Graph-Based Transfer Learning for Managing Brain Signals Variability in NIRS-Based BCIs

摘要: One of the major limitations to use brain-computer interfaces (BCIs) based on near-infrared spectroscopy (NIRS) in realistic interaction settings is long calibration time needed before every order train a subject-specific classifier. way reduce this data collected from other users or previous recording sessions same user as training set. However, brain signals are highly variable and using heterogeneous single classifier may dramatically deteriorate classification performance. This paper proposes transfer learning framework which we model variability feature space bipartite graph. The partitioning graph into sub-graphs allows creating homogeneous groups NIRS sharing similar spatial distributions explanatory variables will be used multiple prediction models that accurately knowledge between sets.