The neural architecture of language: Integrative modeling converges on predictive processing

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Neuroimaging

Background:

• The neuroscience of perception has been advanced by integrative modeling linking computation, brain function, and behavior.
• This approach reveals trends across computational models, offering insights into cognitive and neural mechanisms.

Purpose of the Study:

• To apply integrative modeling to higher-level cognition, specifically human language processing.
• To investigate the predictive processing mechanisms in the human brain during language comprehension.

Main Methods:

• Utilized advanced "transformer" computational models to analyze neural responses to language.
• Employed functional magnetic resonance imaging (fMRI) and electrocorticography (ECoG) for neural data.
• Correlated model performance with neural and behavioral data, assessing generalization across datasets and modalities.

Main Results:

• Transformer models predicted nearly 100% of explainable variance in neural responses to sentences.
• Model performance on next-word prediction strongly correlated with neural and behavioral fits.
• Model architecture significantly influenced the accuracy of neural fits, termed "brain score".

Conclusions:

• Predictive processing is a fundamental mechanism shaping language comprehension in the human brain.
• Integrative modeling provides computationally explicit evidence for cognitive and neural theories of language.
• Advanced AI models serve as powerful tools for understanding complex human cognitive skills.