| name | stst-jepa-eeg-foundation |
| description | STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture for EEG self-supervised learning. Combines latent-prediction objective with auxiliary signal-reconstruction term under spatiotemporal block masks. Pretrained on 47,703 EEG sessions (ages 5-81), achieves MAE=3.06 years for brain-age regression. Native 30-second windows achieve rank-1 on NeuralBench x EEGD leaderboard for sex classification (BA=0.911) and age prediction (r=0.749). Age-prediction residual negatively correlated with cognitive efficiency. |
| version | 1.0.0 |
| author | Hermes Agent |
| license | MIT |
| metadata | {"hermes":{"tags":["eeg","self-supervised-learning","brain-age","JEPA","foundation-model","neural-benchmark","spatiotemporal-masking","EMA-target","cognitive-efficiency"],"category":"ai_collection/collection/skills/neuroscience","arxiv_id":"2607.06629","arxiv_url":"https://arxiv.org/abs/2607.06629","published":"2026-07-07","authors":["Roy Segal","Yoni Svechinsky","Tomer Fekete"],"categories":["cs.LG","q-bio.NC"],"trigger_words":["STST-JEPA","shallow-target JEPA","EEG self-supervised","EEG brain-age","EEG foundation model","NeuralBench","EEGD","latent prediction EEG","spatiotemporal block mask","cognitive efficiency EEG","EMA tokenizer EEG"]}} |
| created | 2026-07-13 |
| updated | 2026-07-13 |
STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture For EEG Self-Supervised Learning
arXiv: 2607.06629 | Published: 2026-07-07 (v2: 2026-07-09) | Authors: Roy Segal, Yoni Svechinsky, Tomer Fekete
Core Innovation
Introduces STST-JEPA, a self-supervised transformer for resting-state and task EEG that combines two objectives:
- Latent-prediction objective: Predicting masked-token representations against an EMA-of-tokenizer target
- Auxiliary signal-reconstruction term: Applied to 30-second multi-channel windows under spatiotemporal block masks
Key Results
Pretraining Scale
- 47,703 EEG sessions spanning ages 5-81
- Data from TUH (Temple University Hospital) and Healthy Brain Network (HBN) corpora
- First EEG foundation model to cover the full pediatric-to-older-adult range
Brain-Age Regression
- MAE = 3.06 years (r = 0.924) on 3,367 held-out sessions
- Baseline: ~10 years MAE (predict-the-mean)
- Lightweight attentive probe on frozen pretrained embeddings
NeuralBench x EEGD Leaderboard (Rank-1)
Using native 30-second windows with light task-specific finetuning of final layers:
| Task | Metric | Score |
|---|
| Sex classification | Balanced accuracy | 0.911 |
| Age prediction | Pearson r | 0.749 |
| Psychopathology composite | Pearson r | 0.215 |
Cognitive Efficiency Correlation
- Age-prediction residual (predicted - chronological age) is negatively correlated with cognitive efficiency across several examined tasks
- This validates brain-age deviation as a meaningful biomarker
Architecture Details
STST-JEPA Design
Shallow-Target refers to predicting against an EMA (Exponential Moving Average) of the tokenizer's representations, rather than predicting raw signal or deep features. This approach:
- Provides stable, smooth targets for learning
- Avoids the collapse issues common in self-supervised learning
