Model-based analysis of brain activity reveals the hierarchy of language in 305 subjects

  • 2021-10-12 15:30:21
  • Charlotte Caucheteux, Alexandre Gramfort, Jean-RĂ©mi King
  • 1


A popular approach to decompose the neural bases of language consists incorrelating, across individuals, the brain responses to different stimuli (e.g.regular speech versus scrambled words, sentences, or paragraphs). Althoughsuccessful, this `model-free' approach necessitates the acquisition of a largeand costly set of neuroimaging data. Here, we show that a model-based approachcan reach equivalent results within subjects exposed to natural stimuli. Wecapitalize on the recently-discovered similarities between deep language modelsand the human brain to compute the mapping between i) the brain responses toregular speech and ii) the activations of deep language models elicited bymodified stimuli (e.g. scrambled words, sentences, or paragraphs). Ourmodel-based approach successfully replicates the seminal study of Lerner et al.(2011), which revealed the hierarchy of language areas by comparing thefunctional-magnetic resonance imaging (fMRI) of seven subjects listening to7min of both regular and scrambled narratives. We further extend and precisethese results to the brain signals of 305 individuals listening to 4.1 hours ofnarrated stories. Overall, this study paves the way for efficient and flexibleanalyses of the brain bases of language.


Quick Read (beta)

loading the full paper ...