Understanding how experience shapes language development through comparisons of large language models and neural representations - Project Summary Language experience plays a critical role in language development across childhood. However, in studying humans, it is difficult to disentangle the effects of language experience from non-linguistic factors that also change across childhood and adolescence. Recent advances in AI have led to the emergence of large language models (LLMs), which not only are exceptional at producing language, but also share remarkably similar representations with the adult human brain. Building on my extensive prior training in cognitive neuroscience and language development, the proposed research will combine precision functional neuroimaging and computational modeling to systematically test the role of language experience in the development of the brain’s language network. Specifically, I will apply model-to-brain similarity approaches that have been developed in work with adults to characterize developmental changes in how language is represented by the human brain, and test whether LLMs trained on age-matched amounts of language data recapitulate developmental shifts in the linguistic representational space. First, I will use functional magnetic resonance imaging (fMRI) to measure brain responses to a diverse set of sentences in young children (N=12, ages 6-8), older children (N=12, ages 10-12), adolescents (N=12, ages 14-16), and two groups of adults (N=8, ages 20-22; and N=8, ages 28-30). Next, I will train an LLM on different amounts of language data, corresponding to the estimated amount of language experience across the lifetime in each of these age groups. Finally, I will test the critical hypothesis that representations from an LLM will best align with neural representations of the corresponding age group. The proposed research will support my fellowship goals of gaining proficiency in the use of computational modeling, especially LLMs, as a tool in cognitive neuroscience research, and continuing to develop expertise in state-of-the-art fMRI data collection and analysis methods. An additional goal during the fellowship period is to develop practical skills to prepare for being a principal investigator. Dr. Evelina Fedorenko’s lab at MIT, and the department of Brain and Cognitive Sciences more broadly, is the optimal place to conduct both the proposed research and my postdoctoral training: on a lab-, department-, and institution-wide scale, there is enthusiasm, expertise, and ample resources to apply AI and modeling approaches to scientific inquiry. I will take advantage of opportunities within MIT and the greater Boston area’s academic environment to attend multidisciplinary talks, engage in coursework with experts particularly in computational tools, and collaborate with other researchers in related fields.
