Incremental Classification for High-Dimensional EEG Manifold Representation Using Bidirectional Dimensionality...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Symmetric positive definite (SPD) manifold in Riemannian space is key for spatial feature extraction in brain-computer interface (BCI) systems using electroencephalogram (EEG) signals.
• High dimensionality of SPD matrices in BCI applications leads to significant computational burden, hindering real-time performance, particularly in dynamic tasks like incremental learning.
• Conventional dimensionality reduction (DR) methods can alter SPD matrix properties and incremental learning approaches often require retaining old data, posing challenges for efficiency and adaptability.

Purpose of the Study:

• To propose a novel dimensionality reduction technique, bidirectional two-dimensional principal component analysis for SPD manifold (B2DPCA-SPD), that preserves the SPD manifold properties.
• To extend B2DPCA-SPD for incremental learning tasks, enabling adaptation to new data without storing historical information.
• To integrate the incremental B2DPCA-SPD with matrix-formed growing neural gas network (MF-GNG) for efficient incremental EEG classification.

Main Methods:

• Developed B2DPCA-SPD to reduce the dimensionality of SPD matrices while ensuring reduced matrices remain on the SPD manifold.
• Extended B2DPCA-SPD to an incremental version, allowing for continuous learning without the need to retain old data.
• Integrated incremental B2DPCA-SPD with MF-GNG for incremental EEG classification, facilitating recalculation of prototype representations.

Main Results:

• The proposed B2DPCA-SPD method significantly reduced computation time by 38.53% and 35.96% on two public EEG datasets.
• Achieved superior classification accuracy compared to conventional methods, outperforming them by 4.21% to 19.59%.
• Demonstrated the effectiveness of the incremental B2DPCA-SPD for dynamic BCI tasks.

Conclusions:

• B2DPCA-SPD effectively reduces SPD matrix dimensionality while preserving essential manifold properties for BCI applications.
• The incremental extension of B2DPCA-SPD offers an efficient solution for dynamic EEG classification tasks, eliminating the need for old data storage.
• The combined B2DPCA-SPD and MF-GNG approach enhances both computational efficiency and classification accuracy in BCI systems.