Reducing the computational cost of a classifier by subtracting the dual-class features

This paper presents a methodology to reduce the computational cost of a dual-class classifier by means of subtracting the features. In this way, the classifier needs to deal with a relatively half length vector containing the difference of the features of the two classes. This approach reduces the cost of the classifier by one-half. The validity of the proposed method is tested with Electroencephalography based BCI system by conducting experiments targeting the right and left sensorimotor cortexes. The adaptive autoregressive modeled features of the EEG was extracted with both recursive least square and least mean square algorithms. The linear discriminant analysis is performed to classify the dual-class features with the classical features' staking approach and with the proposed features' subtracting approach. The results demonstrate the potential of the proposed method.