TL;DR. BrainTaskonomy measures which fMRI domains and tasks help one another, then uses those relationships to organize learning. A joint domain-and-noise curriculum lowers three pretraining errors, while a directed transfer graph guides the allocation of limited supervision.
Should every dataset enter training in the same way?
Combining many fMRI datasets usually starts with a flat mixture: sample data from different domains and train one model. Downstream tasks are then often handled independently. These choices leave the relationships between domains and tasks unused.
BrainTaskonomy asks whether measured learning relationships can improve both stages without changing the backbone. It treats pretraining order and downstream supervision as decisions that can be informed by experiments.
Measure learning relationships, then use them to plan
The framework operates at two stages. During pretraining it chooses an order for learning domains; during adaptation it chooses which tasks to supervise. Both decisions use measured relationships rather than treating all domains or tasks as independent choices.
Method overview. The first part builds a domain-and-timestep curriculum. The second measures single- and higher-order task transfer, then selects supervised sources and routes under a budget. The arrows encode direction, not just similarity. Source paper ↗
Measure domain relationships
A lightweight Brain-DiT proxy estimates difficulty and directed facilitation across ten domains. Facilitation asks whether learning one domain helps the model learn another.
Build a cumulative curriculum
The resulting priorities determine how domains enter pretraining. Domain scheduling is combined with a high-to-low-noise timestep schedule and a joint consolidation stage.
Plan transfer under a budget
Controlled first- and higher-order transfer across 15 tasks forms a directed task graph. Budgeted integer programming selects which source tasks receive supervision and which routes serve each target.
A better curriculum—and a directional map of transfer
The curriculum is compared with uniform sampling over both domains and diffusion timesteps. The reported changes below are relative reductions in pretraining error.
| Pretraining error | Relative reduction |
|---|---|
| v-NMSE | 6.5% |
| PSD-NMSE | 16.3% |
| FC-MSE | 10.5% |
The curriculum improves several views of the signal
The joint schedule lowers v-NMSE, PSD-NMSE and FC-MSE, and shows strong downstream performance across six in- and out-of-domain tasks. The three errors capture different evaluation targets rather than being three versions of downstream accuracy.
Task similarity is not the same as transfer direction
Transfer is asymmetric and target-dependent: a helpful route from one task to another need not work equally well in reverse. In exploratory sealed-test evaluation, budgeted integer programming benefits more when higher-order routes are available than matched random controls do.
A training budget includes decisions about supervision
The work broadens what it means to design a training recipe. Architecture and compute are only part of it; domain order and the choice of supervised tasks also affect what the model learns. Applying this strategy to a new setting starts with measuring its own domain and task relationships.
Read further
Use the paper as a workflow for choosing pretraining curricula and supervised source tasks. Applying the approach to a new task set requires measuring its learning relationships and specifying the supervision budget. No public code or checkpoint link is listed here.
Learning How to Pretrain and What to Transfer in fMRI Foundation Models
Research post by fMRI Atlas, based on the original study linked above.
Read the paper’s 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. We study whether measured learning relations can organize both stages without modifying the backbone. During pretraining, a lightweight Brain-DiT proxy estimates difficulty and directed facilitation across ten fMRI domains, yielding a priority-guided cumulative domain curriculum combined with high-to-low-noise timestep scheduling and joint consolidation. During adaptation, controlled first- and higher-order transfer across fifteen tasks constructs a directed taskonomy, from which budgeted integer programming (BIP) selects directly supervised source tasks and target-specific 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, and shows strong downstream performance across six in- and out-of-domain tasks. The taskonomy reveals asymmetric, target-dependent transfer, while exploratory sealed-test evaluation shows larger descriptive gains for BIP policies when higher-order route spaces are available than for matched random controls. Together, these findings support organizing fMRI pretraining and adaptation by measured learning relations rather than treating domains and tasks as independent flat sets.
BibTeX
@misc{braintaskonomy2026,
title={BrainTaskonomy: Learning How to Pretrain and What to Transfer in fMRI Foundation Models},
author={Xia, Junfeng and Ye, Wenhao and Zhang, Junxiang and Zuo, Jiayu and Wang, Mo and Liu, Quanying},
year={2026},
eprint={2609.10518},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2609.10518}
}