Riemannian Embedding Banks for Common Spatial Patterns with EEG-based SPD Neural Networks

Modeling non-linear data as symmetric positive definite (SPD) matrices on Riemannian manifolds has attracted much attention for various classification tasks. In the context of deep learning, SPD matrix-based networks have been shown to be a promising solution classifying electroencephalogram (EEG) signals, capturing geometry within their structured 2D feature representation. However, existing approaches usually learn spatial-temporal structures in an embedding space all available EEG and optimization procedures rely computationally expensive iterations. Furthermore, these often struggle encode types relationships into single distance metric, resulting loss generality. To address above limitations, we propose Embedding Banks method, which divides problem common spatial patterns learning entire K-subproblems builds one model each subproblem, combined with neural networks. By leveraging concept "separate learn" technology manifold, REB K non-overlapping subsets learns separate metrics geometric instead vector space. Then, learned are grouped neurons network's layer. Experimental results public datasets demonstrate superiority proposed approach signals despite non-stationary nature, increasing convergence speed while maintaining generalization.