Interpretable Representation Learning for Biosignals via Variable Projection

Kovács, Péter (2024) Interpretable Representation Learning for Biosignals via Variable Projection. In: Workshop Biosignals, 2024.02.28-2024.03.01, Göttingen.

Preview

Text WSBiosignals2024_Kovacs2.pdf Available under License Creative Commons Attribution. Download (892kB) | Preview

Abstract

Over the last decade, deep learning (DL) has exhibited a huge advancement in many fields including medical sciences, although they continue to raise several concerns. First, due to the large number of nonlinear model parameters, DL approaches can be considered as black-box methods, where the parameters have no or little physical meaning. Second, training these DL models requires vast amounts of labeled data, which is problematic to collect in medical applications due to the involvement of clinical experts, the measurement costs, and the usually low prevalence of abnormal cases. In order to surpass the limitations of pure data-driven DL approaches, the so-called model-driven machine learning methods have been recently introduced. This concept is an emerging trend in signal processing, which combines the advantages of modelbased methods and DL techniques. In this study, we propose VPNet, a novel model-driven neural network architecture based on variable projection (VP) that is able to learn efficient representations of various types of biosignals, including visually evoked potentials, photoplethysmograms, and electrocardiograms.

Actions (login required)

Edit Item