Year: 2020
Type: dissertation
Abstract: Linguistic communication is a unique characteristic of intelligent behaviour \nthat distinguishes humans from non-human animals. Natural language is \na structured, complex communication system supported by a variety of cognitive \nfunctions, realized by hundreds of millions of neurons in the brain. Artificial \nneural networks typically used in natural language processing (NLP) are often \ndesigned to focus on benchmark performance, where one of the main goals is \nreaching the state-of-the-art performance on a set of language tasks. Although \nthe advances in NLP have been tremendous in the past decade, such networks \nprovide only limited insights into biological mechanisms underlying linguistic \nprocessing in the brain. \n \nIn this thesis, we propose an integrative approach to the study of \ncomputational mechanisms underlying fundamental language processes, spanning \nbiologically plausible neural networks, and learning of basic communicative \nabilities through environmentally grounded behaviour. In doing so, we argue for \nthe usage-based approach to language, where language is supported by a variety \nof cognitive functions and learning mechanisms. Thus, we focus on the three \nfollowing questions: How are basic linguistic units, such as words, represented \nin the brain? Which neural mechanisms operate on those representations in \ncognitive tasks? How can aspects of such representations, such as associative \nsimilarity and structure, be learned in a usage-based framework? \n \nTo answer the first two questions, we build novel, biologically realistic \nmodels of neural function that perform different semantic processing tasks: the \nRemote Associates Test (RAT) and the semantic fluency task. Both tasks have \nbeen used in experimental and clinical environments to study organizational \nprinciples and retrieval mechanisms from semantic memory. The models we propose \nrealize the mental lexicon and cognitive retrieval processes operating on that \nlexicon using associative mechanisms in a biologically plausible manner. We \nargue that such models are the first and only biologically plausible models \nthat propose specific mechanisms as well as reproduce a wide range of human \nbehavioural data on those tasks, further corroborating their plausibility. \n \nTo address the last question, we use an interactive, collaborative agent-based \nreinforcement learning setup in a navigation task where agents learn to \ncommunicate to solve the task. We argue that agents in such a setup learn to \njointly coordinate their actions, and develop a communication protocol that is \noften optimal for the performance on the task, while exhibiting some core \nproperties of language, such as representational similarity structure and \ncompositionality, essential for associative mechanisms underlying cognitive \nrepresentations.
Source: UWSpace (University of Waterloo)
Author: Ivana Kajić
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W3105187925