BINE022/EEGPT

What it solves

EEGPT addresses the challenge of extracting reliable, universal features from EEG (electroencephalogram) signals, which often suffer from low signal-to-noise ratios (SNR) that can degrade the quality of features extracted by traditional mask-based self-supervised learning methods.

How it works

EEGPT is a 10-million-parameter pretrained transformer model. It uses a mask-based dual self-supervised learning method for feature extraction. Unlike standard methods, it incorporates spatio-temporal representation alignment to create self-supervised tasks based on representations with high SNR and rich semantic information rather than relying on raw, noisy signals.

Who it’s for

Researchers and developers working with brain-computer interfaces (BCI) and EEG signal processing who need a robust, pretrained foundation model for various downstream neurological or physiological tasks.

Highlights

• Dual Self-Supervised Learning: Employs a specialized masking strategy to improve feature extraction efficiency.
• Spatio-Temporal Alignment: Specifically designed to avoid poor feature quality caused by low SNR signals.
• Pretrained Weights: Provides pretrained models for universal EEG feature extraction.
• Downstream Versatility: Supports linear probing and fine-tuning for various downstream EEG datasets and tasks.