InceptionSSVEP: A Multi-Scale Convolutional Neural Network for Steady-State Visual Evoked Potential Classification

Abstract

This paper presents a deep learningbased classification model, referred to as InceptionSSVEP, for the steady-state visual evoked potential (SSVEP) based braincomputer interface (BCI). InceptionSSVEP adopts the main concept of Inception network, which is a deep learning model performing well in image classification tasks, to improve the performance of SSVEP classification. A multi-scale convolution structure is utilized in InceptionSSVEP to extract both long-term and short-term features of SSVEP signals, for the purpose of ensuring the comprehensiveness of high-dimensional features in extracted SSVEPs. Moreover, a data enhancement scheme is proposed to overcome the impact of SSVEP data amount limitation on classifier training. Results show that the proposed InceptionSSVEP outperforms other existing methods significantly, and validate that Inception networks have good transferability on SSVEP signals. Reasons for the good performance of InceptionSSVEP are analyzed using deep learning interpretability tools.