FBCNet: A Multi-view Convolutional Neural Network for Brain-Computer Interface

TL;DR

A novel Filter-Bank Convolutional Network (FBCNet) for MI classification that employs a multi-view data representation followed by spatial filtering to extract spectro-spatially discriminative features and proposes a novel Variance layer that effectively aggregates the EEG time-domain information.

Abstract

Lack of adequate training samples and noisy high-dimensional features are key challenges faced by Motor Imagery (MI) decoding algorithms for electroencephalogram (EEG) based Brain-Computer Interface (BCI). To address these challenges, inspired from neuro-physiological signatures of MI, this paper proposes a novel Filter-Bank Convolutional Network (FBCNet) for MI classification. FBCNet employs a multi-view data representation followed by spatial filtering to extract spectro-spatially discriminative features. This multistage approach enables efficient training of the network even when limited training data is available. More significantly, in FBCNet, we propose a novel Variance layer that effectively aggregates the EEG time-domain information. With this design, we compare FBCNet with state-of-the-art (SOTA) BCI algorithm on four MI datasets: The BCI competition IV dataset 2a (BCIC-IV-2a), the OpenBMI dataset, and two large datasets from chronic stroke patients. The results show that, by achieving 76.20% 4-class classification accuracy, FBCNet sets a new SOTA for BCIC-IV-2a dataset. On the other three datasets, FBCNet yields up to 8% higher binary classification accuracies. Additionally, using explainable AI techniques we present one of the first reports about the differences in discriminative EEG features between healthy subjects and stroke patients. Also, the FBCNet source code is available at https://github.com/ravikiran-mane/FBCNet.

Figures (6)

• Figure 1: Proposed network architecture: FBCNet. ($C$: number of EEG channels, $T$: number of time points, $N_b$: number of frequency bands, $m$: number of convolution filters per frequency band, $N_c$: number of output classes)
• Figure 2: Classification accuracy for each subject from OpenBMI Data in 10-fold cross-validation settings (sorted by FBCSP-SVM acc.). It can be observed that deep learning architectures (Deep ConvNet, EEGNet-8,2) performed far better than FBCSP-SVM, the classical approach, for subjects with FBCSP-SVM accuracy $<$70%. On the other end of the spectrum, the performance of deep learning architectures was far worse than that of FBCSP-SVM in subjects with FBCSP-SVM accuracy $>$70%. Here, in contrast, FBCNet matched the performance of the best performing method for most of the subjects resulting in the best subject averaged classification accuracy. A similar trend was observed in other datasets as well.
• Figure 3: The sensitivity of classification algorithms to small training sets. Here, the effect of a small amount of data on the test accuracy is evaluated using the BCIC-IV-2A Data for various algorithms. A fraction of the training data (x-axis) is used to train a model, which is tested on data from an independent test session. It can be observed that the baseline deep learning architectures (Deep ConvNet and EEGNet-8,2) are highly sensitive to the small training sets, whereas the classical approach of FBCSP-SVM is relatively less susceptible. The proposed method (FBCNet) matches the accuracy achieved by deep learning methods in the presence of ample training data while retaining relatively better performance even when the training set is small. The error bars represent a standard mean error.
• Figure 4: FBCNet cross validation classification accuracies with different temporal feature extraction layers (mean$\pm$std). Temporal feature extraction using Variance layer resulted in best classification accuracies across all datasets.
• Figure 5: FBCNet cross-validation classification accuracies with different number of spatial filters per frequency band ($m$) and variance window length ($w$).