模式识别与人工智能之十_第1页
模式识别与人工智能之十_第2页
模式识别与人工智能之十_第3页
模式识别与人工智能之十_第4页
模式识别与人工智能之十_第5页
已阅读5页,还剩60页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

1、Pattern Recognition if it is gender, then Bernoulli distribution,20,EM estimation,Basic setting in EM,21,EM estimation,X is a set of data points: observed data is a parameter vector. EM is a method to find ML where Calculating P(X | ) directly is hard. Calculating P(X,Y|) is much simpler, where Y is

2、 “hidden” data (or “missing” data).,Recall: maximum likelihood,22,EM estimation,Latent variables or missing data,23,EM estimation,Incomplete Data,Complete Data,Missing data,Complete Data Likelihood,EM estimation,Complete Data Likelihood,EM estimation,Log likelihood,In many cases, we cannot find the

3、solution directly. An alternative is to find a sequence:,s.t.,Complete Data Likelihood,EM estimation,Define: : the probability of y : function of y, namely g(y),0,Ln(x) is strictly concave on (0, ),Complete Data Likelihood,EM estimation,We need to choose the maximum lower bound,How to choose Q(y)?,I

4、f is given, Q(y) and P(xi, y) will determine the value of l(),Step1: find Q(y) (Expectation) Step2: according to Q(y), maximize l() by adjusting (Maximization),EM estimation,Expectation Step: how to choose Q(y),According to the Jensen inequality, when g(y) is a constant, the lower bound is maximum.,

5、We know that ( is the probability of y, so,It is easily obtained:,EM estimation,Expectation Step: how to choose Q(y),EM estimation,Maximization Step: maximize l() by adjusting ,EM algorithm is an iterative procedure for maximizing . Assume that after the Iteration the current estimation for is given

6、 by . Since the objective is to maximize , we wish to compute an update estimate such that: Equivalently, we want to maximize the difference,EM estimation,How to guarantee the likelihood function L() is increased at each step?,Denote the hidden random vector by , the total probability is,EM estimati

7、on,How to guarantee the likelihood function L() is increased at each step?,EM estimation,How to guarantee the likelihood function L() is increased at each step?,EM estimation,Our objective is to choose a value of so that is maximized.,Gaussian Mixture Model,35,Single Gaussian,Solution: MLE by maximi

8、zing,Gaussian Mixture Model,36,Multiple Gaussians,Which component does each point belong to? Solution: ?,EM for GMM,37,The aim is to estimate the unknown parameters representing between the Gaussians and the means and covariances of each:,let x=(x1,x2,xn) be a sample of n independent observations fr

9、om a mixture of multiple normal distributions of dimension d, and let z=(z1,z2,zn) be the latent variables that determine the component from which the observation originates.,where,with,EM for GMM,38,The incomplete-data likelihood function is:,The complete-data likelihood function is:,where is an in

10、dicator function and f is the probability density function of a multivariate normal.,EM for GMM,39,This may be rewritten in log likelihood function form:,E-step:,EM for GMM,40,M-step:,Using MLE to find solution of,EM for GMM,41,M-step:,Eg:,SMM:,GMM:,EM for GMM,42,M-step:,Eg:,SMM,GMM,EM for GMM,43,M-

11、step:,Eg:,S.t.,Difference between K-means and EM,44,K-means,Classifier (K-Means),Classification Results x1C(x1) x2C(x2) xiC(xi) ,Cluster Parameters m1 for C1 m2 for C2 mk for Ck,Difference between K-means and EM,45,K-means,x1=r1, g1, b1 x2=r2, g2, b2 xi=ri, gi, bi ,Classification Results (1) C(x1),

12、C(x2), , C(xi),Initial Guess of Cluster Parameters m1 , m2 , , mk,Input (Known),Output (Unknown),Cluster Parameters(1) m1 , m2 , , mk,Classification Results (2) C(x1), C(x2), , C(xi),Cluster Parameters(2) m1 , m2 , , mk,Classification Results (ic) C(x1), C(x2), , C(xi),Cluster Parameters(ic) m1 , m2

13、 , , mk,Difference between K-means and EM,46,K-means,Boot Step: Initialize K clusters: C1, , CK Each Cluster is represented by its mean mj Iteration Step: Estimate the cluster of each data Re-estimate the cluster parameters,Difference between K-means and EM,47,Kmeans -EM,Boot Step: Initialize K clus

14、ters: C1, , CK Iteration Step: Estimate the cluster of each data Re-estimate the cluster parameters,(j, j) and P(Cj) for each cluster j.,For each cluster j,Expectation,Maximization,Difference between K-means and EM,48,Classifier (EM),Classification Results p(C1|x1) p(Cj|x2) p(Cj|xi) ,Cluster Paramet

15、ers (1,1),p(C1) for C1 (2,2),p(C2) for C2 (k,k),p(Ck) for Ck,EM,Difference between K-means and EM,49,EM,x1=r1, g1, b1 x2=r2, g2, b2 xi=ri, gi, bi ,Cluster Parameters (1,1), p(C1) for C1 (2,2), p(C2) for C2 (k,k), p(Ck) for Ck,Input (Known),Output (Unknown),Classification Results p(C1|x1) p(Cj|x2) p(

16、Cj|xi) ,Difference between K-means and EM,50,EM(E-step),x1=r1, g1, b1 x2=r2, g2, b2 xi=ri, gi, bi ,Cluster Parameters (1,1), p(C1) for C1 (2,2), p(C2) for C2 (k,k), p(Ck) for Ck,Input (Known),Output,Classification Results p(C1|x1) p(Cj|x2) p(Cj|xi) ,Estimation,+,Difference between K-means and EM,51,

17、EM (M-step),x1=r1, g1, b1 x2=r2, g2, b2 xi=ri, gi, bi ,Cluster Parameters (1,1), p(C1) for C1 (2,2), p(C2) for C2 (k,k), p(Ck) for Ck,Input (Known),Input (Estimation),Output,+,Classification Results p(C1|x1) p(Cj|x2) p(Cj|xi) ,Difference between K-means and EM,52,EM,Boot Step: Initialize K clusters:

18、 C1, , CK Iteration Step: Expectation Step Maximization Step,(j, j) and P(Cj) for each cluster j.,53,Case Study (1),We have a data set with 1000 2-D data points (three clusters, k=3), how can we cluster it?,54,Case Study (2),Step 1: Initialize the partition Randomly select 3 points as initial class

19、centers; Partition the data with the 3 mean points Step 2: E step Step 3: EM algorithm do M Step E Step until convergence,55,Step 1: Initialize the partition,56,57,58,Image Segmentation using EM,Step 1: Feature Extraction Step 2: Image Segmentation using EM,Symbols,The feature vector for pixel i is called xi. There are going to be K seg

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论