Sampling Methods CMSC 691 UMBC (Some) Learning Techniques - - PowerPoint PPT Presentation

Approximate Inference: Sampling Methods

CMSC 691 UMBC

(Some) Learning Techniques

MAP/MLE: Point estimation, basic EM Variational Inference: Functional Optimization Sampling/Monte Carlo

today

Outline

Monte Carlo methods Sampling Techniques

Uniform sampling Importance Sampling Rejection Sampling Metropolis-Hastings Gibbs sampling

Example: Collapsed Gibbs Sampler for Topic Models

Two Problems for Sampling Methods to Solve

Generate samples from p

𝑞 𝑦 = 𝑣 𝑦 𝑎 , 𝑦 ∈ ℝ𝐸 𝑦1, 𝑦2, … , 𝑦𝑆 samples Q: Why might sampling from p(x) be hard?

Two Problems for Sampling Methods to Solve

Generate samples from p

𝑞 𝑦 = 𝑣 𝑦 𝑎 , 𝑦 ∈ ℝ𝐸 𝑦1, 𝑦2, … , 𝑦𝑆 samples Q: Why might sampling from p(x) be hard? A1: Can we evaluate Z? A2: Can we sample without enumerating? (Correct samples should be where p is big)

Two Problems for Sampling Methods to Solve

Generate samples from p

𝑞 𝑦 = 𝑣 𝑦 𝑎 , 𝑦 ∈ ℝ𝐸 𝑦1, 𝑦2, … , 𝑦𝑆 samples Q: Why might sampling from p(x) be hard? A1: Can we evaluate Z? A2: Can we sample without enumerating? (Correct samples should be where p is big) 𝑣 𝑦 = exp(.4 𝑦 − .4 2 − 0.08𝑦4)

ITILA, Fig 29.1

Two Problems for Sampling Methods to Solve

Generate samples from p Estimate expectation of a function 𝜚

𝑞 𝑦 = 𝑣 𝑦 𝑎 , 𝑦 ∈ ℝ𝐸 𝑦1, 𝑦2, … , 𝑦𝑆 samples Q: Why might sampling from p(x) be hard? A1: Can we evaluate Z? A2: Can we sample without enumerating? (Correct samples should be where p is big) Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

= ∫ 𝑞 𝑦 𝜚 𝑦 𝑒𝑦

Two Problems for Sampling Methods to Solve

Generate samples from p Estimate expectation of a function 𝜚

𝑞 𝑦 = 𝑣 𝑦 𝑎 , 𝑦 ∈ ℝ𝐸 𝑦1, 𝑦2, … , 𝑦𝑆 samples Q: Why might sampling from p(x) be hard? A1: Can we evaluate Z? A2: Can we sample without enumerating? (Correct samples should be where p is big) Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

= ∫ 𝑞 𝑦 𝜚 𝑦 𝑒𝑦

෡ Φ =

1 𝑆 σ𝑠 𝜚 𝑦𝑠

Two Problems for Sampling Methods to Solve

Generate samples from p Estimate expectation of a function 𝜚

If we could sample from p… 𝑞 𝑦 = 𝑣 𝑦 𝑎 , 𝑦 ∈ ℝ𝐸 𝑦1, 𝑦2, … , 𝑦𝑆 samples Q: Why is sampling from p(x) hard? A1: Can we evaluate Z? A2: Can we sample without enumerating? (Correct samples should be where p is big) Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

= ∫ 𝑞 𝑦 𝜚 𝑦 𝑒𝑦

෡ Φ =

1 𝑆 σ𝑠 𝜚 𝑦𝑠 𝔽 ෡ Φ = Φ

consistent estimator

Outline

Monte Carlo methods Sampling Techniques

Uniform sampling Importance Sampling Rejection Sampling Metropolis-Hastings Gibbs sampling

Example: Collapsed Gibbs Sampler for Topic Models

Uniform Sampling

෡ Φ = ෍

𝑠

𝜚 𝑦𝑠 𝑞∗(𝑦𝑠)

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝑦1, 𝑦2, … , 𝑦𝑆

Uniform Sampling

𝑞∗ 𝑦 = 𝑣 𝑦 𝑎∗

෡ Φ = ෍

𝑠

𝜚 𝑦𝑠 𝑞∗(𝑦𝑠)

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal: 𝑎∗ = ෍

𝑠

𝑣(𝑦𝑠)

sample uniformly: 𝑦1, 𝑦2, … , 𝑦𝑆

Uniform Sampling

𝑞∗ 𝑦 = 𝑣 𝑦 𝑎∗

෡ Φ = ෍

𝑠

𝜚 𝑦𝑠 𝑞∗(𝑦𝑠)

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal: 𝑎∗ = ෍

𝑠

𝑣(𝑦𝑠)

sample uniformly: 𝑦1, 𝑦2, … , 𝑦𝑆

this might work if R (the number of samples) sufficiently hits high probability regions

Uniform Sampling

𝑞∗ 𝑦 = 𝑣 𝑦 𝑎∗

෡ Φ = ෍

𝑠

𝜚 𝑦𝑠 𝑞∗(𝑦𝑠)

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal: 𝑎∗ = ෍

𝑠

𝑣(𝑦𝑠)

sample uniformly: 𝑦1, 𝑦2, … , 𝑦𝑆

this might work if R (the number of samples) sufficiently hits high probability regions Ising model example:

• 2H states of high

probability

• 2N states total

Uniform Sampling

𝑞∗ 𝑦 = 𝑣 𝑦 𝑎∗

෡ Φ = ෍

𝑠

𝜚 𝑦𝑠 𝑞∗(𝑦𝑠)

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal: 𝑎∗ = ෍

𝑠

𝑣(𝑦𝑠)

sample uniformly: 𝑦1, 𝑦2, … , 𝑦𝑆

this might work if R (the number of samples) sufficiently hits high probability regions Ising model example:

• 2H states of high

probability

• 2N states total

chance of sample being in high prob. region:

2𝐼 2𝑂

• min. samples needed: ∼ 2𝑂−𝐼

Outline

Monte Carlo methods Sampling Techniques

Uniform sampling Importance Sampling Rejection Sampling Metropolis-Hastings Gibbs sampling

Example: Collapsed Gibbs Sampler for Topic Models

Importance Sampling

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆 approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦

ITILA, Fig 29.5

Importance Sampling

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆 approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦

p(x)

ITILA, Fig 29.5

x where Q(x) > p(x):

• ver-represented

x where Q(x) < p(x): under-represented

Importance Sampling

𝑥 𝑦𝑠 = 𝑣𝑞 𝑦 𝑣𝑟 𝑦

෡ Φ = σ𝑠 𝜚 𝑦𝑠 𝑥(𝑦𝑠) σ𝑠 𝑥 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆 approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦

p(x)

ITILA, Fig 29.5

x where Q(x) > p(x):

• ver-represented

x where Q(x) < p(x): under-represented

Importance Sampling

𝑥 𝑦𝑠 = 𝑣𝑞 𝑦 𝑣𝑟 𝑦

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆 approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦

p(x)

ITILA, Fig 29.5

x where Q(x) > p(x):

• ver-represented

x where Q(x) < p(x): under-represented

෡ Φ = σ𝑠 𝜚 𝑦𝑠 𝑥(𝑦𝑠) σ𝑠 𝑥 𝑦𝑠

Q: How reliable will this estimator be?

Importance Sampling

𝑥 𝑦𝑠 = 𝑣𝑞 𝑦 𝑣𝑟 𝑦

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆 approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦

p(x)

ITILA, Fig 29.5

x where Q(x) > p(x):

• ver-represented

x where Q(x) < p(x): under-represented

෡ Φ = σ𝑠 𝜚 𝑦𝑠 𝑥(𝑦𝑠) σ𝑠 𝑥 𝑦𝑠

Q: How reliable will this estimator be? A: In practice, difficult to say. 𝑥 𝑦𝑠 may not be a good indicator

Importance Sampling

𝑥 𝑦𝑠 = 𝑣𝑞 𝑦 𝑣𝑟 𝑦

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆 approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦

p(x)

ITILA, Fig 29.5

x where Q(x) > p(x):

• ver-represented

x where Q(x) < p(x): under-represented

෡ Φ = σ𝑠 𝜚 𝑦𝑠 𝑥(𝑦𝑠) σ𝑠 𝑥 𝑦𝑠

Q: How reliable will this estimator be? A: In practice, difficult to say. 𝑥 𝑦𝑠 may not be a good indicator Q: How do you choose a good approximating distribution?

Importance Sampling

𝑥 𝑦𝑠 = 𝑣𝑞 𝑦 𝑣𝑟 𝑦

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆 approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦

p(x)

ITILA, Fig 29.5

x where Q(x) > p(x):

• ver-represented

x where Q(x) < p(x): under-represented

෡ Φ = σ𝑠 𝜚 𝑦𝑠 𝑥(𝑦𝑠) σ𝑠 𝑥 𝑦𝑠

Q: How reliable will this estimator be? A: In practice, difficult to say. 𝑥 𝑦𝑠 may not be a good indicator Q: How do you choose a good approximating distribution? A: Task/domain specific

Importance Sampling: Variance Estimator may vary

ITILA, Fig 29.6

true value

q(x): Gaussian q(x): Cauchy distribution iterations

Outline

Monte Carlo methods Sampling Techniques

Uniform sampling Importance Sampling Rejection Sampling Metropolis-Hastings Gibbs sampling

Example: Collapsed Gibbs Sampler for Topic Models

Rejection Sampling

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 , 𝑑 ∗ 𝑣𝑟 > 𝑣𝑞

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦

ITILA, Fig 29.8

Rejection Sampling

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝑨𝑙 ∼ Unif(0, 𝑑 ∗ 𝑣𝑟 𝑦𝑙 )

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦

ITILA, Fig 29.8

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆∗

select tuples

approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 , 𝑑 ∗ 𝑣𝑟 > 𝑣𝑞

Rejection Sampling

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝑨𝑙 ∼ Unif(0, 𝑑 ∗ 𝑣𝑟 𝑦𝑙 )

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦

ITILA, Fig 29.8

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆∗

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦 𝑨

select tuples

if 𝑨𝑙 ≤ 𝑣𝑞 𝑦𝑙 : add 𝑦𝑙 to sampled R points

• therwise: reject it

approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 , 𝑑 ∗ 𝑣𝑟 > 𝑣𝑞

Rejection Sampling

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝑨𝑙 ∼ Unif(0, 𝑑 ∗ 𝑣𝑟 𝑦𝑙 )

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦

ITILA, Fig 29.8

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆∗

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦 𝑨

select tuples

if 𝑨𝑙 ≤ 𝑣𝑞 𝑦𝑙 : add 𝑦𝑙 to sampled R points

• therwise: reject it

approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 , 𝑑 ∗ 𝑣𝑟 > 𝑣𝑞

this produces samples from the p-distribution

Rejection Sampling

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝑨𝑙 ∼ Unif(0, 𝑑 ∗ 𝑣𝑟 𝑦𝑙 )

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦

ITILA, Fig 29.8

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆∗

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦 𝑨

select tuples

if 𝑨𝑙 ≤ 𝑣𝑞 𝑦𝑙 : add 𝑦𝑙 to sampled R points

• therwise: reject it

approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 , 𝑑 ∗ 𝑣𝑟 > 𝑣𝑞

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦 𝑨

Rejection Sampling

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝑨𝑙 ∼ Unif(0, 𝑑 ∗ 𝑣𝑟 𝑦𝑙 )

ITILA, Fig 29.8

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆∗

select tuples

if 𝑨𝑙 ≤ 𝑣𝑞 𝑦𝑙 : add 𝑦𝑙 to sampled R points

• therwise: reject it

Q: How reliable will this estimator be?

approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 , 𝑑 ∗ 𝑣𝑟 > 𝑣𝑞

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦 𝑨

Rejection Sampling

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝑨𝑙 ∼ Unif(0, 𝑑 ∗ 𝑣𝑟 𝑦𝑙 )

ITILA, Fig 29.8

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆∗

select tuples

if 𝑨𝑙 ≤ 𝑣𝑞 𝑦𝑙 : add 𝑦𝑙 to sampled R points

• therwise: reject it

Q: How reliable will this estimator be? A: How well does Q approximate P? Q: How do you choose a good approximating distribution?

approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 , 𝑑 ∗ 𝑣𝑟 > 𝑣𝑞

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦 𝑨

Rejection Sampling

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝑨𝑙 ∼ Unif(0, 𝑑 ∗ 𝑣𝑟 𝑦𝑙 )

ITILA, Fig 29.8

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆∗

select tuples

if 𝑨𝑙 ≤ 𝑣𝑞 𝑦𝑙 : add 𝑦𝑙 to sampled R points

• therwise: reject it

Q: How reliable will this estimator be? A: How well does Q approximate P? Q: How do you choose a good approximating distribution? A: Task/domain specific

approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 , 𝑑 ∗ 𝑣𝑟 > 𝑣𝑞

𝑑 ∗ 𝑣𝑟 𝑦 𝑣𝑞 𝑦 𝑨

Rejection Sampling

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝑨𝑙 ∼ Unif(0, 𝑑 ∗ 𝑣𝑟 𝑦𝑙 )

ITILA, Fig 29.8

sample from Q: 𝑦1, 𝑦2, … , 𝑦𝑆∗

select tuples

if 𝑨𝑙 ≤ 𝑣𝑞 𝑦𝑙 : add 𝑦𝑙 to sampled R points

• therwise: reject it

Q: How reliable will this estimator be? A: How well does Q approximate P? Q: How do you choose a good approximating distribution? A: Task/domain specific

approximating distribution: 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 , 𝑑 ∗ 𝑣𝑟 > 𝑣𝑞 rejection sampling can be difficult to use in high- dimensional spaces 

Outline

Monte Carlo methods Sampling Techniques

Uniform sampling Importance Sampling Rejection Sampling Metropolis-Hastings Gibbs sampling

Example: Collapsed Gibbs Sampler for Topic Models

Markov Chain Monte Carlo

transition kernel

Metropolis-Hastings

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

importance and rejection sampling: a single proposal distribution 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 Metropolis-Hastings (and Gibbs): create a proposal distribution based

• n current state

𝑅 𝑦|𝑦(𝑢) ∝ 𝑣𝑟 𝑦| 𝑦(𝑢)

Metropolis-Hastings

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

importance and rejection sampling: a single proposal distribution 𝑅 𝑦 ∝ 𝑣𝑟 𝑦 Metropolis-Hastings (and Gibbs): create a proposal distribution based

• n current state

𝑅 𝑦|𝑦(𝑢) ∝ 𝑣𝑟 𝑦|𝑦(𝑢)

Q does not need to look similar to P

ITILA, Fig 29.10

Metropolis-Hastings

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝛽𝑙 = 𝑣𝑞(𝑦𝑙) 𝑣𝑞 𝑦(𝑢) 𝑣𝑟 𝑦𝑙 𝑣𝑟 𝑦(𝑢) sample from 𝑹 𝑦|𝑦(𝑢) : 𝑦1, 𝑦2, … , 𝑦𝑆∗ if𝛽𝑙 ≥ 1: add 𝑦𝑙 to sampled R points

• therwise: accept with

probability 𝛽𝑙 transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) ∝ 𝑣𝑟 𝑦|𝑦(𝑢)

Metropolis-Hastings

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝛽𝑙 = 𝑣𝑞(𝑦𝑙) 𝑣𝑞 𝑦(𝑢) 𝑣𝑟 𝑦𝑙 𝑣𝑟 𝑦(𝑢) sample from 𝑹 𝑦|𝑦(𝑢) : 𝑦1, 𝑦2, … , 𝑦𝑆∗ if 𝛽𝑙 ≥ 1: add 𝑦𝑙 to sampled R points

• therwise: accept with

probability 𝛽𝑙 transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) ∝ 𝑣𝑟 𝑦|𝑦(𝑢)

if accepted: 𝑦(𝑢+1) = 𝑦𝑙

• therwise: 𝑦(𝑢+1) = 𝑦(𝑢)

samples are not independent

Metropolis-Hastings

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample uniformly: 𝛽𝑙 = 𝑣𝑞(𝑦𝑙) 𝑣𝑞 𝑦(𝑢) 𝑣𝑟 𝑦𝑙 𝑣𝑟 𝑦(𝑢) sample from 𝑹 𝑦|𝑦(𝑢) : 𝑦1, 𝑦2, … , 𝑦𝑆∗ if 𝛽𝑙 ≥ 1: add 𝑦𝑙 to sampled R points

• therwise: accept with

probability 𝛽𝑙 transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) ∝ 𝑣𝑟 𝑦|𝑦(𝑢)

if accepted: 𝑦(𝑢+1) = 𝑦𝑙

• therwise: 𝑦(𝑢+1) = 𝑦(𝑢)

samples are not independent Metropolis-Hastings can be used effectively in high- dimensional spaces ☺

Outline

Monte Carlo methods Sampling Techniques

Uniform sampling Importance Sampling Rejection Sampling Metropolis-Hastings Gibbs sampling

Example: Collapsed Gibbs Sampler for Topic Models

Gibbs Sampling

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) = 𝑞 𝑦 all other variables)

Next sampled value

• f current variable

Values of all other variables, both new and old

𝑦 𝑗

(𝑢+1) ∼ 𝑞 ⋅ 𝑦 1 𝑢+1 , … , 𝑦 𝑗−1 𝑢+1 , 𝑦 𝑗+1 𝑢

, … , 𝑦 𝑂

𝑢 )

x[i]

Remember: Markov Blanket

x Markov blanket of a node x is its parents, children, and children's parents

𝑞 𝑦𝑗 𝑦𝑘≠𝑗 = 𝑞(𝑦1, … , 𝑦𝑂) ∫ 𝑞 𝑦1, … , 𝑦𝑂 𝑒𝑦𝑗 = ς𝑙 𝑞(𝑦𝑙|𝜌 𝑦𝑙 ) ∫ ς𝑙 𝑞 𝑦𝑙 𝜌 𝑦𝑙 ) 𝑒𝑦𝑗

factor out terms not dependent on xi

factorization

• f graph

= ς𝑙:𝑙=𝑗 or 𝑗∈𝜌 𝑦𝑙 𝑞(𝑦𝑙|𝜌 𝑦𝑙 ) ∫ ς𝑙:𝑙=𝑗 or 𝑗∈𝜌 𝑦𝑙 𝑞 𝑦𝑙 𝜌 𝑦𝑙 ) 𝑒𝑦𝑗

the set of nodes needed to form the complete conditional for a variable xi

Gibbs Sampling

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) = 𝑞 𝑦 MB variables)

Next sampled value of current variable Values of just the Markov blanket variables, both old and old

𝑦 5

(𝑢+1) ∼ 𝑞 ⋅ 𝑦 2 (𝑢+1), 𝑦 3 𝑢 , 𝑦 4 (𝑢+1) , 𝑦 6 (𝑢+1), 𝑦 7 𝑢 , 𝑦 8 (𝑢+1))

x5 x2 x3 x4 x6 x7 x8

Gibbs Sampling

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample (always accept) from 𝑹 𝑦|𝑦(𝑢) : 𝑦1, 𝑦2, … , 𝑦𝑆∗ transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) = 𝑞 𝑦 MB(𝑦(𝑢)))

𝑦(𝑢+1) = 𝑦𝑙

Markov blanket

samples are not independent Gibbs Sampling can be used effectively in high-dimensional spaces ☺

Collapsed Gibbs Sampling (CGS)

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample (always accept) from 𝑹 𝑦|𝑦(𝑢) : 𝑦1, 𝑦2, … , 𝑦𝑆∗ transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) = ∫ 𝑞 𝑦 MB(𝑦 𝑢 )) 𝑒𝑧 = 𝑞 𝑦 MB−y 𝑦𝑢 )

𝑦(𝑢+1) = 𝑦𝑙

integrate out some of Markov blanket

samples are not independent Collapsed Gibbs can be used effectively in high-dimensional spaces ☺

Collapsed Gibbs Sampling

෡ Φ = 1 𝑆 ෍

𝑠

𝜚 𝑦𝑠

Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

sample (always accept) from 𝑹 𝑦|𝑦(𝑢) : 𝑦1, 𝑦2, … , 𝑦𝑆∗ transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) = ∫ 𝑞 𝑦 MB(𝑦 𝑢 )) 𝑒𝑧 = 𝑞 𝑦 MB−y 𝑦𝑢 )

𝑦(𝑢+1) = 𝑦𝑙

integrate out some of Markov blanket

samples are not independent Collapsed Gibbs can be used effectively in high-dimensional spaces ☺ Warning: collapsing changes the Markov blanket

Collapsed Gibbs Sampling Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) = 𝑞 𝑦 select vars in MB)

x5 x2 x3 x4 x6 x7 x8 x9

Let’s integrate out x4

Collapsed Gibbs Sampling Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) = 𝑞 𝑦 select vars in MB)

x5 x2 x3 x4 x6 x7 x8 x9

Let’s integrate out x4

Collapsed Gibbs Sampling Φ = 𝜚 𝑦

𝑞 = 𝔽𝑦∼𝑞 𝜚 𝑦

Goal:

transition kernel/distribution: 𝑅 𝑦|𝑦(𝑢) = 𝑞 𝑦 select vars in MB)

Next sampled value of current variable Values of some of the Markov blanket variables, both old and old

𝑦 5

(𝑢+1) ∼ 𝑞 ⋅ 𝑦 2 (𝑢+1), 𝑦 3 𝑢 , 𝑦 9 (𝑢+1) , 𝑦 6 (𝑢+1), 𝑦 7 𝑢 , 𝑦 8 (𝑢+1))

x5 x2 x3 x4 x6 x7 x8 x9

Let’s integrate out x4

What Are the Trade-offs of CGS?

Benefits

• Collapsing variables

removes variables from the model via integration/marginalization

– Depending on which variables are marginalized out, this could be a drastic reduction – The priors/hyperparams of the collapsed variables still impact the result

• The “steps” are less

incremental Drawbacks

What Are the Trade-offs of CGS?

Benefits

• Collapsing variables

removes variables from the model via integration/marginalization

– Depending on which variables are marginalized out, this could be a drastic reduction – The priors/hyperparams of the collapsed variables still impact the result

• The “steps” are less

incremental Drawbacks

• Collapsing removes

conditional independences in the model

• Math may not be easy
• You may be restricted via

conjugacy/other statistical properties

• You still have the drawbacks
• f sampling

Outline

Monte Carlo methods Sampling Techniques

Uniform sampling Importance Sampling Rejection Sampling Metropolis-Hastings Gibbs sampling

Example: Collapsed Gibbs Sampler for Topic Models

Latent Dirichlet Allocation (Blei et al., 2003)

Per- document (latent) topic usage Per-document (unigram) word counts Per-topic word usage

d

Gibbs Sampler for LDirA

for each document d: resample θd | zd,1,…, zd,Nd for each token i in d: resample zd,i | wd,i , {ψk }, θd for each topic k: resample ψk

Latent Dirichlet Allocation (Blei et al., 2003)

Per- document (latent) topic usage Per-document (unigram) word counts Per-topic word usage

d

integrate these out

Collapsed Gibbs Sampler for LDirA

for each document d: resample θd | zd,1,…, zd,Nd for each token i in d: resample zd,i | wd,i , {ψk }, {z*,-i} for each topic k: resample ψk

Collapsed Gibbs Sampler for LDirA

for each document d: resample θd | zd,1,…, zd,Nd for each token i in d: resample zd,i | wd,i , {ψk }, {z*,-i} for each topic k: resample ψk 𝑞 𝑨𝑒𝑗 𝑨∗,−𝑗) = 𝑞(𝑨∗,∗) 𝑞(𝑨∗,−𝑗)

Sampling: Discrete Observations

Griffiths and Stevers (PNAS, 2004)

Sampling: Discrete Observations

Griffiths and Stevers (PNAS, 2004)

Sampling: Discrete Observations

Griffiths and Stevers (PNAS, 2004)

Sampling: Discrete Observations

Griffiths and Stevers (PNAS, 2004)

Gamma function fact: Γ 𝑦 + 1 = 𝑦Γ(𝑦) maintain count tables

Sampling: Discrete Observations

Griffiths and Stevers (PNAS, 2004)

𝑞 𝑨𝑒𝑗 𝑨∗,−𝑗) = 𝑞(𝑨∗,∗) 𝑞(𝑨∗,−𝑗)

Collapsed Gibbs Sampling goal:

Sampling: Discrete Observations

Griffiths and Stevers (PNAS, 2004)

Gamma function fact: Γ 𝑦 + 1 = 𝑦Γ(𝑦)

𝑞 𝑨𝑒𝑗 𝑨∗,−𝑗) = Γ(σ𝑙 𝛽𝑙) Γ(σ𝑙 𝑑 𝑒, 𝑙 + 𝛽𝑙) ς𝑙 Γ(𝑑 𝑒, 𝑙 + 𝛽𝑙) Γ(𝛽𝑙) Γ(σ𝑙 𝛽𝑙) Γ(σ𝑙 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) ς𝑙 Γ(𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) Γ(𝛽𝑙)

Collapsed Gibbs Sampling goal:

Sampling: Discrete Observations

Griffiths and Stevers (PNAS, 2004)

Gamma function fact: Γ 𝑦 + 1 = 𝑦Γ(𝑦) 𝑞 𝑨𝑒𝑗 𝑨∗,−𝑗) = ς𝑙(𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙)Γ(𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) (σ𝑙 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) Γ(σ𝑙 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) ς𝑙 Γ(𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) Γ(σ𝑙 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙)

Collapsed Gibbs Sampling goal:

Sampling: Discrete Observations

Griffiths and Stevers (PNAS, 2004)

Gamma function fact: Γ 𝑦 + 1 = 𝑦Γ(𝑦) 𝑞 𝑨𝑒𝑗 𝑨∗,−𝑗) = ς𝑙(𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙)Γ(𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) (σ𝑙 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) Γ(σ𝑙 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) ς𝑙 Γ(𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙) Γ(σ𝑙 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙)

Collapsed Gibbs Sampling goal:

Sampling: Discrete Observations

Griffiths and Stevers (PNAS, 2004)

Gamma function fact: Γ 𝑦 + 1 = 𝑦Γ(𝑦)

𝑞 𝑨𝑒𝑗 = 𝑙 𝑨∗,−𝑗) ∝ 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙

Collapsed Gibbs Sampling goal:

maintain count tables

Collapsed Gibbs Sampler for LDirA

for each document d: for each token i in d: resample zd,i | wd,i , {ψk }, {z*,-i} 𝑞 𝑨𝑒𝑗 𝑨∗,−𝑗) =∝ 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙 ∗topic-word counts

Collapsed Gibbs Sampler for LDirA

randomly assign z*,* maintain count tables: c(d,k): document-topic counts c(k,v): topic-word counts for each document d: for each token i in d: unassign topic: zd,i resample zd,i | wd,i , {ψk }, {z*,-i} reassign topic: zd,i

𝑞 𝑨𝑒𝑗 𝑨∗,−𝑗) =∝ 𝑑 𝑒, 𝑙 − 1 + 𝛽𝑙 ∗topic-word counts

decrease counts increase counts

Outline

Monte Carlo methods Sampling Techniques

Uniform sampling Importance Sampling Rejection Sampling Metropolis-Hastings Gibbs sampling

Example: Collapsed Gibbs Sampler for Topic Models