Synthetic fMRI Augmentation via TRIBE v2 Boosts Brain Decoding Accuracy by Up to 68%

Augmenting small fMRI datasets with synthetic data from TRIBE v2 — a large encoding model pretrained on 1000+ hours of fMRI responses to video, audio, and language — improves Top-10 image-retrieval accuracy by up to 68% on the NSD and BOLD5000 datasets. The optimal ratio of synthetic to real data varies by data source and must be tuned per dataset.