MsCNN: A Deep Learning Framework for P300-Based Brain-Computer Interface Speller

In this paper, a novel multiscale convolutional neural network (MsCNN) architecture is proposed for P300 based BCI speller. Major limitation of BCI system is that it requires a large number of data to train the system. However, collection of large amount of data from a single subject makes the task tedious for the subject and this will affect the quality of the acquired electroencephalogram (EEG) signal. To cope with this issues, a MsCNN model with transfer learning (MsCNN-TL) technique is proposed in this work for improvement of the P300 based character recognition performance with limited amount of training data. The multi-resolution deep features are extracted from the fully connected layer of the trained MsCNN-TL architecture. These deep features are optimized using Fisher ratio (F-ratio) based feature selection technique and the optimal features are applied to the ensemble of support vector machines (ESVMs) for P300 detection. The proposed method is tested on BCI competition datasets and it achieves better performance compared to the other state-of-the-art techniques for limited training dataset.