Reproducing kernel Hilbert spaces. Nuno Vasconcelos ECE Department, UCSD

Similar documents
element k Using FEM to Solve Truss Problems

The support vector machine. Nuno Vasconcelos ECE Department, UCSD

Big Data Analytics! Special Topics for Computer Science CSE CSE Mar 31

A New Method for Solving Integer Linear. Programming Problems with Fuzzy Variables

Reproducing kernel Hilbert. Nuno Vasconcelos ECE Department, UCSD

Circuits Op-Amp. Interaction of Circuit Elements. Quick Check How does closing the switch affect V o and I o?

A Note on Equivalences in Measuring Returns to Scale

The Support Vector Machine

This is natural first assumption, unless theory rejects it.

Design of Analog Integrated Circuits

Equilibrium of Stress

Physics 107 HOMEWORK ASSIGNMENT #20

Transient Conduction: Spatial Effects and the Role of Analytical Solutions

Section 3: Detailed Solutions of Word Problems Unit 1: Solving Word Problems by Modeling with Formulas

Generative classification models

14 The Boole/Stone algebra of sets

Shell Stiffness for Diffe ent Modes

C/CS/Phy191 Problem Set 3 Solutions Out: Oct 1, 2008., where ( 00. ), so the overall state of the system is ) ( ( ( ( 00 ± 11 ), Φ ± = 1

Wp/Lmin. Wn/Lmin 2.5V

A Note on the Linear Programming Sensitivity. Analysis of Specification Constraints. in Blending Problems

Chapter 3 Kinematics in Two Dimensions; Vectors

CONVEX COMBINATIONS OF ANALYTIC FUNCTIONS

Support-Vector Machines

Fall 2010 Analysis of Experimental Measurements B. Eisenstein/rev. S. Errede. (n.b. for now, we do not require that k. vectors as a k 1 matrix: ( )

Introduction to Electronic circuits.

V. Electrostatics Lecture 27a: Diffuse charge at electrodes

Large-Scale Data-Dependent Kernel Approximation Appendix

Lyapunov Stability Stability of Equilibrium Points

4DVAR, according to the name, is a four-dimensional variational method.

Homology groups of disks with holes

EE 204 Lecture 25 More Examples on Power Factor and the Reactive Power

CTN 2/23/16. EE 247B/ME 218: Introduction to MEMS Design Lecture 11m2: Mechanics of Materials. Copyright 2016 Regents of the University of California

SIMULATION OF THREE PHASE THREE LEG TRANSFORMER BEHAVIOR UNDER DIFFERENT VOLTAGE SAG TYPES

The soft-margin support vector machine. Nuno Vasconcelos ECE Department, UCSD

Lecture 3: Dual problems and Kernels

Feedback Principle :-

APPENDIX A Some Linear Algebra

READING STATECHART DIAGRAMS

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017

1 GSW Iterative Techniques for y = Ax

Inference in Simple Regression

What regression does. = β + β x : The conditional mean of y given x is β + β x 1 2

Which Separator? Spring 1

A/2 l,k. Problem 1 STRATEGY. KNOWN Resistance of a complete spherical shell: r rk. Inner and outer radii

STAT 309: MATHEMATICAL COMPUTATIONS I FALL 2018 LECTURE 16

The Simple Linear Regression Model: Theory

We call V a Banach space if it is complete, that is to say every Cauchy sequences in V converges to an element of V. (a (k) a (l) n ) (a (l)

What regression does. x is called the regressor The linear form is a natural first assumption, unless theory rejects it. positive or negative.

Chapter 7. Systems 7.1 INTRODUCTION 7.2 MATHEMATICAL MODELING OF LIQUID LEVEL SYSTEMS. Steady State Flow. A. Bazoune

Work, Energy, and Power

Pattern Recognition 2014 Support Vector Machines

IAML: Support Vector Machines

Linear discriminants. Nuno Vasconcelos ECE Department, UCSD

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space.

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

CIRCUIT ANALYSIS II Chapter 1 Sinusoidal Alternating Waveforms and Phasor Concept. Sinusoidal Alternating Waveforms and

Lecture 13: Markov Chain Monte Carlo. Gibbs sampling

Chapter 9 Vector Differential Calculus, Grad, Div, Curl

ÉCOLE POLYTECHNIQUE FÉDÉRALE DE LAUSANNE

The Feynman path integral

1 The limitations of Hartree Fock approximation

Chem 204A, Fall 2004, Mid-term (II)

Approach: (Equilibrium) TD analysis, i.e., conservation eqns., state equations Issues: how to deal with

f t(y)dy f h(x)g(xy) dx fk 4 a. «..

NOTE ON APPELL POLYNOMIALS

Quantum Mechanics I - Session 4

1 Convex Optimization

Example 1. A robot has a mass of 60 kg. How much does that robot weigh sitting on the earth at sea level? Given: m. Find: Relationships: W

Problem 1. Refracting Surface (Modified from Pedrotti 2-2)

ME2142/ME2142E Feedback Control Systems. Modelling of Physical Systems The Transfer Function

DEDICATED TO THE MEMORY OF R.J. WEINSHENK 1. INTRODUCTION

CHAPTER 3 ANALYSIS OF KY BOOST CONVERTER

LECTURE 8-9: THE BAKER-CAMPBELL-HAUSDORFF FORMULA

The Order Relation and Trace Inequalities for. Hermitian Operators

One Dimensional Axial Deformations

CHAPTER 2 Algebraic Expressions and Fundamental Operations

Problem Set 5 Solutions - McQuarrie Problems 3.20 MIT Dr. Anton Van Der Ven

Digital Modems. Lecture 2

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

A proposition is a statement that can be either true (T) or false (F), (but not both).

Expected Value and Variance

Exercises H /OOA> f Wo AJoTHS l^»-l S. m^ttrt /A/ ?C,0&L6M5 INFERENCE FOR DISTRIBUTIONS OF CATEGORICAL DATA. tts^e&n tai-ns 5 2%-cas-hews^, 27%

Final Exam Spring 2014 SOLUTION

_J _J J J J J J J J _. 7 particles in the blue state; 3 particles in the red state: 720 configurations _J J J _J J J J J J J J _

Lecture 3. Ax x i a i. i i

In the OLG model, agents live for two periods. they work and divide their labour income between consumption and

PHYSICS 536 Experiment 12: Applications of the Golden Rules for Negative Feedback

Thermodynamics Partial Outline of Topics

), it produces a response (output function g (x)

SVMs: Duality and Kernel Trick. SVMs as quadratic programs

Perturbation approach applied to the asymptotic study of random operators.

ρ some λ THE INVERSE POWER METHOD (or INVERSE ITERATION) , for , or (more usually) to

Kernels in Support Vector Machines. Based on lectures of Martin Law, University of Michigan

Hypothesis Tests for One Population Mean

IGEE 401 Power Electronic Systems. Solution to Midterm Examination Fall 2004

ENGI 4430 Parametric Vector Functions Page 2-01

Classification learning II

Chapter Summary. Mathematical Induction Strong Induction Recursive Definitions Structural Induction Recursive Algorithms

Solutions to Homework 7, Mathematics 1. 1 x. (arccos x) (arccos x) 1

B. Definition of an exponential

Transcription:

Reprucng ernel Hlbert spaces Nun Vascncels ECE Department UCSD

Classfcatn a classfcatn prblem has tw tpes f varables X -vectr f bservatns features n the wrl Y - state class f the wrl Perceptrn: classfer mplements the lnear ecsn rule h sgn[ g ] wth T g w + b w apprprate when the classes are lnearl separable t eal wth nn-lnear separablt we ntruce a ernel b w g w

Kernel summar. D nt lnearl separable n X appl feature transfrmatn Φ:X Z such that mz >> mx. cmputng Φ t epensve: wrte ur learnng algrthm n t-pruct frm nstea f Φ we nl nee Φ T Φ 3. nstea f cmputng Φ T Φ efne the t-pruct ernel K z Φ T Φ z an cmpute K rectl nte: the matr M K LK L M s calle the ernel r Gram matr 4. frget abut Φ an use Kz frm the start! 3

4 Plnmal ernels ths maes a sgnfcant fference when Kz s easer t cmpute that Φ T Φz e.g. we have seen that whle Kz has cmplet O Φ T Φz s O fr Kz T z we g frm O t O T T T z z z K L L L M R R Φ Φ Φ : wth

Questn n practce: pc a ernel frm a lbrar f nwn ernels we tale abut the lnear ernel Kz T z the Gaussan faml the plnmal faml K z e z σ what f ths s nt g enugh? T + z { L} K z hw I nw f a functn s a t-pruct ernel? 5

Dt-pruct ernels let s start b the efntn Defntn: a mappng : X X R s a t-pruct ernel f an nl f <ΦΦ> where Φ: X H H s a vectr space <..> s a t-pruct n H nte that bth H an <..> can be abstract nt necessarl R n 3 Φ X H 6

Dt pruct vs pstve efnte ernels that s prett abstract. hw I turn t nt smethng I can cmpute? Therem: X s a t-pruct ernel f an nl f t s a pstve efnte ernel ths can be chece let s start b the efntn Defntn: s a pstve efnte ernel n X X f l an {... l } X the Gram matr s pstve efnte. [ K ] 7

Pstve efnte matrces recall that e.g. Lnear Algebra an Applcatns Strang Defntn: each f the fllwng s a necessar an suffcent cntn fr a real smmetrc matr A t be pstve efnte: T A 0 0 all egenvalues f A satsf λ 0 all upper-left submatrces A have nn-negatve etermnant v there s a matr R wth nepenent rws such that A R T R upper left submatrces: A a A a a a a A 3 a a a 3 a a a 3 a a a 3 3 33 L 8

9 Dt pruct vs pstve efnte ernels equvalence between t pruct an pstve efnte autmatcall prves sme smple prpertes t-pruct ernels are nn-negatve functns t-pruct ernels are smmetrc Cauch-Schwarz nequalt fr t-pruct ernels nte that ths s ust X 0 X X cs - 44 4 3 4 4 θ

Dt pruct vs pstve efnte ernels the prf actuall gves nsght n what the ernel es we efne H as the space spanne b lnear cmbnatns f. H m f. f. α. m X e.g. the the space f all lnear cmbnatns f Gaussans when the we apt the Gaussan ernel. σ K. e nte: ths s a functn f the frst argument s fe 0

Dt pruct vs pstve efnte ernels f f. an g. H wth f. m α we efne the t-pruct <..> * n H as fr the Gaussan ernel ths s m '. g. β. ' m m ' α ' * f g β f m m ' ' σ g α β e * a t pruct n H a nn-lnear measure f smlart n X smewhat relate t lelhs

The t-pruct <..> * t s nt har t shw that ths means that.. * wth Φ Φ * Φ: X H. the feature transfrmatn asscate wth the ernel sens the pnts t the functns. furthermre <..> * s tself a t-pruct ernel n H H

The reprucng prpert wth ths efntn f H an <..> * f H <.f.> * f ths s calle the reprucng prpert an analg s t thn f lnear tme-nvarant sstems the t pruct as a cnvlutn. as the Drac elta f. as a sstem nput the equatn abve s the bass f all lnear tme nvarant sstems ther we wll see that t als plas a funamental rle n the ther an use f reprucng Kernel Hlbert Spaces 3

The bg pcture when we use the Gaussan ernel the pnt X s mappe nt the Gaussan G σi H s the space f all functns that are lnear cmbnatns f Gaussans the ernel s a t pruct n H an a nn-lnear smlart n X reprucng prpert n H: analg t lnear sstems X Φ n K e 3 * * σ * H 4

Hlbert spaces pla an mprtant rle n functnal analss we wll ver nfrmal revew H s a vectr space wth a t pruct: ths s nwn as a t-pruct space r a pre-hlbert space the fference t a Hlbert space s mstl techncal Defntn: a Hlbert space s a cmplete t-pruct space what we mean b cmpleteness? Defntn: S s cmplete f all Cauch sequences n S cnverge ths s useful mstl t prve cnvergence results 5

Cauch sequences ust fr cmpleteness n pun ntene Defntn: a sequence {...} n a nrme space s a Cauch sequence f ε n N s.t. n ' n" > n n ε 0 ' n " u can pcture ths as ε n wh mght ths nt cnverge? well the lmt pnt cul be utse the space wh s ths mprtant? cmplete cnvergent Cauch 6

reprucng ernel Hlbert spaces t turn ur pre-hlbert space H nt a Hlbert space we have t cmplete t wth respect t the nrm f f f * * ths s ust ang t H the lmt pnts f all ts Cauch sequences we represent the cmpletn f H b H mre than an Hlbert space H becmes a reprucng ernel Hlbert space RKHS 7

reprucng ernel Hlbert spaces Defntn: Let H be a Hlbert space f functns f: X R. H s a RKHS enwe wth t-pruct <..> * f : X X R such that. spans H.e. { } {α } such that H.<f..> * f f H n summar span {. } f. f. α. the H we bult can easl be transfrme nt a RKHS b ang t t the lmt pnts functns f all Cauch seqs Questn: s there a ne-t-ne mappng between ernels an RKHSs? 8

ernels vs RKHSs answer: RKHS es nee specf ernel unquel prf: assume that ernels an span H usng reprucng prpert <.. > * <..> * hence frm smmetr f t pruct <..> * we must have an frm smmetr f ernel t fllws that snce ths hls fr all an the ernels are the same 9

ernels vs RKHSs hwever t s nt clear that a ernel unquel specfes the RKHS there mght be multple feature transfrms an t-pructs that are cnsstent wth a ernel t stu ths we nee t ntruce Mercer ernels Defntn: a smmetrc mappng : X X R such that f f 0 s a Mercer ernel f s.t. f < * 0

Mercer ernels wh we care abut them? Tw reasns Therem: a ernel s pstve efnte an t-pruct f an nl f t s a Mercer ernel prf that Mercer mples PD: cnser an sequence {... n } X. Let snce f < f s Mercer t fllws frm * that 0 n f s. t. w < f wδ f n w w δ δ

Mercer ernels 0 w w δ δ w w w T Kw snce ths hls fr an w s PD

3 Mercer ernels prf that PD mples Mercer: suppse there s a g s.t. cnser a parttn f X X fne enugh that hence wth u g... g n T K s nt PD an nt a PD ernel the Therem fllws b cntractn < 0 ε g g ε g g g g 0 0 < < Ku u g g T

Mercer ernels the have the fllwng prpert wthut prf Therem: Let : X X R be a Mercer ernel. Then there ests an rthnrmal set f functns an a set f λ 0 such that φ φ δ λ λ φ φ ntutn: thn f ths as a D Furer seres +Parseval ** 4

5 Egenfunctns nte: f we efne the peratr then the functns φ are the egenfunctns f Tf agan a cnnectn t lnear sstems LTI f h- f Tf T λ φ φ φ φ λ φ φ φ λ φ φ δ 4 43 4 4

Mercer ernels the egenfunctn ecmpstn gves us anther wa t esgn the feature transfrmatn Φ : X l λ φ λ φ L T where l s the space f vectrs s.t. a < an the number f nn-zer egenvalues λ clearl Φ T Φ.e. there s a vectr space l ther than H s.t. s a t pruct n that space 6

The pcture ths s a mappng frm R n t R X Φ l Φ 3 much mre le t a mult-laer Perceptrn than befre the ernelze Perceptrn as a neural net Φ. Φ. Φ. 7

In summar Reprucng ernel map Mercer ernel map f. f. α. m H K m m ' α ' * f g β Φ :. H M f g l * f T g < λ φ λ L T Φ : φ where λ φ are the egenvalues an egenfunctns f tw ver fferent pctures f what the ernel es are the tw spaces reall that fferent? 8

RK vs Mercer maps nte that fr H M we are wrtng r Φ λ φ e + L+ λφ r but snce the φ. are rthnrmal there s a - map Γ : l r e span λ { φ.} φ. e an we can wrte Γ Φ λ φ φ.. + L+ λ φ φ. frm ** *** hence. maps nt M span{φ.} 9

RK vs Mercer maps efne the t pruct n M s that then {φ.} s a bass M s a vectr space an functn n M can be wrtten as an φ φ φ φ ** λ f α φ f.. **.e. s a reprucng ernel n M α φ. λ α λφ φ φ 3 δ λ φ. φ δ λ ** ** α φ f 30

RK vs Mercer maps furthermre snce. M an functns f the frm are n M an. g.. α f β. f g ** α. α α α β β β lm l l λ φ. φ l φ φ φ. φ. λ φ φ l l λ λ m l l β. l nte that fg H an ths s the t pruct we ha n H l l m m ** λ m l m φ. φ m m ** ** α β 3

In summar H M an <..> * n H s the same as <..> ** n M Questn: s M H? nee t shw that an f H frm *** an fr an sequence {... } α φ. λ φ φ. + L+ λ f there s an nvertble P then φ α. H an M H. λφ λ φ φ. M M L M.. λφ λ φ φ 4444 44444 3 P φ φ. 3

In summar snce λ > 0 φ φ λ P M L M φ φ 0 444 4444 3 Π 0 λ s nvertble when Π s. If Π s nt nvertble then f there s n sequence fr whch Π s nvertble then the φ cannt be rthnrmal. Hence there must be nvertble P an M H. λ 0 α 0 s.t. αφ 0 α 0 s.t. α φ 33

In summar H M an <..> * n H s the same as <..> ** n M the reprucng ernel map an the Mercer ernel map lea t the same RKHS Reprucng ernel map m H K f. f. α. m m ' α * ' f g β Φ r :. Mercer ernel map H M Φ l f g * f T g < λ φ λ L T M : φ Γ ΦM Φ r Γ : l span{ φ.} r e λ φ. 34

35