• date post

01-Jul-2020
• Category

## Documents

• view

0

0

Embed Size (px)

### Transcript of SINGULAR LEARNING THEORY Part I: Statistical Learningcmnd/programs/mis2013/undergrad/Lin/Powe… ·...

• SINGULAR LEARNING THEORY

Part I: Statistical Learning

Shaowei Lin(Institute for Infocomm Research, Singapore)

21-25 May 2013Motivic Invariants and Singularities Thematic Program

Notre Dame University

• Overview

Probability

Statistics

Bayesian

Regression

Algebraic Geometry

Asymptotic Theory∫

[0,1]2(1−x2y2)N/2dxdy ≈

π8 N

−12 logN−

π8 (

1log 2−2 log 2−

γ)N−12 − 14N

−1logN + 14 (1

log 2 +

1−γ)N−1 −√

2π128 N

−32 logN + · · ·

Statistical Learning

A B

D C

Algebraic Statistics Singular Learning Theory

• Overview

Probability

Statistics

Bayesian

Regression

Part I: Statistical Learning

Part II: Real Log Canonical Thresholds

Part III: Singularities in Graphical Models

• Probability

Probability

• Random Variables

• Discrete·Continuous

• Gaussian

• Basic Concepts

• Independence

Statistics

Bayesian

Regression

• Random Variables

Probability

• Random Variables

• Discrete·Continuous

• Gaussian

• Basic Concepts

• Independence

Statistics

Bayesian

Regression

A probability space (ξ,F ,P) consists of

• a sample space ξ which is the set of all possible outcomes,• a collection♯ F of events, which are subsets of ξ,• an assignment♭ P : F → [0, 1] of probabilities to events

A (real-valued) random variable X : ξ → Rk is

• a function♮ from the sample space to a real vector space.• a measurement of the possible outcomes.• X ∼ P means “X has the distribution given by P ”.♯ σ-algebra: closed under complement, countable union and contains ∅.♭ probability measure: P(∅) = 0, P(ξ) = 1, countable additivity for disjoint events.♮ measurable function: for all x ∈ R, the preimage of {y ∈ Rk : y ≤ x} is in F .

Example . Rolling a fair die.

ξ = { , , , , , } F = {∅, { }, { , }, . . .}X ∈ {1, 2, 3, 4, 5, 6} P(X=1) = 16 , P(X≤3) =

12

• Discrete and Continuous Random Variables

Probability

• Random Variables

• Discrete·Continuous

• Gaussian

• Basic Concepts

• Independence

Statistics

Bayesian

Regression

If ξ is finite, we say X is a discrete random variable.

• probability mass function p(x) = P (X = x), x ∈ Rk.

If ξ is infinite, we define the

• cumulative distribution function (CDF) F (x)

F (x) = P(X≤x), x ∈ Rk.

• probability density function ♯ (PDF) p(y)

F (x) =∫

{y∈Rk:y≤x} p(y)dy, x ∈ Rk.

If the PDF exists, then X is a continuous ♭ random variable.♯ Radon-Nikodym derivative of F (x) with respect to the Lebesgue measure on Rk .♭ We can also define PDFs for discrete variables if we allow the Dirac delta function.

The probability mass/density function is often informally referred toas the distribution of X .

• Gaussian Random Variables

Probability

• Random Variables

• Discrete·Continuous

• Gaussian

• Basic Concepts

• Independence

Statistics

Bayesian

Regression

Example . Multivariate Gaussian distribution X ∼ N (µ,Σ).

X ∈ Rk, mean µ ∈ Rk, covariance Σ ∈ Rk≻0

p(x) =1

(2π detΣ)k/2exp

(

−1

2(x− µ)⊤Σ−1(x− µ)

)

PDFs of univariate Gaussian distributions X ∼ N (µ, σ2)

• Basic Concepts

Probability

• Random Variables

• Discrete·Continuous

• Gaussian

• Basic Concepts

• Independence

Statistics

Bayesian

Regression

The expectation E[X] is the integral of the random variable Xwith respect to its probability measure, i.e. the “average”.

Discrete variables Continuous variables

E[X] =∑

x∈Rk

xP(X=x) E[X] =

Rk

xp(x) dx

The variance E[(X − E[X])2] measures the “spread” of X .

The conditional probability ♯ P(A|B) of two events A,B ∈ F isthe probability that A will occur given that we know B has occurred.

• If P(B) > 0, then P(A|B) = P(A ∩B)/P(B).♯ formal definition depends on the notion of conditional expectation.

Example . Weather forecast.

P(Rain|Thunder)= 0.2/(0.1 + 0.2) = 2/3

Rain No Rain

Thunder 0.2 0.1No Thunder 0.3 0.4

• Independence

Probability

• Random Variables

• Discrete·Continuous

• Gaussian

• Basic Concepts

• Independence

Statistics

Bayesian

Regression

Let X ∈ Rk, Y ∈ Rl, Z ∈ Rm be random variables.

X,Y are independent (X⊥⊥Y ) if

• P(X∈S, Y ∈T ) = P(X∈S)P(Y ∈T )for all measurable subsets S ⊂ Rk, T ⊂ Rl.

• i.e. “ knowing X gives no information about Y ”

X,Y are conditionally independent given Z (X⊥⊥Y | Z) if

• P(X∈S, Y ∈T |Z=z) = P(X∈S|Z=z)P(Y ∈T |Z=z)for all z ∈ Rm and measurable subsets S ⊂ Rk, T ⊂ Rl.

• i.e. “ any dependence between X and Y is due to Z ”

Example . Hidden variables.

Favorite color X ∈ {red, blue}, favorite food Y ∈ {salad, steak}.If X,Y are dependent, one may ask if there is a hidden variable,e.g. gender Z ∈ {female,male}, such that X⊥⊥Y | Z .

• Statistics

Probability

Statistics

• Statistical Model

• Maximum Likelihood

• Kullback-Leibler

• Mixture Models

Bayesian

Regression

• Statistical Model

Probability

Statistics

• Statistical Model

• Maximum Likelihood

• Kullback-Leibler

• Mixture Models

Bayesian

Regression

Let ∆ denote the space of distributions with outcomes ξ.

Model : a family M of probability distributions, i.e. a subset of ∆.

Parametric model : family M of distributions p(·|ω) are indexedby parameters ω in a space Ω, i.e. we have a map Ω → ∆.

Example . Biased coin tosses.

Number of heads in two tosses of coin: H ∈ ξ = {0, 1, 2}Space of distributions:

∆ = {p ∈ R3≥0 : p(0)+p(1)+p(2) = 1}

Probability of getting heads: ω ∈ Ω = [0, 1] ⊂ RParametric model for H :

p(0|ω) = (1− ω)2

p(1|ω) = 2(1− ω)ωp(2|ω) = ω2

implicit equation4 p(0)p(2)− p(1)2 = 0

• Maximum Likelihood

Probability

Statistics

• Statistical Model

• Maximum Likelihood

• Kullback-Leibler

• Mixture Models

Bayesian

Regression

A sample X1, . . . , XN of X is a set of independent, identicallydistributed (i.i.d.) random variables with the same distribution.

Goal: Given a statistical model {p(·|ω) : ω∈Ω} and a sample,find a distribution p(·|ω̂) that best describes the sample.

A statistic f(X1, . . . , XN ) is a function of the sample.An important statistic is the maximum likelihood estimate (MLE).It is a parameter ω̂ that maximizes the likelihood function

L(ω) =

N∏

i=1

p(Xi|ω).

Example . Biased coin tosses.

Suppose the table below summarizes a sample of H of size 100.

H 0 1 2

Count 25 45 30Then, L(ω) = 245ω105(1− ω)95

ω̂ = 105/200.

• Kullback-Leibler Divergence

Probability

Statistics

• Statistical Model

• Maximum Likelihood

• Kullback-Leibler

• Mixture Models

Bayesian

Regression

Let X1, . . . , XN be a sample of a discrete variable X .The empirical distribution is the function

q̂(x) =1

N

N∑

i=1

δ(x−Xi)

where δ(·) is the Kronecker delta function.

The Kullback-Leibler divergence of a distribution p from q is

K(q||p) =∑

x∈Rk

q(x) logq(x)

p(x).

Proposition . ML distributions minimize the KL divergenceof q(·) = p(·|ω) ∈ M from the empirical distribution q̂(·).

K(q̂||q) =∑

x∈Rk

q̂(x) log q̂(x)

︸ ︷︷ ︸

entropy

−1

NlogL(ω)

• Kullback-Leibler Divergence

Probability

Statistics

• Statistical Model

• Maximum Likelihood

• Kullback-Leibler

• Mixture Models

Bayesian

Regression

Let X1, . . . , XN be a sample of a continuous variable X .The empirical distribution is the generalized function

q̂(x) =1

N

N∑

i=1

δ(x−Xi)

where δ(·) is the Dirac delta function.

The Kullback-Leibler divergence of a distribution p from q is

K(q||p) =

Rk

q(x) logq(x)

p(x)dx.

Proposition . ML distributions minimize the KL divergenceof q(·) = p(·|ω) ∈ M from the empirical distribution q̂(·).

K(q̂||q) =

Rk

q̂(x) log q̂(x)dx

︸ ︷︷ ︸

entropy

−1

NlogL(ω)

• Kullback-Leibler Divergence

Probability

Statistics

• Statistical Model

• Maximum Likelihood

• Kullback-Leibler

• Mixture Models

Bayesian

Regression

• Kullback-Leibler Divergence

Probability

Statistics

• Statistical Model

• Maximum Likelihood

• Kullback-Leibler

• Mixture Models

Bayesian

Regression

Example . Population mean and variance.

Let X ∼ N(µ, σ2) be the height of a random Singaporean.Given sample X1, . . . , XN , estimate mean µ and variance σ

2.

Now, the Kullback-Leibler divergence is

K(q̂||q) =1

2σ2N

N∑

i=1

(Xi − µ)2 +

1

Nlog σ + constant.

Differentiating this function gives us the MLE

µ̂ =1

N

N∑

i=1

Xi, σ̂2 =

1

N

N∑

i=1

(Xi − µ̂)2.

MLE for the model mean is the sample mean.MLE for the model variance is the sample variance.

• Mixture Models

Probability

Statistics

• Statistical Model

• Maximum Likelihood

• Kullback-Leibler

• Mixture Models

Bayesian

Regression

A mixture of distributions p1(·), . . . , pm(·) is a convex combination

p(x) =m∑

i=1

αipi(x), x ∈ Rk

i.e. the mixing coefficients αi are nonnegative and sum to one.

Example . Gaussian mixtures.

Mixing univariate Gaussians N (µi, σ2i ), i = 1, . . . ,m, produces

distributions of the form

p(x) =m∑

i=1

αi√

2πσ2i

exp

(

−(x− µi)

2

2σ2i

)

.

This mixture model is therefore described by parameters

ω = (α1, . . . , αm, µ1, . . . , µm, σ1, . . . , σm)

and is frequently used in cluster analysis.

• Bayesian Statistics

Probability

Statistics

Bayesian

• Interpretations

• Bayes’ Rule

• Parameter Estimation

• Model Selection

• Estimating Integrals

• Information Criterion

• Singularities

Regression

• Interpretations of Probability Theory

Probability

Statistics

Bayesian

• Interpretations

• Bayes’ Rule

• Parameter Estimation

• Model Selection

• Estimating Integrals

• Information Criterion

• Singularities

Regression

• Interpretations of Probability Theory

Probability

Statistics

Bayesian

• Interpretations

• Bayes’ Rule

• Parameter Estimation

• Model Selection

• Estimating Integrals

• Information Criterion

• Singularities

Regression

• Bayes’ Rule

Probability

Statistics

Bayesian

• Interpretations

• Bayes’ Rule

• Parameter Estimation

• Model Selection

• Estimating Integrals

• Information Criterion

• Singularities

Regression

Updating our belief of event A based on an observation B.

P(A|B)︸ ︷︷ ︸

posterior (new belief)

=P(B|A)

P(B)P(A)︸ ︷︷ ︸

prior (old belief)

Example . Biased coin toss.

Let θ denote P(heads) of a coin.Determine if the coin is fair (θ = 12) or biased (θ =

34).

Old belief : P(fair) = 0.9Now, suppose we observed a sample with 8 heads and 2 tails.New belief :

P(fair|sample) =P(sample|fair)

P(sample)P(fair)

=(12)

8(12)2(0.9)8

(12)8(12)

2(0.9)8 + (34)8(14)

2(0.1)8≈ 0.584

• Parameter Estimation

Probability

Statistics

Bayesian

• Interpretations

• Bayes’ Rule

• Parameter Estimation

• Model Selection

• Estimating Integrals

• Information Criterion

• Singularities

Regression

Let D = {X1, . . . , XN} denote a sample of X , i.e. “the data”.

Frequentists : Compute maximum likelihood estimate.Bayesians : Treat parameters ω ∈ Ω as random variables

with priors p(ω) that require updating.

Posterior distributionon parameters ω ∈ Ω

p(ω|D) =p(D|ω)p(ω)

Ω p(D|ω)p(ω)dω

Posterior mean

Ωω p(ω|D)dω

Posterior mode argmaxω∈Ω p(ω|D)

Predictive distributionon outcomes x ∈ ξ

p(x|D) =

Ωp(x|ω)p(ω|D)dω

These integrals are difficult to compute or even estimate!

• Model Selection

Probability

Statistics

Bayesian

• Interpretations

• Bayes’ Rule

• Parameter Estimation

• Model Selection

• Estimating Integrals

• Information Criterion

• Singularities

Regression

Which model M1, . . . ,Mm best describes the sample D?

Frequentists : Pick model containing ML distribution.Bayesians : Assign priors p(Mi), p(ω|Mi) and compute

P(Mi|D) =P(D|Mi)P(Mi)

P(D)∝ P(D|Mi)P(Mi).

where P(D|Mi) is the likelihood integral

P(D|Mi) =

N∏

i=1

p(Xi|ω,Mi)p(ω|Mi)dω.

Parameter estimation is a form of model selection!For each ω ∈ Ω, we define a model Mω with one distribution.

• Estimating Integrals

Probability

Statistics

Bayesian

• Interpretations

• Bayes’ Rule

• Parameter Estimation

• Model Selection

• Estimating Integrals

• Information Criterion

• Singularities

Regression

Generally, there are three ways to estimate statistical integrals.

1. Exact methodsCompute a closed form formula for the integral, e.g.Baldoni·Berline·De Loera·Köppe·Vergne, 2010;Lin·Sturmfels·Xu, 2009.

2. Numerical methodsApproximate using Markov Chain Monte Carlo (MCMC)and other sampling techniques.

3. Asymptotic methodsAnalyze how the integral behaves for large samples.Rewrite the likelihood integral as

ZN =

Ωe−Nf(ω)ϕ(ω)dω

where f(ω) = − 1N logL(ω) and ϕ(ω) is the prior on Ω.

• Bayesian Information Criterion

Probability

Statistics

Bayesian

• Interpretations

• Bayes’ Rule

• Parameter Estimation

• Model Selection

• Estimating Integrals

• Information Criterion

• Singularities

Regression

Laplace approximation : If f(ω) is uniquely minimized at MLE ω̂and the Hessian ∂2f(ω̂) is full rank, then asymptotically

− logZN ≈ Nf(ω̂) +dimΩ

2logN +O(1)

as sample size N → ∞.

Bayesian information criterion (BIC): Select model that maximizes

Nf(ω̂) +dimΩ

2logN

N = 1 N = 10

Graphs of e−Nf(ω) for different N . Integral = volume under graph.

• Singularities in Statistical Models

Probability

Statistics

Bayesian

• Interpretations

• Bayes’ Rule

• Parameter Estimation

• Model Selection

• Estimating Integrals

• Information Criterion

• Singularities

Regression

Informally, the model is singular at ω0 ∈ Ω if the Laplaceapproximation fails when the empirical distribution is p(·|ω0).

Formally, if we define the Kullback-Leibler function

K(ω) =

ξp(x|ω0) log

p(x|ω0)

p(x|ω)dx.

then ω0 is a singularity when the Hessian ∂2K(ω0) is not full rank.

Statistical models with hidden variables , e.g. mixture models,often contain many singularities.

• Linear Regression

Probability

Statistics

Bayesian

Regression

• Least Squares

• Sparsity Penalty

• Least Squares

Probability

Statistics

Bayesian

Regression

• Least Squares

• Sparsity Penalty

Suppose we have random variables Y ∈ R, X ∈ Rd that satisfy

Y = ω ·X + ε.

Parameters ω ∈ Rd; noise ε ∈ N (0, 1); data (Yi, Xi), i = 1 .. N .

• Commonly computed quantitiesMLE argminω

∑Ni=1 |Yi − ω ·Xi|

2

Penalized MLE argminω∑N

i=1 |Yi − ω ·Xi|2 + π(ω)

• Commonly used penaltiesLASSO π(ω) = |ω|1 · βBayesian Info Criterion (BIC) π(ω) = |ω|0 · logNAkaike Info Criterion (AIC) π(ω) = |ω|0 · 2

• Commonly asked questionsModel selection (e.g. which factors are important?)Parameter estimation (e.g. how important are the factors?)

• Sparsity Penalty

Probability

Statistics

Bayesian

Regression

• Least Squares

• Sparsity Penalty

The best model is usually selected using a score

argminu∈Ω l(u) + π(u)

where the likelihood l(u) measures the fitting error of the modelwhile the penalty π(u) measures its complexity .

Recently, sparse penalties derived from statistical considerationswere found to be highly effective.

Bayesian info criterion (BIC) ↔ Marginal likelihood integralAkaike info criterion (AIC) ↔ Kullback-Leibler divergence

Compressive sensing ↔ ℓ1-regularization of BIC

Singular learning theory plays an important role in these derivations.

• Paradigms in Statistical Learning

Probability

Statistics

Bayesian

Regression

• Least Squares

• Sparsity Penalty

• Probabilitytheory of random phenomenaStatisticstheory of making sense of dataLearningthe art of prediction using data

• Frequentists : “True distribution, maximum likelihood”Bayesians : “Belief updates, maximum a posteriori”

Interpretations do not affect the correctness of probability theory,but they greatly affect the statistical methodology .

• We often have to balance the complexity of the modelwith its fitting error via some suitable probabilistic criteria.The learning algorithm also needs to be computable to be useful.

• Probability

Statistics

Bayesian

Regression

• Least Squares

• Sparsity Penalty

Thank you!

“Algebraic Methods for Evaluating Integrals in Bayesian Statistics”

http://math.berkeley.edu/~shaowei/swthesis.pdf

(PhD dissertation, May 2011)

http://math.berkeley.edu/~shaowei/swthesis.pdf

• References

Probability

Statistics

Bayesian

Regression

• Least Squares

• Sparsity Penalty

1. V. I. ARNOL’D, S. M. GUSEĬN-ZADE AND A. N. VARCHENKO: Singularities of Differentiable Maps, Vol. II,Birkhäuser, Boston, 1985.

2. V. BALDONI, N. BERLINE, J. A. DE LOERA, M. KÖPPE, M. VERGNE: How to integrate a polynomialover a simplex, Mathematics of Computation 80 (2010) 297–325.

3. A. BRAVO, S. ENCINAS AND O. VILLAMAYOR: A simplified proof of desingularisation and applications,Rev. Math. Iberoamericana 21 (2005) 349–458.

4. D. A. COX, J. B. LITTLE, AND D. O’SHEA: Ideals, Varieties, and Algorithms: An Introduction toComputational Algebraic Geometry and Commutative Algebra, Springer-Verlag, New York, 1997.

5. R. DURRETT: Probability - Theory and Examples (4th edition), Cambridge U. Press, 2010.6. M. EVANS, Z. GILULA AND I. GUTTMAN: Latent class analysis of two-way contingency tables by

Bayesian methods, Biometrika 76 (1989) 557–563.7. H. HIRONAKA: Resolution of singularities of an algebraic variety over a field of characteristic zero I, II,

Ann. of Math. (2) 79 (1964) 109–203.8. S. LIN, B. STURMFELS AND Z. XU: Marginal likelihood integrals for mixtures of independence models,

J. Mach. Learn. Res. 10 (2009) 1611–1631.9. S. LIN: Algebraic methods for evaluating integrals in Bayesian statistics, PhD dissertation, Dept.

Mathematics, UC Berkeley (2011).10. L. MACKEY: Fundamentals - Probability and Statistics, (slides for CS 281A: Statistical Learning Theory

at UC Berkeley, Aug 2009).11. S. WATANABE: Algebraic Geometry and Statistical Learning Theory, Cambridge Monographs on

Applied and Computational Mathematics 25, Cambridge University Press, Cambridge, 2009.

OverviewOverviewProbabilityRandom VariablesDiscrete and Continuous Random VariablesGaussian Random VariablesBasic ConceptsIndependence

StatisticsStatistical ModelMaximum LikelihoodKullback-Leibler DivergenceKullback-Leibler DivergenceKullback-Leibler DivergenceKullback-Leibler DivergenceMixture Models

Bayesian StatisticsInterpretations of Probability TheoryInterpretations of Probability TheoryBayes' RuleParameter EstimationModel SelectionEstimating IntegralsBayesian Information CriterionSingularities in Statistical Models

Linear RegressionLeast SquaresSparsity PenaltyParadigms in Statistical LearningReferences