Understanding the brain with AI-driven explanations and experiments
AI Summary: A recent study published in *Nature Neuroscience* introduces generative causal testing (GCT), a framework developed by researchers from Microsoft and several universities to enhance the interpretability of large language model (LLM) predictions regarding human brain responses to language. GCT translates complex predictive models into concise verbal explanations of cortical responses, such as "food preparation" or "location names." The framework then tests these explanations by having an LLM generate stories aimed at activating specific brain areas, which are subsequently measured in fMRI scans to confirm or refute the proposed hypotheses. This approach addresses the challenge of understanding the underlying mechanisms of brain activity predictions made by LLMs, moving from opaque models to testable scientific theories.