/
CSBStatB Spring  Statistical Learning Theory Lecture  Reproducing Kernel Hilbert Spaces CSBStatB Spring  Statistical Learning Theory Lecture  Reproducing Kernel Hilbert Spaces

CSBStatB Spring Statistical Learning Theory Lecture Reproducing Kernel Hilbert Spaces - PDF document

phoebe-click
phoebe-click . @phoebe-click
Follow
633 views
Uploaded On 2014-12-16

CSBStatB Spring Statistical Learning Theory Lecture Reproducing Kernel Hilbert Spaces - PPT Presentation

1 Hilbert Space and Kernel An inner product uv can be 1 a usual dot product uv 2 a kernel product uv vw where may have in64257nite dimensions However an inner product must satisfy the following conditions 1 Symmetry uv vu uv 8712 X 2 Bilinearity ID: 24763

Hilbert Space and Kernel

Share:

Link:

Embed:

Download Presentation from below link

Download Pdf The PPT/PDF document "CSBStatB Spring Statistical Learning Th..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.


Presentation Transcript

2ReproducingKernelHilbertSpacesAReproducingKernelHilbertSpace(RKHS)isaHilbertspaceHwithareproducingkernelwhosespanisdenseinH.Wecouldequivalentlyde neanRKHSasaHilbertspaceoffunctionswithallevaluationfunctionalsboundedandlinear.Forinstance,theL2spaceisaHilbertspace,butnotanRKHSbecausethedeltafunctionwhichhasthereproducingpropertyf(x)=Zs(x�u)f(u)dudoesnotsatisfythesquareintegrablecondition,thatis,Zs(u)2du61;thusthedeltafunctionisnotinL2.Nowletusde neakernel.De nition.k:XX!Risakernelif1.kissymmetric:k(x;y)=k(y;x).2.kispositivesemi-de nite,i.e.,8x1;x2;:::;xn2X,the"GramMatrix"Kde nedbyKij=k(xi;xj)ispositivesemi-de nite.(AmatrixM2Rnnispositivesemi-de niteif8a2Rn,a0Ma0.)Herearesomepropertiesofakernelthatareworthnoting:1.k(x;x)0.(ThinkabouttheGrammatrixofn=1)2.k(u;v)p k(u;u)k(v;v).(ThisistheCauchy-Schwarzinequality.)Toseewhythesecondpropertyholds,weconsiderthecasewhenn=2:Leta=k(v;v)�k(u;v).TheGrammatrixK=k(u;u)k(u;v)k(v;u)k(v;v)0()a0Ka0()[k(v;v)k(u;u)�k(u;v)2]k(v;v)0.Bythe rstpropertyweknowk(v;v)0,sok(v;v)k(u;u)k(u;v)2.1.2BuildanReproducingKernelHilbertSpace(RKHS)Givenakernelk,de nethe"reproducingkernelfeaturemap":X!RXas:(x)=k(;x)Considerthevectorspace:span(f(x):x2Xg)=ff()=nXi=1 ik(;xi):n2N;xi2X; i2RgForf=Pi ik(;ui)andg=Pi ik(;vi),de nehf;gi=Pi;j i jk(ui;vj).Notethat:hf;k(;x)i=Xi ik(x;ui)=f(x),i.e.,khasthereproducingproperty.Weshowthathf;giisaninnerproductbycheckingthefollowingconditions: 4ReproducingKernelHilbertSpacesTotakeananalogueinthe nitecase,thatis,X=fx1;:::;xng.LetKij=k(xi;xj),andf:X!Rnwithfi=f(xi).Then,Tkf=nXi=1k(;xi)fi8f;f0Kf0)K0)K=Xiviv0iHence,k(xi;xj)=Kij=(VV0)ij=nXk=1kvkivkj=nXk=1k k(xi) k(xj)) k(xi)=(vk)iWesummarizeseveralequivalentconditionsoncontinuous,symmetrickde nedoncompactX:1.EveryGrammatrixispositivesemi-de nite.2.Tkispositivesemi-de nite.3.kcanbeexpressedask(u;v)=Pii i(u) i(v).4.kisthereproducingkernelofanRKHSoffunctionsonX.