Show simple item record

dc.contributor.authorNiroomand, MP
dc.contributor.authorCafolla, Conor
dc.contributor.authorMorgan, John
dc.contributor.authorWales, David
dc.date.accessioned2022-01-28T16:40:07Z
dc.date.available2022-01-28T16:40:07Z
dc.date.issued2022-03-01
dc.date.submitted2021-11-10
dc.identifier.issn2632-2153
dc.identifier.othermlstac49a9
dc.identifier.otherac49a9
dc.identifier.othermlst-100459.r1
dc.identifier.urihttps://www.repository.cam.ac.uk/handle/1810/333218
dc.description.abstract<jats:title>Abstract</jats:title> <jats:p>One of the most common metrics to evaluate neural network classifiers is the area under the receiver operating characteristic curve (AUC). However, optimisation of the AUC as the loss function during network training is not a standard procedure. Here we compare minimising the cross-entropy (CE) loss and optimising the AUC directly. In particular, we analyse the loss function landscape (LFL) of approximate AUC (appAUC) loss functions to discover the organisation of this solution space. We discuss various surrogates for AUC approximation and show their differences. We find that the characteristics of the appAUC landscape are significantly different from the CE landscape. The approximate AUC loss function improves testing AUC, and the appAUC landscape has substantially more minima, but these minima are less robust, with larger average Hessian eigenvalues. We provide a theoretical foundation to explain these results. To generalise our results, we lastly provide an overview of how the LFL can help to guide loss function analysis and selection.</jats:p>
dc.description.sponsorshipEPSRC Downing College, Cambridge Interdisciplinary Institute for Artificial Intelligence at 3iA Cote d'Azur
dc.languageen
dc.publisherIOP Publishing
dc.subjectPaper
dc.subjectarea under the curve
dc.subjectloss function landscape
dc.subjectbasin hopping
dc.subjectalternative loss function
dc.subjectloss function
dc.titleCharacterising the area under the curve loss function landscape
dc.typeArticle
dc.date.updated2022-01-28T16:40:06Z
prism.issueIdentifier1
prism.publicationNameMachine Learning: Science and Technology
prism.volume3
dc.identifier.doi10.17863/CAM.80641
dcterms.dateAccepted2022-01-07
rioxxterms.versionofrecord10.1088/2632-2153/ac49a9
rioxxterms.versionVoR
rioxxterms.licenseref.urihttp://creativecommons.org/licenses/by/4.0
dc.contributor.orcidNiroomand, MP [0000-0002-7189-0456]
dc.contributor.orcidCafolla, Conor [0000-0003-2021-974X]
dc.contributor.orcidMorgan, John [0000-0002-9157-9278]
dc.contributor.orcidWales, David [0000-0002-3555-6645]
dc.identifier.eissn2632-2153
pubs.funder-project-idEngineering and Physical Sciences Research Council (EP/N035003/1)
cam.issuedOnline2022-01-21


Files in this item

Thumbnail
Thumbnail

This item appears in the following Collection(s)

Show simple item record