Multiband Convolutional Riemannian Network With Band-Wise Riemannian Triplet Loss for Motor Imagery Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery classification is crucial for brain-computer interfaces (BCIs).
• Riemannian geometry offers superior performance over traditional methods like Common Spatial Patterns (CSP).
• Deep learning generalization of Riemannian approaches remains underexplored.

Purpose of the Study:

• To propose a novel motor imagery classification algorithm.
• To enhance classification performance and reduce overfitting in Riemannian networks.
• To develop a state-of-the-art multiband convolutional Riemannian network.

Main Methods:

• Utilizing an overlapping multiscale multiband convolutional Riemannian network.
• Implementing band-wise Riemannian triplet loss for regularization.
• Employing convolutional layers before subband covariance matrix computation to mitigate overfitting.

Main Results:

• The proposed method demonstrated improved classification performance on public datasets (BCI Competition IV 2a, OpenBMI).
• Achieved state-of-the-art classification accuracy compared to existing Riemannian networks.
• Effectively reduced overfitting issues inherent in Riemannian network architectures.

Conclusions:

• The novel convolutional Riemannian network with band-wise Riemannian triplet loss significantly advances motor imagery classification.
• The method offers a robust solution for BCI applications requiring high accuracy.
• This work highlights the potential of deep learning to generalize Riemannian approaches effectively.