Performance Evaluation of Compressed Deep CNN for Motor Imagery Classification using EEG.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise in classifying Motor Imagery (MI) from Electroencephalography (EEG) data.
• Deploying CNN-based Brain-Computer Interfaces (BCIs) on resource-constrained mobile and embedded devices presents significant computational and memory challenges.

Purpose of the Study:

• To investigate the effectiveness of magnitude-based weight pruning for reducing the parameters of pre-trained CNN classifiers for MI tasks.
• To assess the impact of pruning on classification performance and model compression for practical BCI applications.

Main Methods:

• Utilized a magnitude-based weight pruning technique to reduce the number of parameters in a pre-trained CNN model.
• Evaluated the pruned model on an open-source Korea University dataset comprising EEG recordings from 54 healthy subjects performing left- and right-hand MI tasks.
• Assessed subject-independent model performance, compression ratio, and classification accuracy post-pruning.

Main Results:

• The subject-independent CNN model achieved 90% sparsity through pruning, resulting in a compression ratio of 4.77×.
• Classification accuracy was maintained at 84.44%, with only a 0.02% loss compared to the baseline model.
• Demonstrated that significant model compression is achievable without substantial performance degradation.

Conclusions:

• Magnitude-based weight pruning is a viable technique for creating compact deep CNN-based BCIs.
• The proposed method enables the development of efficient BCIs suitable for deployment on mobile and embedded systems with limited resources.
• This approach facilitates the practical application of advanced deep learning models in real-world BCI scenarios.