An algorithm for movement related potentials feature extraction based on transfer learning

Abstract

Movement related potentials (MRPs) are utilized as features in many motor related brain-computer interfaces (BCIs). MRP feature extraction is challenging since multi-channel brain signals are high dimensional and often contains various artifacts. The discriminative spatial pattern (DSP) algorithm successfully improves the signal-to-noise ratio of MRPs. However, abundant labeled training data are required for DSP to learn reliable spatial filters for each subject respectively. This is inconvenient for the applications of BCIs. In this paper, we propose a regularized DSP (RDSP) algorithm for MRP feature extraction, which does not need any labeled training data for a new subject. The regularization function of RDSP is built on empirical maximum mean discrepancy (MMD) to reduce the differences not only in marginal distribution but also in conditional distribution between subjects. RDSP transfers the common discriminative spatial filters across subjects and updates them iteratively by semi-supervised learning. Experiment results on BCI competition datasets show the effectiveness of RDSP.