Nathan Safir
Institute for Artificial Intelligence, University of Georgia, Athens, GA 30602 USA

Meekail Zain
Department of Computer Science, University of Georgia, Athens, GA 30602 USA

Curtis Godwin
Institute for Artificial Intelligence, University of Georgia, Athens, GA 30602 USA

Eric Miller
Institute for Artificial Intelligence, University of Georgia, Athens, GA 30602 USA

Bella Humphrey
Department of Computer Science, University of Georgia, Athens, GA 30602 USA

Shannon P Quinn
Department of Computer Science, University of Georgia, Athens, GA 30602 USA
Department of Cellular Biology, University of Georgia, Athens, GA 30602 USA

Abstract

Deep metric learning (DML) methods generally do not incorporate unlabelled data. We propose borrowing components of the variational autoencoder (VAE) methodology to extend DML methods to train on semi-supervised datasets. We experimentally evaluate the atomic benefits to the perform- ing DML on the VAE latent space such as the enhanced ability to train using unlabelled data and to induce bias given prior knowledge. We find that jointly training DML with an autoencoder and VAE may be potentially helpful for some semi-suprevised datasets, but that a training routine of alternating between the DML loss and an additional unsupervised loss across epochs is generally unviable.

Keywords

Variational Autoencoders, Metric Learning, Deep Learning, Representation Learning, Generative Models

DOI

10.25080/majora-212e5952-022

Bibtex entry

Full text PDF

Copyright © 2022 Nathan Safir et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.