Unsupervised domain adaptation with synchronized self-training for EEG-based motor imagery recognition (2025)
Open in webOpen in zoteroOpen pdf
1 Abstract
Robust decoding performance is urgently desired for the widespread applications of brain-computer interface (BCI) systems. In general, most EEG decoding models are carefully designed for specific subjects, each with independent training, and their accuracy relies on extensive annotated data and timeconsuming calibration. Such a situation can not meet for the demand of practical applications, especially in the field of BCIbased rehabilitation, where is often confronted with problems in acquiring large amounts of data from patients. To address these issues, we explore a solution, named Synchronized Self-Training Domain Adaptation (SSTDA), for EEG-based motor imagery classification. Specifically, SSTDA framework leverage the signals from label-rich source domain and adopt the self-training scheme in unlabelled target domain for simultaneous training of a more robust classifier. On one hand, the discriminative representations learning is empowered by mapping raw EEG signals into a latent space with a feature extractor. On the other hand, a domain-shared representation space can be learned by optimizing feature extractor with source and target samples jointly, in a self-training manner. To validate the efficacy of the proposed method, we conduct extensive experiments on two public motor imagery datasets, including Dataset IIa of BCI Competition IV and High Gamma dataset. Our work achieved average classification performance of 60.15% and 78.50% in the intersubject task, respectively. Meanwhile, the performance on intersession are better than those of previous state-of-the-art methods. All the results demonstrate that our SSTDA method is capable of learning discriminative and domain-invariant representations for improving decoding performance and outperforming the other algorithms.
2 NOTES
Quotes:
Quote
Quote here.