrease in the ISC. We speculate that this shift from high to low ISC might reflect individualized patterns of activity that support inter-individual variability in learning strategies as well as the fact that in late stages of learning greater neural real estate is available for contemporaneous, non-task-related processing.  In contrast to the ISC, the ISFC gradually increased between the first and second day and subsequently decreased in day 3, thereafter remaining relatively constant. This empirically observed trajectory of ISFC is reminiscent of the typical shape of a traditional learning curve \cite{rescorla1971variation}, suggesting that functional connectivity may comprise an general adaptive mechanism for learning \cite{baeg2007learning, fatima2016dynamic, patel2013functional}
Importantly, we found that putative cognitive systems displayed differential patterns of learning-induced ISFC and ISC. The high ratio of ISFC-to-ISC observed in the sensorimotor system could be attributed to common inputs to the sensory system for the processing of stimulus information. The complimentary increase in the ISFC of the visual and sensorimotor systems may result from trained motor coordination of hand and finger movements to press the button for task response \cite{bassett2011dynamic,toni2001learning}. The resulting changes suggest a role in visual and sensorimotor systems in supporting visual identification of objects, interpretation of value, and motor coordination of hand-finger movement to guide more accurate and efficient decision making after the critical learning period. The lower ISC in the later stage of learning could be ascribed to higher individual differences in time courses of BOLD signals after adaptation to the task and once the procedure is over learned and mastered in the early stage \cite{bastian2008understanding,keller2017stimulus}. Interestingly, we find differential involvement of the default mode over the course of value learning: connectivity within the default mode system becomes more generalized across subjects, and activity within the default mode system becomes more individualized across subjects. Although the default mode system is often viewed as a task-negative system that is typically more active during resting state processes, here we report evidence that the architecture of this system is highly conserved across learning.

Functional network drivers of value learning

Finally, it is important to consider the question of whether, and to what degree, intersubject correlations in activity and connectivity support or hamper learning. We find that ISFC, but not ISC, was significantly correlated with task accuracy, suggesting that behavior might be improved when the functional network reorganizes according to subject-general constraints on neural processes, but that stimulus induced activity and learning related activity constraints are not relevant to learning performance. While individual variation in stimulus-induced activity might be responsible for modulating subject-specific behavior during value learning tasks \cite{gerraty2014transfer,laird2011behavioral,yamashita2015predicting}, the extent to which the activity deviated from the group level constraints did not hamper or help performance. 
Our observation that the relationship between ISFC and task accuracy was particularly localized in the sensorimotor, DMN, and subcortical regions including putamen, thalamus, and caudate agrees with previous findings that these regions are critical to various forms of reward-based learning including value learning \cite{haruno2004neural}. The putamen and caudate nucleus are part of the dorsal striatum, and together play an important role in decision making, a cognitive capability that requires determining value \cite{delgado2008role}. In particular, the positive relationship between task accuracy and caudate ISFC supports a role for the dorsal striatum in modulating behavior by interacting with other brain systems during feedback-related learning \cite{seger2005roles}. Our results suggest that interregional coupling of brain activity with the dorsal striatum may be a promising focal point of future work in the identification of target brain areas that contribute to individual deficits in learning more generally.

Methodological Considerations

The number of subjects analyzed here is smaller than large-scale data collections \cite{Van2013}. However, given the four day experimental involving the same subjects, this dataset nevertheless represents a rich opportunity to analyze the temporal evolution of constraints on activity and connectivity during learning. We hope that future large-scale data collections will include similar experiments. We chose a simple and intuitive method to measure the similarity of brain activity and connectivity across subjects, but certainly more sophisticated methods could be developed and will likely uncover results that are complimentary to the results found here. Finally, there are many ways to parcellate the brain into nodes for network analyses. While many parcellations maximize the similarity of functional connectivity within each parcel \cite{Schaefer2017}, given that we compared activity to connectivity, we chose an anatomical parcellation so as not to bias our results towards activity or connectivity. Moreover, this atlas has been used extensively in prior literature.

Conclusions

Brain activity and functional connectivity display specific temporal constrains across individuals during distinct stages of learning the value of novel objects. Activity is most similar across subjects during stimulus encoding in learning (day one), while the similarity of functional connectivity peaks during the early stage of learning (day two). While constraints on both activity and functional connectivity decrease in later stages of learning (days three and four), functional connectivity remains more constrained than activity. Critically, how constrained an individual's connectivity is to the group level is predictive of task accuracy.