Driving and suppressing the human language network using large language models.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Linguistics

Background:

• Transformer models, such as Generative Pre-trained Transformer (GPT), exhibit human-like language generation capabilities.
• These models have shown predictive power regarding human brain responses to linguistic stimuli.

Purpose of the Study:

• To investigate if a GPT-based encoding model can predict the magnitude of brain responses to diverse sentences using functional magnetic resonance imaging (fMRI).
• To utilize the model to identify novel sentences that can either drive or suppress activity within the human language network.
• To analyze the characteristics of model-selected sentences that influence neural activity.

Main Methods:

• Collected fMRI data measuring brain responses to 1,000 diverse sentences.
• Developed and applied a GPT-based encoding model to predict brain response magnitudes.
• Selected new sentences predicted to modulate language network activity.
• Validated the effect of model-selected sentences on neural activity in new participants.

Main Results:

• The GPT-based encoding model successfully predicted the magnitude of fMRI-measured brain responses to sentences.
• Model-selected novel sentences demonstrated a significant ability to drive and suppress activity in human language areas.
• Sentence surprisal and well-formedness were identified as key factors influencing the strength of language network responses.

Conclusions:

• Neural network models, like GPT, can accurately mimic human language processing.
• These models offer a non-invasive method to control neural activity in higher-level cortical areas, specifically the human language network.
• The findings provide a foundation for using AI to understand and modulate complex cognitive functions.