Developing machine-learning methods for the analysis of electromagnetic brain activity

Traditionally, analysis of electromagnetic brain activity focuses on modeling the data-generating process and identifying which components in the measured signal are associated with experimental manipulations. At the beginning of XXI century, machine-learning based approaches aiming to infer brain states from the measurements started to gain increasing popularity. These methods rely on extracting complex multivariate patterns allowing to predict experimental conditions from the measurements. This thesis summarizes how such prediction-based methods can be applied to measurements of electromagnetic brain activity in a way that allows to advance our understanding of the underlying neural processes. Because these techniques belong to a class of inverse probability problems and do not model the data-generating process directly, interpreting the learning outcomes in terms of the underlying neurophysiological processes is not straightforward. Instead, predictive models allow testing the generalization properties of brain activity across e.g. experimental tasks (Publication I) and individuals (Publication II), as well as employ model comparison techniques to gain additional insights about the statistical properties of the data- generating process indirectly i.e. by comparing models with different structural constraints (Publication II). Moreover, projecting relevant model parameters learned from the data back into the input space can provide additional insights into the data-generating process and thus complement traditional approaches. These approaches are implemented in an open-sourceacademic software described in Publication III.