Using machine learning to reveal neural mechanisms of word-learning across development

点击翻译按钮获取中文摘要

Immediately after birth, infants are immersed in a symphony of sounds that they will need to identify as they mature. Human language is a major part of this early soundscape, and within the first few months of life, infants have learned to parse out some of these sounds in language-specific categories. How infant learners can do this so early in life remains unsolved. A long-term goal of my research is to characterize the nature of early sound-based representations and the learning mechanisms that form and change these representations. Recent advances in neuroimaging research have revolutionized our ability to characterize the neural instantiation of individuals' linguistic representations. In the next five years, my team and I therefore pursue four objectives: Objective 1: Recruit machine-learning analysis tools to characterize the electro-encephalographic (EEG) signature of known and unknown words in adults Objective 2: Adapt this analysis and paradigm toolkit to developmental populations Objective 3: Examine the transformation of word representations as they are acquired via statistical learning, a mechanism believed to subserve early word-learning. Objective 4: Link the EEG signatures to behavioural measures of individual differences. Three sets of experiments address these goals. Experiment Set 1 (machine-learning analysis of adult EEG) addresses Objective 1 by using machine-learning classifiers to identify and compare the spatio-temporal EEG signatures of brain activity associated with words and non-words. It also addresses Objective 4 by correlating behavioural measures (e.g., response times to words vs nonwords) to machine-learning classifier performance. This set of experiments will increase our knowledge of mature word and non-word neural representations and will provide a proof-of-principle of the machine-learning analysis toolkit before it is applied to infant and child brain recordings. Experiment Set 2 adapts these paradigms and analysis tools to 5-7 year old children and 12- to 18-month-old infants (Objective 2). This will reveal where in the spatio-temporal EEG signature there is stability and where there is divergence in the organization of word and non-word representations across development. Experiment Set 3 addresses Objective 3 by teaching infants, children, and adults new words, which will allow us to examine how learners' representations transform given specific types of learning experience. Machine-learning analysis techniques have revolutionized neuroimaging research, but have only very recently been applied to developmental populations. We expect that our program of research will improve existing machine-learning neuroimaging analysis tools, particularly with respect to their application in developmental populations. Crucially, this work will improve our basic understanding of how human brains develop the fundamental ability to process and classify complex streams of sounds into meaningful words.