Researchers at FAU have proven that the human brain predicts the probability of word sequences, similar to the processes used in AI language models.
Even while listening, the brain attempts to anticipate the next words. This is the conclusion reached by a current study conducted by an interdisciplinary team of researchers led by PD Dr. Patrick Krauss, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), and PD Dr. Achim Schilling, Heidelberg University. The researchers combined three methods: a natural listening situation, high resolution measurements of brain activity, and an AI language model as reference. The higher the probability of a certain word occurring in the relevant context, the weaker the neural reaction during processing. At the same time, the data indicate a rise in pre-onset activity before the word begins, suggesting the brain works with predictions.
Are humans born with innate grammatical scaffolding, or does language develop on the basis of use and experience? This is a question that is still debated by the various linguistic schools of thought. Recently, powerful AI language models (Large Language Models, LLMs), which process language by predicting subsequent words, have fueled this debate.
“In our study, we combined the natural, continuous language of an audio book with simultaneous electroencephalography and magnetoencephalography measurements and compared the brain activity of the participants directly with the predictive probabilities of large language models, using a temporal resolution of mere milliseconds,” explains Dr. Patrick Krauss.
Are the brain’s predictions measurable?
The measurements indicate that the brain becomes active before the word actually starts. The neural reaction was less pronounced the higher the probability of a word occurring in the relevant context. In contrast, unexpected words triggered stronger neural responses. “This allowed us to prove that the brain actively predicts language. These predictions can be measured and follow similar patterns to modern language models,” explains Dr. Patrick Krauss.
Language models are based on artificial neural networks. They are mathematical information processing units with an architecture based on the human brain. While biological nervous systems work with electrical or chemical signals, language models, or rather their algorithms, calculate numerical values.
“We were particularly surprised that the brain and language models not only show similar predictions. It is also appearing increasingly likely that both systems organize language internally in a comparable way,” says Patrick Krauss.
Do our brains and AI work on similar principles?
The results of the study corroborate key assumptions in cognitive neuroscience and, at the same time, deliver an explanation why AI language models are so effective in a number of applications.
“The fact that the brain and language models come to similar results does not automatically mean that they work in the same way. However, it may suggest that they follow similar information processing principles,” emphasizes Achim Schilling. “The exciting question is why two so different systems share such identical ways of organizing language – and where the boundaries of this convergence lie,” adds Dr. Patrick Krauss.
What is next in the pipeline?
As a next step, the research team would like to find out whether the principles they have discovered are robust and whether they can be transferred to specific applications. “Once we have a better understanding of how the brain and language models represent and predict language, this may in the long term lead to new approaches for diagnosis, personalized therapies, brain-computer interfaces or more transparent AI.”
Current research on AI and language:
Contact:
PD Dr. Patrick Krauss
Head of Cognitive Computational Neuroscience (CCN) at the Pattern Recognition Lab at FAU
patrick.krauss@uk-erlangen.de
PD Dr. Achim Schilling
Head of the Neuro-AI and brain-computer-interfaces working group at the Mannheim Center for Neuromodulation and Neuroprosthetics (MCNN) from Heidelberg University
achim.schilling@bgu-ludwigshafen.de