ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 OptimizationoverZonotopes andTrainingSupportVectorMachines MarshallBern XeroxPaloAltoResearchCtr. DavidEppstein Univ.ofCalifornia,Irvine Dept.ofInformationandComputerScience
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 SupportVectorMachines(SVM) Machinelearningtechniqueforclassifjcationproblems i.e.givenalargenumberoflabeledyes/noinstances, predictyes/novalueofadditionalinstances Liftdatavaluestomoderate-orhigh-dimensionalEuclideanspace maybeimplicit,using“kernelfunctions”toreplacedotproducts Findhyperplaneseparatingliftedyesandnoinstances dependingononlyfew“supportvectors” Predictfuturevaluesbyliftingandusingsamehyperplane
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 SupportVectorMachines(SVM) Machinelearningtechniqueforclassifjcationproblems i.e.givenalargenumberoflabeledyes/noinstances, predictyes/novalueofadditionalinstances Liftdatavaluestomoderate-orhigh-dimensionalEuclideanspace maybeimplicit,using“kernelfunctions”toreplacedotproducts Findhyperplaneseparatingliftedyesandnoinstances dependingononlyfew“supportvectors” Predictfuturevaluesbyliftingandusingsamehyperplane Mathematicaloptimizationproblem Usinglinearorconvexprogrammingalgorithms
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 DirectionsofSVMResearch ApplySVMtechniquestomachinelearningapplications CompareSVMtechniquestootherclassifjers ModifySVMtoproducebetterclassifjers DeriveeffjcientpracticalalgorithmsforSVMoptimization Dotheoreticalanalysisofhyperplaneseparationalgorithms
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 DirectionsofSVMResearch ApplySVMtechniquestomachinelearningapplications CompareSVMtechniquestootherclassifjers ModifySVMtoproducebetterclassifjers DeriveeffjcientpracticalalgorithmsforSVMoptimization Dotheoreticalanalysisofhyperplaneseparationalgorithms Ourinterests
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 DirectionsofSVMResearch ApplySVMtechniquestomachinelearningapplications CompareSVMtechniquestootherclassifjers ModifySVMtoproducebetterclassifjers DeriveeffjcientpracticalalgorithmsforSVMoptimization Dotheoreticalanalysisofhyperplaneseparationalgorithms Thistalk
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 Isn’titjustlinearprogramming? fjnd v , c defjningseparatinghyperplane v·x + c =0 satisfyingconstraints v·Y i + c ≥ 0,foryes-instances, v·N i + c ≤ 0forno-instances FromcomputationalgeometryweknowLPiseffjcientwhen n >> d No,because... Manyfeasiblesolutions,needtochooseone “maximummarginclassifjer”leadstoquadraticprogram,stillnotsohard Use“softmarginclassifjer”toavoiddependenceonoutliers blowsupdimensionfrom d to n + d ifexpressedasLP sowantalgorithmsthatstayinlowdimension
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 Maximummarginclassifjer Choosehyperplaneatmaximumdistancefrombothconvexhulls Workswell(butsodomanyotherchoices)whensetswell-separated Whensetsoverlap,distancefromhullsisnegative Maximummarginunpopularinthiscase duetosensitivedependenceonthemostextremepoints(outliers)
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 SoftConvexHull Idea:shrinkthetwoconvexhullssotheyarewellseparated Usualhull:sum a i p i ,0 ≤ a i ≤ 1,sum a i =1 Centroid:sum a i p i ,0 ≤ a i ≤ 1/ n ,sum a i =1 Softconvexhull:sum a i p i ,0 ≤ a i ≤ µ ,sum a i =1 Chooseparameter1/ n ≤ µ ≤ 1toshrinkhulltowardscentroid Resultisa“centroidpolytope”[Bernetal.,ESA‘95]: weightedaverageofpointswhereweightsvaryininterval[0, µ ] Formedbyintersectingzonotopesum a i p i ,0 ≤ a i ≤ µ withhyperplanesum a i =1
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 SoftConvexHulls µ = µ = µ = µ = 5/12 1/3 1/2 3/4 x 1 x 2 x 3
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 SoftMarginClassifjers If µ islarge,optimalseparatinghyperplanedependsonlyonfew“supportvectors” ratherthanonentiredataset If µ issmall,softhullswillbewellseparated Choose µ automaticallytolargestvalueforwhichhullsareseparated Geometrically:fjndlowestpointinintersectionoftwozonotopes or... Choose µ empirically(e.g.bycross-validatingtofjndbestclassifjer) Performmaximummarginclassifjcationforchosenvalue Geometrically:fjndclosestpointsintwozonotopecross-sections Ourtechniquesapplytobothproblems
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 Zonotopes: Minkowskisumsoflinesegments Chooseonepointfromeachsegment,addthecoordinates Typically Θ ( n d –1 )facets correspondingtohyperplanearrangementin d –1dimensions
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 OptimizationoverZonotopes Givenacollectionofzonotopesgeneratedby n linesegments andgivenalinearobjectivefunction f fjndthepoint x intheintersectionofthezonotopesminimizing f ( x ) Likelinearprogrammingwithzonotopeinsteadofhalfspaceconstraints CouldbeturnedintoanexplicitLPbutnumberofconstraintsblowsup Thissolvesautomaticchoiceof µ ,fjxed- µ variantissimilar Goals: scalablealgorithm(linearornear-linearin n ) lowdependenceon d —typicalCGalg.isexponential,wepreferpolynomial
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 Optimizationoveronezonotope Givenzonotopeandlinearfunction,whatisbestvertex? Veryeasy:optimizeindependentlyovereachlinesegment Zonotopeintersecthyperplanealmostaseasy:fractionalknapsack (solvedbyagreedyalgorithm) Buthowtoextendtomorethanonezonotope?
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 EllipsoidMethod Generaltechniqueforlinearorconvexoptimization Notverypractical Convertsseparationintooptimization Needsasinputa“separationoracle” thattestsifapointisinfeasibleregion, ifnotfjndshyperplaneseparatingitfromfeasibleregion Dually,convertsoptimizationintoseparation Separationonaconvexset=optimizationonitspolar,viceversa Cansolveseparationproblemusingasinputan“optimizationoracle“ thatfjndsextremevertexforalinearobjectivefunction
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 Zonotopeoptimizationalgorithm Useellipsoidtoconvertsingle-zonotopeoptimizationtoseparation Multi-zonotopeseparationsolvedbytestingeachzonotopeindependently Useellipsoidagaintoconvertseparationtomulti-zonotopeoptimization Analysis: Twolevelsofrecursivecallsinellipsoidmethods Eachlevelmultipliestimebypoly( d ,precision) Requiredprecisioncanbeshowntobesmall:polylog( n )timesinitialprecision Noblowupindependenceonn Totaltime:O( n poly( d ,log n ,precision))
ZonotopesandSVM D.Eppstein,UCIrvine,WADS2001 Conclusions CansolveSVMoptimizationintimeO( n polylog) Scalable(near-lineardependenceon n ) Polynomialdependenceon d Alternatives? Typicalcomputationalgeometryapproach:parametricsearch Convertsdecisionproblemintooptimization,similarlytoellipsoid soagainneedtwolevelsofrecursion SeemstoleadtoO(npolylog),nodependenceonprecision butexponentialdependenceondimension Whataboutapracticalpolynomialtimealgorithm?
Recommend
More recommend