Understanding natural language involves the engagement of specific left-lateralized frontal and temporal brain regions, forming what is commonly referred to as the ‘language network.’ Despite significant progress, many aspects of the representations and algorithms supporting language comprehension in this network remain elusive.
Recent advancements in computing power, the availability of extensive text corpora, and breakthroughs in machine learning have paved the way for significant progress in artificial intelligence, particularly in language-related tasks. This progress has prompted researchers to explore the use of large language models (LLMs) as potential models for human language processing.
In a new breakthrough paper Driving and suppressing the human language network using large language models, a research team from Massachusetts Institute of Technology, MIT-IBM Watson AI Lab, University of Minnesota and Harvard University leverages a GPT-based encoding model to identify sentences predicted to elicit specific responses within the human language network.
The study had two primary objectives:
- To subject the new class of models, LLMs, to a rigorous evaluation as models of language processing.
- To gain an intuitive-level understanding of language processing by characterizing the stimulus properties that drive or suppress responses in the language network across a diverse range of linguistic input and associated brain responses.
The team developed an encoding model to predict brain responses to arbitrary sentences in the language network. The model utilized last-token sentence embeddings from GPT2-XL and was trained on 1,000 diverse, corpus-extracted sentences from five participants. The model achieved a commendable prediction performance of r=0.38 on held-out sentences within the baseline set.
To ensure the robustness of the encoding model, the team verified its predictivity on held-out sentences using different procedures for obtaining sentence embeddings and even incorporating embeddings from a distinct LLM architecture. The model demonstrated consistent high predictivity performance, affirming its reliability.
The encoding model showcased impressive predictivity performance on anatomically defined language regions, providing a non-invasive means of controlling neural activity in areas associated with higher-level cognition. Notably, the brain-aligned Transformer model GPT2-XL successfully drove and suppressed responses in the language network of new individuals.
This groundbreaking research not only validates the potential of LLMs as accurate models for human language processing but also introduces a paradigm shift in non-invasive neural activity control. The ability to influence neural responses in areas linked to higher-level cognition holds profound implications for both neuroscientific research and practical applications, marking a significant milestone in the intersection of artificial intelligence and neuroscience.
The paper Driving and suppressing the human language network using large language models on Nature Human Behaviour.
Author: Hecate He | Editor: Chain Zhang
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.