Machine Learning ! ! ! ! !

Srihari

Basic Sampling Methods

Sargur Srihari
[email protected]

1
Machine Learning ! ! ! ! ! Srihari

Topics
1.  Motivation
Intractability in ML
How sampling can help
2.  Ancestral Sampling
Using BNs
3.  Transforming a Uniform Distribution
4.  Rejection Sampling
5.  Importance Sampling
6.  Sampling-Importance-Resampling
2
Machine Learning ! ! ! ! ! Srihari

1. Motivation
•  When exact inference is intractable, we need
some form of approximation
–  True of probabilistic models of practical significance
•  Inference methods based on numerical sampling
are known as Monte Carlo techniques
•  Most situations will require evaluating
expectations of unobserved variables, e.g., to
make predictions
–  Rather than the posterior distribution

3
Machine Learning ! ! ! ! ! Srihari

Common Task
Illustration

•  Find expectation of some

function f(z)
–  with respect to a probability
distribution p(z)
–  Components of z can be
discrete, continuous or In Bayesian regression,
combination From sum and product rules
p(t | x)= ∫ p(t|x,w)p(w)dw
•  In case of continuous we
where p(t |x,w)~N(y(x,w),β-1)
wish to evaluate
E[ f ] = ∫ f (z) p(z)dz In general such expectations
are too complex to be evaluated
analytically
•  In discrete case, integral
replaces by summation 4
Machine Learning ! ! ! ! ! Srihari

Sampling: main idea

•  Obtain set of samples z(l) where i =1,.., L
•  Drawn independently from distribution p(z)
•  Allows expectation E[ f ] = ∫ f (z) p(z)dz to be approximated by
L
1
fˆ = ∑ f ( z (l ) )
L i =1 Called an estimator

E[ fˆ ] = E[ f ]
•  Then , i.e., estimator has the correct mean
1 "
–  And var[ fˆ ] = E #( f − E( f )) $%
2
which is the variance of the estimator
L
•  Accuracy independent of dimensionality of z
–  High accuracy can be achieved with few (10 or twenty samples)
•  However samples may not be independent
–  Effective sample size may be smaller than apparent sample size
–  In example f(z) is small when p(z) is high and vice versa
•  Expectation may be dominated by regions of small probability thereby requiring
large sample sizes

5
 Machine Learning ! ! ! ! ! Srihari

2. Ancestral Sampling
•  If joint distribution is represented by a BN
–  no observed variables
–  a straightforward method exists
M
•  Distribution is specified by p( z ) = ∏ p(z i | pa i )
i =1
–  where zi are set of variables associated with node i and
–  pai are set of variables associated with node parents of node i
•  To obtain samples from joint
–  we make one pass through set of variables in order z1,..zM
sampling from conditional distribution p(z|pai)
•  After one pass through the graph we obtain one sample
•  Frequency of different values defines the distribution
–  E.g., allowing us to determine marginals
P(L, S) = ∑ P(D, I, G, L, S)= ∑ P(D)P(I )P(G | D, I )P(L | G)P(S | I )
D,I,G D,I,G 6
Machine Learning ! ! ! ! ! Srihari

Ancestral sampling with some

nodes instantiated
•  Directed graph where some nodes instantiated
with observed values
P(L = l 0 , s = s1 ) =
∑ P(D)P(I )P(G | D, I ) ×
D,I,G

P(L = l 0 | G)P(S = s1 | I )

•  Called Logic sampling

•  Use ancestral sampling, except
–  When sample is obtained for an observed value:
•  if they agree then sample value is retained and proceed
to next variable
•  If they don’t agree, whole sample is discarded
7
 Machine Learning ! ! ! ! ! Srihari

Properties of Logic Sampling

P(L = l 0 , s = s1 ) =
∑ P(D)P(I )P(G | D, I ) ×
D,I,G

P(L = l 0 | G)P(S = s1 | I )

•  Samples correctly from posterior distribution

–  Corresponds to sampling from joint distribution of
hidden and data variables
•  But prob. of accepting sample decreases as
–  no of variables increase and
–  number of states that variables can take increases
•  A special case of Importance Sampling
–  Rarely used in practice 8
Machine Learning ! ! ! ! ! Srihari

Undirected Graphs
•  No one-pass sampling strategy even from
prior distribution with no observed
variables
•  Computationally expensive methods such
as Gibbs sampling must be used
–  Start with a sample
–  Replace first variable conditioned on rest of
values, then next variable, etc

9
Machine Learning ! ! ! ! ! Srihari

3. Basic Sampling Algorithms

•  Simple strategies for generating random
samples from a given distribution
–  They will be pseudo random numbers that pass a test
for randomness
•  Assume that algorithm is provided with a
pseudo-random generator for uniform
distribution over (0,1)
•  For standard distributions we can use
transformation method of generating non-
uniformly distributed random numbers
10
Machine Learning ! ! ! ! ! Srihari

Transformation Method
•  Goal: generate random numbers from a
simple non-uniform distribution
•  Assume we have a source of uniformly
distributed random numbers
•  z is uniformly distributed over (0,1), i.e.,
p(z) =1 in the interval
p(z)
0 1
z

11
 Machine Learning ! ! ! ! ! Srihari

Transformation for Standard Distributions

f(z)

p(z) p(y)
0 1
z y = f(z)

•  If we transform values of z using f() such that y=f(z)

•  Distribution of y is governed by dz
p( y ) = p( z ) (1)
dy
•  Goal is to choose f(z) such that values of y have distribution p(y)
•  Integrating (1) above
–  Since p(z)=1 and integral of dz/dy wrt y is z
y
z = h(y) ≡ ∫ p( ŷ) dŷ
−∞
–  which is an indefinite integral of p(y)
•  Thus y = h-1(z)
•  So we have to transform uniformly distributed random numbers
12
–  Using function which is inverse of the indefinite integral of the distribution
 Machine Learning ! ! ! ! ! Srihari

Geometry of Transformation
•  We are interested in generating r.v. s from p(y)
–  non-uniform random variables

•  h(y) is indefinite integral of desired p(y)

•  z ~ U(0,1) is transformed using y = h-1(z)
•  Results in y being distributed as p(y)
13
Machine Learning ! ! ! ! ! Srihari

Transformations for Exponential & Cauchy

•  We need samples from the Exponential Distribution
p( y) = λ exp(−λy )
where 0 ≤ y <∞
y
–  In this case the indefinite integral is z = h(y) = ∫ p(y )d y
0

=1− exp(− λ y)
–  If we transform using y = − λ ln(1− z)−1

–  Then y will have an exponential distribution

•  We need samples from the Cauchy Distribution
–  Cauchy Distribution 1 1
p( y ) =
π 1+ y2
–  Inverse of the indefinite integral can be expressed as a
“tan” function 14
Machine Learning ! ! ! ! ! Srihari

Generalization of Transformation Method

•  Single variable transformation

dz
p(y) = p(z)
dy
–  Where z is uniformly distributed over (0,1)

•  Multiple variable transformation
d(z1 ,..., zM )
p(y1 ,...., yM ) = p(z1 ,..., zM )
d(y1 ,..., yM )

15
Machine Learning ! ! ! ! ! Srihari

Transformation Method for Gaussian

•  Box-Muller method for Gaussian
–  Example of a bivariate Gaussian
•  First generate uniform distribution in unit
circle
–  Generate pairs of uniformly distributed random
numbers
z1 , z2 ∈(−1, 1)
•  Can be done from U(0,1) using zà2z-1
–  Discard each pair unless z12+z22<1
–  Leads to uniform distribution of points inside
unit circle with
1
p(z1 , z2 ) =
π 16
Machine Learning ! ! ! ! ! Srihari

Generating a Gaussian
•  For each pair z1,z2 evaluate the quantities
1/ 2 1/ 2
⎛ − 2 ln z1 ⎞ ⎛ − 2 ln z2 ⎞
y1 = z1 ⎜ 2 ⎟ y 2 = z 2⎜ 2 ⎟
⎝ r ⎠ ⎝ r ⎠
where r 2 = z12 + z 22
–  Then y1 and y2 are independent Gaussians with zero
mean and variance
•  For arbitrary mean and covariance matrix
If y  N(0,1) then σ y + µ has N( µ, σ 2 )
•  In multivariate case
If components are independent and N ( 0,1)
then y = µ + Lz
will have N( µ, Σ) 
17
where Σ = LLT , is called the Cholesky decomposition
Machine Learning ! ! ! ! ! Srihari

Limitation of Transformation Method

•  Need to first calculate and then invert
indefinite integral of the desired distribution
•  Feasible only for a small number of
distributions
•  Need alternate approaches
•  Rejection sampling and Importance sampling
applicable to univariate distributions only
–  But useful as components in more general
strategies
18
 Machine Learning ! ! ! ! ! Srihari

4. Rejection Sampling
•  Allows sampling from a relatively complex distribution
•  Consider univariate, then extend to several variables
•  Wish to sample from distribution p(z)
–  Not a simple standard distribution
–  Sampling from p(z) is difficult
•  Suppose we are able to easily evaluate p(z) for any
given value of z, upto a normalizing constant Z
1
p(z) = p (z) Unnormalized
Zp
–  where p (z) can readily be evaluated but Zp is unknown
–  e.g., p (z) is a mixture of Gaussians
–  Note that we may know the mixture distribution but we need
19
samples to generate expectations

Machine Learning ! ! ! ! ! Srihari

Rejection sampling: Proposal distribution

•  Samples are drawn from simple distribution,

called proposal distribution q(z)
•  Introduce constant k whose value is such that
kq(z) ≥ p(z) for all z
–  Called comparison function

20
 Machine Learning ! ! ! ! ! Srihari

Rejection Sampling Intuition

•  Samples are drawn from
simple distribution q(z)
•  Rejected if they fall in grey
area
–  Between un-normalized
distribution p~(z) and scaled
distribution kq(z)
•  Resulting samples are
distributed according to p(z)
which is the normalized
version of p~(z)
21
 Machine Learning ! ! ! ! ! Srihari

Determining if sample is shaded area

•  Generate two random numbers
–  z0 from q(z)
–  u0 from uniform distribution [0,kq(z0)]

•  This pair has uniform distribution under the

curve of function kq(z)
•  If u0 > p(z0 ) the pair is rejected otherwise it
is retained
•  Remaining pairs have a uniform distribution
under the curve of p(z) and hence the
corresponding z values are distributed
according to p(z) as desired 22
Machine Learning ! ! ! ! ! Srihari

Example of Rejection Sampling

•  Task of sampling from Gamma
b a z a −1 exp(−bz )
Gam( z | a, b) =
Γ( a ) Scaled
Cauchy
•  Since Gamma is roughly bell- Gamma
shaped, proposal distribution is
is Cauchy
•  Cauchy has to be slightly
generalized
–  To ensure it is nowhere smaller
than Gamma

23
Machine Learning ! ! ! ! ! Srihari

Adaptive Rejection Sampling

•  When difficult to find suitable
analytic distribution p(z)
•  Straight-forward when p(z) is
log concave
–  When ln p(z) has derivatives that
are non-increasing functions of z
–  Function ln p(z) and its gradient
are evaluated at initial set of grid
points
–  Intersections are used to
construct envelope
•  A sequence of linear functions
24
Machine Learning ! ! ! ! ! Srihari

Dimensionality and Rejection Sampling

•  Gaussian example
Proposal distribution q(z) is Gaussian
•  Acceptance rate is Scaled version is kq(z)
ratio of volumes
under p(z) and
kq(z)
–  diminishes
exponentially with
True distribution p(z)
dimensionality

25
 5. Importance Sampling
Machine Learning ! ! ! ! ! Srihari

•  Principal reason for sampling p(z) is evaluating

expectation of some f(z) In Bayesian regression,

p(t | x)= ∫ p(t|x,w)p(w)dw

E[ f ] = ∫ f (z) p(z)dz where
p(t |x,w)~N(y(x,w),β-1)

•  Given samples z(l), l=1,..,L, from p(z), the finite

sum approximation is 1 L (l )
f̂ =
L
∑ f (z )
i=1

•  But drawing samples p(z) may be impractical

•  Importance sampling uses:
–  a proposal distribution– like rejection sampling
•  But all samples are retained
–  Assumes that for any z, p(z) can be evaluated
26
 Machine Learning ! ! ! ! ! Srihari

Determining Importance weights

•  Samples {z(l)} are drawn from simpler dist. q(z)
Proposal
E[ f ] = ∫ f (z)p(z)dz distribution

p(z)
= ∫ f (z) q(z)dz
q(z)
1 L p(z(l ) ) (l )
= ∑ (l )
f (z ) Unlike rejection sampling
L l=1 q(z )
All of the samples are retained

•  Samples are weighted by ratios rl= p(z(l)) / q(z(l))

–  Known as importance weights
•  Which corrects the bias introduced by wrong distribution
27
Machine Learning ! ! ! ! ! Srihari

Likelihood-weighted Sampling
•  Importance sampling of graphical model using
ancestral sampling
•  Weight of sample z given evidence variables E

r(z) = ∏ p(zi | pai )

zi ∈E

28
 Machine Learning ! ! ! ! ! Srihari

6. Sampling-Importance Re-sampling (SIR)

•  Rejection sampling depends on Rejection Sampling

suitable value of k where kq(z)> p(k) Proposal q(z) is Gaussian.

Scaled version is kq(z)
–  For many pairs of distributions p(z) and
q(z) it is impractical to determine value
of k
–  If it is sufficiently large to guarantee a
bound then impractically small
acceptance rates
•  Method makes use of sampling True distribution p(z)

distribution q(z) but avoids having to

determine k
•  Name of method: Sampling
proposal distribution followed by
determining importance weights 29
and then resampling
 Machine Learning ! ! ! ! ! Srihari

SIR Method
•  Two stages
•  Stage 1: L samples z(1),..,z(L) are drawn from q(z)
•  Stage 2: Weights w1,..,wL are constructed
–  As in importance sampling wl= p(z(l)) / q(z(l))
•  Finally a second set of L samples are drawn
from the discrete distribution {z(1),..,z(L) } with
probabilities given by {w1,..,wL}
•  If moments wrt distribution p(z) are needed,
use:
E[ f ] = ∫ f (z)p(z)dz
L
=∑ wl f (z(l ) )
l=1
30