A multi-scale EEGNet for cross-subject RSVP-based BCI system

Abstract

In the cross-subject classification task, a subject-agnostic model is trained for the classification task of other subjects, according to the prior knowledge from EEG data of some subjects. It is one of the challenges for ERP classification in the RSVP-based BCI system. So far, convolutional neural networks (CNNs) for RSVP classification only use a fixed-size kernel for each layer to extract features in the temporal domain, which limits the ability of the network to detect ERP. In this work, a multi-scale EEGNet model (MS-EEGNet) for cross-subject RSVP classification task was proposed, which adopted parallel convolution layers with multi-scale kernels to extract discrimination information in the temporal domain, and increased the robustness of the model. The proposed model was used for the BCI Controlled Robot Contest in the World Robot Contest 2022 and achieved good results. The UAR of the A and B datasets got 0.493 and 0.528, respectively. Compared with other CNN algorithms including EEGNet and PLNet, the proposed model had better classification performance.