Summary: New research shows that modern artificial intelligence models that predict the next word in a sentence mirror activity patterns seen in human language areas of the brain, suggesting prediction may be a core mechanism of human language processing.
Source: MIT
Recent advances in artificial intelligence have produced language models that are exceptionally good at forecasting the next word in a sequence. These predictive models power features like autocomplete in search engines and messaging apps, and their capabilities have grown to include tasks that appear to require real comprehension, such as answering questions, summarizing documents, and continuing narratives.
Although these systems were engineered to optimize prediction without any explicit attempt to emulate human neural processing, a new study from MIT neuroscientists indicates that their functional behavior resembles the way human brains handle language. The results suggest that next-word prediction might be a central computation used by language-processing circuits in the brain.
The study compared many types of language models and found that those optimized for next-word prediction showed the strongest alignment with human neural responses. Models trained for other language tasks did not match brain activity as closely, strengthening the case that prediction plays a significant role in human language understanding.
“The better the model is at predicting the next word, the more closely it matches human brain activity,” says Nancy Kanwisher, Walter A. Rosenblith Professor of Cognitive Neuroscience and a member of MIT’s McGovern Institute for Brain Research and the Center for Brains, Minds, and Machines (CBMM). “It’s striking how well these models fit—this gives indirect evidence that the human language system may be driven by predictions about what comes next.”
The study’s senior authors include Joshua Tenenbaum, professor of computational cognitive science at MIT and member of CBMM and CSAIL, and Evelina Fedorenko, Frederick A. and Carole J. Middleton Career Development Associate Professor of Neuroscience and member of the McGovern Institute. The paper appears in Proceedings of the National Academy of Sciences. Martin Schrimpf, an MIT graduate student affiliated with CBMM, is the lead author.
How the comparison was made
The most successful next-word prediction systems are deep neural networks: computational models organized in layers of interconnected units that learn statistical relationships from vast amounts of text. Over the past decade, similar deep networks have produced computer vision systems that match primate object recognition performance and whose internal organization parallels the primate visual cortex.
Applying comparable methods to language, the MIT researchers evaluated 43 distinct language models. The set included models optimized for next-word prediction—among them GPT-3—and models built for other tasks like sentence completion or masked-word prediction. For each model, the researchers recorded the activation patterns of internal network units as the model processed the same language stimuli used in human experiments.
Human neural data came from three experimental paradigms: people listening to natural stories, participants reading sentences presented one at a time, and participants reading sentences where words were revealed sequentially. The team analyzed both functional MRI (fMRI) signals and intracranial electrocorticographic recordings collected from patients undergoing surgery for epilepsy.
When they compared model activations with human brain responses, the researchers found that models that excel at next-word prediction produced activation patterns that closely matched the neural signals from language regions. Those same models also correlated with behavioral measures, such as how quickly people read text in the experiments.
“Models that best predict neural responses also tend to best predict human behavioral responses like reading times, and both of these relationships are driven by the models’ next-word prediction performance,” Schrimpf explains. “This triangular relationship ties together model performance, neural fit, and behavior.”
What makes predictive models special
A defining feature of many high-performing predictive models is the forward, one-way transformer architecture. This design lets a model use long stretches of preceding text—sometimes hundreds of words—to predict upcoming words. That capacity to integrate extended context distinguishes these models from earlier approaches that relied on much shorter histories.
While neuroscientists have not yet identified brain circuits that implement transformer-like computation, the study’s findings align with long-standing theoretical ideas that prediction is central to real-time language comprehension. “Language unfolds in time, and the brain must continually anticipate and interpret incoming input to keep up,” Tenenbaum says.
Going forward, the researchers plan to create modified versions of these language models to test which architectural elements are essential for matching human neural data. They will explore how small changes affect both task performance and similarity to brain activity.
Fedorenko calls the results transformative for her research. “I did not expect that we would soon have computationally explicit models that align closely enough with the brain to use them as tools for understanding how language is processed,” she says. That alignment opens a path to using these models as experimental probes into the computations performed by human language systems.
The team also aims to integrate high-performing language models with other computational systems developed in Tenenbaum’s lab that represent perceptual and physical knowledge. Combining language prediction with models of perception and physical reasoning could yield more comprehensive accounts of how the brain supports complex cognitive abilities and how more general forms of intelligence emerge.
Funding: The study received support from a Takeda Fellowship; the MIT Shoemaker Fellowship; the Semiconductor Research Corporation; the MIT Media Lab Consortia; the MIT Singleton Fellowship; the MIT Presidential Graduate Fellowship; the Friends of the McGovern Institute Fellowship; the MIT Center for Brains, Minds, and Machines through the National Science Foundation; the National Institutes of Health; MIT’s Department of Brain and Cognitive Sciences; and the McGovern Institute.
Additional authors include Idan Blank (PhD ’16) and graduate students Greta Tuckute, Carina Kauf, and Eghbal Hosseini.
About this AI and language research news
Author: Sarah McDonnell
Source: MIT
Contact: Sarah McDonnell – MIT
Image: The image is in the public domain
Original Research: The findings will appear in PNAS