| name | nerve-fc-bilinear-tokenization |
| description | NERVE: Network-Aware Bilinear Tokenization for Brain Functional Connectivity representation learning. Self-supervised learning framework that redefines FC tokenization via structured bilinear factorization, aligning with large-scale brain network organization. Activation: NERVE, brain FC tokenization, functional connectivity representation, bilinear tokenization, network-aware MAE, brain network MAE. |
| category | neuroscience |
NERVE: Network-Aware Bilinear Tokenization for Brain FC
arXiv: 2605.14048v2 (2026-05-15)
Authors: Leo Milecki, Qingyu Hu, Bahram Jafrasteh, Mert R. Sabuncu, Qingyu Zhao
Affiliation: Weill Cornell Medicine, Cornell University
Code: https://github.com/leomlck/NERVE
Keywords: rs-fMRI, Brain Functional Connectivity, Deep Learning, Masked Autoencoder, Self-Supervised Learning
Overview
NERVE (Network-Aware Representations of Brain Functional Connectivity via Bilinear Tokenization) is a self-supervised learning framework that addresses the fundamental question of how to tokenize brain functional connectivity (FC) matrices in a way that aligns with the intrinsic modular organization of large-scale brain networks.
Core Innovation
The key insight is that FC matrices should be tokenized based on brain network organization, not treated as structurally homogeneous elements. NERVE partitions FC matrices into patches of intra- and inter-network connectivity blocks, then embeds these heterogeneous patches through a novel structured bilinear factorization.
Technical Details
Problem Statement
Existing approaches for FC representation learning (e.g., Masked Autoencoders adapted from computer vision) use heuristic tokenization schemes that:
- Treat FC matrices as structurally homogeneous
- Ignore large-scale brain network organization
- Use region-centric or graph-based schemes that don't reflect functional modularity
NERVE Architecture
-
Network-Aware Patch Definition: FC matrices are partitioned into patches based on brain network pairs (e.g., Default Mode Network × Salience Network), creating heterogeneous-sized blocks that correspond to distinct functional roles.
-
Structured Bilinear Factorization: Each FC patch is embedded via a bilinear decomposition:
- Preserves network identity
- Reduces parameter complexity from quadratic O(N²) to linear O(N) scaling in the number of networks
- Captures both within-network and between-network connectivity patterns
-
Masked Autoencoder Training: Patches are masked and reconstructed from visible tokens, learning robust representations without labeled data.
-
Downstream Prediction: Learned representations are used for behavioral and psychopathology prediction tasks.
Key Design Choices
| Component | NERVE Approach | Baseline Approach |
|---|
| Tokenization | Network-pair blocks | Fixed-size patches / regions |
| Embedding | Bilinear factorization | Linear projection |
| Parameter Scaling | Linear in # networks | Quadratic in # regions |
| Domain Prior | Brain network organization | Structurally agnostic |
Evaluation Results
Datasets
- ABCD (Adolescent Brain Cognitive Development)
- PNC (Philadelphia Neurodevelopmental Cohort)
- CCNP (Columbia Center for Computational Neuroimaging Pediatric)
Performance
- Outperforms structurally agnostic MAE variants and graph-based self-supervised baselines
- More stable and transferable representations, particularly in cross-cohort evaluation
- Ablation studies confirm bilinear network embedding and anatomically grounded parcellation are critical for performance
Practical Applications
1. Brain-Behavior Prediction
Use NERVE representations to predict:
- Cognitive abilities
- Behavioral outcomes
- Psychopathology risk
2. Cross-Cohort Generalization
NERVE's network-aware design enables better transfer learning across different neuroimaging datasets and populations.
3. Developmental Neuroscience
Particularly effective for studying brain development across age groups due to robust representation learning.
Implementation Guide
Setup
git clone https://github.com/leomlck/NERVE
cd NERVE
Typical Workflow
- Data Preparation: Prepare FC matrices from rs-fMRI data using a standardized parcellation
- Network Assignment: Assign regions to canonical brain networks (e.g., Yeo 7-network or 17-network parcellation)
- Bilinear Tokenization: Partition FC matrices into network-pair blocks
- MAE Pre-training: Train masked autoencoder on unlabeled FC data
- Fine-tuning: Use learned representations for downstream prediction tasks
Key Concepts
Functional Connectivity (FC)
Temporal correlation between spatially distributed brain regions measured via rs-fMRI. Used to study individual differences in brain organization.
Masked Autoencoder (MAE)
Self-supervised learning framework that:
- Partitions input into tokens
- Masks a subset of tokens
- Reconstructs masked content from visible tokens
- Learns robust representations without labels
Bilinear Factorization
Decomposition that captures interactions between two sets of factors while reducing parameter complexity. In NERVE, it preserves network identity while enabling efficient embedding of heterogeneous FC patches.
Research Implications
-
Domain-Specific Inductive Biases Matter: Simply applying computer vision techniques to neuroimaging data is suboptimal. Incorporating brain network organization significantly improves performance.
-
Self-Supervised Learning for Connectomics: MAE-based approaches hold promise for learning from the vast amounts of unlabeled rs-fMRI data.
-
Cross-Cohort Transfer: Network-aware representations generalize better across different populations and scanning protocols.
Related Concepts
- Resting-state fMRI analysis
- Brain network parcellation (Yeo networks, Schaefer atlas)
- Self-supervised learning for neuroimaging
- Graph neural networks for brain connectivity
- Developmental neuroscience
- Psychopathology prediction
Activation Triggers
- NERVE tokenization
- brain FC representation
- functional connectivity MAE
- bilinear factorization brain
- network-aware brain ML
- rs-fMRI deep learning
- cross-cohort brain prediction
- brain behavior prediction
- psychopathology ML
- connectomics representation learning
References
- Milecki, L., Hu, Q., Jafrasteh, B., Sabuncu, M.R., Zhao, Q. (2026). "Network-Aware Bilinear Tokenization for Brain Functional Connectivity Representation Learning." arXiv:2605.14048v2