Neural Network Sparsity in Brain-Body-Machine Interfaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Brain-body-machine interfaces (BMIs) translate brain signals for individuals with severe motor impairments.
• Electroencephalography (EEG) is a cost-effective method for capturing brain signals, serving as a proxy for user intent.
• Dense neural networks, while effective, present computational challenges for real-time BMI applications.

Purpose of the Study:

• To investigate the efficacy of sparsity in neural networks for electroencephalography (EEG)-based motor classification.
• To reduce computational expense in BMI systems without compromising performance.
• To compare the performance of sparse neural networks against dense networks for EEG signal processing.

Main Methods:

• Utilized two sparsity-inducing algorithms: weight pruning and sparse evolutionary training.
• Compared sparse neural networks with a densely connected neural network.
• Evaluated performance under three distinct experimental conditions for EEG-based motor classification.

Main Results:

• Sparse neural networks demonstrated superior performance accuracy and generalization compared to dense networks for EEG-based motor classification.
• Sparse evolutionary training yielded the highest and most consistent performance across all experimental conditions.
• Introducing sparsity is a viable strategy for efficient EEG-based control in BMI systems.

Conclusions:

• Sparse neural networks offer a computationally efficient approach for EEG-based control in brain-body-machine interfaces.
• These findings have promising applications in assistive technologies and rehabilitation, promoting independence for individuals with motor impairments.
• Sparsity in neural networks represents a significant advancement towards more accessible and realizable BMI technologies.