An empirical suggestion for collaborative learning in motor imagery-based BCIs

In modern Brain-Computer Interfaces (BCIs), it usually requires the so-called calibration session to adapt a BCI model, e.g., spatial filter and classifier, to a target subject before use, due to high intra- and inter-subject variability in brain signals. From a practical perspective, this is one of the main challenges that should be resolved, thus motivating to use information from other subjects via collaborative learning. In this study, we analyze the effects of utilizing data from other subjects and identify whether generic patterns, which are informative for general BCI, exist by conducting experiments on the BCI Competition IV-IIa dataset. Based on our two experiments of naïve inter-subject BCI and generic pattern-guided inter-subject BCI, we suggest utilizing 1) categorical information of training samples and 2) samples of generic patterns for generalization of a BCI model.