Statistical Learning Dong Liu Dept EEIS USTC Chapter
- Slides: 61
Statistical Learning Dong Liu Dept. EEIS, USTC
Chapter 9. Probabilistic Graphical Model 1. 2. 3. 4. 5. Generative and Bayesian Naïve Bayesian network Markov random field Belief propagation 2020/10/7 Chap 9. Probabilistic Graphical Model 2
Generative vs. Discriminative • In generative methods to learning, estimate • So we “reconstruct” the joint distribution of • We are able to “generate” new data • In discriminative methods to learning, estimate simpler • • or even We only model the relationship between For example: Linear regression is Generalized logistic regression is • For classification problems, generative is usually more difficult than discriminative, e. g. writing versus reading 2020/10/7 Chap 9. Probabilistic Graphical Model 3
Generative and Bayesian • • Naïve Bayes assumes conditional independence • Probabilistic graphical models simplify the joint distribution by factorization 2020/10/7 Chap 9. Probabilistic Graphical Model 4
Chapter 9. Probabilistic Graphical Model 1. 2. 3. 4. 5. Generative and Bayesian Naïve Bayesian network Markov random field Belief propagation 2020/10/7 Chap 9. Probabilistic Graphical Model 5
Naïve Bayes • In the Bayesian approach, we need • Naïve Bayes (NB) estimates directly from data, e. g. calculates empirical distributions (histograms) • For high-dimensional data, NB uses a simple (conditional) independence assumption 2020/10/7 Chap 9. Probabilistic Graphical Model 6
Example 1/2 • Dataset: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 x 1 1 1 2 2 2 3 3 3 x 2 s m m s s s m m l l l m m l l y -1 -1 1 1 -1 -1 -1 1 1 1 -1 • Empirical distributions: x 1=1 x 1=2 x 1=3 y=1 2/9 3/9 4/9 y=-1 3/6 2/6 1/6 2020/10/7 x 2=s x 2=m x 2=l y=1 1/9 4/9 y=-1 3/6 2/6 1/6 Chap 9. Probabilistic Graphical Model y=1 y=-1 9/15 6/15 7
Example 2/2 • Given • We can calculate • Empirical distributions: x 1=1 x 1=2 x 1=3 y=1 2/9 3/9 4/9 y=-1 3/6 2/6 1/6 2020/10/7 x 2=s x 2=m x 2=l y=1 1/9 4/9 y=-1 3/6 2/6 1/6 Chap 9. Probabilistic Graphical Model y=1 y=-1 9/15 6/15 8
Example: Dealing with missing values 1/2 • Dataset: 1 2 3 4 5 6 7 8 9 10 11 x 1 1 1 2 2 2 3 x 2 s m m s s s m m l l l y -1 -1 +1 +1 -1 -1 -1 +1 +1 • Empirical distributions: x 1=1 x 1=2 x 1=3 y=+1 2/6 3/6 1/6 y=-1 3/5 2/5 0/5 2020/10/7 x 2=s x 2=m x 2=l y=+1 1/6 2/6 3/6 y=-1 3/5 2/5 0/5 Chap 9. Probabilistic Graphical Model y=+1 y=-1 6/11 5/11 9
Example: Dealing with missing values 2/2 • Laplace smoothing is to avoid 0 by smooth the empirical distributions • Smoothed empirical distributions ( ): x 1=1 x 1=2 x 1=3 y=+1 3/9 4/9 2/9 y=-1 4/8 3/8 1/8 2020/10/7 x 2=s x 2=m x 2=l y=+1 2/9 3/9 4/9 y=-1 4/8 3/8 1/8 Chap 9. Probabilistic Graphical Model y=+1 y=-1 7/13 6/13 10
Why Laplace smoothing? 1/2 • Consider a categorical distribution is the indicator function • Given a set of samples, how to estimate the parameter • The likelihood function is • Using ML estimation, we have 2020/10/7 Chap 9. Probabilistic Graphical Model 11
Why Laplace smoothing? 2/2 • Using the Bayesian approach, we set prior Dirichlet distribution • Then the posterior is • Using the MAP estimation • This interprets Laplace smoothing if 2020/10/7 Chap 9. Probabilistic Graphical Model 12
Notes • Naïve Bayes is especially suitable for discrete variables, so it is more used for classification 2020/10/7 Chap 9. Probabilistic Graphical Model 13
Chapter 9. Probabilistic Graphical Model 1. 2. 3. 4. 5. Generative and Bayesian Naïve Bayesian network Markov random field Belief propagation 2020/10/7 Chap 9. Probabilistic Graphical Model 14
Bayesian networks • Naïve Bayes may be too simple, consider conditional distributions instead • Draw a graph to help understand • This is a directed acyclic graph (DAG) • Such graphs represent Bayesian networks, or directed graphical models 2020/10/7 Chap 9. Probabilistic Graphical Model 15
Why Bayesian networks? • To simplify the model, we assume partial conditionally independence, which can be captured by Bayesian networks • Example 2020/10/7 Chap 9. Probabilistic Graphical Model 16
Example: linear regression • • Given a set of inputs Plate, stands for recurrence 2020/10/7 Chap 9. Probabilistic Graphical Model 17
Example: linear regression • Consider • And our observations are and Point, stands for non-random variables 2020/10/7 Chap 9. Probabilistic Graphical Model Filled, stands for observed 18
Example: linear regression • Given new input , use the model to estimate Predictive distribution 2020/10/7 Chap 9. Probabilistic Graphical Model 19
Conditional independence (CI) • 2020/10/7 Chap 9. Probabilistic Graphical Model 20
CI case 1 • Common parent: unknown -> unblock, known -> block 2020/10/7 Chap 9. Probabilistic Graphical Model 21
CI case 2 • In-the-chain: unknown -> unblock, known -> block 2020/10/7 Chap 9. Probabilistic Graphical Model 22
CI case 3 • Common child: unknown -> block, known -> unblock 2020/10/7 Chap 9. Probabilistic Graphical Model 23
Example: Gold after Door The door I guess The door with gold ? The door MC opens 2020/10/7 Chap 9. Probabilistic Graphical Model 24
Example: Gold after Door • The key is to build the graphical model The door I guess The door with gold The door I guess The door MC opens 2020/10/7 The door with gold The door MC opens Chap 9. Probabilistic Graphical Model 25
D-separation • 2020/10/7 Chap 9. Probabilistic Graphical Model 26
Example of D-separation 2020/10/7 Chap 9. Probabilistic Graphical Model 27
Example of D-separation • 2020/10/7 Chap 9. Probabilistic Graphical Model 28
Markov blanket • Use CI, we can prove All other random variables where neighbors include: • Parents • Children’s other parents • Such a neighborhood is called Markov blanket 2020/10/7 Chap 9. Probabilistic Graphical Model 29
Chapter 9. Probabilistic Graphical Model 1. 2. 3. 4. 5. Generative and Bayesian Naïve Bayesian network Markov random field Belief propagation 2020/10/7 Chap 9. Probabilistic Graphical Model 30
Markov random fields (MRF) • We know some random variables are correlated, but don’t know exactly how they are dependent • Markov random fields, or undirected graphical models, can help to represent the correlations • Unlike Bayesian network, Markov random fields don’t have direct conditional distributions 2020/10/7 Chap 9. Probabilistic Graphical Model 31
Conditional independence and Markov blanket • 2020/10/7 Chap 9. Probabilistic Graphical Model 32
Define distributions on MRF • 2020/10/7 Chap 9. Probabilistic Graphical Model 33
Clique and maximal clique • Cliques with 2 nodes: • 1&2, 1&3, 2&4, 3&4 • Cliques with 3 nodes: (also maximal cliques): • 1&2&3, 2&3&4 • No bigger clique 2020/10/7 Chap 9. Probabilistic Graphical Model 34
Define joint distribution • The joint distribution can be defined only on cliques (or equivalently on maximal cliques) Potential functions • In practice, we often use the exponential family Energy functions 2020/10/7 Chap 9. Probabilistic Graphical Model 35
Ising model • 2020/10/7 Chap 9. Probabilistic Graphical Model 36
S. Geman and D. Geman, “Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images”, IEEE Trans. PAMI, vol. 6: 721 -741. 1984 Example: image denoising • Assume a binary image and we observed a noisy image Original and observed images Note: the images are binary, we just display 1 as blue and -1 as yellow 2020/10/7 Chap 9. Probabilistic Graphical Model 37
Example: MRF • 2020/10/7 Chap 9. Probabilistic Graphical Model 38
Example: clique and energy functions • More energy, less probability 2020/10/7 Chap 9. Probabilistic Graphical Model 39
Example: MAP, ICM • 2020/10/7 Chap 9. Probabilistic Graphical Model 40
Example: results Noisy image 2020/10/7 Restored by ICM Chap 9. Probabilistic Graphical Model 41
BN vs. MRF Bayesian networks Markov Random Fields • Directed graphical models • Modeling conditional probabilities • CI: D-separation • Markov blanket: parents, children’s other parents • Undirected graphical models • Modeling correlations, need to define probability functions • CI: simple connection • Markov blanket: directly connected nodes 2020/10/7 Chap 9. Probabilistic Graphical Model 42
Convert BN to MRF (1) • Directly dependent -> correlated 2020/10/7 Chap 9. Probabilistic Graphical Model 43
Convert BN to MRF (2) • For each node, different parents are also correlated 2020/10/7 Chap 9. Probabilistic Graphical Model 44
Convert BN to MRF - Moralization • After conversion, the Markov blanket of each node keeps unchanged; this conversion is known as moralization 2020/10/7 Chap 9. Probabilistic Graphical Model 45
Convert BN to MRF • Often cause information loss 2020/10/7 Chap 9. Probabilistic Graphical Model 46
Convert MRF to BN? • One MRF can correspond to several BNs • But MRF may not correspond to BN 2020/10/7 Chap 9. Probabilistic Graphical Model 47
Chapter 9. Probabilistic Graphical Model 1. 2. 3. 4. 5. Generative and Bayesian Naïve Bayesian network Markov random field Belief propagation 2020/10/7 Chap 9. Probabilistic Graphical Model 48
Inference in graphical models • Inference of unobserved variables • Given a set of observed variables, graphical models can help to infer the posterior distributions of target variables • Need integral to eliminate “neither observed nor target” variables • Inference of conditional distributions • Inference of graph structure 2020/10/7 Chap 9. Probabilistic Graphical Model 49
Toy example • 2020/10/7 Chap 9. Probabilistic Graphical Model 50
Factor graph • The joint distribution is product of factors • Each factor is connected to related variables • Each variable is connected to related factors 2020/10/7 Chap 9. Probabilistic Graphical Model 51
Sum-product algorithm (1) • Two kinds of messages: variable to factor, factor to variable • Message generation 2020/10/7 Chap 9. Probabilistic Graphical Model 52
Sum-product algorithm (2) • Message gathering and forwarding • Finally 2020/10/7 Chap 9. Probabilistic Graphical Model 53
Example of sum-product algorithm • Message passing rules: 2020/10/7 Chap 9. Probabilistic Graphical Model 54
Example of sum-product algorithm • 2020/10/7 Chap 9. Probabilistic Graphical Model 55
Toy example • 2020/10/7 Face Computer Plant Lab 0. 075 0. 15 0. 025 Dorm 0. 15 0. 1 0 Campus 0. 25 0 0. 25 Ellipse Rec Irregular Face 0. 8 0. 05 0. 15 Computer 0. 05 0. 9 0. 05 Plant 0. 2 0. 6 Yellow Black Green Face 0. 8 0. 15 0. 05 Computer 0. 2 0. 5 0. 3 Plant 0. 1 0. 45 Chap 9. Probabilistic Graphical Model 56
Toy example: Given Color • Face Computer Plant Lab 0. 075 0. 15 0. 025 Dorm 0. 15 0. 1 0 Campus 0. 25 0 0. 25 Ellipse Rec Irregular Face 0. 8 0. 05 0. 15 Computer 0. 05 0. 9 0. 05 Plant 0. 2 0. 6 Black 2020/10/7 Face 0. 15 Computer 0. 5 Plant 0. 45 Chap 9. Probabilistic Graphical Model 57
Toy example: Given Location and Color • Face Computer Plant 0. 075 0. 15 0. 025 Ellipse Rec Irregular Face 0. 8 0. 05 0. 15 Computer 0. 05 0. 9 0. 05 Plant 0. 2 0. 6 Lab Black 2020/10/7 Face 0. 15 Computer 0. 5 Plant 0. 45 Chap 9. Probabilistic Graphical Model 58
Remarks on sum-product algorithm • For continuous variables the sum-product algorithm is still valid, replacing the probability distribution with PDF • If the factor graph is a tree (i. e. no loop), the sum-product algorithm is exact • If the factor graph contains loops, loopy belief propagation can be used • Need to decide a message passing schedule • Not always converge • In many cases, we retreat to approximate inference 2020/10/7 Chap 9. Probabilistic Graphical Model 59
Chapter summary Dictionary Toolbox • Bayesian network • Factor graph • Graphical model • Markov blanket • Markov random field (MRF) • Belief propagation, aka sumproduct algorithm, aka message passing, loopy ~ • D-separation • Iterative conditional modes (ICM) • Laplace smoothing • Moralization • Naïve Bayes 2020/10/7 Chap 5. Non-Parametric Supervised Learning 60
Home exercises 2020/10/7 Chap 1. Linear Regression 61
- Dong liu ustc
- Dong liu ustc
- Dong liu ustc
- Ding dong, ding dong christmas bells are ringing
- Alex liu cecilia liu
- Líu líu lo lo ta ca hát say sưa
- Ustc repeg
- Lunc and ustc
- Mail.ustc.edu.n
- Ustc search
- Cuadro comparativo entre e-learning b-learning y m-learning
- Chapter 9 thinking and language
- Statistical mechanics of deep learning
- Dept nmr spectroscopy
- Florida dept of agriculture and consumer services
- Finance department organizational chart
- Worcester ma building department
- Dept. name of organization (of affiliation)
- Mn dept of education
- Dept of finance and administration
- Dept. name of organization (of affiliation)
- Ohio employment first
- Affiliation poster
- Vaginal dept
- Gome dept
- Gome dept
- Nyttofunktion
- Gome dept
- Hoe dept
- Battalion chief interview questions
- Oviposition
- Dept of education
- Florida dept of agriculture and consumer services
- Florida dept of agriculture and consumer services
- Dept a
- Central islip fire department
- Dss rowan county
- Dept of education
- Tabella nmr
- Pt dept logistik
- Nys dept of homeland security
- La dept of revenue
- Geaux biz login
- Oxford dept of continuing education
- Nebraska dept of agriculture
- Iit
- Dept ind onegov
- Albany county dept of social services
- What is the basic idea behind the bow wow theory
- Peter dong
- Dong nao jin
- Dòng thông tin trong logistics
- Dong quai nedir
- Jae dong noh
- Sơ đồ mạch điện chiều dòng điện
- Erika dong
- Changyu dong
- Dong a university
- Hawmin
- Disdoard
- Zhen dong berkeley
- Lan dong