A speedup SVM decision method for online EEG processing in motor imagery BCI

Abstract

In BCI research community, support vector machine (SVM) is an effective method for motor imagery (MI)-based electroencephalographic (EEG) classification. However, the computation of decision function during SVM classification stage for a new EEG trial is time-consuming due to the large number of support vectors (SV). This paper proposes a new method to reduce the number of support vectors so that speed up SVM decision. The method first obtains all the support vectors by classical SVM. Then, γ-index measuring the average distance between each support vector and its nearest neighbors is evaluated. Thirdly, the support vector with smallest γ-index is selected. And then iteratively re-weight γ-index and select only a few support vectors to represent all the support vectors. Our experiments show only 10%-30% of the support vectors can be used to speed up the decision while loss in generalization performance remains acceptable.