Abstract
The neuroscience of perception has recently been revolutionized with an integrative reverse-engineering approach in which computation, brain function, and behavior are linked across many different datasets and many computational models. We here present a first systematic study taking this approach into higher-level cognition: human language processing, our species’ signature cognitive skill. We find that the most powerful ‘transformer’ networks predict neural responses at nearly 100% and generalize across different datasets and data types (fMRI, ECoG). Across models, significant correlations are observed among all three metrics of performance: neural fit, fit to behavioral responses, and accuracy on the next-word prediction task (but not other language tasks), consistent with the long-standing hypothesis that the brain’s language system is optimized for predictive processing. Model architectures with initial weights further perform surprisingly similar to final trained models, suggesting that inherent structure – and not just experience with language – crucially contributes to a model’s match to the brain.
Competing Interest Statement
The authors have declared no competing interest.
