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

TL;DR

This paper introduces BrainTaskonomy, a method for pretraining and transferring knowledge in fMRI foundation models. It organizes pretraining domains and downstream tasks using measured learning relations, improving performance on various tasks. The approach reduces errors and shows strong downstream performance.

Why It Matters

The proposed method can improve the accuracy and efficiency of fMRI analysis, enabling better understanding of brain function and behavior. This can have significant implications for neuroscience research and clinical applications. By optimizing pretraining and transfer, BrainTaskonomy can facilitate the development of more effective brain-computer interfaces and neuroimaging techniques.

Key Findings

• priority-guided cumulative domain curriculum improves pretraining
• directed taskonomy enables efficient transfer across tasks
• budgeted integer programming selects optimal source tasks and routes