Convolutional Monge Mapping Normalization for studying on sleep knowledge
Authors: Théo Gnassounou, Rémi Flamary, Alexandre Gramfort
Summary: In lots of machine studying functions on indicators and biomedical knowledge, particularly electroencephalogram (EEG), one main problem is the variability of the info throughout topics, classes, and {hardware} gadgets. On this work, we suggest a brand new methodology known as Convolutional Monge Mapping Normalization (CMMN), which consists in filtering the indicators in an effort to adapt their energy spectrum density (PSD) to a Wasserstein barycenter estimated on coaching knowledge. CMMN depends on novel closed-form options for optimum transport mappings and barycenters and gives particular person take a look at time adaptation to new knowledge with no need to retrain a prediction mannequin. Numerical experiments on sleep EEG knowledge present that CMMN results in important and constant efficiency features unbiased from the neural community structure when adapting between topics, classes, and even datasets collected with completely different {hardware}. Notably our efficiency acquire is on par with way more numerically intensive Area Adaptation (DA) strategies and can be utilized at the side of these for even higher performances