Brains and algorithms partially converge in natural language processing

Recent advancements in deep learning algorithms, particularly language transformers, have demonstrated remarkable capabilities in processing natural language. This study investigates the computational principles underlying the similarity between these algorithms and human brain activity during sentence processing. By analyzing fMRI and MEG data from 102 subjects reading Dutch sentences, and comparing these responses to the activations of a wide variety of deep language models, the researchers found that the similarity primarily depends on the algorithm's ability to predict words from context. This similarity also reveals the hierarchical development of perceptual, lexical, and compositional representations within cortical regions. The study concludes that modern language algorithms partially converge towards brain-like solutions, suggesting a promising path to understanding the foundations of natural language processing.

Charlotte Caucheteux, Jean-Rémi King