Year: 2022
Type: report
Abstract: We explore the challenge of uncertainty evaluation in machine learning regression problems. We begin by outlining how regression problems typically arise in a metrology setting. We then perform a numerical investigation of three methods for performing uncertainty-aware machine learning regression, namely Gaussian Processes (GPs), Monte-Carlo Dropout and Deep Ensembles. Based upon our investigations, we identify a list of metrology requirements that methods for uncertainty evaluation in machine learning regression must satisfy, and we evaluate the three methods against these requirements. We conclude by describing how the requirements give rise naturally to several open challenges which constitute an agenda for future work.
DOI: 10.47120/npl.ms34
Authors: A Thompson, K Jagan, A Sundar, R Khatry, J Donlevy, S Thomas, P Harris
Institution: National Physical Laboratory
Language: English
Subfield: Statistics, Probability and Uncertainty
Field: Decision Sciences
Domain: Social Sciences
Open Access status: gold
Funder: Government of the United Kingdom
OpenAlex ID: W4206335083