Transfer Learning through Kernel Alignment : Application to Adversary Data Shifts in Automatic Sleep Staging

This doctoral project aims at improving an automatic sleep staging system by taking into account inter-and-intra-individual variabilities, the latter having adversary effects on the classification. We focus on the detection of Rapid-Eye Movement periods during sleep. The core of our research is transfer learning and the selection of suitable detector(s) among a set, allowing the individualisation of the analysis by the exploitation of the observed data properties. We focus on the application of kernel alignment methods, firstly through the use of kernel-target alignment, studied here in a dual way, i.e. the kernel is fixed and the criterion is optimised with respect to the sought target labels. In a second step, we introduced kernel-cross alignment, allowing to take more efficiently advantage of the information contained in the training data. The ideas developed in the framework of this work have been extended to automatically selecting one or more efficient training sets for a given test set. The contributions of this work are both methodological and algorithmic, general in scope, but also focused on the application.