Language in the brain and algorithms.

Jean-Rémi KING : Chercheur à l’Ecole Normale Supérieure de Paris et Meta AI.

It took less than five years for AI to revolutionize the field of natural language processing. Can the resulting algorithms help us understand this complex cognitive ability? Here, we develop state-of-the-art language models and compare them to the adult human brain. To do so, we are integrating large-scale datasets of magneto/electroencephalography, functional magnetic resonance imaging, and intracranial recordings while individuals read or listen to natural sentences. Our analyses demonstrate systematic similarities between AI models and the brain, which allow us to (1) efficiently decode speech from brain activity and (2) delineate the location and temporal course of the hierarchy of language representations. Overall, these studies show how the interface between AI and neuroscience sheds new light on the computations at the heart of human language.