Brain-computer interface using deep neural network and its application to mobile robot control

Abstract

Functional near-infrared spectroscopic (fNIRS) systems have recently attracted considerable attention for their potential in the domain of brain-computer interfaces (BCIs). This study presents a method for brain activity classification using signals obtained through fNIRS and suggests strategies to optimize the classifier for real-time control applications. A deep neural network (DNN) classifies differing brain activity signals from the pre-frontal cortex that were generated by pre-defined activities. Optimization of the DNN showed that varying the number of neurons per layer does not adversely affect the classification accuracy past a certain size and using a dropout method during training further improves the classification accuracy of the DNN. In the offline classification trials, the DNN achieved an accuracy of 82% for the two-class (activity vs rest) classification. A control system for a mobile robot is conceived to explore the practical application of BCIs. The components of the input vector to the DNN were altered and a post-processing step was added to the output of the DNN to use an fNIRS-based BCI for real-time data classification. Trials with online data classification depicted the plausibility of using DNNs for real-time control with fNIRS-based BCIs; however, the maximum classification accuracy of the system is 66%, which renders it impractical for real-time application.