A Multi-view CNN with Novel Variance Layer for Motor Imagery Brain Computer Interface.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Classifying motor imagery (MI) from electroencephalography (EEG) signals is crucial for Brain-Computer Interfaces (BCI).
• Existing deep learning models face challenges in accurately and robustly classifying MI signals.

Purpose of the Study:

• To propose a novel, neuro-physiologically inspired convolutional neural network (CNN) named Filter-Bank Convolutional Network (FBCNet) for enhanced MI classification.
• To capture neurophysiological signatures of MI for improved BCI performance.

Main Methods:

• FBCNet utilizes bandpass-filtering to create multi-view EEG representations.
• Spatially discriminative patterns are learned via CNN layers, and temporal information is aggregated using a variance layer.
• Classification into MI classes is performed using a fully connected layer.

Main Results:

• FBCNet achieved over 6.7% higher accuracy than state-of-the-art deep learning architectures on a public dataset.
• The model requires less than 1% of the learning parameters compared to existing methods.
• Feature visualization demonstrated FBCNet's superiority in learning interpretable and generalizable discriminative features.

Conclusions:

• FBCNet offers a highly accurate and efficient solution for motor imagery classification in BCI.
• The neuro-physiologically inspired design enables superior learning of relevant EEG features.
• The source code is provided for result reproducibility and further research.