Subject-independent Classification on Brain-Computer Interface using Autonomous Deep Learning for finger movement...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Subject-independent classification in Brain-Computer Interfaces (BCIs) is challenging due to data variability.
• Current methods often rely on large, multi-subject datasets for training.
• Adapting BCI systems to new users without extensive retraining remains a significant hurdle.

Purpose of the Study:

• To introduce Autonomous Deep Learning (ADL) as a streaming online learning method for subject-independent BCI classification.
• To evaluate ADL's effectiveness in classifying electroencephalography (EEG) signals for individual finger movements.
• To compare ADL's performance against established machine learning models like Random Forest (RF) and Convolutional Neural Networks (CNN).

Main Methods:

• Utilized Common Spatial Patterns (CSP) extracted from EEG signals as input for the ADL model.
• Implemented ADL, a deep learning architecture capable of self-structuring and adapting to streaming input.
• Conducted 5-fold cross-validation across four subjects for subject-independent classification tasks.

Main Results:

• ADL achieved a classification accuracy of approximately 77% for subject-independent finger classification.
• ADL significantly outperformed Random Forest (53%) and Convolutional Neural Network (72%) in subject-independent classification.
• ADL demonstrated stable training and testing accuracies, unlike CNN which experienced significant degradation.

Conclusions:

• Autonomous Deep Learning (ADL) shows significant promise for addressing the challenge of subject-independent classification in BCIs.
• ADL's adaptive nature allows for robust performance across different subjects and changing input data.
• The findings suggest ADL is a viable and effective machine learning approach for real-world BCI applications.