FNIRS DATA CLASSIFICATION FOR BRAIN COMPUTER INTERFACE USING DEEP LEARNING

Abstract

Brain Computer Interfaces (BCIs) translate recorded brain data directly to machine commands that can be used to control external devices. They are composed of three different functions i.e. recording of data from the brain, processing of data to recognize the intention of the subject and translation of the data into appropriate command for the machine being controlled. Functional near-infrared spectroscopy (fNIRS) is among one of the brain signal recording techniques which uses near-infrared spectroscopy (NIRS) for functional neuroimaging. It uses near-infrared light wavelengths (between 650 and 1000 nm) to measure the optical absorption changes of brain tissues. Use of fNIRS for BCIs limited because of slow hemodynamic response to stimulus, blood flow in scalp and undeveloped techniques for classifying signals. In this thesis we train Convolutional Neural Networks to classify fNIRS signals for BCIs. These networks classify raw signals with more than 95% testing accuracy for cognitive and imagery tasks with upto five separate categories in less than two milliseconds, (dependent on the processing power available), thus showing promising improvement in current classification efficiency.