CoCoT: Contrastive-pretrained multiscale convolutional transformer for EEG decoding

Researchers introduce CoCoT, a novel contrastive-pretrained foundation model designed to address the limitations of masked reconstruction approaches in non-invasive electroencephalogram (EEG) decoding. The architecture combines multiscale temporal convolution input layers with Transformer encoder blocks to better handle high-noise data and information confined to narrow frequency bands.

• CoCoT matches or beats state-of-the-art reconstruction-pretrained EEG models on extensive benchmark decoding tasks with heterogeneous electrode configurations.
• Models trained from scratch outperform previous single-task decoding models and rival pretrained models, demonstrating high flexibility and data efficiency.
• Systematic ablations confirm the viability of contrastive learning for building EEG foundation models and highlight key architectural design considerations.

The study demonstrates that contrastive pretraining is a viable strategy for EEG foundation models, offering an effective alternative to traditional masked reconstruction methods.