Show simple item record

dc.contributor.authorCipolla, Robertoen
dc.contributor.authorGal, Yen
dc.contributor.authorKendall, Alexen
dc.date.accessioned2018-07-16T15:42:35Z
dc.date.available2018-07-16T15:42:35Z
dc.date.issued2018-12-14en
dc.identifier.isbn9781538664209en
dc.identifier.issn1063-6919
dc.identifier.urihttps://www.repository.cam.ac.uk/handle/1810/278142
dc.description.abstractNumerous deep learning applications benefit from multi-task learning with multiple regression and classification objectives. In this paper we make the observation that the performance of such systems is strongly dependent on the relative weighting between each task's loss. Tuning these weights by hand is a difficult and expensive process, making multi-task learning prohibitive in practice. We propose a principled approach to multi-task deep learning which weighs multiple loss functions by considering the homoscedastic uncertainty of each task. This allows us to simultaneously learn various quantities with different units or scales in both classification and regression settings. We demonstrate our model learning per-pixel depth regression, semantic and instance segmentation from a monocular input image. Perhaps surprisingly, we show our model can learn multi-task weightings and outperform separate models trained individually on each task.
dc.publisherIEEE
dc.titleMulti-task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semanticsen
dc.typeConference Object
prism.endingPage7491
prism.publicationDate2018en
prism.publicationNameProceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognitionen
prism.startingPage7482
dc.identifier.doi10.17863/CAM.25486
dcterms.dateAccepted2018-02-19en
rioxxterms.versionofrecord10.1109/CVPR.2018.00781en
rioxxterms.versionAM*
rioxxterms.licenseref.urihttp://www.rioxx.net/licenses/all-rights-reserveden
rioxxterms.licenseref.startdate2018-12-14en
dc.contributor.orcidCipolla, Roberto [0000-0002-8999-2151]
dc.contributor.orcidKendall, Alex [0000-0003-1904-5885]
rioxxterms.typeConference Paper/Proceeding/Abstracten
rioxxterms.freetoread.startdate2019-07-16


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record