Resumen
As brain-computer interfaces (BCI) must provide reliable ways for end users to accomplish a specific task, methods to secure the best possible translation of the intention of the users are constantly being explored. In this paper, we propose and test a number of convolutional neural network (CNN) structures to identify and classify single-trial P300 in electroencephalogram (EEG) readings of an auditory BCI. The recorded data correspond to nine subjects in a series of experiment sessions in which auditory stimuli following the oddball paradigm were presented via earphones from six different virtual directions at time intervals of 200, 300, 400 and 500 ms. Using three different approaches for the pooling process, we report the average accuracy for 18 CNN structures. The results obtained for most of the CNN models show clear improvement over past studies in similar contexts, as well as over other commonly-used classifiers. We found that the models that consider data from the time and space domains and those that overlap in the pooling process usually offer better results regardless of the number of layers. Additionally, patterns of improvement with single-layered CNN models can be observed.