Deep learning based on Batch Normalization for P300 signal detection

Abstract

Detecting P300 signals from electroencephalography (EEG) is the key to establishing a P300 speller, which is a type of braincomputer interface (BCI) system based on the oddball paradigm that allows users to type messages simply by controlling eye-gazes. The convolutional neural network (CNN) is an approach that has achieved good P300 detection performances. However, the standard CNN may be prone to overfitting and the convergence may be slow. To address these issues, we develop a novel CNN, termed BN3, for detecting P300 signals, where Batch Normalization is introduced in the input and convolutional layers to alleviate over-fitting, and the rectified linear unit (ReLU) is employed in the convolutional layers to accelerate training. Since our model is fully data-driven, it is capable of automatically capturing the discriminative spatio-temporal features of the P300 signal. The results obtained on previous BCI competition P300 data sets show that BN3 both achieves the state-of-the-art character recognition performance and that it outperforms existing detection approaches with small flashing epoch numbers. BN3 can be used to improve the character recognition performance in P300 speller systems.