Transcription of An Introduction to Locally Linear Embedding
1 Research180 ParkAve, problemsininformationprocessinginvolve (LLE),anunsu-pervisedlearningalgorithmth atcomputeslowdimensional, , LLEmapsitsinputsintoa singleglobalcoordinatesystemoflowerdimen sionality, anditsoptimizations thoughcapableofgeneratinghighlynonlinear embeddings problemsinstatisticalpatternrecognitionb eginwiththepreprocessingofmultidimension alsignals, , thegoalofpreprocessingis someformofdimensionalityreduction:to com-pressthesignalsin sizeandto popularformsofdimensionalityreductionare themethodsofprincipalcom-ponentanalysis( PCA)[1]andmultidimensionalscaling(MDS)[2 ]. BothPCAandMDSareeigenvectormethodsdesign edto modellinearvariabilitiesin , (ormetric)MDS, thesedistancescorrespondtoEuclideandista nces, ,andtheiroptimizationsdonotinvolve , , weintroducedaneigenvectormethod calledlocallylinearembedding(LLE) fortheproblemofnonlineardimensionalityre duction[4].
2 ,thedimen-sionalityreductionbyLLEsucceed sinidentifyingtheunderlyingstructureofth emanifold, PCAandMDS,ouralgorithmis sim-pletoimplement,anditsoptimizationsdo notinvolve ,however, it is [5, 6], whichclusterthedataandperformPCAwithinea chcluster, donotaddresstheproblemconsideredhere namely, howtomaphighdimensionaldataintoa , wereviewtheLLEalgorithminitsmostbasicfor mandillustrateapotentialapplicationtoaud iovisualspeechsynthesis[3].-10105-10123( A)-10105-10123(B)-2-1012-2-1012(C)Figure 1:Theproblemofnonlineardimensionalityred uction,asillustratedforthreedimensionald ata(B)sampledfroma two dimensionalmanifold(A).Anun-supervisedle arningalgorithmmustdiscovertheglobalinte rnalcoordinatesofthemanifoldwithoutsigna lsthatexplicitlyindicatehowthedatashould beembed-dedintwo (C) , , is real-valuedvectors , eachofdimensionality , sufficientdata(suchthatthemanifoldis well-sampled),weexpecteachdatapointandit sneigh-borstolieonorclosetoa ,oneidentifies nearestneighborsperdatapoint,asmeasuredb yEuclideandis-tance.
3 (Alternatively, onecanidentifyneighborsbychoosingallpoin tswithinaballoffixedradius,orbyusingmore sophisticatedrulesbasedonlocalmetrics.)R econstructionerrorsarethenmeasuredbythec ostfunction: (1) summarizethecontributionofthe thdatapointtothe computetheweights , weminimizethecostfunc-tionsubjecttotwo constraints:first,thateachdatapoint is reconstructedonlyfromitsneighbors,enforc ing if doesnotbelongtothisset;second,thattherow softheweightmatrixsumtoone: . Thereasonforthesum-to-oneconstraintwillb ecomeclearshortly. Theoptimalweights subjecttotheseconstraintsarefoundbysolvi nga leastsquaresproblem, animportantsymmetry:forany particulardatapoint,they areinvarianttorotations,rescalings, (1); consequenceofthissymmetryis thatthereconstructionweightscharacterize intrinsicgeometricpropertiesofeachneighb orhood,asopposedtopropertiesthatdependon a smoothnonlinearmanifoldofdimensionality!
4 #" . To a goodapproximation,then,thereexistsa linearmapping consistingofa translation,rotation,andrescaling ,thereconstructionweights , thesameweights thatreconstructthe thdatapointin dimensionsshouldalsoreconstructitsembedd edmanifoldcoordinatesin!dimensions.(Info rmally, imaginetakinga pairofscissors,cuttingoutlocallylinearpa tchesoftheunderlyingmanifold, doneina ,thetransplantationofeachpatchinvolvesno morethana translation,rotation,andrescalingofitsda ta, ,whenthepatcharrivesatitslowdimensionald estination,weexpectthesameweightstorecon structeachdatapointfromitsneighbors.)LLE constructsa neighborhoodpreservingmappingbasedonthea bove ,eachhighdimensionalobservation is mappedtoalow dimensionalvector $ donebychoosing!-dimensionalcoordinates $ tominimizetheembeddingcostfunction:% $ & $ ' $ )((2)Thiscostfunction like thepreviousone isbasedonlocallylinearreconstructionerro rs,butherewefixtheweights whileoptimizingthecoordinates $.
5 TheembeddingcostinEq.(2)definesa quadraticforminthevectors $ . Subjecttoconstraintsthatmake theproblemwell-posed,it canbeminimizedbysolvingasparse +*, eigenvectorproblem,whosebottom! independentoftheweightsforotherdatapoint s theem-beddingcoordinatesarecomputedbyan -*. eigensolver, a globaloperationthatcouplesalldatapointsi n ;thisis donesimplybycomputingthebottomeigenvecto rsfromeq.(2)oneata fairlystraightforward,asthealgorithmhaso nlyonefreeparameter:thenumberofneighbors perdatapoint, . , . thatbestreconstructeachdatapoint fromitsneighbors,minimizingthecostineq.( 1) $/ bestreconstructedbytheweights , minimizingthequadraticformineq.(2) :SummaryoftheLLEalgorithm,mappinghighdim ensionaldatapoints, , tolow dimensionalembeddingvectors, $0.
6 Arechosen,theoptimalweights andcoordinates$ andfindsglobalminimaofthereconstructiona ndembeddingcostsinEqs.(1)and(2).Asdiscus sedinAppendixA,intheunusualcasewherethen eighborsoutnumbertheinputdimensionality 213 , theleastsquaresproblemforfindingtheweigh tsdoesnothave a uniquesolution,anda regularizationterm forexample,onethatpenalizesthesquaredmag nitudesoftheweights , ,takesasinputthe highdimensionalvec-tors, . Inmany settings,however, theusermaynothave accesstodataofthisform, ,describedinAppendixC, , matricesofpairwisedistancescanbeanalyzed byLLEjustaseasilyasMDS[2];infactonlya smallfractionofallpos-siblepairwisedista nces(representingdistancesbetweenneighbo ringpointsandtheirrespective neighbors) , forexample,theinputto LLEconsisted 546 7 datapointssampledoff thealgorithm,using 8 :9neighborsperdatapoint,successfullyunra veledtheunderlyingtwo showsanothertwo dimensionalmanifold,thisonelivingina ,wegeneratedexamples showninthemiddlepanelofthefigure bytranslatingtheimageofa singlefaceacrossa two-dimensionalmanifoldparameterizedbyth eface s <;74= grayscaleimages,witheachimagecontaininga97>*96 facesu-perimposedona?
7 ;*? ,themanifoldoftranslatedfacesis highlynonlinearinthehighdimensional( @6 7 A;) vectorspaceofpixel showsthefirsttwo componentsdiscoveredbyLLE,with ,thetopportionshowsthefirsttwo clearthatthemanifoldstructureinthisexamp leis , inadditiontotheseexamples,forwhichthetru emanifoldstructurewasknown,wealsoapplied LLEtoimagesoflipsusedintheanimationoftal kingheads[3]. Ourdatabasecontained D>?>7>color(RGB)imagesoflipsat E A>*> ( 59AG69= :4) is showthefirsttwo componentsdiscovered,respectively, byPCAandLLE(with H :4).Ifthelipimagesdescribeda nearlylinearmanifold,thesetwo methodswouldyieldsimilarresults;thus, somewhatuniformdistributionaboutitsmean, thelocallylinearembeddinghasa distinctlyspiny structure,withthetipsofthespinescorrespo ndingto is worthnotingthatmany hill-climbingmethodsforautoencoderneural networks[7, 8], self-organizingmaps[9], andlatentvariablemodels[10] donothave thesameguaranteesofglobaloptimalityorcon -vergence;they alsotendtoinvolve many morefreeparameters,suchaslearningrates,c onvergencecriteria, ,computingnearestneighborsscales(inthewo rstcase)asI , orlinearlyintheinputdimensionality, , andquadraticallyinthenumberofdatapoints.
8 Formany6 Figure3:TheresultsofPCA(top)andLLE(botto m),appliedtoimagesofa singlefacetranslatedacrossa thecornersofitstwo dimensionalembedding,whilePCAfailstopres erve :Imagesoflipsmappedintotheembeddingspace describedbythefirsttwo coordinatesofPCA(top)andLLE(bottom).Repr esentative ,however andespeciallyfordatadistributedona thinsubman-ifoldoftheobservationspace constructionssuchasK-Dtreescanbeusedtoco mputetheneighborsinI 8 JLK7M time[13].InStep2,computingtherecon-struc tionweightsscalesasI ON ; thisis thenumberofoperationsrequiredtosolve a * ,computingthebottomeigenvectorsscalesasI ! , linearlyinthenumberofembeddingdimensions ,!, andquadraticallyinthenumberofdatapoints, . Methodsforsparseeigenproblems[14], however, canbeusedtoreducethecomplexitytosub-quad raticin.
9 Notethatasmoredimensionsareaddedtotheemb eddingspace,theexistingonesdonotchange,s othatLLEdoesnothave sizeN generalprincipleofmanifoldlearning,eluci datedbyTenenbaumetal[11], thatoverlappinglocalneighborhoods collectively analyzed canpro-videinformationaboutglobalgeometr y. Many virtuesofLLEaresharedbytheIsomapalgorith m[11], anextensionofMDSinwhichembeddingsareopti mizedtopreserve geodesic distancesbetweenpairsofdatapoints; avoidstheneedtosolve accumulateverysparsematrices, howtolearna parametricmappingbetweentheobservationan dembeddingspaces, touseP $ LLEbroadlyusefulinmany particulardatapoint Swith nearestneighbors T andreconstructionweightsU canwritethereconstructionerroras:V S U T W U S T YXU UX7Z) YX (3)9whereinthefirstidentity, wehave exploitedthefactthattheweightssumtoone,a ndinthesecondidentity, wehave introducedthelocalcovariancematrix,Z) YX + S T [\ S TX ((4)Thiserrorcanbeminimizedinclosedform, usinga Lagrangemultipliertoenforcetheconstraint that U ].
10 Intermsoftheinverselocalcovariancematrix ,theoptimalweightsaregivenby:U XZ#^`_ aX cbedZ^`_bfd((5)Thesolution,aswrittenineq .(5), ,a moreefficientwaytominimizetheerrorissimp lytosolve thelinearsystemofequations, Z) aXUX , andthentorescaletheweightssothatthey sumto one(whichyieldsthesameresult).Byconstruc tion,thelocalcovariancematrixineq.(4)iss ymmetricandsemipositive singularornearlysingular asarises,forexample,whentherearemoreneig hborsthaninputdimensions( 21g ), orwhenthedatapointsarenotingeneralpositi on itcanbeconditioned(beforesolvingthesyste m)byaddinga smallmultipleoftheidentitymatrix,Z) aX,hiZ) aXkjml)n porq YX (6)wheren is smallcomparedtothetraceofZ. $0 arefoundbyminimizingthecostfunction,eq.( 2),forfixedweights :sutwvx% $ y $ / 2 F $ )((7)Notethatthecostdefinesa quadraticform,% $ {z $ [ $ 10involvinginnerproductsoftheembeddingve ctorsandthe *| matrixz:z q j X X X} (8)whereq is 1 if and0 performedsubjecttoconstraintsthatmake is clearthatthecoordinates $ canbetranslatedbya constantdisplace-mentwithoutaffectingthe cost,% $.]
