33 Pages

semaine1

Course: IFT 6141, Fall 2009
School: Neumont
Rating:
 
 
 
 
 

Word Count: 5611

Document Preview

2008 IFT6141 Meunier Reconnaissance des Formes Semaine #1 Professeur: Jean MEUNIER meunier@iro.umontreal.ca Bureau: 2163 D.I.R.O. Au menu cette semaine! Plan de cours Introduction la R. de formes Mthodes bayesiennes (Chap. 2) Estimation des paramtres (Chap. 3) Le problme de la dimensionalit (Chap. 3) 2 Meunier 2008 1 Meunier 2008 Plan du cours Rfrence en rserve la bibliothque: Duda, Hart, Stork,...

Register Now

Unformatted Document Excerpt

Coursehero >> Utah >> Neumont >> IFT 6141

Course Hero has millions of student submitted documents similar to the one
below including study guides, practice problems, reference materials, practice exams, textbook help and tutor support.

Course Hero has millions of student submitted documents similar to the one below including study guides, practice problems, reference materials, practice exams, textbook help and tutor support.
2008 IFT6141 Meunier Reconnaissance des Formes Semaine #1 Professeur: Jean MEUNIER meunier@iro.umontreal.ca Bureau: 2163 D.I.R.O. Au menu cette semaine! Plan de cours Introduction la R. de formes Mthodes bayesiennes (Chap. 2) Estimation des paramtres (Chap. 3) Le problme de la dimensionalit (Chap. 3) 2 Meunier 2008 1 Meunier 2008 Plan du cours Rfrence en rserve la bibliothque: Duda, Hart, Stork, "Pattern Classification", 2e dition, Wiley-Interscience, 2001 Site web associ: http://rii.ricoh.com/~stork/DHS.html Nombreux autres ouvrages la bibliothque Notes de cours sur: http://www.iro.umontreal.ca/~meunier/IFT6141/ Meunier 2008 3 Dfinition Selon le grand dictionnaire terminologique de la langue franaise: Branche de l'informatique qui tudie les techniques permettant un ordinateur d'identifier des formes des structures ou des configurations. Comment? En se basant sur certaines de leurs caractristiques. Meunier 2008 4 2 Meunier 2008 Introduction la RF Prtraitement: IFT6150 Rehaussement et restauration de limage prise par la camra Segmentation Extraction de caractristiques: IFT6150 et IFT6410 Pour rduire le quantit de donnes (e.g. 1024 x 1024 x 8 bits) Proprits de lobjet qui serviront la classification Classification: IFT6141 Algorithme qui va valuer les vidences qui lui sont prsentes et prendra une dcision finale Bar ou loup (de mer) Meunier 2008 5 2001 John Wiley & Sons, Inc. Applications de la RF Robotique/industrie Assemblage (reconnaissance de pices) Contrle de qualit (pices, fruits ou autres) Vhicule autonome etc. Mto (tempte, ouragan) Identification et suivi des cultures/forts/rserves deau Cartographie, analyse des ressources terrestres (ptrole) Pollution Analyse de lECG ou EEG pour fin de diagnostic Analyse dimages mdicales (tumeurs, cellules cancreuses) Identification de squences dADN Guidage de missile (reconnaissance dune cible et du terrain) Reconnaissance arienne (espionnage) Reconnaissance de texte par ordinateur (OCR) Analyse de document Reconnaissance de la parole Identification des empreintes digitales, iris, mains, signatures Reconnaissance de visage et de la parole Classification de signaux sismiques Tldtection Mdecine Application militaire Bureautique Scurit Meunier 2008 6 3 Meunier 2008 Plan du cours Contenu: Mthode de Bayes (Chap. 1 et 2), Estimation des paramtres (Chap. 3) Rduction de la dimensionnalit (chap. 3) Mthodes non-paramtriques (Chap. 4) Fonctions discriminantes linaires (Chap. 5) Rseaux de neurones (Chap. 6) Mthodes stochastiques (Chap. 7). Mthodes syntaxiques et structurelles (Chap. 8) Mthode non-supervises (Chap. 10) Remarques indpendantes de lalgorithme (Chap. 9) Attributs en traitement de signaux et dimages Applications en reconnaissance audio Applications en vision par ordinateur 7 Meunier 2008 Plan du cours Horaire (dbutant le mardi 8 janvier): Mardi 10h30 12h30 Z-345 Jeudi 10h30 12h30 Z-337 Thorie de la RdeF jusquau 14 fvrier Projet #1 (examen): nonc le 18 fvrier, date limite de remise: le 7 mars 22h00 par courriel (Note: Semaine de lecture (FAS): 3 mars 7 mars) Applications: du 11 au 27 mars Projet #2: date limite pour le rapport: le 11 avril 22h00 par courriel et prsentation les 15 et 17 avril xx xx xx Meunier 2008 8 4 Meunier 2008 valuation Examen-Projet #1: 40 % Travail impos raliser en 3 semaines. Application de la RdeF avec implmentation. Rapport crit (format libre) faire seul Projet #2: 60% (dont 20% pour un expos oral) Projet de RdeF raliser en 5 semaines. Application avec implmentation au choix. Rapport crit sous forme dun article format IEEE* double colonne (max 8 pages) Document powerpoint de 20 30 diapositives (prsentation orale de 30 minutes) faire seul Remarques: - Donner toutes les rfrences (articles, livres et sites internets). Vous pouvez utiliser du code disponible sur Internet mais clairement mentionner les rfrences. Pas de rfrence = plagiat = 0. - Tous les autres lments du projet (donnes utilises, code source, code excutable, etc.) devront tre accessibles par Internet ou remis autrement (par ex: CD). *voir par exemple le site http://www.computerrobotvision.org/submit.html Article rdig en franais ou en anglais pour le cours. Meunier 2008 9 Introduction la RF Prtraitement: IFT6150 Rehaussement et restauration de limage prise par la camra Segmentation Extraction de caractristiques: IFT6150 et IFT6410 Pour rduire le quantit de donnes (e.g. 1024 x 1024 x 8 bits) Proprits de lobjet qui serviront la classification Classification: IFT6141 Algorithme qui va valuer les vidences qui lui sont prsentes et prendra une dcision finale Bar ou loup (de mer) Meunier 2008 10 2001 John Wiley & Sons, Inc. 5 Meunier 2008 Introduction la RF Exemple de caractristiques: Longueur luminosit (environnement contrl) http://www.gastromer.ch/fr/produits.htm Meunier 2008 11 2001 John Wiley & Sons, Inc. Introduction la RF Peut-on amliorer les rsultats avec plus dune caractristique? Le seuil devient une droite ou une courbe (puis une surface) Si on ajoute dautres caractristiques est-ce quon peut encore amliorer? linfini? Pourquoi ne pas adopter un modle (courbe) plus complexe? Erreur = 0 sur lensemble dentranement Quelle sera lerreur pour de nouveaux tests (poissons)? Par cur vs. Gnralisation 12 2001 John Wiley & Sons, Inc. Meunier 2008 6 Meunier 2008 RF: Mthode simpliste #1 2 classes saumons: 1 bars: 2 Aucune mesure Aucun prtraitement Aucune extraction de caractristiques Classification: Base sur les probabilits a priori P(i) Choisir 1 si P(1) > P(2) sinon 2 Erreur: Min[P(1), P(2)] Meunier 2008 13 2001 John Wiley & Sons, Inc. RF: mthode simpliste #2 2 classes saumons: 1 bars: 2 Extraction de caractristiques: Une mesure x est faite O classer? Classification: Base sur la vraisemblance ou densit de probabilit de x pour chacune des classe Choisir 1 si p(x|1) > p(x|2) sinon 2 Erreur: Min[p(x|1), p(x|2)] / [p(x|1)+ p(x|2)] Quarrive-t-il si les saumons ne comptent que pour 1 poisson sur 10 dans les prises? =1 i Meunier 2008 14 2001 John Wiley & Sons, Inc. 7 Meunier 2008 Bayes Un classificateur Bayesien tient compte des deux facteurs prcdents pour calculer la Probabilit a posteriori La probabilit jointe quune observation provienne dune classe i avec comme caractristique la valeur x est donne par: p (i , x ) = P (i ) p ( x i ) = p ( x) P (i x) Do le thorme de Bayes: P (i x) = P (i ) p ( x i ) p( x) 15 p( x ) = P (1 ) p( x 1 ) + P (2 ) p( x 2 ) Meunier 2008 Bayes 2 classes saumons: 1 bars: 2 Extraction de caractristiques: Une mesure x est faite P(1)p(x|1) Classification: Base sur les probabilits a posteriori p(i|x) Calcules partir des vraisemblances p(x|i) et des probabilits a priori P(i) Choisir 1 si P(1| x) > P(2 | x) sinon 2 ou Choisir 1 si p(x|1)P(1) > p(x|2)P(2) sinon 2 P(2)p(x|2) Erreur: Min[P(1| x), P(2 | x)] Meunier 2008 16 2001 John Wiley & Sons, Inc. 8 Meunier 2008 Compliquons un peu les choses: Notion de Risque 2 classes saumons: 1 bars: 2 Classificateur erreur minimale? 0 ij = 1 i= j i j Extraction de caractristiques: Une mesure x est faite Classification: Base sur la minimisation du risque conditionnel dune classification Calcul partir des probabilits a posteriori Plus gnral R(1| x) = 11P(1| x) + 12P(2 | x) R(2| x) = 21P(1| x) + 22P(2 | x) Choisir 1 si R(1| x) < R(2| x) sinon 2 11 risque de classer 12 risque de classer 22 risque de classer 21 risque de classer un saumon dans la classe saumon (faible) un bar dans la classe saumon (lev) un bar dans la classe bar (faible) un saumon dans la classe bar (moyen) Meunier 2008 Pour satisfaire le consommateur (diffrent si on veut simplement maximiser les revenus sans gard aux consommateurs) 17 Fonctions discriminantes Fonctions gi(x) tel que: gi(x) > gj(x) j i si x i Par exemple: P(i) p(x|i) P(i| x) p(x | i) P(i) -R(i| x) ln p(x | i) + ln P(i) max Meunier 2008 18 2001 John Wiley & Sons, Inc. 9 Meunier 2008 Modle pour p(x|)? x devient un vecteur x plutt quun scalaire Quelle est la forme de p(x | i)? La distribution normale ou gaussienne est de loin le modle prfr Proprit mathmatique et statistique unique Fourier (G) = G Transformation linaire (G) = G Projection (G) = G Simple 2 paramtres dans le cas 1-D d+d(d+1)/2 pour d dimensions Thorme de la limite centrale Somme dun grand nombre de variables alatoires indpendantes donne une distribution gaussienne Donc idale pour modliser p(x | i) Meunier 2008 ( p( x) = 1 e 2 2 1 x 2 ) 19 2001 John Wiley & Sons, Inc. Normale ou gaussienne en d-D p ( x) = 1 i ( 2 ) d e 1 ( xi )T i1 ( x i ) 2 Transformation linaire (G) = G Mahalanobis2 =cte Meunier 2008 20 2001 John Wiley & Sons, Inc. 10 Meunier 2008 Fonctions discriminantes gaussiennes Cas de 2 classes On peut tester 0.175 g(x) = g1(x) g2(x) 0.15 Si g(x) > 0 alors 1 0.125 0.1 sinon 2 0.075 g(x)=0 g1(x) = g2(x) 0.05 0.025 correspond la frontire de -2 dcision Meunier 2008 g(x)=0 => x = seuil p(x) P(1)p(x|1) P(2)p(x|2) 2 4 6 8 10 21 Frontires de dcision 1D Si les valeurs de la caractristique x pour chaque classe 1 et 2 suivent une loi normale 1 x i 2 ( ) P ( i ) 1 g i ( x ) = ln [P ( i ) p ( x i ) ] = ln P ( i ) e 2 i = ln 2 i 2 i x i + 1 2 i 2 La frontire de dcision est: g ( x) = g1 ( x) g 2 ( x) = 0 P(1 ) ln 2 1 Meunier 2008 P (2 ) ln 2 2 1 x 1 1 x 2 2 + 2 =0 1 2 22 2 2 11 Meunier 2008 Frontires de dcision 1D Cas Gnral: 2 solutions P(i) 2 2 P(1 ) ln 2 1 P( 2 ) ln 2 2 1 x 1 1 x 2 2 + 2 =0 2 1 forme ax 2 + bx + c = 0 Meunier 2008 23 2001 John Wiley & Sons, Inc. Frontires de dcision 1D Cas 1 = 2 : 1 solution 0 = C + ( x 1 ) 2 ( x 2 ) 2 0 = C ' 2( 1 2 ) x x = C" = seuil unique Cas avec en plus P(1)=P(2) ( x 1 ) 2 = ( x 2 ) 2 x= 1 + 2 2 24 2001 John Wiley & Sons, Inc. Meunier 2008 12 Meunier 2008 Frontires de dcision: d-D Si nous avons c classes et d caractristiques, nous pouvons reprsenter les moyennes des attributs de chaque classe i par un vecteur de moyenne 1 i = d Meunier 2008 Les variances et covariances des caractristiques de chaque classe i sont reprsentes par une matrice de covariance 12 12 1d 2 12 2 2d i = 2 d 1d Cette matrice est symtrique La variance de chaque attribut 25 est sur la diagonale Frontires de dcision: d-D La fonction discriminante a la forme P(i ) p( x i ) = P(i ) | i | ( 2) d e 1 ( xi )T i1 ( xi ) 2 En prenant le logarithme et en multipliant par 2 nous obtenons 2ln P(i ) ln | i | d ln2 ( x i )T i1 ( x i ) Meunier 2008 26 13 Meunier 2008 Frontires de dcision: d-D La frontire entre 2 classes i et j est dfinie par 2 ln P ( i ) ln | i | ( x i ) t i1 ( x i ) = 2 ln P ( j ) ln | j | ( x j ) t 1 ( x j ) j Sachant que: 1 1 1 ( x i )t i1 ( x i ) = xt i x 2t i x + t i i i i On obtient la forme: Meunier 2008 x t Ax + b t x + c = 0 27 Frontires de dcision: d-D g ( x ) = x t Ax + b t x + c = 0 A = 1 i1 j b = 2 1 j + 2 i1 i j c = 2 ln |j | P ( i ) + ln + Tj 1 j T i1 i j i P ( j ) | i | Quadratique! Surfaces de dcision sont des hyperquadriques hyperplans, paires dhyperplans, hypershres, hyperellipsodes, hyperparabolodes, hyperhyperbolodes Meunier 2008 2-D 28 14 Meunier 2008 Cas particulier: i = A=0 b t x + c = 0 Hyperplan b = 2 1 j + 2 1 i = 2 1 ( i j ) c = 2 ln P ( i ) + tj 1 j t 1 i i P( j ) Cas particulier: i = 2I b = 22 ( i j ) Hyperplan au vecteur ij c = 2 ln P ( i ) 1 t t + i i P ( j ) 2 j j Cas particulier: i = 2I et P(i) = P(j) b= 2 2 ( i j ) c= i + j 2 2 1 t t j j i i bt x + c = 0 x = + (i j ) Hyperplan au vecteur ij et situ juste entre les deux Meunier 2008 Sergios Theodoridis 2003 29 Exemple: Calcul de la frontire g (x) = xt Ax + bt x + c = 0 A = 1 i1 j b = 2 1 j + 2 i1 i j c = 2 ln |j | P(i ) + ln P( j ) | i | x2 = 3.514 1.125 x1 + 0.1875 x1 2 + Tj 1 j T i1 i j i Meunier 2008 2001 John Wiley & Sons, Inc. 30 15 Meunier 2008 Indpendance des caractristiques: Bayes Naf La fonction discriminante a souvent la forme P() p(x ) = P() | | (2 )d e 1 ( x)T 1 (x) 2 Indpendance P ( ) p ( x ) = P( ) d e k d 1 xk k 2 k =1 k 2 (2 ) Meunier 2008 d /2 k =1 31 Fonctions discriminantes pour plusieurs classes? Entre deux classes la frontire est une hyperquadrique Toutefois les rgions et frontires dfinitives en tenant compte de toutes les classes sont beaucoup plus complexes gi(x) > gj(x) j i si x i On maximise la fonction discriminante Meunier 2008 2001 John Wiley & Sons, Inc. 32 16 Meunier 2008 Probabilit derreurs? 2001 John Wiley & Sons, Inc. P ( erreur ) = P ( 1 ) p(x xR 2 1 ) dx + P ( 2 ) p( x xR1 2 ) dx Meunier 2008 33 Probabilit derreurs? Erreur c pour classes est dduite de la probabilit dtre OK (+ simple calculer) P (OK ) = P ( i ) i =1 c p ( x )dx i xRi c i =1 xRi Des bornes faciles calculer existent P (erreur ) = 1 P (OK ) = 1 P ( i ) = P ( i ) i =1 Meunier 2008 p ( x )dx i c p ( x )dx i xRi Pour chaque , plusieurs rgions Rj Ri; difficile et long calculer 34 17 Meunier 2008 Au menu cette semaine! Introduction la R. de formes Plan de cours Mthodes bayesiennes (Chap. 2) Estimation des paramtres (Chap. 3) Le problme de la dimensionalit 35 Meunier 2008 Estimation des paramtres? Classifier = Choisir la + grande fonction discriminante. Les fonctions discriminantes font intervenir en gnral les probabilits a priori et les vraisemblances (Bayes) g i ( x) P(i x) = P(i ) p (x i ) p ( x) P(i ) p (x i ) Les probabilits a priori ne sont en gnral pas trop compliques dterminer, par ex. partir dun chantillon D des poissons pchs on dtermine P(bar) et P(saumon) Quen est-il des vraisemblances? => Distribution statistiques des attributs. partir des donnes D? Meunier 2008 36 18 Meunier 2008 Estimation par maximisation de la vraisemblance (ML) On maximise pour chaque classe n Moyenne inconnue (variance connue ou fixe) p ( X ) = p ( x k ) k =1 avec X = {x1 , x 2 ,..., x k ,..., x n } ou l () = ln ( p ( X ) ) = ln p ( x k ) k =1 n On essaie de trouver la meilleure distribution Solution: tel que p(X) = 0 ou l = 0 Meunier 2008 37 Cas o p(xk|) est gaussien p( x | ) = 1 ( x ) 2 1 e 22 1 2 2 1 ln p( x | ) = 1 ln 2 2 2 2 ( x 1 ) 2 2 1 n n ( xk 1 ) l = ln p( xk | ) = 1 2 1 2 = 0 + 2 2 ( xk 1 ) k =1 k =1 2 2 2 n 1 = = 1 xk k =1 n et 2 = 2 s 2 = 1 ( xk )2 n k =1 n Meunier 2008 38 19 Meunier 2008 Cas o p(xk|) est gaussien Biais ML non-biais pour 1 1 1 E [ ] = E xi = E[xi ] = = n n n ML est biais pour 2 car: 1 1 n 1 2 E s 2 = E ( xi ) 2 = . 2 = ( xi ) 2 n n n 1 [] De mme pour d-D: 1n xk n k =1 n = 1 ( x k )( x k )t n k =1 n C = 1 ( x k )( x k )t n 1 k =1 = Non biais si on utilisait dans la formule au lieu dun estim ML x qui contient xi dans sa somme... xx 2 = E[(x E[x])2 ] Sinon cas n=1 donnerait = 0 au lieu de infini Meunier 2008 39 Estimation avec maximum a posteriori (MAP) On maximise p(D| ) p ( X ) = p () p ( X ) p ( X) n p () p ( X ) = p () p (x k ) k =1 Sergios Theodoridis 2003 On considre comme un vecteur alatoire MAP=ML lorsque p() est uniforme On doit connatre p() : information priori sur Meunier 2008 40 20 Meunier 2008 Exemple 1D: loi normale avec connu 1er tape: On calcule p( |D) = p(|D) On connat (sans les donnes) p(x| ) N(, ) p() N(0, 0) n p ( xk | ) = 1 2 2 1 (x k 2 2 e )2 p ( | D) p ( ) p(xk | ) k =1 1 2 0 2 p( ) = Meunier 2008 e 1 ( )2 0 2 2 0 41 loi normale avec connu 2e tape: maximisation ln[ p ( | D)] = 0 N 1 1 ( 0 ) 2 + 2 ( xk ) 2 = 0 2 2 0 2 k =1 n 0 2 2 n + n = 2 2 n + 2 n + 2 0 0 0 n = ? n = 0? 0 = ? (p() est uniforme) Meunier 2008 n = 1n xk n k =1 42 21 Meunier 2008 2 exemples de lvolution de p(|D) p(|D) est une gaussienne de moyenne n et de variance n 2 = 2 0 2 2 n 0 + 2 Meunier 2008 43 Exemple en imagerie numrique: Format PPM (Portable Pixel Map) P3 # Created by Paint Shop #ligne #colonne 64 64 Image 8 bits 255 255 0 0 255 0 0 208 207 208 207 255 208 207 255 000000000000 000000000000 000000000000 000000000000 000000000000 000000000000 000000000000 208 207 255 208 207 255 Meunier 2008 Pro 5 entte 255 00 00 00 00 00 00 00 00 208 208 207 255 0000 0000 0000 0000 0000 0000 0000 0 208 207 255 207 255 Jasmine.ppm Note: P4 si le fichier est binaire (bytes ou raw) 44 22 Meunier 2008 Exemple Image couleur 64 x 64 3 rgions principales: 1: maison blanche 2: vgtation 3: ciel 10 x 10 = 100 chantillons Caractristiques: R, V et B Meunier 2008 extract( MB) 45 Exemple Verdure = 28.47 9.46 29.17 Verdure = 79.364 476.049 258.88 29.148 79.364 49.762 49.762 258.88 185.241 Segmentation Maison = 142.34 124.99 157.87 Maison = 3.005 103 2.57 103 2.28 103 3 3 3 2.57 10 2.252 10 2.036 10 2.28 103 2.036 103 1.892 103 Ciel = 139.98 94.47 198.12 Ciel = 31.849 5.019 16.104 5.019 16.104 10.66 0.992 0.992 25.706 Variance leve maison verdure ciel Meunier 2008 Verdure := 4 1 Ciel := 4 1 Maison := 2 1 46 23 Meunier 2008 EM (Expectation-Maximization) Comment trouver les paramtres dun modle statistique dans le cas dun mlange ou mixture de distributions? Par exemple: 2 p ( x | ) = 1 p1 ( x | 1 , 12 ) + 2 p2 ( x | 2 , 2 ) = ( 1 , 2 , 1 , 12 , 2 , 22 ) Solution: approche itrative EM Meunier 2008 47 EM (Expectation-Maximization) Mixture de c distributions: p (x ) = P ( j =1 c j ) p (x j ) = j =1 c j p (x j ) On veut maximiser la vraisemblance: ln ( p ( X ) ) = ln p ( x k ) = ln j p ( x k j ) k =1 k =1 j =1 n n c Difficile maximiser on ne sait pas quelle distribution j est associ chacune des donnes xk (si on le savait ce serait trivial en procdant une distribution la fois) cela va mener un systme dquations non-linaires couples insoluble analytiquement...essayons voir avec un mlange de gaussiennes! Meunier 2008 48 24 Meunier 2008 EM (Expectation-Maximization) =0 quations nonlinaires couples insolubles analytiquement 0= d d j c n c ln j N ( x k j , j ) + j 1 k =1 j =1 j =1 j = nj 1 n j N (x k j , j ) 1n = kj = c n k =1 n k =1 n i N (x k i , i ) i =1 n (avec = -n) 0= c d ln j N (x k j , j ) d j k =1 j =1 n 1 j = kj x k n j k =1 kj : Contribution la classe j de lchantillon k nj : Contribution la classe j de tous les chantillons 0= Meunier 2008 c d n ln j N (x k j , j ) d j k =1 j =1 1n T j = kj (x k j )(x k j ) n j k =1 Mthode point fixe? 49 Mthode point fixe (exemple) x = g(x) xn+1 = g(xn) x = g ( x) = 3 x2 Estimations initiales: x = 4 Tolrance sur x: 0.0005 et Maximum d'itrations:100 Itration 1 2 3 4 5 6 7 8 9 10 11 4. 1.5 -6. -0.375 -1.26316 -0.919355 -1.02762 -0.990876 -1.00305 -0.998984 -1.00034 x 1.5 -6. -0.375 -1.26316 -0.919355 -1.02762 -0.990876 -1.00305 -0.998984 -1.00034 -0.999887 g(x) 2.5 7.5 5.625 0.888158 0.343803 0.108269 0.0367484 0.0121747 0.00406649 0.00135458 0.00045162 |x| Meunier 2008 50 25 Meunier 2008 EM Initialisation: ( j , j j ) = ( j , j j )0 Itrer en deux tapes tape E: kj = c i =1 j N (x k j , j ) i k N (x i , i ) kj : Probabilit que lchantillon k appartienne la classe j nj : Contribution de tous les chantillons la classe j tape M: j = j = j = n 1n kj = nj n k =1 1 nj 1 nj k =1 n k =1 n kj xk j )(x k j ) T (x kj k Meunier 2008 51 EM : Gnralisation Initialisation: =0 Itrer en deux tapes tape E: Estimer les Probabilits kj kj = P ( j | x k , i ) = E [zkj | X, i ] = p ( zkj = 1 | X, i ) E z kj | X, i = 0 p( zkj = 0 | X, i ) + 1 p( z kj = 1 | X, i ) = p ( z kj = 1 | X, i ) [ ] Introduction dune variable cache z. zkj = 1 si xk appartient la classe j, zkj = 0 sinon tape M: Maximiser lEZ[-] de la vraisemblance Q(, i ) = EZ [ln p( X, Z | ) | X, i ] Q(, i ) = 0 i +1 qui a une solution analytique cette fois! Meunier 2008 p ( X, Z | ) = j kj N ( xk | j , j ) k =1 j =1 n c n c z z kj ln p( X , Z | ) = z kj {ln j + ln N ( xk | j , j )} k =1 j =1 Q (, i ) = EZ [ln p( X, Z | ) | X, i ] = = EZ [ z kj | X, i ]{ln j + ln N ( xk | j , j )} k =1 j =1 n c kj tape E 52 26 Meunier 2008 EM Exemple 1 ( i 100) 2 ( i 200) 2 theorie1 := i 1 2 40 100 2 10000 3 4 e 2 40 2 theorie2 := i 10000 e 4 2 2 30 1 2 30 2 histogrammei theorie1i theorie2i 50 0 0 100 i 200 300 Note: On utilise 10000 points (chantillons...

Find millions of documents on Course Hero - Study Guides, Lecture Notes, Reference Materials, Practice Exams and more. Course Hero has millions of course specific materials providing students with the best way to expand their education.

Below is a small sample set of documents:

Neumont - IFT - 6141
Meunier 2007Au menu cette semaine Le problme de la dimensionnalit (Chap. 3 suite et fin) Rduction du nombre dattributs par projection: PCA, LDA, Fisher Par slectionMthodes non-paramtriques (Chap. 4) Introduction: Les histogrammes Fentre
Neumont - IFT - 6141
Au menu cette semaineFonctions discriminantes linaires (Chap. 5) Fonctions discriminantes linaires Machines linaires Comment faire du linaire avec du non-linaire Le perceptron Machine vecteurs de support (SVM)Les rseaux de neurones (Chap. 6
Neumont - IFT - 6141
Au menu cette semaineLes rseaux de neurones (Chap. 6 suite) Paramtres pour la rtropropagation Rseau convolution Cascade et lagage Rseaux rcurrents Elman HopfieldMthodes stochastiques (Chap. 7 dbut) Apprentissage et rseau de Boltzmann Recu
Neumont - IFT - 6141
IFT6141 Reconnaissance des Formes Semaine #5Professeur: Jean MEUNIER meunier@iro.umontreal.ca Bureau: 2163 D.I.R.O.Au menu cette semaine Rseaux Baysiens Donnes squentielles Modle de Markov Modle de Markov cach Filtre prdictif bayesian: Kalman
Neumont - IFT - 6141
IFT6141 Reconnaissance des Formes Semaine #6Professeur: Jean MEUNIER meunier@iro.umontreal.ca Bureau: 2163 D.I.R.O.Au menu cette semaineAttributs non-mtriques (Chap. 8) Arbre de dcision Reconnaissance de formes structurelle Chanes de caractres
Neumont - IFT - 6141
IFT6141 Reconnaissance des Formes Semaine #7Professeur: Jean MEUNIER meunier@iro.umontreal.ca Bureau: 2163 D.I.R.O.Plan du coursContenu: Mthode de Bayes (Chap. 1 et 2), Estimation des paramtres (Chap. 3) Rduction de la dimensionnali
Neumont - IFT - 6141
IFT6141 Reconnaissance des Formes Semaine #8Professeur: Jean MEUNIER meunier@iro.umontreal.ca Bureau: 2163 D.I.R.O.Au menu aujourdhuiReconnaissance Base sur des attributs Base sur un modle de forme Base sur un modle dapparenceMeunier 2008 2
Neumont - IFT - 6141
IFT6141 Reconnaissance des Formes Semaine #9Professeur: Jean MEUNIER meunier@iro.umontreal.ca Bureau: 2163 D.I.R.O.Au menu aujourdhuiApplications de la RdeF Dtection et reconnaissance de visages humains Reconnaissance dactivits humainesMeunier
Neumont - IFT - 6141
IFT6141 Reconnaissance des Formes Semaine #10Professeur: Jean MEUNIER meunier@iro.umontreal.ca Bureau: 2163 D.I.R.O.Au menu aujourdhuiApplications de la RdeF Reconnaissance automatique de la parole (RAP) Introduction Extraction des attributs
MD University College - INSS - 690
Digital ImageryAcquisition and Distribution with a Focus on the Department of DefenseBy Dennis J. Boucher A thesis submitted in partial fulfillment of the requirements for the degree of Masters of Science Management Information Systems Bowie State
MD University College - INSS - 690
IP Telephony: Convergence Among NetworksLarry Jennings Bowie State INSS 690 Term 1 01/02Table of ContentsTable of ContentsTable of Figures Table of Figures Abstract Abstract Regulatory Bodies Internet Engineering Task Force (IETF) International T
MD University College - INSS - 690
The Training Profession The Times of TodayI. II. Introduction The Three Factors Affecting Training a. Intellectual Leadership b. Shortage of Skilled Employees c. Impact of Training An Inward Fundamental View a. Teaching b. Learning Cracks in the Fou
MD University College - INSS - 690
Firewalls:A Security Gate to the Internetby James V. McGovern (jvm64@hotmail.com) University of Maryland - European Division Bowie State UniversityGraduate Program in Management Information Systems INSS 690 - Term I Submitted: October 1, 2001--
MD University College - INSS - 690
WIRELESS TECHNOLOGY WITHIN A MILITARY TREATMENT FACILITYBy Andrew M AlmstromA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of Science in Management Information SystemsBowie S
MD University College - INSS - 690
Implementation of VoIPIS IMPLEMETATION OF VOICE OVER INTERNET PROTOCOL (VOIP) MORE ECONOMICAL FOR BUSINESSES WITH LARGE CALL CENTERS? by Zachary A. BarnesA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements
MD University College - INSS - 690
MD University College - INSS - 690
COMPUTING AT THE SPEED OF LIGHT, IS VON NEUMANN WRONG? by Micah Bell A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science: Management Information Systems Bowie State University 2002iBowie State Univers
MD University College - INSS - 690
Dot Bomb1Dot Bomb:Why Some Worked and Others Did NotBy William M. Bennett A thesis submitted for INSS 690 Professional Seminar for completion of the Masters Of Science Management Information Systems Bowie State University Term 3 2001 2002Dot
MD University College - INSS - 690
WIRELESS TECHNOLOGY AND ALTERNATE FUEL SOURCES TECHNOLOGIES TO ENHANCE OR REPLACE BATTERIES by Raymond J. BergerA Graduate Research Report Submitted for INSS 690 in Parital Fulfillment of the Requirements of the Degree of Master of Science in Manag
MD University College - INSS - 690
Program Management System Analysis1Construction Management Systems and the International Environment: A Program Management Analysis William Bersing Professional Seminar INSS 690 University of Maryland University College Dr. Gae Holladay 23 Februa
MD University College - INSS - 690
INSS 690 Research Paper Term 2002-3 ByVictor Buonamia 10 March 2002Table of Contents I. II. III. IV. V. Abstraction Introduction What is a VPN? Why use a VPN? VPN Implementation Figure 1 VPN VI. Elements of A VPN Connection Figure 2. Components o
MD University College - INSS - 690
TOTAL ASSET VISIBILITY-AIR FORCE (TAV-AF): JUSTIFYING THE CAPITAL INVESTMENT IN RADIO FREQUENCY IDENTIFICATION (RFID) TECHNOLOGY ON U.S. AIR FORCE (USAF) SMART WEAPONS By Donald R. BurnsA Graduate Research Study Submitted for INSS 690 in Partial Fu
MD University College - INSS - 690
COMPARING WEB LANGUAGES IN THEORY AND PRACTICE by Kristofer J. CarlsonA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of Science in Management Information SystemsBowie State Uni
MD University College - INSS - 690
Evaluation of Delawares 1 Running head: EVALUATION OF DELAWARES POST ADOPTION SERVICEAn Evaluation of Delawares Post Adoption Service Program: Services Currently Offered to Adoptive Families Katina L. Clarke Bowie State UniversityEvaluation of De
MD University College - INSS - 690
Broadband Satellite Communication and the Global Information InfrastructureBowie State University INSS 690 Professional Seminar Prof. John G. Meinke Written by: Kevin T. CurryABSTRACT The purpose of this paper is to discuss broadband technology a
MD University College - INSS - 690
USING RADIO FREQUENCY IDENTIFICATION (RFID) TECHNOLOGY IN HUMANS IN THE UNITED STATES FOR TOTAL CONTROL by David B. SmithA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of Science
MD University College - INSS - 690
DATA OVERLOAD: HOW MUCH IS TOO MUCH? by Sean W. Detmers swdetmers@yahoo.comA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of Science in Management Information SystemsBowie Stat
MD University College - INSS - 690
DSL or Cable Modems, Which is Better? By George B. Moses1Table of ContentsI. II.Abstract HistoryIII. What is DSL? IV. V. VI. What is a Cable Modem? Concepts of operation TechnologyVII. What are the usages for DSL and cable modems technolo
MD University College - INSS - 690
Training: Key to Effective Enterprise Resource Planning ImplementationPrepared for: Mr. John Meinke INSS - 690 University of Maryland EuropeBy John Felter johnfelter@yahoo.com March 9, 2002Training: Key to ERP2Training: Key to Effective En
MD University College - INSS - 690
Consolidating Servers: Windows Server20031Consolidating Servers with Microsoft Windows Server 2003 on Intel-based Platforms Bryan Frame Bowie State UniversityConsolidating Servers: Windows Server20032Abstract Server consolidation method
MD University College - INSS - 690
E-Medicine: CHCS II Trends and CapabilitiesGeralyn Essick INSS 690A b s t r a c t: Computerized patient records (CPR) have been around for many years. Their value as a medical information system has been substantiated in several documented resear
MD University College - INSS - 690
VOICE OVER INTERNET PROTOCOL: SECURE OR NOT RECOMMENDATIONS TO THE BUSINESS AND PRIVATE SECTOR by Ronald P. Gagner, Jr.A Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of Science i
MD University College - INSS - 690
PARALLEL COMPUTING FOR MAGNETIC RESONANCE IMAGING (MRI) By Victor Q. GarciaA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of Science in Management Information SystemsBowie Stat
MD University College - INSS - 690
E-Medicine: CHCS II Trends and CapabilitiesGeralyn Essick INSS 690A b s t r a c t: Computerized patient records (CPR) have been around for many years. Their value as a medical information system has been substantiated in several documented resear
MD University College - INSS - 690
Computer Security: The Attack and Defense(By: Bryant D. Glando) INSS 690, Dr. John Meinke THESIS No matter what the system is, there is someone who will want to attack it and discover the secrets that lie within. Therefore to prevent the loss of val
MD University College - INSS - 690
AN EXECUTIVE DECISION: WINDOWS OS VS. LINUX OS VS. HYBRID by J. Renee GrayA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of Science in Management Information SystemsBowie State
MD University College - INSS - 690
OVERCOMING THE TECHNICAL CHALLENGES OF PROVIDING DISTANCE EDUCATION TO DEVELOPING COUNTRIES By Reginald J. Haines Reggie_Haines@Hotmail.comA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements for the Degree o
MD University College - INSS - 690
Deductive Database 1 Running head: BOTTOM-UP IS THE TRENDDatalog Bottom-up is the Trend in the Deductive Database Evaluation StrategyYurek K. Hinz INSS 690University of Maryland, 2002Deductive Database 2 Table of contentsAbstract Introducti
MD University College - INSS - 690
INSS 690 PROFESSIONAL SEMINARUNIVERSITY OF MARYLAND Education Center, SHAPE Technology Enhanced Format, Term 4 - 2001-2002 COURSE DESCRIPTION: (3 semester hours) Prerequisites: Advancement to candidacy in the MIS program and successful completion of
MD University College - INSS - 690
BEST PRACTICES IN INFORMATION ASSURANCE AND INFORMATION TECHNOLOGY NETWORKING IN ORGANIZATIONS THAT HAVE TWO DEPARTMENTS by David Johnson, Jr.A Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree
MD University College - INSS - 690
Wireless Local Area Network, 1Wireless Advantages versus DisadvantagesWireless Local Area Network (WLAN) Advantages vs. DisadvantagesMike M. KhayatINNS 690, Professional Seminar Mr. John Meinke March 12, 2002Wireless Local Area Network. 2
MD University College - INSS - 690
Quantum Computing1Running head: QUANTUM COMPUTINGQuantum Computing: A Primer Jimmy Kilduff Bowie State University in association with University of Maryland, University CollegeQuantum Computing Abstract Quantum computing is a fascinating and
MD University College - INSS - 690
HIPAA COMPLIANCE AND ELECTRONIC MEDICAL RECORDS: ARE BOTH POSSIBLE? by Mark A. KnitzA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of Science in Management Information SystemsB
MD University College - INSS - 690
IT IN HEALTHCARE: THE EMRIT APPLICATIONS IN HEALTHCARE: THE ELECTRONIC MEDICAL RECORDby Rodney L. Koeller A thesis submitted in partial fulfillment of the requirements for the degree of Masters of Science in Management Information Systems Univers
MD University College - INSS - 690
USING INTERNET BROWSERS TO SAFELY ACCESS INFORMATION AND SERVICES ON THE WORLD WIDE WEB by Edward D Lavieri Jr. elavieri@acm.orgA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of
MD University College - INSS - 690
Biometrics for Secure Identity Verification: Trends and DevelopmentsBy Joseph W. LewisA Thesis Presented in Partial Fulfillment of the Requirements for INSS 690 Professional SeminarMaster of Science in Management Information SystemsUniversity
MD University College - INSS - 690
SERVICE-ORIENTED ARCHITECTURE: ACHIEVING INTEROPERABILITY IN GOVERNMENT WITH XML AND WEB SERVICES by Joaquin A. Marquez Jr. jay@marquez-online.comA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements for the D
MD University College - INSS - 690
KNOWLEDGE MANAGEMENT THE AIR FORCE WAY AHEAD by Joseph E. MartinA Graduate Research Report Submitted for INSS 690 In Partial Fulfillment of the Requirements of the Degree of Master of Science in Management Information SystemsBowie State Universi
MD University College - INSS - 690
MUNICIPAL WIRELESS METROPOLITAN AREA NETWORKSby Jesse McGuire mcguirejesse@hotmail.comA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements for the Degree of Master of Science in Management Information Syste
MD University College - INSS - 690
UNIVERSITY OF MARYLAND UNIVERSITY COLLEGE INSS690 PROFESSIONAL SEMINARTHE EXTENSIBLE MARKUP LANGUAGE (XML): THE TREND IN WEB-BASED APPLICATIONSWANDA Y. McLEAN 10 MARCH 2002TABLE OF CONTENTSI. Abstract II. In the Beginning a. SMGL (Standard Ge
MD University College - INSS - 690
MD University College - INSS - 690
University of Maryland INSS 690 John Mitchell Date Due: October 7, 2001 Date Submitted: October 7, 2001Table of ContentsIntroduction. 1 Preliminary Thesis . 2 Purpose Of Project . 3 What is Linux?. 3 The Linux System . 4 A Brief History of Linux.
MD University College - INSS - 690
Impact of Standards on Wireless Computing 1Impact of Standards on Wireless Computing Chris A. Payton INSS 690 Mr. J. Meinke Bowie State UniversityImpact of Standards on Wireless Computing 2 Table of ContentsTitle page Table of Contents Abstract
MD University College - INSS - 690
CORPORATE ISSUES SURROUNDING THE TRANSITIONING TO INTERNET PROTOCOL VERSION 6 (IPV6) FROM INTERNET PROTOCOL VERSION 4 (IPV4) by Christopher T. RatliffA Graduate Research Report Submitted for INSS 690 In Partial Fulfillment of the Requirements of th
MD University College - INSS - 690
PERSONAL WIRELESS COMMUNICATIONS by Daniel RodrinA Graduate Research Report Submitted for INSS 690 In Partial Fulfillment of the Requirements of the Degree of Master of Science in Management Information SystemsBowie State University Maryland in E
MD University College - INSS - 690
MD University College - INSS - 690
Home Networking A Comparison of Modern TechnologiesKevin W. Spurling University of Maryland European Division INSS 690 TERM I, 2001/2002 September 15, 2001Table of ContentsAbstract .. 1 Introduction . 2 Wired Home Network Technologies.. 4 Ethern
MD University College - INSS - 690
DEPARTMENT OF DEFENSES USE OF RFID TECHNOLOGY FOR IN-TRANSIT VISIBILITY, ASSET VISIBILITY, AND ITS RETURN ON INVESTMENT FINAL DRAFT by Anthony Tony Stoneking tony_ret@yahoo.comA Graduate Research Proposal Submitted for INSS 690 in Partial Fulfillme
MD University College - INSS - 690
Information Technology (IT) Positive Impact on Public School Curriculums for K-12 in the Next 5 years.INSS-690 MIS UMUC 6 OCT 2001 Charles E. TilleyTable of Contents1. Introduction 2. Serious Problems Facing the American School System 3. Learni
MD University College - INSS - 690
PERVASIVE COMPUTING:noitamrofnI ot ytilibisseccA fo noitulovE ehTBy Corwin Chris Walks University of Maryland, European Division Bowie State University INSS 690 Professor, Dr. John Meinke1 Table of Contents Abstract Introduction General Overview
MD University College - INSS - 690
TRENDS IN BIOMOLECULAR COMPUTING: RECOMMENDATIONS TO THE GOVERNMENT SECTOR AND IT PROFESSIONALS by Robert A. WalzA Graduate Research Report Submitted for INSS 690 in Partial Fulfillment of the Requirements of the Degree of Master of Science in Mana