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

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