55
Announcements Spring Courses Somewhat Relevant to Machine Learning 5314: Algorithms for molecular bio (who’s teaching?) 5446: Chaotic dynamics (Bradley) 5454: Algorithms (Frangillo) 5502: Data mining (Lv) 5753: Computer performance modeling (Grunwald) 7000-006: Geospatial data analysis (Caleb Phillips) 7000-008: Human-robot interaction (Dan Szafir) 7000-009: Data analytics: Systems algorithms and applications (Lv) 7000-021: Bioinformatics (Robin Dowell-Dean) Homework Importance sampling via likelihood weighting

Spring 2014 Course

  • Upload
    suchin

  • View
    34

  • Download
    0

Embed Size (px)

DESCRIPTION

Spring 2014 Course. PSYC 5835 - Thinking Proseminar – Matt Jones - PowerPoint PPT Presentation

Citation preview

Page 1: Spring 2014 Course

Announcements

Spring Courses Somewhat Relevant to Machine Learning  5314: Algorithms for molecular bio (who’s teaching?)  5446: Chaotic dynamics (Bradley)  5454: Algorithms (Frangillo)  5502: Data mining (Lv)  5753: Computer performance modeling (Grunwald)  7000-006: Geospatial data analysis (Caleb Phillips)  7000-008: Human-robot interaction (Dan Szafir)  7000-009: Data analytics: Systems algorithms and applications (Lv)  7000-021: Bioinformatics (Robin Dowell-Dean)

Homework

Importance sampling vialikelihood weighting

Page 2: Spring 2014 Course

Learning In Bayesian Networks:Missing Data And Hidden Variables

Page 3: Spring 2014 Course

Missing Vs. Hidden Variables

Missing

often known but absent for certain data points

missing at random or missing based on value  e.g., netflix ratings

Hidden

never observed but essential for predicting visible variables  e.g., human memory state

a.k.a. latent variables

Page 4: Spring 2014 Course

Quiz

“Semisupervised learning” concerns learning where additional input examples are available, but labels are not. According to the model below, will partial data (either X or Y) inform the model parameters?

X known?Y known?

X Y

θy|xθx θy|~x

XX Y

Page 5: Spring 2014 Course

X Y

θy|xθx θy|~x

X Y

Page 6: Spring 2014 Course

Missing Data: Exact Inference In Bayes Net

Y: observed variablesZ: unobserved variables

How do we do parameter updates for θi in this case?

If Xi and Pai are observed, then situation is straightforward (e.g., like single-coin toss case).

If Xi or any Pai are missing, need to marginalize over Z

E.g., Xi ~ Categorical(θij)

Note: posterior is a Dirichlet mixture

Dirichlet

# values of Xi

Specific value of Xi

Dirichlet

X = {Y,Z}

parameter vector for Xiwith parent configuration j

Page 7: Spring 2014 Course

Missing Data: Gibbs Sampling

Given a set of observed incomplete data, D = {y1, ..., yN}

1. Fill in arbitrary values for unobserved variables for each case Dc

2. For each unobserved variable zi in case n, sample:

3. evaluate posterior density on complete data Dc’

4. repeat steps 2 and 3, and compute mean of posterior density

Page 8: Spring 2014 Course

Missing Data: Gaussian Approximation

Approximateas a multivariate Gaussian.

 Appropriate if sample size |D| is large, which is also the case when Monte Carlo is inefficient

1. find the MAP configuration by maximizing g(.)

2. approximate using 2nd degree Taylor polynomial

3. leads to approximate result that is Gaussian

~

negative Hessian of g(.) eval at

~

Page 9: Spring 2014 Course

Missing Data: Further Approximations

As the data sample size increases, Gaussian peak becomes sharper, so can make predictions

based on the MAP configuration can ignore priors (diminishing importance) -> max likelihood

How to do ML estimation Expectation Maximization Gradient methods

Page 10: Spring 2014 Course

Expectation Maximization

Scheme for picking values of missing data and hidden variables that maximizes data likelihood

E.g., population of Laughing Goat

baby stroller, diapers, lycra pants

backpack, saggy pants

baby stroller, diapers

backpack, computer, saggy pants

diapers, lycra

computer, saggy pants

backpack, saggy pants

Page 11: Spring 2014 Course

Expectation Maximization Formally

V: visible variables

H: hidden variables

θ: model parameters Model

P(V,H|θ) Goal

Learn model parameters θ in the absence of H Approach

Find θ that maximizes P(V|θ)

Page 12: Spring 2014 Course

EM Algorithm (Barber, Chapter 11)

Page 13: Spring 2014 Course

EM Algorithm

Guaranteed to find local optimum of θ Sketch of proof

Bound on marginal likelihood

  equality only when q(h|v)=p(h|v,θ)

E-step: for fixed θ, find q(h|v) that maximizes RHS

M-step: for fixed q, find θ that maximizes RHS

if each step maximizes RHS, it’s also improving LHS  technically, it’s not lowering LHS

Page 14: Spring 2014 Course

Barber Example

Contours are of the lower boundNote alternating steps along θ and q axes

note that steps are not gradient steps and can be largeChoice of initial θ determines local likelihood optimum

Page 15: Spring 2014 Course

Clustering: K-Means Vs. EM

K means

1. choose some initial values of μk

2. assign each data point to the closest cluster

3. recalculate the μk to be the means of the set of points assigned to cluster k

4. iterate to step 2

Page 16: Spring 2014 Course

K-means Clustering

From C. Bishop, Pattern Recognition and Machine Learning

Page 17: Spring 2014 Course

K-means Clustering

Page 18: Spring 2014 Course

K-means Clustering

Page 19: Spring 2014 Course

K-means Clustering

Page 20: Spring 2014 Course

Clustering: K-Means Vs. EM

K means

1. choose some initial values of μk

2. assign each data point to the closest cluster

3. recalculate the μk to be the means of the set of points assigned to cluster k

4. iterate to step 2

Page 21: Spring 2014 Course

Clustering: K-Means Vs. EM

EM

1. choose some initial values of μk

2. probabilistically assign each data point to clusters1. P(Z=k|μ)

3. recalculate the μk to be the weighted mean of the set of points

1. weight by P(Z=k|μ)

4. iterate to step 2

Page 22: Spring 2014 Course

EM for Gaussian Mixtures

Page 23: Spring 2014 Course

EM for Gaussian Mixtures

Page 24: Spring 2014 Course

EM for Gaussian Mixtures

Page 25: Spring 2014 Course

Variational Bayes

Generalization of EM

also deals with missing data and hidden variablesProduces posterior on parameters

not just ML solutionBasic (0th order) idea

do EM to obtain estimates of p(θ) rather than θ directly

Page 26: Spring 2014 Course

Variational Bayes

Assume factorized approximation of joint hidden and parameter posterior:

Find marginals that make this approximation as close as possible.

Advantage?

Bayesian Occam’s razor: vaguely specified parameter is a simpler model -> reduces overfitting

Page 27: Spring 2014 Course

Gradient Methods

Useful for continuous parameters θMake small incremental steps to maximize the likelihood

Gradient update:

swap

Page 28: Spring 2014 Course

All Learning Methods Apply ToArbitrary Local Distribution Functions

Local distribution function performs either Probabilistic classification (discrete RVs) Probabilistic regression (continuous RVs)

Complete flexibility in specifying local distribution fn Analytical function (e.g., homework 5) Look up table Logistic regression Neural net Etc.

LOCAL DISTRIBUTION FUNCTION

Page 29: Spring 2014 Course

Summary Of Learning Section

Given model structure and probabilities,inferring latent variables

Given model structure,learning model probabilities Complete data Missing data

Learning model structure

Page 30: Spring 2014 Course

Learning Model Structure

Page 31: Spring 2014 Course

Learning Structure and Parameters

The principleTreat network structure, Sh, as a discrete RV

Calculate structure posterior

Integrate over uncertainty in structure to predict

The practiceComputing marginal likelihood, p(D|Sh), can be difficult.

Learning structure can be impractical due to the large number of hypotheses (more than exponential in # of nodes)

Page 32: Spring 2014 Course

source: www.bayesnets.com

Page 33: Spring 2014 Course

Approach to Structure Learning

model selection  find a good model, and treat it as the correct model

selective model averaging  select a manageable number of candidate models and pretend that these models are exhaustive

Experimentally, both of these approaches produce good results.

  i.e., good generalization

Page 34: Spring 2014 Course
Page 35: Spring 2014 Course

SLIDES STOLEN FROM DAVID HECKERMAN

Page 36: Spring 2014 Course
Page 37: Spring 2014 Course

Interpretation of Marginal Likelihood

Using chain rule for probabilities

Maximizing marginal likelihood also maximizes sequential prediction ability!

Relation to leave-one-out cross validation

Problems with cross validation can overfit the data, possibly because of interchanges (each item is used for

training and for testing each other item) has a hard time dealing with temporal sequence data

Page 38: Spring 2014 Course

Coin Example

Page 39: Spring 2014 Course

αh, αt, #h, and #t all indexed by these conditions

Page 40: Spring 2014 Course

# parent config

# nodes

# node states

Page 41: Spring 2014 Course

Computation of Marginal Likelihood

Efficient closed form solution if

no missing data (including no hidden variables)

mutual independence of parameters θ

local distribution functions from the exponential family (binomial, Poisson, gamma, Gaussian, etc.)

conjugate priors

Page 42: Spring 2014 Course

Computation of Marginal Likelihood

Approximation techniques must be used otherwise.E.g., for missing data can use Gibbs sampling or Gaussian approximation described earlier.

  Bayes theorem

1. Evaluate numerator directly, estimate denominator using Gibbs sampling

  2. For large amounts of data, numerator can be approximated by a multivariate Gaussian

Page 43: Spring 2014 Course

Structure Priors

Hypothesis equivalence  identify equivalence class of a given network structure

All possible structures equally likelyPartial specification: required and prohibited arcs(based on causal knowledge)

Ordering of variables + independence assumptions  ordering based on e.g., temporal precedence  presence or absence of arcs are mutually independent ->n(n-1)/2 priors

p(m) ~ similarity(m, prior Belief Net)

Page 44: Spring 2014 Course

Parameter Priors

all uniform: Beta(1,1)use a prior Belief Net

parameters dependonly on local structure

Page 45: Spring 2014 Course

Model Search

Finding the belief net structure with highest score among those structures with at most k parents is NP-hard for k > 1 (Chickering, 1995)

Sequential search add, remove, reverse arcs ensure no directed cycles efficient in that changes to arcs affect only

some components of p(D|M)

Heuristic methods greedy greedy with restarts MCMC / simulated annealing

Page 46: Spring 2014 Course
Page 47: Spring 2014 Course
Page 48: Spring 2014 Course

two most likely structures

Page 49: Spring 2014 Course
Page 50: Spring 2014 Course

2x1010

Page 51: Spring 2014 Course
Page 52: Spring 2014 Course
Page 53: Spring 2014 Course
Page 54: Spring 2014 Course
Page 55: Spring 2014 Course