| name | cocot-eeg-contrastive-foundation |
| description | Contrastive pretraining methodology for EEG foundation models using multiscale convolutional Transformer architecture. Demonstrates contrastive learning as a superior alternative to masked reconstruction pretraining for EEG, which has high noise and narrow-band information. Achieves SOTA on heterogeneous electrode configurations. Activation: EEG contrastive learning, CoCoT, EEG foundation model, masked reconstruction, multiscale temporal convolution, self-supervised EEG, electrode heterogeneous, EEGFM, EEG decoding, LaBraM alternative, REVE alternative |
| tags | ["eeg","foundation-model","contrastive-learning","self-supervised","neuroscience","deep-learning"] |
| metadata | {"arxiv_id":"2607.09543","published":"2026-07-10","authors":"Gabriel Mahuas, Victoria Shevchenko, Ugo Tanielian, Yassir Bendou, Richard Gao","conference":"arXiv"} |
CoCoT-EEG: Contrastive-Pretrained Multiscale Convolutional Transformer for EEG Decoding
Core Innovation
CoCoT-EEG demonstrates that contrastive learning is a superior pretraining strategy for EEG foundation models compared to the dominant masked reconstruction (MAE) approach. The key insight: EEG has high noise amplitude and information confined to limited dimensions (narrow frequency bands), making reconstruction-based objectives suboptimal.
Why Reconstruction Fails for EEG
- High noise amplitude — Masked autoencoders waste capacity reconstructing noise
- Narrow-band information — Only specific frequency bands carry task-relevant signal
- Heterogeneous electrode configurations — Different datasets have different montages, making token-level reconstruction hard to generalize
Architecture
Multiscale Temporal Convolution Input Layers
- Parallel convolutions at multiple temporal scales capture different frequency bands
- Acts as a learnable filter bank replacing hand-crafted spectral features
- Each scale focuses on a different temporal resolution
Transformer Encoder Blocks
- Standard Transformer architecture applied after multiscale convolution
- Processes the filtered representations for downstream task learning
- Benefits from both local (convolution) and global (attention) processing
Pretraining Strategy
Contrastive Learning for EEG
- Positive pairs: augmented views of same EEG segment
- Negative pairs: different EEG segments
- Forces model to learn invariant, task-relevant representations
- Avoids the noise-reconstruction trap of MAE
EEG-Specific Augmentations
- Time warping, channel dropout, frequency masking
- Preserves task-relevant structure while creating diverse views
Key Results
- CoCoT pretrained matches or beats reconstruction-pretrained SOTA models (LaBraM, CBraMod, CSBrain, LUNA, REVE)
- CoCoT from scratch outperforms previous single-task decoding models and rivals pretrained models
- Works well with heterogeneous electrode configurations — no montage-specific tuning needed
- Demonstrates the viability of contrastive learning as a general EEG FM strategy
Comparison: Contrastive vs Reconstruction Pretraining
| Aspect | Reconstruction (MAE) | Contrastive (CoCoT) |
|---|
| Noise sensitivity | High (reconstructs noise) | Low (learns invariance) |
| Band efficiency | Poor (all bands equal) | Good (task-relevant only) |
| Electrode heterogeneity | Poor | Good |
| Data efficiency | Requires massive data | Works from scratch |
| Architecture flexibility | Fixed tokenization | Multiscale convolution |
Applications
- EEG foundation model pretraining
- Brain-computer interface (BCI) decoding
- Cross-dataset EEG generalization
- Clinical EEG analysis with limited labeled data
- Heterogeneous electrode montage handling
Pitfalls
- Augmentation design is critical: Poor augmentations destroy task-relevant EEG structure. Must preserve frequency-band information that downstream tasks need.
- Negative sample selection: Random negatives may not be hard enough for EEG where segments from same subject/session are similar.
- Multiscale filter design: Number and size of convolutional scales should match the frequency bands of interest for the target domain.
- Not a silver bullet: While contrastive outperforms MAE on average, the gap varies by task — motor imagery may still benefit from reconstruction.
- From-scratch performance: While CoCoT from scratch rivals pretrained models on some tasks, large-scale pretraining still provides the best cross-domain generalization.
Related Skills
reve-eeg-foundation — REVE: EEG FM with 4D positional encoding (reconstruction-based)
eeg-channel-adaptation-benchmark — Channel adaptation for EEG FMs
eeg-test-time-adaptation-benchmark — TTA benchmark for EEG FMs
cross-subject-eeg-decoding — Cross-subject EEG decoding survey
kast-brain-autoregressive — KAST-BAR: EEG foundation with brain topology
References
- arXiv: 2607.09543 (2026-07-10)
- Authors: Mahuas, Shevchenko, Tanielian, Bendou, Gao
- Affiliations: Sigma Nova (Paris), Goethe University Frankfurt