An Approach to Use Convolutional Neural Network Features in Eye-Brain-Computer-Interface

We propose an approach to use the features formed by a convolutional neural network (CNN) trained on big data for classification of electroencephalograms (EEG) in the eye-brain-computer interface (EBCI) working on short (500 ms) gaze fixations. The multidimensional EEG signals are represented as 3D-images that makes possible to apply them to CNN input. The features for EEG classifier are selected from first fully connected CNN layer’s outputs. It is shown that most of them are useless for classification but at the same time, there were a relatively small number of CNN-features with a good separating ability. Their use together with the EEG amplitude features improved the sensitivity of a linear binary classifier applied to an EEG dataset obtained in an EBCI experiment (when participants played a specially designed game EyeLines) by more than 30% at a fixed specificity of 90%. The obtained result demonstrates the efficiency of the features formed by the CNN trained on big data even with respect to the essentially different classification task.