Flexible Patched Brain Transformer model for EEG decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Non-invasive human brain decoding presents significant challenges.
• Advancements in machine learning offer potential solutions for improving electroencephalography (EEG) decoding.

Purpose of the Study:

• To enhance EEG decoding by developing novel machine learning methods.
• To introduce the attention-based Patched Brain Transformer model for improved brain decoding.

Main Methods:

• Development of the attention-based Patched Brain Transformer model.
• Investigation of data augmentation and pre-training strategies.
• Architectural inspection for training behavior insights.
• Comparison with state-of-the-art models on motor imagery datasets.

Main Results:

• The Patched Brain Transformer model demonstrates superior performance compared to existing methods.
• Achieved high performance with a significantly reduced number of parameters.
• Effective multi-participant classification using supervised pre-training and time-shift data augmentation.

Conclusions:

• The proposed Patched Brain Transformer model offers a flexible and efficient approach to EEG decoding.
• Supervised pre-training and data augmentation are crucial for enhancing model performance.
• This advancement has implications for developing more effective brain-computer interfaces.