[Paper Review] FBCNet: A Multi-view Convolutional Neural Network for Brain-Computer Interface

Ravikiran Mane, Effie Chew+6 more|arXiv (Cornell University)|Mar 17, 2021

EEG and Brain-Computer Interfaces6 references123 citations

TL;DR

FBCNet introduces a multi-view spectro-spatial CNN with a novel Variance layer for motor-imagery EEG classification, achieving state-of-the-art results on BCIC-IV-2A and strong gains on other datasets, including stroke patients; source code is available.

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.

Motivation & Objective

• Motivate MI-EEG classification challenges due to limited training data and noisy high-dimensional features.
• Propose a compact CNN architecture that encodes neurophysiological priors of motor imagery.
• Introduce a Variance layer to effectively extract temporal EEG information.
• Demonstrate improved classification accuracy across healthy and stroke populations.
• Provide interpretable insights into discriminative EEG features via explainable AI techniques.

Proposed method

• Develop multi-view EEG representation by spectrally filtering raw EEG into nine narrow bands (4-8 Hz to 36-40 Hz).
• Apply a Depthwise Convolution spatial block to learn spectro-spatial patterns with a fixed number of spatial filters per band.
• Introduce a Variance layer to compute temporal variance over non-overlapping windows (w) to reduce feature dimensionality and emphasize ERD/ERS-related fluctuations.
• Use a fully connected layer with a softmax activation for final class probabilities, with weight normalization for regularization.
• Train end-to-end with Adam optimizer in a two-stage scheme (training+validation, then full training) and employ early stopping based on validation performance.

Experimental results

Research questions

• RQ1 Can a neurophysiologically informed multi-view CNN improve MI-EEG classification with limited training data?
• RQ2 Does the proposed Variance layer provide superior temporal feature extraction compared to standard pooling?
• RQ3 How does FBCNet perform across healthy and chronic-stroke populations compared to classical ML and other CNN architectures?
• RQ4 What insights can be gained about discriminative EEG features in healthy vs. stroke subjects using explainable AI?
• RQ5 Is FBCNet robust to small training datasets and inter-subject variability?

Key findings

• FBCNet achieves 76.20% 4-class accuracy on BCIC-IV-2A in Hold-Out testing, setting a new state-of-the-art.
• On OpenBMI and stroke datasets, FBCNet provides up to 8% higher binary accuracies than competitive baselines.
• Across datasets, FBCNet consistently outperforms FBCSP-SVM, Deep ConvNet, and EEGNet-8,2 in CV and HO analyses.
• Variance layer yields the best temporal feature extraction compared to average or max pooling across all datasets.
• FBCNet shows greater utility for subjects with lower baseline accuracies and improves usability for stroke patients (more subjects achieving >70% accuracy).
• Explainable AI (DeepLIFT) reveals differences in discriminative EEG features between healthy and stroke populations.