BrainTaskonomy: Learning How to Pretrain and What to Transfer in fMRI Foundation Models

Pretraining fMRI foundation models has traditionally relied on flat mixtures of heterogeneous data and isolated task adaptation. BrainTaskonomy introduces a structured alternative without modifying model backbones. By using a lightweight diffusion proxy across ten fMRI domains, it establishes a priority-guided curriculum paired with noise-level scheduling, reducing key reconstruction and connectivity errors by up to 16.3%. Downstream, transfer dynamics among fifteen tasks are mapped into a directed graph where budgeted integer programming selects optimal transfer routes, replacing brute-force data aggregation with a systematic taxonomy of learning relations.