bridge-ai-neuro

RABBiT: Foundation model for zero-shot and few-shot prediction of speech-elicited fMRI brain responses

Researchers introduce RABBiT, a compact audio-to-fMRI encoder foundation model designed to predict brain responses to natural speech with zero or few participant-specific data. Evaluated on 324 participants across multiple unseen fMRI datasets, RABBiT outperforms the current state-of-the-art fMRI foundation model and group-average baselines in auditory and language-selective brain regions. With only 10 minutes of participant data, parameter-efficient fine-tuning further improves performance substantially over per-participant linear models. Key innovations include learned region-specific attention and a decomposition of brain responses into shared and subject-specific components.