A Convolution Network of Multi-Windows Spatial-Temporal Feature Analysis for Single-trial EEG Classification in RSVP Task

It is a challenge to reducing the calibration time of the brain-computer interfaces (BCI) system in the rapid serial visual presentation (RSVP) paradigm. However, the short calibration time can cause the problems, such as small training data, the extremely low signal-to-noise ratio of event-related potentials (ERPs), and inter-trial variability of ERPs, which will increase classification difficulty. In this work, a novel convolution network of multi-windows spatial-temporal features analysis was proposed to alleviate the temporal variability and improve the classification performance for single-trial EEG data. According to the phase-locked information of ERPs, the single-trial was split as the input of the network using the sliding window method. The network adopted three depthwise convolution layers to learn the spatiotemporal features in different windows. The separable convolution was utilized to extract the global features of all windows. Compared with several state-of-the-art algorithms using RSVP datasets of 12 subjects, the proposed network had better classification performance and the online application potential of RSVP-BCI.