Turning brain prediction models into testable explanations (www.microsoft.com)

Researchers from Microsoft and several universities have developed a groundbreaking method called generative causal testing (GCT) that transforms complex language prediction models into understandable explanations. While large language models (LLMs) have excelled at predicting how the human brain responds to language, they often operate as black boxes with millions of parameters that are difficult to interpret. GCT addresses this issue by generating concise verbal hypotheses about what specific brain regions respond to, such as "food preparation" or "location names," and then testing these hypotheses by presenting new, targeted narrative stimuli to subjects in fMRI scanners. GCT not only confirmed known selectivity in the brain but also elucidated previously ambiguous areas by demonstrating that distinct brain regions have unique responses to carefully crafted stimuli. For instance, it differentiated roles within adjacent place-processing regions and identified new "micro-regions" in the prefrontal cortex that respond to very specific concepts like dialogue and measurements. The method promises to enhance our understanding of neuroscience by linking AI-driven predictions to scientifically testable theories, potentially accelerating research in various fields where complex models outpace our comprehension.