Unsupervised learning for brain–computer interfaces based on event-related potentials

Machine learning has become a core component in brain-computer interfaces (BCIs). Unfortunately, the use of machine learning typically requires the collection of subject specific labelled data. This process is time-consuming and not productive from a user's point of view, as during calibration the user has to follow given instructions and cannot make own decisions. Only after calibration, the user is able to use the BCI freely. In this chapter, we describe how the supervised calibration process can be circumvented by unsupervised learning in which the decoder is trained while the user is utilising the system. We discuss three variations. First, expectation-maximisation (EM)-based training, which works wells empirically but can sometimes be unstable. Learning from label proportions (LLP)-based training, which is guaranteed to converge to the optimal solution, but learns more slowly. Third, a hybrid approach combining the stability of LLP with the speed of learning of EM in a highly efficient and effective approach that can readily replace supervised decoders for event-related potential BCI.