ABSTRACT
Classification and regression in brain–computer interfaces (BCIs) suffer from performance variations across subjects as well as across sessions within individual subjects due to various technical and biological reasons. Such variations limit the amount of data that can be used to train machine learning models, rendering BCIs highly subject specific and requiring tedious calibration and retraining before each usage in order to achieve optimal performance. This chapter presents state-of-the-art approaches to calibration-free and improved decoding through transfer learning techniques from multiple subjects and sessions. Within the feature space, a generalized framework for the popular common spatial patterns is described in terms of covariance regularization with multiple subjects. Recent methods using Riemannian features offer invariance to spatial characteristics between subjects and sessions. Finally, a probabilistic Bayesian multitask learning framework is presented that works on top of feature spaces and learns prior distribution over decision rules of different subjects and sessions. The mathematical details are left in for more technically inclined readers, but intuitions and (when applicable) code bases are referenced for the more pragmatic developer.
