Compact CNN with Dynamic Window for SSVEP-based BCIs

Abstract

Conventional recognition methods for the detection of steady-state visual evoked potential (SSVEP) usually use the strategy of fixed window. These methods would cost a lot of time finding an optimal window length by iterating various window lengths. Additionally, due to the instability and complexity of the electroencephalogram (EEG), the optimal window length is not static and is interfered by surroundings. Dynamic window methods were proposed to solve above problems, which could dynamically choose an optimal window length and output results at an appropriate time. Filter bank canonical correlation analysis with dynamic window (FBCCA-DW) is a dynamical training-free method which combines FBCCA method and a designed dynamic window strategy. FBCCA-DW achieves better performance than other similar methods. However, FBCCA method only extracts the components of SSVEP at the fundamental and harmonic frequencies, while some other small but essential information would be lost. Compact convolutional neural network (EEGNet) is a subject-independent training method which can extract more features from raw SSVEP data and outperforms FBCCA. In this study, EEGNet with dynamic window (EEGNet-DW) is proposed to make use of EEGNet and the dynamic window strategy. We evaluate EEGNet-DW across subjects on a public SSVEP dataset. We show that the EEGNet-DW, with an information transfer rate (ITR) of 147.6 bits/min, performed better than EEGNet with fixed window (EEGNet-FW) and FBCCA-DW. Our proposed method shows promise for real-time SSVEP-based BCI applications.