Short course A vademecum of statistical pattern recognition techniques with applications to image and video analysis. Agenda
|
|
- Hugo Miles
- 6 years ago
- Views:
Transcription
1 Short course A vademecum of statistica attern recognition techniques with aications to image and video anaysis Lecture 3 Density estimation Massimo Piccardi University of Technoogy, Sydney, Austraia Massimo Piccardi, UTS Agenda Density estimation Likeihood function Maximum-ikeihood density estimation Gaussian mixture modes on-arametric (Kerne Density Estimation) Exame aers Massimo Piccardi, UTS 2
2 Density estimation A the reviewed cassification criteria use dfs Accurate modeing of dfs from sets of sames is therefore a fundamenta task in Bayesian cassification (robabiity) density (function) estimation Parametric: Gaussians, Gaussian mixtures etc on-arametric: histogram, k-nearest neighbours, KDE, meanshift etc Gaussian mixtures and KDE in the foowing B: sames modes: density estimation modes sames: saming Massimo Piccardi, UTS 3 Density estimation Let us have a set of sames, X = {x i }, i=.., that are a generated from the same distribution and indeendenty of one another (i.i.d. sames) Density estimation is erformed by choosing an aroriate df mode (for exame, Gaussian) and fitting its arameters, θ, to the set of sames The robabiity of the entire set, X, is then given by: ( X θ ) = ( θ ) i= x i Massimo Piccardi, UTS 4
3 Likeihood function We define a function that we ca the ikeihood of the arameters given the set of sames, L(θ X): ( θ X ) ( X θ ) L = ease note that L is a function of θ given X, whereas is a function of X given θ. Often, L(θ X) is noted just as L(θ). The og-ikeihood, LL(θ X), is often used instead of L for these main reasons: og (or n) is a monoticay increasing function of its argument: maxima of the argument are maxima aso of the og og of roduct = sum of og: removes roduct oerator avoids numerica underfow during the evauation of L: = Σ = -000 Massimo Piccardi, UTS 5 ML density estimation Find arameters θ maximising L(θ) or, equivaenty, LL(θ) It is ossibe to undertake direct maximisation by differentiation: comute dl/dθ = 0 and sove for θ If cosed-form soutions are nor ossibe, iterative methods can be used instead (e.g. ewton-rahson) dl/dθ θ n+ is a better aroximation than θ n for the zero of dl/dθ θ n+ θ n Massimo Piccardi, UTS 6
4 ML for the Gaussian Easy case for ML density estimation: Gaussian df. Its arameters are θ = {µ, Σ} µ Σ ML ML = = i= ( xi µ ML )( xi µ ML ) i= x i T The variance is a so-caed biased estimate; it shoud be mutiied by /(-) to unbias; yet, the correction is negigibe for any reasonabe Massimo Piccardi, UTS 7 The EM aroach Direct maximisation of the og-ikeihood is often inconvenient or just difficut A very ouar aternative for ML estimation is given by the Exectation-Maximization (EM) aroach In the EM aroach, instead of maximising function L(θ) (or LL(θ)), we maximise another function, Q(θ) The aroach works since obtaining a maximum for Q(θ) guarantees a maximum or at east an increase in LL(θ) over an initiay arbitrary choice of θ The aroach was roosed by Demster, Laird, Rubin (DLR) in Maximum Likeihood from Incomete Data via the EM Agorithm, JSTOR, 977 a 22-age aer with 23 reviewers Read H. Tagare, A Gente Introduction to the EM agorithm. Part I: Theory, htt://noode.med.yae.edu/hdtag/ubs/em.s for an easy introduction Massimo Piccardi, UTS 8
5 The EM aroach EM osits the existence of atent variabes (n.b it aso works for missing data), Y, and sets a different target for maximisation: Q ( θ, θ ) = E[ n ( X,Y θ ) X, θ ] = n ( X, y θ ) ( y X, θ )dy L(θ) is therefore caed the incomete data ikeihood, whie Q(θ,θ ) is the exected vaue of the comete data ogikeihood, n (X,Y θ), on conditiona robabiity (y X,θ ) For the aroach to make sense, Q(θ,θ ) must be such that its maximisation is easier than that of L(θ) = Massimo Piccardi, UTS 9 The EM aroach Finding an exression for Q(θ,θ ) requires an exression for og (X,y θ) and for density (y X, θ ) y is an instance of a the atent variabes in the mode We assume that the x i are aso indeendent given y; hence, og (X,y θ) = og i (x i,y θ) = Σ i og (x i,y θ) An exression for (x i,y θ) must be simer than for (x i θ); moreover, for exonentia densities, the og goes An exression for (y X, θ ) may be hard to find Q(θ,θ ) must then be differentiated in θ and maxima found Massimo Piccardi, UTS 0
6 The EM aroach Q(θ,θ ) must be evauated iterativey unti convergence Each iteration is guaranteed to increase the incomete og-ikeihood, LL(θ X) - that is what we want The maximum we find in the arameter sace uon convergence is a oca one! Its osition deends on the choice of the initia θ Each iteration consists of two stes: the E ste, where we comute the udated (y X, θ ) the M ste, where better θ are chosen by differentiation of Q(θ,θ ) Massimo Piccardi, UTS The EM agorithm. Choose an initia θ 2. E ste: comute (y X,θ ) 3. M ste: comute Q(θ,θ ) and find its maxima, θ new 4. Check for convergence of either L or θ; if not, θ θ new and return to ste 2 Massimo Piccardi, UTS 2
7 The EM agorithm courtesy of Prof. Ricardo Gutierrez-Osuna, Texas A&M University Massimo Piccardi, UTS 3 Gaussian mixture mode (GMM) A Gaussian mixture mode (GMM) is a mode with a finite number, M, of Gaussian comonents Each -th comonent has a robabiity in the mixture, or weight, α, and its own arameters, µ, Σ, =..M As a generative mode, a GMM can be samed ike this (ancestra saming): first, draw one vaue out of M according to the discrete distribution given by the α ; this icks the comonent second, draw a same from a Gaussian distribution of arameters µ, Σ (for instance, with Box-Muer and Choesky) Massimo Piccardi, UTS 4
8 Gaussian mixture mode (GMM) The df of a GMM is given by: M ( x) = α ( x µ Σ ) =, GMMs are very usefu and ouar modes since they can estimate mutimoda distributions accuratey Massimo Piccardi, UTS 5 GMM: size of arameters With D-dimensiona data, the GMM arameters size are as: for each weight, α (aka P(ω ) or π ): a scaar for each mean, µ : a D x vector for each covariance matrix, Σ : a D x D symmetric matrix with D(D+)/2 dof, if fu D, if diagona, if sherica At times, the covariances are chosen to be the same for a comonents Massimo Piccardi, UTS 6
9 The ikeihood function for a GMM: an exame Just an exame to disay the ikeihood function for a sime GMM We can easiy visuaise functions of 2 arameters, therefore we choose the foowing sime mode: ( x) =. 3 ( x µ, σ =. 6) ( x µ, σ ) = where the ony arameters are µ and µ 2 µ and µ 2 are made vary in range in 0.5 stes Massimo Piccardi, UTS 7 The data 300 uni-dimensiona sames Their histogram: Massimo Piccardi, UTS 8
10 Likeihood surface Two maxima found, at: µ = 0.5, µ 2 = -4.5 (ikeihood: ) µ = -4, µ 2 = (ikeihood: ) Massimo Piccardi, UTS 9 Quaity of fitting One maximum is ceary better than the other maximum at µ = 0.5, µ 2 = -4.5 maximum at µ = -4, µ 2 = µ 2 µ µ µ 2 Massimo Piccardi, UTS 20
11 EM for GMM EM is the main too to find arameters for a GMM with maximum ikeihood EM for GMMs assumes that, for each x i same, there exists a atent discrete r.v., y i, whose vaue is the index, ={..M}, of the Gaussian comonent resonsibe for generating that same It is assumed that each x i deends ony on its y i Therefore: ( xi, yi ) = ( xi yi ) ( yi ) = ( xi µ y, Σ ) i y α i yi ote that (x i, y i ) is much simer than (x i ), as we wanted: M ( xi ) = α ( xi µ, Σ ) = Massimo Piccardi, UTS 2 EM for GMM An exression for (y X, θ ) is derived by assuming that each y i deends ony on x i : For a GMM, Q(θ,θ ) becomes [Bimes 98]: Q + ( y X, θ ) = ( yi xi, θ ) M ( θ, θ ) = n ( α ) ( xi, θ ) M = i= n = i= ( ( x µ, Σ )) ( x, θ ) i i= Q(θ,θ ) is then differentiated to find the maxima; a constraint, Σ =..M α =, needs to be added to find meaningfu weights, α i + Massimo Piccardi, UTS 22
12 EM for GMM: re-estimation formuas E ste: M ste: α G ( ) ( xi µ, Σ ) y = = i xi, θ M αk G( xi µ k, Σk ) α new i= k= ( x θ ) = i, i new i= = µ Σ new = i= x ( x, θ ) ( x, θ ) i i new new T ( xi µ )( xi µ ) ( xi, θ ) i= i= ( x, θ ) i (aka resonsibiity) Massimo Piccardi, UTS 23 EM for mixture modes: caveats Singuarities may arise during training: one comonent modes one datum ony, tighty its covariance tends to 0 (x θ ) tends to (x θ) aso tends to we have reached a maximum of the ikeihood; yet, the mode s arametrisation is not usefu Common trick: add some esion to Σ The oca maximum we reach uon convergence may vary heaviy with the initia arameters If the data are drawn in a sequence, there exist faster, onine/incrementa versions of the re-estimation formuas Massimo Piccardi, UTS 24
13 ML and MAP density estimation ML density estimation: θ ML = arg max θ ( ( X θ )) MAP density estimation: think of arameters θ as r.v. themseves, aowing some rior distribution (θ): θ MAP = arg max θ ( ( θ X ) ( X θ ) ( θ )) MAP is usefu to favour certain vaues of θ B: Do not confuse ML/MAP density estimation with ML/MAP cassification Massimo Piccardi, UTS 25 The evidence function Both ML and MAP are oint estimates of the arameters. It is aso ossibe to marginaise θ: ( X ) ( X θ ) ( θ ) = dθ The margina ikeihood above is known as the evidence function It can be used to comare different modes i.e. different choices of (X θ), (θ) Massimo Piccardi, UTS 26
14 on-arametric estimators GMMs beong to the genera category of arametric density estimators A different aroach to density estimation can be taken by choosing modes with a minima number of arameters Widesread aroaches incude: histograms k-nearest neighbours (k) kerne density estimation (KDE) mean-shift vector Massimo Piccardi, UTS 27 Histogram With the histogram, the data sace is divided in a reguar grid (each eement is caed a bin) (x) is uniform within each bin and given by: (number of sames in the bin)/(tota number of sames) Limitations: shar/non-smooth estimate deends on the size of the bins deends on the aignment of the grid number of bins grows exonentiay with D Usefu for visuaization in or 2D Massimo Piccardi, UTS 28
15 Generic non-arametric estimation (x) ~ k/v, where V is the voume surrounding x k is the number of sames in V is the tota number of sames it rovides a good estimate if is arge, k grows with and V is sma enough for (x) to be constant Two main aroaches, KDE and k in KDE, V is fixed and k comuted from data set in k, k is fixed and V comuted from data set The mean shift vector aroach is a ost rocessing of KDE that finds the distribution s modes exicity; ends u in a mixture mode, but in a non-arametric manner Massimo Piccardi, UTS 29 Kerne Density Estimation A kerne function, K(u) (aka Parzen window), is fit centred on each same Tyica kernes (in D): K u = u 2 Uniform ( ) Triange K ( u) = ( u ) u Eanechnikov Gaussian K K ( u) = 3 ( u 2 ) u 4 ( ) = ( ) 2 u 2π 2 ex u 2 Massimo Piccardi, UTS 30
16 KDE df Kernes are then a added u and sum normaised; this is the KDE df (in D dimensions): h ( x) = D i= x x K h h is caed the bandwidth Kernes are tyicay radiay symmetric, so there is ony one scaar arameter, h, aso in D dimensions; it equates to a sherica covariance matrix Given x, evauation of (x) is comutationay heavy: high run-time execution time i Massimo Piccardi, UTS 3 KDE bandwidth How to choose the bandwidth? The EM agorithm woud ead to a useess soution: h = 0 A seudo-ikeihood can be used in ace of the standard ikeihood: when evauating (x i ) in L(θ), eave the kerne centred on it out; in this way, same x i has to be exained by its cosest neighbours Many other methods to estimate the bandwidth: Maxima Smoothing Princie, Least Squares Cross Vaidation, Biased Cross Vaidation, Smoothed Cross Vaidation, many! B. A. Turach. Bandwidth Seection in Kerne Density Estimation: A Review. Technica Reort Université Cathoique de Louvain, Begium, 993 Massimo Piccardi, UTS 32
17 GMM vs KDE GMM GMM M ( x) = α G( x µ, Σ ) = KDE (Gaussian kerne) ( x) = G( x, Σ) KDE x i i= one comonent er observation! ony one Σ for a comonents equa weights Simiarity ony notationa centred in the observation Massimo Piccardi, UTS 33 GMM vs KDE GMM KDE (Gaussian kerne) Massimo Piccardi, UTS 34
18 GMM vs KDE: exame Exame with 3 modes in 2D Parameters in GMM with different constraints on covariance; how mode changes fu: * * 3 = 7 arameters diagona: * * 2 = 4 arameters sherica: * * = arameters shared fu: * 2 + * 3 = arameters shared diagona: * 2 + * 2 = 0 arameters shared sherica: * 2 + * = 9 arameters Sherica mode not as restrictive for KDE sherica kerne: arameter Massimo Piccardi, UTS 35 The data Massimo Piccardi, UTS 36
19 y y GMM, fu covariance 5 df(obj,[x,y]) x og-ikeihood = Massimo Piccardi, UTS 37 GMM, diagona covariance 5 df(obj,[x,y]) x og-ikeihood = Massimo Piccardi, UTS 38
20 y y GMM, shared fu covariance 5 df(obj,[x,y]) x og-ikeihood = Massimo Piccardi, UTS 39 GMM, shared diagona covariance 5 df(obj,[x,y]) x og-ikeihood = Massimo Piccardi, UTS 40
21 y KDE, sherica Gaussian kerne 5 df(obj,[x,y]) x ony free arameter! og-ikeihood comares with GMM Massimo Piccardi, UTS 4 Exame aers Aication: background subtraction Extracting moving objects in a video Mixture modes: C. Stauffer and W.E.L. Grimson, Adative background mixture modes for rea-time tracking, Proc. IEEE CVPR 999, cites on Googe Schoar (9 ov 08). Many modification aers have foowed. KDE: A. Egamma, D. Harwood, and L.S. Davis, on-arametric mode for background subtraction, Proc. ECCV 2000, Massimo Piccardi, UTS 42
22 Additiona materias A reference for KDE bandwidth seection: B. A. Turach. Bandwidth Seection in Kerne Density Estimation: A Review. Technica Reort Université Cathoique de Louvain, Begium, 993 Massimo Piccardi, UTS 43
CS229 Lecture notes. Andrew Ng
CS229 Lecture notes Andrew Ng Part IX The EM agorithm In the previous set of notes, we taked about the EM agorithm as appied to fitting a mixture of Gaussians. In this set of notes, we give a broader view
More informationASummaryofGaussianProcesses Coryn A.L. Bailer-Jones
ASummaryofGaussianProcesses Coryn A.L. Baier-Jones Cavendish Laboratory University of Cambridge caj@mrao.cam.ac.uk Introduction A genera prediction probem can be posed as foows. We consider that the variabe
More information6.434J/16.391J Statistics for Engineers and Scientists May 4 MIT, Spring 2006 Handout #17. Solution 7
6.434J/16.391J Statistics for Engineers and Scientists May 4 MIT, Spring 2006 Handout #17 Soution 7 Probem 1: Generating Random Variabes Each part of this probem requires impementation in MATLAB. For the
More informationThe EM Algorithm applied to determining new limit points of Mahler measures
Contro and Cybernetics vo. 39 (2010) No. 4 The EM Agorithm appied to determining new imit points of Maher measures by Souad E Otmani, Georges Rhin and Jean-Marc Sac-Épée Université Pau Veraine-Metz, LMAM,
More information1 Introdution. l=1 k l, as in Figure 1.
Likeihood ratio test for the hyer-bock matrix shericity covariance structure characterization of the exact distribution deveoment of near-exact distributions for the test statistic Bárbara R Correia a,
More informationFluids Lecture 3 Notes
Fids Lectre 3 Notes 1. 2- Aerodynamic Forces and oments 2. Center of Pressre 3. Nondimensiona Coefficients Reading: Anderson 1.5 1.6 Aerodynamics Forces and oments Srface force distribtion The fid fowing
More informationA proposed nonparametric mixture density estimation using B-spline functions
A proposed nonparametric mixture density estimation using B-spine functions Atizez Hadrich a,b, Mourad Zribi a, Afif Masmoudi b a Laboratoire d Informatique Signa et Image de a Côte d Opae (LISIC-EA 4491),
More informationStochastic Variational Inference with Gradient Linearization
Stochastic Variationa Inference with Gradient Linearization Suppementa Materia Tobias Pötz * Anne S Wannenwetsch Stefan Roth Department of Computer Science, TU Darmstadt Preface In this suppementa materia,
More informationON THE INVARIANCE OF NONINFORMATIVE PRIORS. BY GAURI SANKAR DATTA 1 AND MALAY GHOSH 2 University of Georgia and University of Florida
he Annas of Statistics 1996, Vo 4, No 1, 141159 ON HE INVARIANCE OF NONINFORMAIVE PRIORS BY GAURI SANKAR DAA 1 AND MALAY GHOSH University of Georgia and University of Forida Jeffreys rior, one of the widey
More information7. CREST-TO-TROUGH WAVE HEIGHT DISTRIBUTION
7. CREST-TO-TROUGH WAVE HEIGHT DISTRIBUTION 7.1. Introduction In Chater 5, it has been mentioned that, in the wide sectrum case, the assumtion of H η does not hod even in the narrow case (considering that
More informationDiversity Gain Region for MIMO Fading Broadcast Channels
ITW4, San Antonio, Texas, October 4 9, 4 Diversity Gain Region for MIMO Fading Broadcast Channes Lihua Weng, Achieas Anastasoouos, and S. Sandee Pradhan Eectrica Engineering and Comuter Science Det. University
More informationStatistical Astronomy
Lectures for the 7 th IAU ISYA Ifrane, nd 3 rd Juy 4 p ( x y, I) p( y x, I) p( x, I) p( y, I) Statistica Astronomy Martin Hendry, Dept of Physics and Astronomy University of Gasgow, UK http://www.astro.ga.ac.uk/users/martin/isya/
More informationExpectation-Maximization for Estimating Parameters for a Mixture of Poissons
Expectation-Maximization for Estimating Parameters for a Mixture of Poissons Brandon Maone Department of Computer Science University of Hesini February 18, 2014 Abstract This document derives, in excrutiating
More informationGalois covers of type (p,, p), vanishing cycles formula, and the existence of torsor structures.
Gaois covers of tye,, ), vanishing cyces formua, and the existence of torsor structures. Mohamed Saïdi & Nichoas Wiiams Abstract In this artice we rove a oca Riemman-Hurwitz formua which comares the dimensions
More informationFractional Power Control for Decentralized Wireless Networks
Fractiona Power Contro for Decentraized Wireess Networks Nihar Jinda, Steven Weber, Jeffrey G. Andrews Abstract We roose and anayze a new aradigm for ower contro in decentraized wireess networks, termed
More informationTHE GENERALIZED VARIANCE OF A STATIONARY AUTOREGRESSIVE PROCESS. T. W. ANDERSON and RAUL P. MENTZ TECHNICAL REPORT NO.
THE GENERALIZED VARIANCE OF A STATIONARY AUTOREGRESSIVE PROCESS BY / T. W. ANDERSON and RAUL P. MENTZ TECHNICAL REPORT NO. 27 OCTOBER 1976 PREPARED UNDER CONTRACT N00014-75-C-0442 (NR-042-034) OFFICE OF
More informationMath 124B January 31, 2012
Math 124B January 31, 212 Viktor Grigoryan 7 Inhomogeneous boundary vaue probems Having studied the theory of Fourier series, with which we successfuy soved boundary vaue probems for the homogeneous heat
More informationUnitary Space-Time Pulse Position Modulation for Differential Unipolar MIMO IR-UWB Communications
1 Unitary Sace-Time Puse Position Moduation for Differentia Unioar MIMO IR-UWB Communications Chadi Abou-Rjeiy, Senior Member IEEE Abstract In this aer, we resent a genera technique for constructing minima-deay
More informationRadial Basis Functions: L p -approximation orders with scattered centres
Radia Basis Functions: L -aroximation orders with scattered centres Martin D. Buhmann and Amos Ron Abstract. In this aer we generaize severa resuts on uniform aroximation orders with radia basis functions
More informationII. PROBLEM. A. Description. For the space of audio signals
CS229 - Fina Report Speech Recording based Language Recognition (Natura Language) Leopod Cambier - cambier; Matan Leibovich - matane; Cindy Orozco Bohorquez - orozcocc ABSTRACT We construct a rea time
More informationRadial Basis Function Networks: Algorithms
Radial Basis Function Networks: Algorithms Introduction to Neural Networks : Lecture 13 John A. Bullinaria, 2004 1. The RBF Maing 2. The RBF Network Architecture 3. Comutational Power of RBF Networks 4.
More informationBayesian Learning. You hear a which which could equally be Thanks or Tanks, which would you go with?
Bayesian Learning A powerfu and growing approach in machine earning We use it in our own decision making a the time You hear a which which coud equay be Thanks or Tanks, which woud you go with? Combine
More informationSTA 216 Project: Spline Approach to Discrete Survival Analysis
: Spine Approach to Discrete Surviva Anaysis November 4, 005 1 Introduction Athough continuous surviva anaysis differs much from the discrete surviva anaysis, there is certain ink between the two modeing
More informationMATH 172: MOTIVATION FOR FOURIER SERIES: SEPARATION OF VARIABLES
MATH 172: MOTIVATION FOR FOURIER SERIES: SEPARATION OF VARIABLES Separation of variabes is a method to sove certain PDEs which have a warped product structure. First, on R n, a inear PDE of order m is
More informationMaximum Likelihood Estimation. only training data is available to design a classifier
Introduction to Pattern Recognition [ Part 5 ] Mahdi Vasighi Introduction Bayesian Decision Theory shows that we could design an optimal classifier if we knew: P( i ) : priors p(x i ) : class-conditional
More informationTHE THREE POINT STEINER PROBLEM ON THE FLAT TORUS: THE MINIMAL LUNE CASE
THE THREE POINT STEINER PROBLEM ON THE FLAT TORUS: THE MINIMAL LUNE CASE KATIE L. MAY AND MELISSA A. MITCHELL Abstract. We show how to identify the minima path network connecting three fixed points on
More informationAnalysis of Multiband Astronomical Images using Multiscale Tools
Anaysis of Mutiband Astronomica Images using Mutiscae Toos A.Bijaoui A.Guennec Ch.Benoist & E.Seza Cassioée Laboratory Côte d Azur Observatory BP42229 06304 Nice Cedex 04 France Outines Astronomica Imaging
More informationAnalysis of Emerson s Multiple Model Interpolation Estimation Algorithms: The MIMO Case
Technica Report PC-04-00 Anaysis of Emerson s Mutipe Mode Interpoation Estimation Agorithms: The MIMO Case João P. Hespanha Dae E. Seborg University of Caifornia, Santa Barbara February 0, 004 Anaysis
More informationIntroduction to LMTO method
1 Introduction to MTO method 24 February 2011; V172 P.Ravindran, FME-course on Ab initio Modeing of soar ce Materias 24 February 2011 Introduction to MTO method Ab initio Eectronic Structure Cacuations
More informationIEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. X, NO. Y, MONTH
IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. X, NO. Y, MONTH 24 Dynamic On-Chi Therma Sensor Caibration Using Performance Counters Shiting (Justin) Lu, Russe Tessier,
More informationTHE APPENDIX FOR THE PAPER: INCENTIVE-AWARE JOB ALLOCATION FOR ONLINE SOCIAL CLOUDS. Appendix A 1
TE APPENDIX FOR TE PAPER: INCENTIVE-AWARE JO ALLOCATION FOR ONLINE SOCIAL CLOUDS Yu Zhang, Mihaea van der Schaar Aendix A 1 1) Proof of Proosition 1 Given the SCP, each suier s decision robem can be formuated
More informationA Variance Minimization Criterion to Active Learning on Graphs
A Variance inimization Criterion to Active Learning on Grahs ing Ji Deartment of Comuter Science University of Iinois at Urbana-Chamaign Jiawei Han Deartment of Comuter Science University of Iinois at
More informationAutomobile Prices in Market Equilibrium. Berry, Pakes and Levinsohn
Automobie Prices in Market Equiibrium Berry, Pakes and Levinsohn Empirica Anaysis of demand and suppy in a differentiated products market: equiibrium in the U.S. automobie market. Oigopoistic Differentiated
More informationFRST Multivariate Statistics. Multivariate Discriminant Analysis (MDA)
1 FRST 531 -- Mutivariate Statistics Mutivariate Discriminant Anaysis (MDA) Purpose: 1. To predict which group (Y) an observation beongs to based on the characteristics of p predictor (X) variabes, using
More informationBALANCING REGULAR MATRIX PENCILS
BALANCING REGULAR MATRIX PENCILS DAMIEN LEMONNIER AND PAUL VAN DOOREN Abstract. In this paper we present a new diagona baancing technique for reguar matrix pencis λb A, which aims at reducing the sensitivity
More informationA novel Parameter Estimation method based on FRFT in Bistatic MIMO Radar System. Li Li1, a
6th Internationa Conference on Mechatronics Materias Biotechnoogy and Environment (ICMMBE 6 A nove Parameter Estimation method based on FRF in Bistatic MIMO Radar System i i a Information Engineering Coege
More informationNonlinear Analysis of Spatial Trusses
Noninear Anaysis of Spatia Trusses João Barrigó October 14 Abstract The present work addresses the noninear behavior of space trusses A formuation for geometrica noninear anaysis is presented, which incudes
More informationPattern selection and control via localized feedback
PHYSICAL REVIEW E 72, 066208 2005 Pattern seection and contro via ocaized feedback Andreas Hande Deartment of Bioogy, Emory University, Atanta, Georgia 30322, USA Roman O. Grigoriev Schoo of Physics, Georgia
More informationFactorial Markov Random Fields
Factoria Markov Random Fieds Junhwan Kim and Ramin Zabih Comuter Science Deartment Corne University Ithaca, NY 14853 Abstract. In this aer we roose an extension to the standard Markov Random Fied (MRF)
More informationThe Planck high-l temperature and polarization CMB power spectra and likelihood
The Panck high- temperature and poarization CMB power spectra and ikeihood Franz Esner on behaf of the Panck coaboration 1/12/14 Overview Updates since 213 Part 1: Power spectra Anaysis masks The Panck
More informationStatistics for Applications. Chapter 7: Regression 1/43
Statistics for Appications Chapter 7: Regression 1/43 Heuristics of the inear regression (1) Consider a coud of i.i.d. random points (X i,y i ),i =1,...,n : 2/43 Heuristics of the inear regression (2)
More informationInductive Bias: How to generalize on novel data. CS Inductive Bias 1
Inductive Bias: How to generaize on nove data CS 478 - Inductive Bias 1 Overfitting Noise vs. Exceptions CS 478 - Inductive Bias 2 Non-Linear Tasks Linear Regression wi not generaize we to the task beow
More informationAALBORG UNIVERSITY. The distribution of communication cost for a mobile service scenario. Jesper Møller and Man Lung Yiu. R June 2009
AALBORG UNIVERSITY The distribution of communication cost for a mobie service scenario by Jesper Møer and Man Lung Yiu R-29-11 June 29 Department of Mathematica Sciences Aaborg University Fredrik Bajers
More informationA. Distribution of the test statistic
A. Distribution of the test statistic In the sequentia test, we first compute the test statistic from a mini-batch of size m. If a decision cannot be made with this statistic, we keep increasing the mini-batch
More informationGauss Law. 2. Gauss s Law: connects charge and field 3. Applications of Gauss s Law
Gauss Law 1. Review on 1) Couomb s Law (charge and force) 2) Eectric Fied (fied and force) 2. Gauss s Law: connects charge and fied 3. Appications of Gauss s Law Couomb s Law and Eectric Fied Couomb s
More informationON CERTAIN SUMS INVOLVING THE LEGENDRE SYMBOL. Borislav Karaivanov Sigma Space Inc., Lanham, Maryland
#A14 INTEGERS 16 (2016) ON CERTAIN SUMS INVOLVING THE LEGENDRE SYMBOL Borisav Karaivanov Sigma Sace Inc., Lanham, Maryand borisav.karaivanov@sigmasace.com Tzvetain S. Vassiev Deartment of Comuter Science
More informationTHINKING IN PYRAMIDS
ECS 178 Course Notes THINKING IN PYRAMIDS Kenneth I. Joy Institute for Data Anaysis and Visuaization Department of Computer Science University of Caifornia, Davis Overview It is frequenty usefu to think
More informationON SEQUENCES WITHOUT GEOMETRIC PROGRESSIONS
MATHEMATICS OF COMPUTATION VOLUME 00, NUMBER 00 Xxxx 19xx, PAGES 000 000 ON SEQUENCES WITHOUT GEOMETRIC PROGRESSIONS BRIENNE E. BROWN AND DANIEL M. GORDON Abstract. Severa aers have investigated sequences
More informationAppendix of the Paper The Role of No-Arbitrage on Forecasting: Lessons from a Parametric Term Structure Model
Appendix of the Paper The Roe of No-Arbitrage on Forecasting: Lessons from a Parametric Term Structure Mode Caio Ameida cameida@fgv.br José Vicente jose.vaentim@bcb.gov.br June 008 1 Introduction In this
More information6 Wave Equation on an Interval: Separation of Variables
6 Wave Equation on an Interva: Separation of Variabes 6.1 Dirichet Boundary Conditions Ref: Strauss, Chapter 4 We now use the separation of variabes technique to study the wave equation on a finite interva.
More informationA Brief Introduction to Markov Chains and Hidden Markov Models
A Brief Introduction to Markov Chains and Hidden Markov Modes Aen B MacKenzie Notes for December 1, 3, &8, 2015 Discrete-Time Markov Chains You may reca that when we first introduced random processes,
More informationECON 4130 Supplementary Exercises 1-4
HG Set. 0 ECON 430 Sulementary Exercises - 4 Exercise Quantiles (ercentiles). Let X be a continuous random variable (rv.) with df f( x ) and cdf F( x ). For 0< < we define -th quantile (or 00-th ercentile),
More information221B Lecture Notes Notes on Spherical Bessel Functions
Definitions B Lecture Notes Notes on Spherica Besse Functions We woud ike to sove the free Schrödinger equation [ h d r R(r) = h k R(r). () m r dr r m R(r) is the radia wave function ψ( x) = R(r)Y m (θ,
More informationExplicit overall risk minimization transductive bound
1 Expicit overa risk minimization transductive bound Sergio Decherchi, Paoo Gastado, Sandro Ridea, Rodofo Zunino Dept. of Biophysica and Eectronic Engineering (DIBE), Genoa University Via Opera Pia 11a,
More informationIEOR E4570: Machine Learning for OR&FE Spring 2015 c 2015 by Martin Haugh. The EM Algorithm
IEOR E4570: Machine Learning for OR&FE Spring 205 c 205 by Martin Haugh The EM Algorithm The EM algorithm is used for obtaining maximum likelihood estimates of parameters when some of the data is missing.
More informationAkaike Information Criterion for ANOVA Model with a Simple Order Restriction
Akaike Information Criterion for ANOVA Mode with a Simpe Order Restriction Yu Inatsu * Department of Mathematics, Graduate Schoo of Science, Hiroshima University ABSTRACT In this paper, we consider Akaike
More informationLecture 17 - The Secrets we have Swept Under the Rug
Lecture 17 - The Secrets we have Swept Under the Rug Today s ectures examines some of the uirky features of eectrostatics that we have negected up unti this point A Puzze... Let s go back to the basics
More informationShort course A vademecum of statistical pattern recognition techniques with applications to image and video analysis. Agenda
Short course A vademecum of statistical pattern recognition techniques with applications to image and video analysis Lecture Recalls of probability theory Massimo Piccardi University of Technology, Sydney,
More informationPartial permutation decoding for MacDonald codes
Partia permutation decoding for MacDonad codes J.D. Key Department of Mathematics and Appied Mathematics University of the Western Cape 7535 Bevie, South Africa P. Seneviratne Department of Mathematics
More informationExpressing Priorities and External Probabilities in Process Algebra via Mixed Open/Closed Systems
Reace this fie with rentcsmacro.sty for your meeting, or with entcsmacro.sty for your meeting. Both can be found at the ENTCS Macro Home age. Exressing riorities and Externa robabiities in rocess Agebra
More informationAppendix for Stochastic Gradient Monomial Gamma Sampler
3 4 5 6 7 8 9 3 4 5 6 7 8 9 3 4 5 6 7 8 9 3 3 3 33 34 35 36 37 38 39 4 4 4 43 44 45 46 47 48 49 5 5 5 53 54 Appendix for Stochastic Gradient Monomia Gamma Samper A The Main Theorem We provide the foowing
More informationPHASE retrieval aims at recovering a complex-valued signal
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 65, NO., NOVEMBER 15, 017 5957 Undersamed Sarse Phase Retrieva via Majorization Minimization Tianyu Qiu and Danie P. Paomar, Feow, IEEE Abstract In the undersamed
More informationHILBERT? What is HILBERT? Matlab Implementation of Adaptive 2D BEM. Dirk Praetorius. Features of HILBERT
Söerhaus-Workshop 2009 October 16, 2009 What is HILBERT? HILBERT Matab Impementation of Adaptive 2D BEM joint work with M. Aurada, M. Ebner, S. Ferraz-Leite, P. Godenits, M. Karkuik, M. Mayr Hibert Is
More information22.615, MHD Theory of Fusion Systems Prof. Freidberg Lecture 2: The Moment Equations
.615, MHD Theory of Fusion ystems Prof. Freidberg Lecture : The Moment Equations Botzmann-Maxwe Equations 1. Reca that the genera couped Botzmann-Maxwe equations can be written as f q + v + E + v B f =
More informationHomework #04 Answers and Hints (MATH4052 Partial Differential Equations)
Homework #4 Answers and Hints (MATH452 Partia Differentia Equations) Probem 1 (Page 89, Q2) Consider a meta rod ( < x < ), insuated aong its sides but not at its ends, which is initiay at temperature =
More informationShort course A vademecum of statistical pattern recognition techniques with applications to image and video analysis. Agenda
Short course A vademecum of statistical attern recognition techniques with alications to image and video analysis Lecture 6 The Kalman filter. Particle filters Massimo Piccardi University of Technology,
More informationLIKELIHOOD RATIO TEST FOR THE HYPER- BLOCK MATRIX SPHERICITY COVARIANCE STRUCTURE CHARACTERIZATION OF THE EXACT
LIKELIHOOD RATIO TEST FOR THE HYPER- BLOCK MATRIX SPHERICITY COVARIACE STRUCTURE CHARACTERIZATIO OF THE EXACT DISTRIBUTIO AD DEVELOPMET OF EAR-EXACT DISTRIBUTIOS FOR THE TEST STATISTIC Authors: Bárbara
More informationPerformance Comparison of K-Means and Expectation Maximization with Gaussian Mixture Models for Clustering EE6540 Final Project
Performance Comparison of K-Means and Expectation Maximization with Gaussian Mixture Models for Clustering EE6540 Final Project Devin Cornell & Sushruth Sastry May 2015 1 Abstract In this article, we explore
More informationSparse Representations in Unions of Bases
330 IEEE TRANSACTIONS ON INFORMATION THEORY, VO 49, NO, DECEMBER 003 rates taen so high that further increasing them roduced no visibe changes in the figure As can be seen, the a;b obtained in that way
More informationHYDROGEN ATOM SELECTION RULES TRANSITION RATES
DOING PHYSICS WITH MATLAB QUANTUM PHYSICS Ian Cooper Schoo of Physics, University of Sydney ian.cooper@sydney.edu.au HYDROGEN ATOM SELECTION RULES TRANSITION RATES DOWNLOAD DIRECTORY FOR MATLAB SCRIPTS
More informationFirst-Order Corrections to Gutzwiller s Trace Formula for Systems with Discrete Symmetries
c 26 Noninear Phenomena in Compex Systems First-Order Corrections to Gutzwier s Trace Formua for Systems with Discrete Symmetries Hoger Cartarius, Jörg Main, and Günter Wunner Institut für Theoretische
More informationSome Measures for Asymmetry of Distributions
Some Measures for Asymmetry of Distributions Georgi N. Boshnakov First version: 31 January 2006 Research Report No. 5, 2006, Probabiity and Statistics Group Schoo of Mathematics, The University of Manchester
More informationSoft Clustering on Graphs
Soft Custering on Graphs Kai Yu 1, Shipeng Yu 2, Voker Tresp 1 1 Siemens AG, Corporate Technoogy 2 Institute for Computer Science, University of Munich kai.yu@siemens.com, voker.tresp@siemens.com spyu@dbs.informatik.uni-muenchen.de
More informationBesicovitch and other generalizations of Bohr s almost periodic functions
Besicovitch and other generaizations of Bohr s amost eriodic functions Kevin Nowand We discuss severa casses of amost eriodic functions which generaize the uniformy continuous amost eriodic (a..) functions
More informationExpressing Priorities, External Probabilities and Time in Process Algebra via Mixed Open/Closed Systems
Exressing riorities, Externa robabiities and Time in rocess Agebra via Mixed Oen/Cosed Systems M. Bravetti Technica Reort UBLCS-2007-18 June 2007 Deartment of Comuter Science University of Boogna Mura
More informationV.B The Cluster Expansion
V.B The Custer Expansion For short range interactions, speciay with a hard core, it is much better to repace the expansion parameter V( q ) by f(q ) = exp ( βv( q )) 1, which is obtained by summing over
More informationSVM: Terminology 1(6) SVM: Terminology 2(6)
Andrew Kusiak Inteigent Systems Laboratory 39 Seamans Center he University of Iowa Iowa City, IA 54-57 SVM he maxima margin cassifier is simiar to the perceptron: It aso assumes that the data points are
More informationOn estimation of limiting variance of partial sums of functions of associated random variables
isid/ms/06/0 October 0, 06 htt://www.isid.ac.in/ statmath/index.h?modue=prerint On estimation of imiting variance of artia sums of functions of associated random variabes Mansi Garg and Isha Dewan Indian
More informationResearch Article SVM-Based Spectrum Mobility Prediction Scheme in Mobile Cognitive Radio Networks
e Scientific Word Journa, Artice ID 395212, 11 ages htt://dx.doi.org/10.1155/2014/395212 Research Artice SVM-Based Sectrum Mobiity Prediction Scheme in Mobie Cognitive Radio Networks Yao Wang, 1,2 Zhongzhao
More informationMARKOV CHAINS AND MARKOV DECISION THEORY. Contents
MARKOV CHAINS AND MARKOV DECISION THEORY ARINDRIMA DATTA Abstract. In this paper, we begin with a forma introduction to probabiity and expain the concept of random variabes and stochastic processes. After
More informationCollaborative Calibration of On-Chip Thermal Sensors Using Performance Counters
Coaborative Caibration of On-Chi Therma s Using Performance Counters Shiting (Justin) Lu, Russe Tessier, and Wayne Bureson University of Massachusetts Deartment of Eectrica and Comuter Engineering Amherst,
More informationCryptanalysis of PKP: A New Approach
Cryptanaysis of PKP: A New Approach Éiane Jaumes and Antoine Joux DCSSI 18, rue du Dr. Zamenhoff F-92131 Issy-es-Mx Cedex France eiane.jaumes@wanadoo.fr Antoine.Joux@ens.fr Abstract. Quite recenty, in
More informationAppendix for Stochastic Gradient Monomial Gamma Sampler
Appendix for Stochastic Gradient Monomia Gamma Samper A The Main Theorem We provide the foowing theorem to characterize the stationary distribution of the stochastic process with SDEs in (3) Theorem 3
More informationAnt Colony Algorithms for Constructing Bayesian Multi-net Classifiers
Ant Coony Agorithms for Constructing Bayesian Muti-net Cassifiers Khaid M. Saama and Aex A. Freitas Schoo of Computing, University of Kent, Canterbury, UK. {kms39,a.a.freitas}@kent.ac.uk December 5, 2013
More information2M2. Fourier Series Prof Bill Lionheart
M. Fourier Series Prof Bi Lionheart 1. The Fourier series of the periodic function f(x) with period has the form f(x) = a 0 + ( a n cos πnx + b n sin πnx ). Here the rea numbers a n, b n are caed the Fourier
More informationOn parameter estimation in deformable models
Downloaded from orbitdtudk on: Dec 7, 07 On arameter estimation in deformable models Fisker, Rune; Carstensen, Jens Michael Published in: Proceedings of the 4th International Conference on Pattern Recognition
More informationStatistical Learning Theory: A Primer
Internationa Journa of Computer Vision 38(), 9 3, 2000 c 2000 uwer Academic Pubishers. Manufactured in The Netherands. Statistica Learning Theory: A Primer THEODOROS EVGENIOU, MASSIMILIANO PONTIL AND TOMASO
More informationModel-based Clustering by Probabilistic Self-organizing Maps
EEE TRANATN N NERA NETRK, V. XX, N., ode-based ustering by robabiistic ef-organizing aps hih-ian heng, Hsin-hia u, ember, EEE, and Hsin-in ang, ember, EEE Abstract n this paper, we consider the earning
More informationc 2007 Society for Industrial and Applied Mathematics
SIAM REVIEW Vo. 49,No. 1,pp. 111 1 c 7 Society for Industria and Appied Mathematics Domino Waves C. J. Efthimiou M. D. Johnson Abstract. Motivated by a proposa of Daykin [Probem 71-19*, SIAM Rev., 13 (1971),
More informationAnalysis of rounded data in mixture normal model
Stat Papers (2012) 53:895 914 DOI 10.1007/s00362-011-0395-0 REGULAR ARTICLE Anaysis of rounded data in mixture norma mode Ningning Zhao Zhidong Bai Received: 13 August 2010 / Revised: 9 June 2011 / Pubished
More informationUniprocessor Feasibility of Sporadic Tasks with Constrained Deadlines is Strongly conp-complete
Uniprocessor Feasibiity of Sporadic Tasks with Constrained Deadines is Strongy conp-compete Pontus Ekberg and Wang Yi Uppsaa University, Sweden Emai: {pontus.ekberg yi}@it.uu.se Abstract Deciding the feasibiity
More informationLecture Note 3: Stationary Iterative Methods
MATH 5330: Computationa Methods of Linear Agebra Lecture Note 3: Stationary Iterative Methods Xianyi Zeng Department of Mathematica Sciences, UTEP Stationary Iterative Methods The Gaussian eimination (or
More informationL11: Pattern recognition principles
L11: Pattern recognition principles Bayesian decision theory Statistical classifiers Dimensionality reduction Clustering This lecture is partly based on [Huang, Acero and Hon, 2001, ch. 4] Introduction
More informationEfficient Approximate Leave-One-Out Cross-Validation for Kernel Logistic Regression
Machine Learning manuscript No. (wi be inserted by the editor) Efficient Approximate Leave-One-Out Cross-Vaidation for Kerne Logistic Regression Gavin C. Cawey, Nicoa L. C. Tabot Schoo of Computing Sciences
More informationTesting distributional assumptions using a continuum of moments
Testing distributiona assumtions using a continuum of moments Dante Amengua CEMFI Marine Carrasco Université de Montréa Enrique Sentana CEMFI
More informationPolar Snakes: a fast and robust parametric active contour model
Poar Snakes: a fast and robust parametric active contour mode Christophe Coewet To cite this version: Christophe Coewet. Poar Snakes: a fast and robust parametric active contour mode. IEEE Int. Conf. on
More information4 Separation of Variables
4 Separation of Variabes In this chapter we describe a cassica technique for constructing forma soutions to inear boundary vaue probems. The soution of three cassica (paraboic, hyperboic and eiptic) PDE
More informationCombining Logistic Regression with Kriging for Mapping the Risk of Occurrence of Unexploded Ordnance (UXO)
Combining Logistic Regression with Kriging for Maing the Risk of Occurrence of Unexloded Ordnance (UXO) H. Saito (), P. Goovaerts (), S. A. McKenna (2) Environmental and Water Resources Engineering, Deartment
More informationPHYS 110B - HW #1 Fall 2005, Solutions by David Pace Equations referenced as Eq. # are from Griffiths Problem statements are paraphrased
PHYS 110B - HW #1 Fa 2005, Soutions by David Pace Equations referenced as Eq. # are from Griffiths Probem statements are paraphrased [1.] Probem 6.8 from Griffiths A ong cyinder has radius R and a magnetization
More informationActive Learning & Experimental Design
Active Learning & Experimenta Design Danie Ting Heaviy modified, of course, by Lye Ungar Origina Sides by Barbara Engehardt and Aex Shyr Lye Ungar, University of Pennsyvania Motivation u Data coection
More information