AnIncremental Learning Algorithm Basedon Support Vector DomainClassifier
Yinggang Zhao,Qinming He
2006-01-01
Abstract:category (forexample, samples withcategory labelyi=1), yettheother partofsample's category isunknown, then we Incremental learning technique isusually usedtosolvecandesign newtypeofclassifier basedonSVDD named large-scale problem. Wefirstly gaveamodif edsupport vectorsupport vector domain classifier (SVDC).Thisnewclassifier machine (SVM)classification method--support vectoronlyneedtodescribe thedatawithknowncategory, then domainclassifer (SVDC), thenanincremental learning obtaining thedescription boundary ofthisclass ofdata. algorithm basedonSVDCwasproposed. Thebasic ideaof Finally, we canclassify theunknownbinary-class data thisincremental algorithm istoobtain theinitial targetaccording totheobtained boundary. concepts using SVDCduring thetraining procedure andthen Inthis paper ourincremental learning algorithm isbased update these target concepts byanupdating model. Difierent on SVDC, andthisalgorithm ismotivated by the fromtheexisted incremental learning approaches, inour person-learning procedure. Whenlearning a complicated algorithm, themodelupdating procedure equals tosolve a concept, people usually obtain ainitial concept byusing part quadratic programming (QP)problem, andtheupdated model ofuseful information, thenupdate theobtained concept by still ownstheproperty ofspars solution. Compared withotherutilizing new information. Intermofourincremental existed incremental learningalgorithms, theinverse procedure algorithm basedonSVDC,itfirstly utilize partofdata ofouralgorithm (i.e. decreasing learning) iseasytoconduct(memoryspace permitting), thenobtain aconcept (namely the without extra computation. Experiment results showour parameter ofobtained decision hypersurface) by SVDC algorithm iseffective andfeasible. learning algorithm, finally according totheinformation of decision hypersurface acquired inlaststep, update the parameter ofdecision hypersurface gained inlast step utilizing