cocot-eeg-contrastive-foundation

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

1. High noise amplitude — Masked autoencoders waste capacity reconstructing noise
2. Narrow-band information — Only specific frequency bands carry task-relevant signal