From sequences to cognitive structures : neurocomputational mechanisms

Abstract

Understanding how the brain forms representations of structured information distributed in time is a challenging neuroscientific endeavour, necessitating computationally and neurobiologically informed study. Human neuroimaging evidence demonstrates engagement of a fronto-temporal network, including ventrolateral prefrontal cortex (vlPFC), during language comprehension. Corresponding regions are engaged when processing dependencies between word-like items in Artificial Grammar (AG) paradigms. However, the neurocomputations supporting dependency processing and sequential structure-building are poorly understood. This work aimed to clarify these processes in humans, integrating behavioural, electrophysiological and computational evidence. I devised a novel auditory AG task to assess simultaneous learning of dependencies between adjacent and non-adjacent items, incorporating learning aids including prosody, feedback, delineated sequence boundaries, staged pre-exposure, and variable intervening items. Behavioural data obtained in 50 healthy adults revealed strongly bimodal performance despite these cues. Notably, however, reaction times revealed sensitivity to the grammar even in low performers. Behavioural and intracranial electrode data was subsequently obtained in 12 neurosurgical patients performing this task. Despite chance behavioural performance, time- and time-frequency domain electrophysiological analysis revealed selective responsiveness to sequence grammaticality in regions including vlPFC. I developed a novel neurocomputational model (VS-BIND: “Vector-symbolic Sequencing of Binding INstantiating Dependencies”), triangulating evidence to clarify putative mechanisms in the fronto-temporal language network. I then undertook multivariate analyses on the AG task neural data, revealing responses compatible with the presence of ordinal codes in vlPFC, consistent with VS-BIND. I also developed a novel method of causal analysis on multivariate patterns, representational Granger causality, capable of detecting flow of distinct representations within the brain. This alluded to top-down transmission of syntactic predictions during the AG task, from vlPFC to auditory cortex, largely in the opposite direction to stimulus encodings, consistent with predictive coding accounts. It finally suggested roles for the temporoparietal junction and frontal operculum during grammaticality processing, congruent with prior literature. This work provides novel insights into the neurocomputational basis of cognitive structure-building, generating hypotheses for future study, and potentially contributing to AI and translational efforts.