Convolutional neural network architecture and input volume matrix design for ERP classifications in a tactile P300-based Brain-Computer Interface

In the presented study we conduct the off-line ERP classification using the convolutional neural network (CNN) classifier for somatosensory ERP intervals acquired in the full-body tactile P300-based Brain-Computer Interface paradigm (fbBCI). The main objective of the study is to enhance fbBCI stimulus pattern classification accuracies by applying the CNN classifier. A 60 × 60 squared input volume transformed by one-dimensional somatosensory ERP intervals in each electrode channel is input to the convolutional architecture for a filter training. The flattened activation maps are evaluated by a multilayer perceptron with one-hidden-layer in order to calculate classification accuracy results. The proposed method reveals that the CNN classifier model can achieve a non-personal-training ERP classification with the fbBCI paradigm, scoring 100 % classification accuracy results for all the participated ten users.