Researchers introduce RABBiT, a compact audio-to-fMRI encoder designed for accurate zero- and few-shot prediction of brain activity in response to natural speech. Evaluated on 324 participants across multiple unseen datasets, the model surpasses current state-of-the-art foundation models and group-average predictions.
- RABBiT enables accurate zero-shot prediction of fMRI responses in auditory and language-selective regions.
- With only 10 minutes of participant-specific data, parameter-efficient tuning substantially outperforms per-participant linear models.
- Performance relies on learned region-specific attention and a decomposition of brain responses into shared and subject-specific components.
- The model's structured representations support interpretability while eliminating the need for extensive per-participant data fitting.
By enabling scalable population-level analyses of language in the human brain, RABBiT addresses the challenge of building computational models that generalize across diverse experimental stimuli and individuals.