Continual Learning of a Transformer-Based Deep Learning Classifier Using an Initial Model from Action Observation EEG Data to Online Motor Imagery Classification

Abstract

The motor imagery (MI)-based brain computer interface (BCI) is an intuitive interface that enables users to communicate with external environments through their minds. However, current MI-BCI systems ask naïve subjects to perform unfamiliar MI tasks with simple textual instruction or a visual/auditory cue. The unclear instruction for MI execution not only results in large inter-subject variability in the measured EEG patterns but also causes the difficulty of grouping cross-subject data for big-data training. In this study, we designed an BCI training method in a virtual reality (VR) environment. Subjects wore a head-mounted device (HMD) and executed action observation (AO) concurrently with MI (i.e., AO + MI) in VR environments. EEG signals recorded in AO + MI task were used to train an initial model, and the initial model was continually improved by the provision of EEG data in the following BCI training sessions. We recruited five healthy subjects, and each subject was requested to participate in three kinds of tasks, including an AO + MI task, an MI task, and the task of MI with visual feedback (MI-FB) three times. This study adopted a transformer- based spatial-temporal network (TSTN) to decode the user’s MI intentions. In contrast to other convolutional neural network (CNN) or recurrent neural network (RNN) approaches, the TSTN extracts spatial and temporal features, and applies attention mechanisms along spatial and temporal dimensions to perceive the global dependencies. The mean detection accuracies of TSTN were 0.63, 0.68, 0.75, and 0.77 in the MI, first MI-FB, second MI-FB, and third MI-FB sessions, respectively. This study demonstrated the AO + MI gave an easier way for subjects to conform their imagery actions, and the BCI performance was improved with the continual learning of the MI-FB training process.