Ensemble learning based classification for BCI applications

Abstract

This paper reports the use of combinations of multiple learning models, a type of structure called ensemble system in the classification of movement imagination, an interesting topic in the Brain Computer Interface (BCI) field. The considered models are the well-known Multilayer Perceptron (MLP), perhaps the most often used Neural Network (NN) based models. The proposed method has been applied on the Graze data set 2b of the BCI competition IV. The data was pre-processed by selecting the most appropriate frequency range, correcting the most relevant artifacts such as EOG interference and source decorrelation by using PCA. Experimental results show that an ensemble system of MLP models can outperform a single MLP model if adequate training strategies are employed. For the current database and in the absence of other changes in the system the ensemble strategy outperforms conventional classification, both using MLPs in about 2% when a linear combination of the outputs of the ensemble components was adopted and in about 4% when a non-linear combination was used.