Subject-Adaptive EEG Decoding via Filter-Bank Neural Architecture Search for BCI Applications.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Individual differences present a major hurdle in brain-computer interface (BCI) research.
• Existing universally applicable network architectures are impractical due to human brain variability.

Purpose of the Study:

• To introduce Filter-Bank Neural Architecture Search (FBNAS), an automated EEG decoding framework.
• To address individual differences in BCI by customizing network architecture design.

Main Methods:

• FBNAS employs three temporal cells to process diverse EEG frequencies using dilated convolutions.
• A multi-path neural architecture search (NAS) algorithm optimizes architectures for multi-scale feature extraction.
• The framework was benchmarked on three EEG datasets (BCIC-IV-2a, OpenBMI, SEED) across two BCI paradigms.

Main Results:

• FBNAS achieved superior cross-session decoding accuracies: 79.78% (BCIC-IV-2a), 70.66% (OpenBMI), and 68.38% (SEED).
• The proposed method outperformed six state-of-the-art deep learning algorithms.
• FBNAS demonstrated effective customization of decoding models for individual brain patterns.

Conclusions:

• FBNAS successfully addresses individual differences in BCI, significantly enhancing decoding performance.
• The study shifts BCI model design from expert-driven to a machine-aided approach.
• Automated, personalized network architecture design is crucial for advancing BCI technology.