MAML-EEG: A Meta-learning Strategy Based Domain Generalization Framework for Unseen Subject Motor Imagery Classification

Abstract Since the acquisition of electroencephalogram (EEG) is time-consuming and costly, the target subjects is usually unknown in the practical application of motor imagery (MI) based brain-computer interface (BCI) system. Due to the heterogeneity of individuals, the distribution of MI signals among subjects is quite different, which makes it difficult for traditional deep learning methods to achieve cross-unknown subjects in practical applications. We propose a meta-learning (ML) based method which alleviates this problem by designing models that generalize well to new test subjects. Specifically, our framwork simulates train/test subject shift during training stage by randomly sampling partial samples of partial subjects to synthesize a virtual meta-task set within each batch. The final meta-optimization improves the test performance of multiple subjects while reducing the meta-training loss of each subject. We evaluated our method on cross-unknown subjects MI classification, as well obtaining its advanced performances on two calssic available MI datasets.