Neural Network Sparsity in Brain-Body-Machine Interfaces.

Brain-body-machine interfaces acquire, process, and translate brain signals for individuals with severe motor impairments to communicate and control the assistive technology that supports their daily life activities. Electroencephalography (EEG) is a standard approach for acquiring such brain signals due to its low cost and high temporal resolution. EEG signals can be thought of as a proxy for the user's intent. One established method for translating this intent into inferences and actions are neural networks. However, densely connected neural networks can be computationally expensive-a problem for real-time, deployed brain-body-machine interface systems. In this paper we investigate the use of sparsity in neural networks for EEG-based motor classification, with the goal of reducing the number of neuronal connections without sacrificing a system's performance. We compare two sparsity-inducing algorithms, weight pruning and sparse evolutionary training, with a dense neural network under three experimental conditions. Overall, our results show that sparse neural networks can achieve higher performance accuracy and generalization than their densely-connected counterparts for an EEG-based classification task. We found that sparse evolutionary training achieves the highest and most stable performance across all experiments. Introducing sparsity into the network is a viable option for efficient EEG-based control, with promising applications in a range of related rehabilitation and assistive technologies. This brings us closer to helping individuals with severe motor impairments reclaim independence through more computationally realizable methods of interacting with their technology and the world around them.