NERVE: Network-Aware Tokenization for Brain FC Learning
NERVE, which stands for Network-Aware Representations of Brain Functional Connectivity via Bilinear Tokenization, is a cutting-edge self-supervised learning framework designed to address tokenization issues in masked autoencoders tied to resting-state brain functional connectivity. Unlike standard approaches that concentrate on specific regions or rely on graph techniques, NERVE breaks down functional connectivity matrices into segments that capture both intra- and inter-network connections, aligning with the brain's modular framework. This approach allows for flexibility in patch sizes and functional roles, ultimately improving representation learning.
Key facts
- NERVE is a self-supervised learning framework for brain functional connectivity.
- It uses bilinear tokenization to partition FC matrices into network-based patches.
- Patches correspond to intra- and inter-network connectivity blocks.
- The method addresses limitations of region-centric and graph-based tokenization.
- It aligns with the intrinsic modular organization of large-scale brain networks.
- NERVE handles heterogeneous patch sizes and distinct functional roles.
- The framework is designed for masked autoencoders in resting-state FC analysis.
- The paper is available on arXiv with ID 2605.14048.
Entities
Institutions
- arXiv