Professional Documents
Culture Documents
HannahPiper
IndependentComponentAnalysisforBlindSourceSignalSeparation
Introduction
Independentcomponentanalysis(ICA)isanimportantcomputationaltechniqueusedtodecompose
nongaussian,statisticallyindependentsignalsintoadditivesubcomponents.ICAdecomposesa
multidimensionalsignalintocomponentsthataremaximallyindependentintermsoftheirnegativeentropy
andmaximallyindependentstatistically.Inourpaper,wediscussthemathematicsbehindthecomputation
ofIndependentComponentsandinadditionwewillpresentarealworldexampletoillustrateits
effectiveness.WedemonstratehowICAcanbeusedtoblindlyseparatetwosignals,withproperties
describedabove,intoindependentcomponents.Wewilldemonstratehowdifferentvoicesinanaudiosignal
canbeseparatedoutusingICA.
BlindSourceSeparation
Fig.1SchematicoftheCocktailPartyProblem
IndependentComponentAnalysis(ICA)isoftenexplainedincoursesusingthecanonicalBlindSource
SeparationproblemknownasthetheCocktailPartyProblem[Fig.1].TheCocktailPartyProblemtermwas
coinedbyColinCherryin1953todescribetheproblemofseparatingoneperson'svoiceamidstmanyother
peoplehavingaconversationatacocktailparty.OnlyinthelasttwodecadeshastheICAtechniqueevolved
enoughtomaketheCocktailPartyProblematractableproblem.TheCocktailPartyProblem(CPP)isan
instancefromalargerdomainofproblemsaimingatseparatingmixedsourcestheseproblemsaretermed
blindsourceseparation(BSS)problems.CPP,becauseoftheassumptionsimpliedintheproblem,isa
problemthatarguablybelongstoasubsetofBSSproblemsthatcanbesolvedusingICA.Manyproblems
fromtheBSSdomaincannotbesolved,however,thisisanactiveareaofresearchwithinthecomputer
science,appliedmathandstatisticscommunities.
CPPisagreatexamplebecauseofthephysicalunderpinningsofsoundi.e.soundaddslinearly.The
recordedamplitudeofasoundwavebyamicrophoneistheadditivelinearsumofthepressurewaves
emanatingfromthesources.
ProblemIntuition
TobuildintuitiontowardsICA,westartbylookingatsomedatarecordedbytwomicrophones.Lets1bea
signalgeneratedbythemusicands2isgeneratedbyaperson'svoice.Thesignalsaresimultaneously
recordedbytwomicrophones.Thetwomicrophonesrecordaunique,linearsummationofthetwosounds.
Thelinearweightsforeachmicrophonearea1,b1anda2,b2,respectively.Theweightsreflectthe
proximityofthesignalsourcetotherespectivemicrophones.Ourgoalistorecovertheoriginalsourcess1
ands2.
WecanplotthedataandvisualizeittogetanintuitivefeelforwhatanIndependentComponent(IC)is.We
plotthesimultaneouslyrecordedamplitudesfrommicrophones1and2onthexandyaxis,respectively.If
therewasonlyonesignalsources(music),allsampleswouldlieonthevectora=(a1,a2),wherea1,a2
signifytheproximityofthemusictomicrophones1and2,respectively.Similarly,themiddleplotdepictsall
datapointsarisingiftheonlysignalbeingrecordedwasthevoice.Ifboththesignalswerepresent,given
thatthesignalisadditive,wewouldexpecttoseetheplotintherightpanelthelinearsumofboth
sources.
Fig.2VisualizationofIndependentComponentsandResultingAdditiveSignal
Thevectorsaandbaretheindependentcomponentsofthedata.Independentcomponentanalysisallows
ustoextracttheindependentcomponentsfromthedatagivenonlythesumofthecomponents.Inthe
followingsectionswewilldiscussthetechniquebywhichwecandecomposealinearlyadditivesignalinto
itsICs.
Nowthatwehavelookedattheprocessofgettingtothemixedsignalfromtheindependentcomponents,
wecannowdiscusstheextractionoftheICsgivenonlythesumoftheICs.Recallthatsoundisidealforthe
CPPbecausethesignalrecordedbythemicrophoneislinearlysummed.
Inthefollowingfigure(Fig.3),welookattherawdataaswellastheindependentcomponentsoftheraw
data.NotethatthedirectionoflargestvarianceisnotalwaysthesameasthatoftheICthisisan
importantdistinctiontonotefromthetechniquePCA.
MathematicalMethods
Fig.3ICsofselectedrawdata
Therawdataxisassumedtobealinearmixtureoftheunderlyingsources.ThisholdstruefortheCPP(as
thesignal)isalinearmixture.
xisoftheform:
Aisknownasthemixturematrix.Itisanunknown,invertible,squarematrixthatlinearlymixesthe
componentsofthesources.
Inthesimpleexamplediscussedabove,withtwosources(musicandvoice)thetheentriesoftheAmatrix
aretheweightsgiventothesignals(proximitytothesource).Thiscanbewrittenas:
NotethatinanICAproblemwecannotsimplyfilltheentriesoftheAmatrixusingtheweightsbecausethe
weightsareunknowninfact,thegoalofICAistofindthemixingmatrixA.Giventherelationshipbetween
A,sandx,giventherawdatax,andifweknowA,wecanseparatethesources.Thiscanbe
mathematicallyinterpretedasthefollowing:
Notetheuseofshat.ThisisbecausewecanonlyapproximatethesourcesignalasWandsareunknown
noclosedformsolutionexists.Theproblemisoftheformab=constant,where,aandbareunknown.This
anunderconstrainedproblembecausethenumberofunknownsexceedsthenumberofobservations.But,
notallhopeislost,itturnsoutthatthestatisticsoftheobserveddataxinconjunctionwithinformation
theorycanbeusedtofindasolutionforthisproblem.
InsteadoftryingtosolveforbothAands.WefirstdetermineA.Wedecomposeourmixturematrixusing
singularvaluedecomposition(SVD)
Geometrically,SVDdecomposesthematrixintointothreelinearoperations[Theanimationonthewikipage
[3]isdoesanexcellentjobconveyingthegeometricalinterpretationofthedecompositioncomponentsV,
andU]
ArotationV
Astretchalongtheaxes
AsecondrotationU.
WeestimateAandWfromtheSVD.Weknowthattheinverseofarotationmatrixisitstransposeandthat
becauseAisinvertible,theinverseofexists.WecanwriteWas:
WesolveforWbyexaminingthecovarianceofthedataxinordertocalculateUand,andusethenotion
ofstatisticalindependenceofstosolveforV.
ComputingUand:
Covarianceofamatrixistheexpectedvalueoftheouterproductofindividualdatapoints<xx>.Recall,for
ICAoftheCPPwemadetheassumptionthatthesourcesarestatisticallyindependent.Thisimpliesthat
covarianceofthesourcesiswhitenedi.e.thesourcesignalsareuncorrelatedandallsourceshavevariance
1.Thisalsoimpliesthatthecovarianceistheidentitymatrix.
Given,
Frombasiclinearalgebraicproperties,thecovarianceofthedatacanbeexpressedintermsofthe
underlyingsourcesas:
Recall,thatweassumedthatifthesiswhitenedthisimplies:
Also,itisknownthatthefollowingpropertyholdsfororthogonalmatrices:
Thisgivesus:
Thistakeonthefamiliarofacanonicaleigenvectoreigenvaluedecomposition:
ThecovarianceofthedatainexpressedintermsofadiagonalmatrixDbetweentwoorthogonalmatricesE.
DisadiagonalmatrixofassociatedeigenvaluesandEisanorthogonalmatrix.Diagonalizationofa
symmetricmatrixintoeigenvectorsisauniquesolutioni.e.nootherbasiscandiagonalizeasymmetric
matrix.
Derivedearlier:
Bysuperpositiongivesup:
DaretheeigenvaluesandEtheeigenvectorsofthecovarianceofthedatax.Allweneedisthefirstrotation
matrixV.
ComputingV:
Wecanwhitenourobserveddatax.Thisgivesus:
Thereforwecanwriteourapproximationofsas:
WeexploitthestatisticsofindependencetoidentifyV.Thisstepisthereasonoursourcesignalsneedto
bestatisticallyindependent.Thecovariancematrixisameasureofthelineardependencebetweenallpairs
ofentries.Itcapturesthesecondordercorrelationsinthedata.Tworandomvariablesaandbare
independentif:
ThesubclassofBSSPwearetryingtosolveallowsustoassumethatallsourcesarestatistically
independentthisgivesus:
WeneedtofindarotationmatrixVsuchthatshatisstatisticallyindependent.Itcanbeeasilyshownthat
thewhiteningtransformofthedataremovesallsecondordercorrelations,theunknownmatrixVmustthen
minimizehigherorderdependenciesbetweenthesources.Thisisequivalenttoincreasingthe
independence.HencetocomputeVweuseadifferentmetrictomeasuredependencebetweenpairof
signalsources.Thismetricisderivedfromthefieldofinformationtheory.
Themetric,mutualinformation,measuresthedepartureoftwovariablesfromstatisticalindependence.
Multiinformation,ageneralizationofmutualinformation,measuresthestatisticaldependencebetween
multiplevariablesandisgivenby:
I(y)isacontinuousnonnegativefunctionthatthatonlygives0ifallvariablesarestatisticallyindependent.
ThemultiinformationisafunctionoftheentropyHofthedata.Whichisgivenby:
Themultiinformationisthedifferencebetweenthesumofentropiesofthemarginaldistributionsandthe
entropyofthejointdistribution.Whichis:
Theproblemthenturnsintoanoptimizationproblemoftheform:
Thisminimizationcanbeperformedusingmanypopularoptimizationalgorithmssuchasgradientdescent
orsimulatedannealing.
Conclusion
AlthoughmanyaspectsofICA,PCA,andBlindSourceSeperationwerediscussedthroughoutthispaper,
thereisstillamultitudeofresearchlefttocompletelyunderstandthesetopics.TheCocktailPartyProblem
isjustoneofthemanyexamplesthatshowstheimportanceofBSS.It'samazingtothinkthatasimple
mathematicalmethod,suchastheuseofeigenvectorsandmatrices,canbeutilizedtosolvesuchcomplex
processesofsignalanalysis.
References
[1]http://www.mit.edu/~gari/teaching/6.555/LECTURE_NOTES/ch15_bss.pdf
[2]http://www.cis.hut.fi/projects/ica/cocktail/cocktail_en.cgi
[3]http://en.wikipedia.org/wiki/Singular_value_decomposition
[4]Shlens,Jonathon."ATutorialonIndependentComponentAnalysis."arXivpreprintarXiv:1404.2986
(2014).