Neural Determinants of Statistical Learning. Bridging Language, Attention and Reward
Proficient language use requires knowledge of a language’s most basic components, words and rules. This knowledge not only allows us to express a vast number of ideas, wishes and wants, but also to understand others’ in real-time communication. Indeed, fluid conversation relies on our ability to predict upcoming information, such as the next word or sentence ending, which is facilitated by the fact that words and rules constitute probabilistic linguistic patterns. Interestingly, humans and other mammalian species show a remarkable sensitivity to patterned regularities in their sensory environment. Provided the appropriate cognitive machinery for speech processing, this ability, often dubbed statistical learning, suggests a route through which knowledge of words and basic rules may be acquired.
This dissertation investigates statistical learning as applied to the learning of words and simple language rules. From a neurocognitive perspective, the outputs of statistical learning are posited to have three main neural determinants, namely modality-specific processing networks, domain-general learning mechanisms able to produce statistical learning, and modulatory systems, such as attention, that can influence learning at any given time-point. More specifically, therefore, this dissertation investigates the statistical learning of words and rules by dedicating a study to each of these neural determinants, with the aim of providing a more integrated account of this extraordinary cognitive feat.
Study 1 examines the neural mechanism underlying modality-specific processes relevant for statistical learning. The combined results from behavioral testing, magneto-encephalography, and diffusion-weighted magnetic resonance imaging, reveal the importance of the spontaneous synchronization of auditory and speech motor cortical areas to speech inputs for the integration of auditory and motor representations and subsequent statistical learning. Even more strikingly, we demonstrate that approximately half the population shows a poor spontaneous synchronization, with the absence of synchrony having clear neurophysiological and neuroanatomical correlates with important consequences for the statistical learning of words.
Study 2 then investigates the implication and role of attentional mechanisms in the statistical learning of simple language rules from speech. Our results provide novel evidence for a dynamic engagement of distinct cortical networks for stimulus-driven and goal-directed attention during learning, effectively resolving long-standing controversies that characterize classic theoretical accounts. Specifically, through behavioral testing and functional neuroimaging, the early learning stages are shown to involve the interplay of stimulus-driven attention and prediction-based learning mechanisms as supported by left fronto-parietal cortical regions overlapping both attention and language networks. Later learning stages, in contrast, are associated with the engagement of a more dorsal and bilateral cortical network supporting the goal-directed attention to rule elements. Inducing cortical interference in this network via repetitive transcranial magnetic stimulation, we were also able to specify a role for goal-directed attention in both later learning and rule generalization.
Finally, by the combination of computational modelling and functional neuroimaging, Study 3 provokingly places the statistical learning of rules as an instance of reinforcement learning. Indeed, participants’ incidental rule-learning behavior conformed to prediction learning as captured by a classic model of reinforcement learning. Moreover, gradual prediction learning was strongly related to activity in the striatum, thus substantiating its implication in statistical learning while further specifying its role.
Altogether, by deconstructing the process of statistical learning into its main neural determinants, this dissertation provides a more detailed understanding of the complexity and dynamics of the phenomenon, with novel insights not only into what and where processes occur but also into how the different parts interact with each other over time to produce online statistical learning.