Thesis defense Kristijan Armeni (Donders series 498)

30 March 2021

Promotor: prof. P. Hagoort
Co-promotor: dr. J.M. Schoffelen

On model-based neurobiology of language comprehension: neuronal oscillations, processing memory, and prediction.

In contextually rich language comprehension settings listeners can rely on past context and to generate predictions about the upcoming words. Neuroscientific theories propose that some cognitive processes related to prediction might be observed in the brain as neuronal oscillations -- rhythmic firing of a large pool of neurons. To perform such feats of prediction, neural circuits must implement some form of processing memory to maintain the past information and use it while processing the new input. In this thesis, we investigated aspects of the neurobiology of language in naturalistic language comprehension using magnetoencephalography (MEG) and computational models of prediction and memory. We begin by reviewing the use of computational linguistic techniques in cognitive neuroscience. In chapter 2, we report an MEG study on oscillatory predictive processing. We show that slow theta-band dynamics are increased in unpredictable contexts, likely reflecting lexical computations, and that faster beta-band dynamics are increased in more predictable contexts possibly reflecting context maintenance. In chapter 3, we describe an MEG dataset for building artificial neural networks of brain dynamics. We provide a brief validation analysis showing accurate and robust localization of MEG dynamics in primary auditory areas. In chapter 4, we used recurrent neural networks (RNNs) to model MEG dynamics on dataset from chapter 3. We show that RNNs reliably predict the MEG signal in the known higher-level areas of the language network. In chapter 5, we investigated a neuronal substrate for memory: neuronal spike-rate adaptation. We show that making neurons more adaptive resulted in progressively higher performance in a working memory task. The results suggest neuronal adaptation plays an important role as memory mechanism in language. In conclusion, we briefly discuss on the potential challenges of model-based approaches in cognitive neuroscience of language and call for a stronger focus on biologically motivated architectures.