Dynamic network coding of working-memory domains and working-memory processes

Abstract

The classic mapping of distinct aspects of working memory (WM) to mutually exclusive brain areas is at odds with the distributed processing mechanisms proposed by contemporary network science theory. Here, we use machine-learning to determine how aspects of WM are dynamically coded in the human brain. Using cross-validation across independent fMRI studies, we demonstrate that stimulus domains (spatial, number and fractal) and WM processes (encode, maintain, probe) are classifiable with high accuracy from the patterns of network activity and connectivity that they evoke. This is the case even when focusing on ‘multiple demands’ brain regions, which are active across all WM conditions. Contrary to early neuropsychological perspectives, these aspects of WM do not map exclusively to brain areas or processing streams; however, the mappings from that literature form salient features within the corresponding multivariate connectivity patterns. Furthermore, connectivity patterns provide the most precise basis for classification and become fine-tuned as maintenance load increases. These results accord with a network-coding mechanism, where the same brain regions support diverse WM demands by adopting different connectivity states.

Contrasting hierarchical and multiple-demand accounts of frontal lobe functional organisation during task-switching

“…This latter notion of domain generality accords well with the observation of functional dissociations within MD, given the capacity of multivariate machine learning algorithms to identify tasks based on both activation and connectivity patterns therein (Crittenden, Mitchell and Duncan, 2016;Soreq, Leech and Hampshire, 2019). Relatedly, we have argued that it may be inappropriate to impose strong functional dissociations on what may essentially be a multivariate system that supports tasks through a many-to-many mapping system (Hampshire and Sharp, 2015;Lorenz, Hampshire and Leech, 2017).…”

“…Our in-house developed watershed transform (Grant et al, 2018;Soreq, Leech and Hampshire, 2019;Hampshire et al, 2020) was used to segment functional activation maps into discrete clusters in a data-driven manner. This common segmentation considers a 3D statistical volume (e.g., activation map) as a multi-dimensional surface where high and low intensities represent elevations.…”

Optimisation of functional network resources when learning behavioural strategies for performing complex tasks

“…Some studies have attempted to address this by developing paradigms where individuals randomly volunteer when to task-switch 10,[13][14][15] . However, these paradigms fail to capture the natural inclination to apply structure to behaviour 16,17 and to optimise behavioural strategies through learning [18][19][20][21][22][23] .…”

“…However, such one-to-one mappings may not capture the dynamic brain mechanisms that support cognition and organise behaviour [37][38][39][40] . Conversely, the contemporary network science view proposes that cognitive controls are fundamentally non-localisable because they are supported by dynamic interactions that occur across distributed networks of brain regions 20,22,[41][42][43][44][45][46][47][48][49][50] . Indeed, distributed activations, including aPFC, can occur during task-switching 8,[51][52][53][54][55][56] , instruction based learning (IBL) 20,21,26 , relational integration 46 and even during tasks that lack overt requirements for abstraction or hierarchy in cognitive processing 57 .…”

“…More broadly, network neuroscience studies of cognitive control demonstrate that when engaging with a task there is a substantial reorganization of the brain's functional connectome 49,60,61 , characterised by the transition from a state of high-activation to a low-activation highconnectivity state [18][19][20][21][22][23]62 , and this transition rate may depend on learning. Critically, the established task-state expresses a stable 44,62 connectivity pattern that is specific enough to the task 22,47,50,[63][64][65] , cognitive load 22 , performance 66 and clinical abnormalities [67][68][69] to be decoded with machine learning.…”