Statistics 352: Spatial statistics. Jonathan Taylor. Department of Statistics. Models for discrete data. Stanford University.

Transcription

1 352: 352: Models for discrete data April 28, / 33

2 Models for discrete data 352: Outline Dependent discrete data. Image data (binary). Ising models. Simulation: Gibbs sampling. Denoising. 2 / 33

3 SIDS (sudden infant death syndrome) in North Carolina 352: 36.5 N 36 N 35.5 N 35 N 34.5 N N N 36 N 35.5 N 35 N 34.5 N 84 W 82 W 80 W 78 W 76 W 34 N / 33

4 Discrete data 352: Description Observations: Incidence of SIDS: {Z i : i is a county in North Carolina}. Births in county: {n i : i is a county in North Carolina}. Natural models: Z i Poisson(λ i ) Z i Binomial(n i, p i ) λ i = exp(x i β): spatial? other features? logit(p i ) = x i β? 4 / 33

5 Discrete data 352: Models How do we introduce spatial dependence in the Z i s? If the Z i s are Gaussian, all we need is a covariance function... more complicated for Poisson, Binomial. One approach: Markov random fields (a.k.a. graphical models) 5 / 33

6 Discrete data 352: Models How do we introduce spatial dependence in the Z i s? If the Z i s are Gaussian, all we need is a covariance function... more complicated for Poisson, Binomial. One approach: Markov random fields (a.k.a. graphical models) Instead of defining a general MRF, we ll start with image models. 6 / 33

7 Simulation of an Ising model 352: 7 / 33

8 Binary images 352: Ising model Let L = {(i, j) : 1 i n 1 ; 1 j n 2 } be our lattice of pixels. Binary images Z are elements of { 1, 1} L. Let i j be the nearest neighbours (with periodic boundary conditions). Given an inverse temperature β = 1/T P(Z = z) exp β z i z j (i,j):i j = e β P (i,j):i j z i z j P w { 1,1} L eβ (i,j):i j w i w j 8 / 33

9 Binary images 352: Ising model To a physicist H(z) = β z i z j (i,j):i j is the potential energy of the system. Another representation: H(z) = β ((z i z j ) 2 2) = βz Lz + C (i,j):i j where L is the graph Laplacian. 9 / 33

10 Binary images 352: Ising model Yet another representation z i z j = #{j : i j, z j = z i } #{j : i j, z j z i } j:i j = 2#{j : i j, z j = z i } 4 The set {i : j i, z j z i } can be thought of as the boundary of the black/white interface of z Leads to an interpretation #{j : i j, z j = z i } = boundary length of z. i 10 / 33

11 Binary images 352: Ising model Conditional distributions are simple P(z i = 1 z j, j i) e βz P i j:i j z j = P(z i = 1 z j, j i) Full joint distribution requires partition function w { 1,1} L e β which is complicated... P (i,j):i j w i w j Simulation of the Ising model is (relatively) easy 11 / 33

12 Gibbs sampler (Geman & Geman (1984)) 352: Algorithm def simulate(initial, beta, niter=1000): Z = initial.copy() for k in range(niter): for i in L: s = sum([z[j] for j in nbrs(i, L)]) odds = exp(2*beta*s) p = odds / (1 + odds) Z[i] = bernoulli(p) return Z 12 / 33

13 Gibbs sampler 352: Convergence For whatever initial configuration initial, as niter Z niter niter Z β where Z β is a realization of the Ising model. In fact, as a process (Z i ) 1 i niter is a Markov chain that has stationary distribution Z β. This is the basis of most of the MCMC literature... We ll see the Gibbs sampler again for more general MRFs. 13 / 33

14 Ising models 352: Adding an external field We can also add a mean to the Ising model through an external field P(Z = z) exp β z i z j + α i z i i (i,j):i j 14 / 33

15 Gibbs sampler with an external field 352: Algorithm def simulate(initial, beta, alpha, niter=1000): Z = initial.copy() for k in range(niter): for i in L: s = sum([z[j] for j in nbrs(i, L)]) odds = exp(2*beta*s + 2*alpha[i]) p = odds / (1 + odds) Z[i] = bernoulli(p) return Z 15 / 33

16 Denoising 352: Model Suppose we observe a noisy image Y { 1, 1} L based on a noise-free image Z { 1, 1} L A plausible choice for noise independent bit-flips P(Y i = y i Z = z) = P(Y i = y i Z i = z i ) { q y i = z i = 1 q y i z i Goal: recover Z, the noise-free image. 16 / 33

17 Noisy image q = : 17 / 33

18 Denoising 352: Model If Z were continuous, we might put some smoothness penalty on Z... Recall the Laplacian interpretation of the potential in the Ising model H(z) = βz Lz Suggests the following Ẑ β = argmin log L(Z Y ) β Z { 1,1} L = argmin Z { 1,1} L i L (i,j):i j Z i Y i logit(q) β z i z j (i,j):i j z i z j 18 / 33

19 Denoising 352: Model Our suggested estimator is the mode of an Ising model with parameter β and field α i = logit(q)y i... In Bayesian terms, if our prior for Z is Ising with parameter β, the posterior is Ising with a field dependent on the observations Y. If it s Bayesian, we can sample from the posterior using Gibbs sampler. 19 / 33

20 Denoising 352: Estimating Z Based on outputs of the Gibbs sampler (samples from the posterior), we can compute Ẑ i = 1 Zi >0 where niter Z = is the posterior mean after throwing away l samples. j=l Z j 20 / 33

21 Posterior mean, q = : 21 / 33

22 Denoising 352: Computing the MAP Our original estimator was the MAP for this prior. Finding the MAP is a combinatorial optimization problem: { 1, 1} L possible configurations to search through. A general approach is based on simulated annealing. (Geman & Geman, 1984). 22 / 33

23 Denoising 352: Simulated annealing Basic observation: Ẑβ(Y ) is also the mode of P T,β,Y (z) exp 1 β z i z j logit(q) Y i z i T i L (i,j):i j BUT, for T > 1, P T,β,Y is more sharply peaked than P 1,β,Y. Depends on a choice of temperature schedule... To prove things, one often has to assume T iter = log(iter). 23 / 33

24 Simulated annealing 352: Algorithm def simulate(initial, beta, alpha, temps): Z = initial.copy() for t in temps: for i in L: s = sum([z[j] for j in nbrs(i, L)]) odds = exp((2*beta*s + 2*alpha[i])/t) p = odds / (1 + odds) Z[i] = bernoulli(p) return Z 24 / 33

25 Denoising 352: Additive Gaussian noise Let the new noisy data be Y i = Z i + ɛ i with ɛ i IID N(0, σ 2 ) The field is now α i = Y i z i and i i (i.e. a new term in the neighbourhood relation). 25 / 33

26 Additive noise 352: 26 / 33

27 Additive noise 352: 27 / 33

28 Additive noise 352: 28 / 33

29 Denoising 352: Mixture model The additive noise example can be thought of as a two-class mixture model. Suggests using LDA (or QDA if variances unequal) to classify. Problem: the mixing proportion is / 33

30 Mixture density 352: 30 / 33

31 FDR curve 352: 31 / 33

32 Thresholded 352: 32 / 33

33 Where this leaves us 352: Markov random fields MRFs generally have distributions like an Ising model. Gibbs sampler can be used to simulate MRFs. Simulated annealing can also generally be used to find MAP (i.e. modes of an MRF). Bayesian image analysis is a huge field. 33 / 33