Regularization using similarities of signals observed in nearby sensors for feature extraction of brain signals

In order to solve uncertainty of spatial weights learned with small amount of training samples for feature extraction from brain signals, a regularization using similarity of signals observed in sensors that are located near each other is proposed. Deriving the regularization is begun defining a distance between the sensors. Under the distance, the proposed regularization works so that the spatial weights extracts similar signals in the nearby sensors. The proposed regularization is applied to the well known common spatial pattern (CSP) method that finds spatial weights for EEG based brain machine interface. In the classification experiment using a dataset of EEG signals during motor imagery, the proposed method achieved maximum improvement by 28% in the classification accuracy over the standard CSP in a setting of even when only five samples are used.