BrainTaskonomy: Learning How to Pretrain and What to Transfer in fMRI Foundation Models
Field: Computer Vision / Neuroimaging Authors: Junfeng Xia, Wenhao Ye, Junxiang Zhang, Jiayu Zuo, Mo Wang, Quanying Liu Published: 2026-09-09 arXiv: 2609.10518
Abstract
fMRI foundation models increasingly aggregate heterogeneous data across brain states, cohorts, and acquisition settings, yet pretraining domains are commonly treated as a flat mixture and downstream tasks are adapted independently. This work studies whether measured learning relations can organize both stages without modifying the backbone.
Method
Pretraining stage
- A lightweight Brain-DiT proxy estimates difficulty and directed facilitation across ten fMRI domains.
- These estimates yield a priority-guided cumulative domain curriculum.
- Combined with high-to-low-noise timestep scheduling and joint consolidation.
- Controlled first- and higher-order transfer across fifteen tasks constructs a directed taskonomy.
- Budgeted integer programming (BIP) selects directly supervised source tasks and target-specific transfer routes.
- The joint priority-domain and high-to-low-timestep curriculum reduces v-NMSE, PSD-NMSE, and FC-MSE by 6.5%, 16.3%, and 10.5%, respectively, relative to uniform sampling over both dimensions.
- Strong downstream performance across six in-domain and out-of-domain tasks.
- The taskonomy reveals asymmetric, target-dependent transfer.
- Exploratory sealed-test evaluation shows larger descriptive gains for BIP policies when higher-order route spaces are available, compared to matched random controls.
Adaptation stage
Results
Conclusion
Organizing fMRI pretraining and adaptation by measured learning relations outperforms treating domains and tasks as independent flat sets.
---
*Auto-collected on 2026-09-11.*