Convolutional Neural Networks and Feature Selection for BCI with Multiresolution Analysis

Abstract

Classification in high-dimensional feature spaces is a difficult task, often hindered by the curse of dimensionality. This is the case of Motor Imagery tasks involving Brain-Computer Interfaces through electroencephalography, where the number of available patterns is limited, making more noticeable the effect of the high dimensionality on the generalization capabilities of the models. This paper tackles classification in that particular setting, drawing a comparison between an explicit feature selection procedure using evolutionary computation and an implicit feature selection using Convolutional Neural Networks. These two alternatives are also compared to a Support Vector Machine approach that serves as a baseline quality threshold. According to the experiments performed in this paper, Convolutional Neural Networks are able to produce promising results when compared to the Support Vector Machine models and, after a partial hyperparameter optimization stage, also to previous work on the same dataset. Furthermore, this raises the issue of the trade-off between computational cost and classification accuracy, which is briefly discussed when assessing the quality of the results in relation to existing work.