BEYOND SINGLE-SOURCE COGNITIVE TASKONOMY: MULTI-SOURCE TASK RELATIONS THROUGH FMRI TRANSFER LEARNING

This study extends the framework of fMRI cognitive taskonomy by moving beyond simple one-to-one task transfer to analyze many-to-one relationships and global task-allocation strategies. Using 23 task states from the Human Connectome Project (HCP), the authors trained over 1,100 models to quantify how representations learned from one or more source tasks can support the reconstruction of a target task under limited data conditions. To determine which tasks should receive direct supervision when resources are constrained, the researchers employed Boolean Integer Programming (BIP), which optimizes task selection based on the global coverage provided by source-target transfer dependencies.

The researchers found that single-source transfer is highly directional and structured by experimental paradigm. Motor tasks, in particular, form a tightly coupled cluster that facilitates strong within-paradigm transfer but offers limited support for non-motor tasks. When moving to multi-source transfer, the study revealed that the effectiveness of a source set depends on its specific composition, indicating that many-to-one relations cannot be fully predicted by averaging pairwise transfer strengths. Finally, the BIP analysis identified that 0-back and 2-back working-memory tasks are consistently prioritized for direct supervision. This priority arises because these tasks act as effective 'hubs' for covering other cognitive states, reflecting the integration of perceptual, attentional, and executive processes within working memory.

Understanding how cognitive tasks share neural processes is essential for efficient neuroimaging research. By identifying which tasks serve as the most effective sources for transfer learning, this work provides a roadmap for researchers to optimize data collection. Instead of collecting large amounts of data for every task, researchers can prioritize direct supervision for high-utility tasks—like working memory—and rely on transfer learning to model other, less central tasks, thereby reducing the overall burden of data acquisition.