Brain-Aligned Language Models for Long-Range Language Understanding and Neuroscientific Insight

Obiettivo

The BrainAlign project aims to revolutionize next-generation artificial intelligence (AI) models by aligning them closely with the way the human brain understands language. While AI systems for language understanding and generation have undergone much progress in recent years thanks to language models, these systems still face significant challenges, such as understanding human intent. Moreover, the successes have mostly stemmed from tremendous increases in model size, and continuing this trend demands unrealistic amounts of data, compute power, and energy.

One way forward is to look to the only system we trust to truly understand complex language: the human brain. Insights from brain functions have long inspired AI, but these insights took years to consolidate and even longer to transfer to AI. For brain functions that are uniquely human, such as understanding complex natural language, the lack of suitable animal model organisms limits the mechanistic insights that can be applied to AI.

The BrainAlign project presents a novel, data-driven solution that will develop brain-aligned language models by forcing their internal processing to closely reflect information sampled directly from the human brain, as humans read and listen to large amounts of every-day language. By integrating machine learning techniques with human neuroimaging and behavioral data from novel experimental paradigms, BrainAlign will develop next-generation models with a deeper, human-like understanding of language. Additionally, innovative interpretability methods will allow these models to serve as model organisms, revealing mechanisms that mirror human brain processing of language and massively enhancing our scientific knowledge of language in the brain.