The 'Sandwich' meta-framework for architecture agnostic deep privacy-preserving transfer learning for non-invasive...

Area of Science:

• Neuroscience
• Machine Learning
• Computer Science

Background:

• Electroencephalography (EEG) brainwave decoding is crucial for analyzing neural activity and developing brain-computer interfaces.
• Training machine learning models on EEG data faces challenges due to data variability and privacy concerns.
• Existing methods struggle to effectively utilize diverse and private EEG datasets.

Purpose of the Study:

• To develop a unified approach integrating transfer learning and federated learning to address EEG data variability and privacy concerns.
• To introduce a novel deep privacy-preserving meta-framework named 'Sandwich' for enhanced EEG signal decoding.
• To improve the performance and applicability of machine learning in analyzing non-invasive neural activity.

Main Methods:

• Developed the 'Sandwich' meta-framework, a novel deep privacy-preserving approach combining transfer learning and federated learning.
• The framework features federated networks for input-level data set differences, a shared network for common rule learning, and individual classifiers for specific tasks.
• Implemented and evaluated the 'Sandwich' architecture on the BEETL motor imagery challenge dataset, comparing it against baseline models.

Main Results:

• The 'Sandwich' framework demonstrated superior performance compared to baseline models like Shallow ConvNet and EEGInception.
• The best-performing model, Inception-SD-Deepset, achieved a 9% performance improvement over existing methods.
• Evaluations on heterogeneous EEG datasets confirmed the framework's effectiveness in handling data variability.

Conclusions:

• The 'Sandwich' framework represents a significant advancement in federated deep transfer learning for diverse time-series data, particularly EEG.
• It effectively addresses data variability and privacy concerns, enabling the use of larger, heterogeneous datasets.
• The model-agnostic nature of 'Sandwich' suggests potential for large-scale brainwave decoding and other time-series analysis tasks.