UC San Diego

We develop and test three deep-learning recurrent convolutional architectures for

learning to recognize single trial EEG event-related potentials for P300 brain-computer

interfaces (BCI)s. The existing classifiers for P300 detection don't preserve the spatiotemporal

structure of the data, thereby losing local spatial and temporal patterns in the

data. The proposed models respect the spatial and temporal nature of the EEG signals.

A three-dimensional convolutional neural network (3D-CNN) is used in concert

with a two-dimensional convolutional neural network (2D-CNN) and LSTM to capture

the spatiotemporal patterns in the EEG signals. Moreover, a transfer learning based approach is applied while training the subjects. One advantage of the neural network

solution is that it provides a natural way to share a lower-level feature space between

subjects while adapting the classifier that works on that feature space. We compare the

deep neural networks with the standard methods for P300 BCI classification.