Manifold embedded instance selection to suppress negative transfer in motor imagery-based brain–computer interface

Abstract

Individual distribution discrepancy poses significant challenges to cross-subject electroencephalography (EEG) signal decoding. Although transfer learning has emerged as an effective approach to minimize this distribution gap, EEG signals’ low voltage amplitude makes them vulnerable to noise and results in abnormal, low-quality samples that cause negative transfer phenomenon. This phenomenon, in turn, undermines the efficacy of transfer learning and impedes brain-computer interface (BCI) applications. To overcome this challenge, we introduce the manifold embedded instance selection (MEIS) algorithm, which addresses negative transfer. The MEIS algorithm operates in two ways: converting raw EEG matrices into manifold embedded vectors that maintain sample discriminability, and designing an evaluator to assess the transferability of samples and filter out negative transfer samples from the source domain. When faced with a large number of source domains, our proposed method employs domain similarity estimation to determine the most beneficial subset of source domains for the target domain. The effectiveness of the proposed method is confirmed via offline and simulated online motor imagery-based BCI experiments where it demonstrates superior accuracy compared to other advanced techniques. Additionally, the MEIS algorithm significantly reduces both training time and the number of required training samples without affecting the model’s performance. The code is available at https://github.com/ZilinL/MEIS.