Reconstructing the cascade of language processing in the brain using the internal computations of transformer language models
Sreejan Kumar, Theodore Sumers, Ariel Goldstein, Uri Hasson, Kenneth Norman, Thomas Griffiths, Robert Hawkins, Samuel Nastase, Princeton University, United States; Takateru Yamakoshi, University of Tokyo, Japan
Session:
Posters 2 Poster
Location:
Pacific Ballroom H-O
Presentation Time:
Fri, 26 Aug, 19:30 - 21:30 Pacific Time (UTC -8)
Abstract:
Piecing together the meaning of a narrative requires understanding not only the individual words but also the intricate relationships between them. How does the brain construct this kind of rich, contextual meaning from natural language? In this paper, we deconstruct the circuit computations of a transformer-based language model (BERT) and analyze the "transformations" at each layer, alongside the more commonly studied "embeddings." Using functional MRI data acquired while participants listened to naturalistic spoken stories, we find that these transformations capture a hierarchy of linguistic computations across cortex and that the emergent syntactic computations performed by individual attention heads correlate with brain activity predictions in specific cortical regions, consistent with a shared pattern of functional specialization. These heads fall along gradients in a low-dimensional cortical space corresponding to different layers and contextual distances.
