Performance study of Neural Structured Learning using Riemannian Features for BCI Classification

Abstract

Riemannian Geometry-based features have been among the most promising electroencephalography(EEG) classification methods in recent years. However, these features can be classified using many machine learning(ML) algorithms. When compared against the standard methods, deep learning-based approaches are successful in classification accuracy and transfer learning. In this paper, we attempt to study Neural structured learning(NSL) to develop robust and regularized neural network models that preserve the similarity structure of the input EEG signals for a more reliable Brain-Computer Interface(BCI) classification. In this study, we have used the state-of-the-art Euclidean Tangent Space features projected from the Riemannian Covariance features of EEG to train the standard feedforward neural nets while incorporating the NSL module. It creates a similarity graph among the input samples and minimizes a graph regularization loss to maintain the neighbor structure. The proposed approach is evaluated on the standard 4-class Dataset 2a from BCI competition 2008. The results show that the proposed model improves accuracy compared to the base model without graph regularization. Surprisingly, it requires very few training samples to achieve almost state-of-the-art accuracy for some subjects using a mere two hidden layered neural network.