Skip navigation
Skip navigation

A hybrid loss for multiclass and structured prediction

Shi, Qinfeng; Reid, Mark; Caetano, Tiberio; van den Hengel, Anton; Wang, Zhenhua

Description

We propose a novel hybrid loss for multiclass and structured prediction problems that is a convex combination of a log loss for Conditional Random Fields (CRFs) and a multiclass hinge loss for Support Vector Machines (SVMs). We provide a sufficient condition for when the hybrid loss is Fisher consistent for classification. This condition depends on a measure of dominance between labels - specifically, the gap between the probabilities of the best label and the second best label. We also prove...[Show more]

dc.contributor.authorShi, Qinfeng
dc.contributor.authorReid, Mark
dc.contributor.authorCaetano, Tiberio
dc.contributor.authorvan den Hengel, Anton
dc.contributor.authorWang, Zhenhua
dc.date.accessioned2015-12-13T22:33:14Z
dc.identifier.issn0162-8828
dc.identifier.urihttp://hdl.handle.net/1885/75928
dc.description.abstractWe propose a novel hybrid loss for multiclass and structured prediction problems that is a convex combination of a log loss for Conditional Random Fields (CRFs) and a multiclass hinge loss for Support Vector Machines (SVMs). We provide a sufficient condition for when the hybrid loss is Fisher consistent for classification. This condition depends on a measure of dominance between labels - specifically, the gap between the probabilities of the best label and the second best label. We also prove Fisher consistency is necessary for parametric consistency when learning models such as CRFs. We demonstrate empirically that the hybrid loss typically performs least as well as - and often better than - both of its constituent losses on a variety of tasks, such as human action recognition. In doing so we also provide an empirical comparison of the efficacy of probabilistic and margin based approaches to multiclass and structured prediction.
dc.publisherInstitute of Electrical and Electronics Engineers (IEEE Inc)
dc.sourceIEEE Transactions on Pattern Analysis and Machine Intelligence
dc.titleA hybrid loss for multiclass and structured prediction
dc.typeJournal article
local.description.notesImported from ARIES
local.identifier.citationvolume37
dc.date.issued2015
local.identifier.absfor080100 - ARTIFICIAL INTELLIGENCE AND IMAGE PROCESSING
local.identifier.ariespublicationU3488905xPUB4866
local.type.statusPublished Version
local.contributor.affiliationShi, Qinfeng, University of Adelaide
local.contributor.affiliationReid, Mark, College of Engineering and Computer Science, ANU
local.contributor.affiliationCaetano, Tiberio, College of Engineering and Computer Science, ANU
local.contributor.affiliationvan den Hengel, Anton, University of Adelaide
local.contributor.affiliationWang, Zhenhua, University of Adelaide
local.description.embargo2037-12-31
local.bibliographicCitation.issue1
local.bibliographicCitation.startpage2
local.bibliographicCitation.lastpage12
local.identifier.doi10.1109/TPAMI.2014.2306414
local.identifier.absseo970108 - Expanding Knowledge in the Information and Computing Sciences
dc.date.updated2015-12-11T09:15:03Z
local.identifier.scopusID2-s2.0-84916910915
CollectionsANU Research Publications

Download

File Description SizeFormat Image
01_Shi_A_hybrid_loss_for_multiclass_2015.pdf889.67 kBAdobe PDF    Request a copy


Items in Open Research are protected by copyright, with all rights reserved, unless otherwise indicated.

Updated:  19 May 2020/ Responsible Officer:  University Librarian/ Page Contact:  Library Systems & Web Coordinator