Review of deep representation learning techniques for brain-computer interfaces.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) utilize electroencephalogram (EEG) signals.
• Deep learning offers advanced methods for EEG signal representation and decoding.
• Representation learning is crucial for improving BCI performance.

Purpose of the Study:

• To review and analyze deep representation learning techniques applied to BCI decoding.
• To synthesize empirical findings on the state-of-the-art in EEG-based BCI.
• To identify trends, motivations, and challenges in deep representation learning for BCIs.

Main Methods:

• Systematic review of 81 articles on deep representation learning for BCI.
• Categorization of studies based on deep learning techniques, motivations, and representation characterization.
• Analysis of trends in autoencoders and self-supervised learning (SSL).

Main Results:

• Autoencoders are the most prevalent deep learning technique (31 articles).
• Self-supervised learning (SSL) is a growing area, with most SSL studies published recently (10 in 2022+).
• Most studies use representation learning for transfer learning, with fewer focusing on robustness, invariances, or data structure.

Conclusions:

• There is a need for standard foundation models for EEG signal decoding using SSL.
• Further research into characterizing learned representations is essential.
• Development of specialized benchmarks and datasets is critical for advancing foundation models in BCIs.