Poster A68, Tuesday, August 20, 2019, 10:15 am – 12:00 pm, Restaurant Hall
The network architecture for natural language in the brain is not fixed
Sarah Aliko1,2, Florin Gheorghiu1, Jeremy I Skipper1;1Experimental Psychology, University College London, 2London Interdisciplinary Doctoral Program
Introduction. What is the network architecture of spoken language comprehension and how is it integrated into whole brain or global network dynamics? Most current models of the neurobiology of language assume a modularity and domain specificity with terms like ‘the language network’ or ‘dual-streams’ and a corresponding set of static brain regions. These models are based on data that ‘assume the antecedent’ in that they derive from unnatural language stimuli and tasks or ‘language localisers’ and rely on averaging and task-subtractions. We take a step back to observe the organisation of natural language in the context of the whole brain. Under these conditions, we propose that language is situated in a core-periphery global network architecture, involving a set of densely connected core regions and a set of sparsely but well-connected peripheral regions. We predict that the core is not comprised of traditional language regions with those, rather, occupying a dynamic and spatially unfixed periphery. Methods. Thirty participants watched two full-length movies (‘500 Days of Summer’ and ‘Citizenfour’), during 1.5T functional magnetic resonance imaging. Movies were labelled for speech (among other things) using machine learning approaches. After preprocessing, dynamic functional connectivity was conducted in running windows of one minute at one second (TR) steps. Voxel-wise graph theoretical measures were sorted into windows of high (M=11.68) and low (M=2.19) numbers of words. Spatial and temporal independent component analysis (ICA) were used to locate stable states. Results were also correlated with ‘gold-standard’ meta-analyses from the neurosynth database. Results. The global network structure had a core-periphery architecture, with primary motor/auditory/visual core regions for high-speech, and primary visual and prefrontal regions for low-speech. Consistent with a core-periphery arrangement, networks changed dynamically over time, with hundreds of possible temporal states in the brain but no stable language networks. Confirming this, only spatial ICA resulted in language-related labels from meta-analyses, and this at low correlation values (max r < 0.35). Furthermore, we found that there were 2-9 network communities during high-word windows originating from the ‘language’ meta-analysis, and 6-16 in the rest of the brain. On average, 99% of ‘language’ communities were part of a community outside of those regions, activating 57% of the rest of the brain in varying spatial configurations. Discussion. The neurobiology of natural language comprehension is situated in a global core-periphery network architecture whose core does not involve traditional language regions. The periphery involves many dynamically-reconfiguring communities that only partially overlap and extend well beyond those regions. Indeed, these communities encompass most of the rest of the brain in a pattern that is not spatio-temporally fixed. We suggest this core-periphery arrangement is the kind of architecture required to support a complex and dynamic behaviour like language, composed of many (putative) subcomponents (e.g., phonology, semantics, etc.), each of which are arguably contextually determined. As these are not the usual conditions under which the neurobiology of language has previously been examined, it is perhaps unsurprising that the idea of fixed language regions emerged.
Themes: Speech Perception, Computational Approaches
Method: Functional Imaging