Knowledge-guided EEG Representation Learning.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Self-supervised learning (SSL) excels in multimedia but faces challenges in biosignal analysis due to data scarcity and domain differences.
• Existing SSL objectives may not effectively capture the unique characteristics of biosignals like electroencephalography (EEG).
• There is a need for adapted SSL methods to leverage unlabeled biosignal data for improved inference tasks.

Purpose of the Study:

• To develop a parameter-efficient, self-supervised learning model for EEG signal analysis.
• To introduce a novel knowledge-guided pre-training objective tailored for EEG idiosyncrasies.
• To enhance representation learning and downstream task performance using unlabeled EEG data.

Main Methods:

• Proposed a state space-based deep learning architecture for EEG self-supervised learning.
• Developed a novel knowledge-guided pre-training objective specifically for EEG signals.
• Evaluated the model's performance on embedding representation learning and exemplary downstream tasks.

Main Results:

• The proposed model demonstrated robust performance and remarkable parameter efficiency.
• Achieved improved embedding representation learning compared to prior works.
• The novel objective significantly reduced the pre-training data required for equivalent performance.

Conclusions:

• The developed self-supervised model effectively adapts SSL to EEG analysis, outperforming previous methods.
• The knowledge-guided pre-training objective enhances learning efficiency and reduces data requirements.
• This approach holds promise for improving various inference tasks on biosignals by leveraging unlabeled data.