icml2004 tutorial on bayesian methods for machine learning

Bayesian Methods for Machine Learning

Zoubin Ghahramani

Gatsby Computational Neuroscience Unit
University College London, UK

Center for Automated Learning and Discovery
Carnegie Mellon University, USA

zoubin@gatsby.ucl.ac.uk
http://www.gatsby.ucl.ac.uk

International Conference on Machine Learning
Tutorial July 2004

Plan

• Introduce Foundations

• The Intractability Problem

• Approximation Tools

• Advanced Topics

• Limitations and Discussion

Detailed Plan

• Introduce Foundations • Approximation Tools
– Some canonical problems: classiﬁcation, – Laplace’s Approximation
regression, density estimation, coin toss – Bayesian Information Criterion (BIC)
– Representing beliefs and the Cox axioms – Variational Approximations
– The Dutch Book Theorem – Expectation Propagation
– Asymptotic Certainty and Consensus – MCMC
– Occam’s Razor and Marginal Likelihoods – Exact Sampling
– Choosing Priors • Advanced Topics
∗ Objective Priors: – Feature Selection and ARD
Noninformative, Jeﬀreys, Reference – Bayesian Discriminative Learning (BPM vs SVM)
∗ Subjective Priors – From Parametric to Nonparametric Methods
∗ Hierarchical Priors ∗ Gaussian Processes
∗ Empirical Priors ∗ Dirichlet Process Mixtures
∗ Conjugate Priors ∗ Other Non-parametric Bayesian Methods
• The Intractability Problem – Bayesian Decision Theory and Active Learning
– Bayesian Semi-supervised Learning
– Reconciling Bayesian and Frequentist Views
– Limitations and Criticisms of Bayesian Methods
– Discussion

Some Canonical Problems

• Coin Toss

• Linear Classiﬁcation

• Polynomial Regression

• Clustering with Gaussian Mixtures (Density Estimation)

Coin Toss

Data: D = (H T H H H T T . . .)
def
Parameters: θ = Probability of heads

P (H|θ) = θ
P (T |θ) = 1 − θ

Goal: To infer θ from the data and predict future outcomes P (H|D).

Linear Classiﬁcation

Data: D = {(x(n), y (n))} for n = 1, . . . , N x
data points o
x x
x x x o
x (n)
∈ R D
x
x
o
o
y (n)
∈ {+1, −1} x xx o o o
o

Parameters: θ ∈ RD+1

D
 (n)
1 if θd xd + θ0 ≥ 0

P (y (n) = +1|θ, x(n)) =
 d=1
0 otherwise


Goal: To infer θ from the data and to predict future labels P (y|D, x)

Polynomial Regression

50

Data: D = {(x(n), y (n))} for n = 1, . . . , N 40

30

R
20
(n)
x ∈

y
10

y (n) ∈ R 0

−10

−20
−2 0 2 4 6 8 10 12
x

Parameters: θ = (a0, . . . , am, σ)

Model:
(n) (n) (n) 2 (n) m
y = a0 + a1x + a2x . . . + amx +
where
∼ N (0, σ 2)

Goal: To infer θ from the data and to predict future outputs P (y|D, x, m)

Clustering with Gaussian Mixtures
(Density Estimation)

Data: D = {x(n)} for n = 1, . . . , N

x(n) ∈ RD

Parameters: θ = (µ(1), Σ(1)) . . . , (µ(m), Σ(m)), π

Model:
m
x(n) ∼ πi pi(x(n))
i=1
where
pi(x(n)) = N (µ(i), Σ(i))

Goal: To infer θ from the data and predict the density p(x|D, m)

Basic Rules of Probability

P (x) probability of x
P (x|θ) conditional probability of x given θ
P (x, θ) joint probability of x and θ

P (x, θ) = P (x)P (θ|x) = P (θ)P (x|θ)

Bayes Rule:
P (x|θ)P (θ)
P (θ|x) =
P (x)
Marginalization
P (x) = P (x, θ) dθ

Warning: I will not be obsessively careful in my use of p and P for probability density and probability
distribution. Should be obvious from context.

Questions

• Why be Bayesian?

• Where does the prior come from?

• How do we do these integrals?

Representing Beliefs (Artiﬁcial Intelligence)

Consider a robot. In order to behave intelligently
the robot should be able to represent beliefs about
propositions in the world:

“my charging station is at location (x,y,z)”

“my rangeﬁnder is malfunctioning”

“that stormtrooper is hostile”

We want to represent the strength of these beliefs numerically in the brain of the
robot, and we want to know what rules (calculus) we should use to manipulate
those beliefs.

Representing Beliefs II
Let’s use b(x) to represent the stength of belief in (plausibility of) proposition x.

0 ≤ b(x) ≤ 1
b(x) = 0 x is deﬁnitely not true
b(x) = 1 x is deﬁnitely true
b(x|y) strength of belief that x is true given that we know y is true

Cox Axioms (Desiderata):
• Strengths of belief (degrees of plausibility) are represented by real numbers
• Qualitative correspondence with common sense
• Consistency
– If a conclusion can be reasoned in more than one way, then every way should
lead to the same answer.
– The robot always takes into account all relevant evidence.
– Equivalent states of knowledge are represented by equivalent plausibility
assignments.
Consequence: Belief functions (e.g. b(x), b(x|y), b(x, y)) must satisfy the rules of
probability theory, including Bayes rule. (see Jaynes, Probability Theory: The Logic
of Science)

The Dutch Book Theorem

Assume you are willing to accept bets with odds proportional to the stength of your
beliefs. That is, b(x) = 0.9 implies that you will accept a bet:

x is true win ≥ $1
x is false lose $9

Then, unless your beliefs satisfy the rules of probability theory, including Bayes rule,
there exists a set of simultaneous bets (called a “Dutch Book”) which you are
willing to accept, and for which you are guaranteed to lose money, no matter
what the outcome.

The only way to guard against Dutch Books to to ensure that your beliefs are
coherent: i.e. satisfy the rules of probability.

Asymptotic Certainty

Assume that data set Dn, consisting of n data points, was generated from some
true θ∗, then under some regularity conditions, as long as p(θ∗) > 0

lim p(θ|Dn) = δ(θ − θ∗)
n→∞

In the unrealizable case, where data was generated from some p∗(x) which cannot
be modelled by any θ, then the posterior will converge to

ˆ
lim p(θ|Dn) = δ(θ − θ)
n→∞

ˆ
where θ minimizes KL(p∗(x), p(x|θ)):

ˆ ∗ p∗(x)
θ = argmin p (x) log dx = argmax p∗(x) log p(x|θ) dx
θ p(x|θ) θ

Warning: careful with the regularity conditions, these are just sketches of the theoretical results

Asymptotic Consensus

Consider two Bayesians with diﬀerent priors, p1(θ) and p2(θ),
who observe the same data D.

Assume both Bayesians agree on the set of possible and impossible values of θ:

{θ : p1(θ) > 0} = {θ : p2(θ) > 0}

Then, in the limit of n → ∞, the posteriors, p1(θ|Dn) and p2(θ|Dn) will converge
(in uniform distance between distibutions ρ(P1, P2) = sup|P1(E) − P2(E)|)
E

coin toss demo: bayescoin

Bayesian Occam’s Razor and Model Comparison

Compare model classes, e.g. m and m , using posterior probabilities given D:
p(D|m) p(m)
p(m|D) = , p(D|m) = p(D|θ, m) p(θ|m) dθ
p(D)

Interpretation of the Marginal Likelihood (“evidence”): The probability that
randomly selected parameters from the prior would generate D.

too simple
Model classes that are too simple are
unlikely to generate the data set.

P(D|m)
Model classes that are too complex can
generate many possible data sets, so "just right"
again, they are unlikely to generate that too complex
particular data set at random.
D
All possible data sets of size n

Model structure and overﬁtting:
A simple example: polynomial regression

M=0 M=1 M=2 M=3

40 40 40 40

20 20 20 20

0 0 0 0

−20 −20 −20 −20
0 5 10 0 5 10 0 5 10 0 5 10
M=4 M=5 M=6 M=7

40 40 40 40

20 20 20 20

0 0 0 0

−20 −20 −20 −20
0 5 10 0 5 10 0 5 10 0 5 10

Bayesian Model Comparison: Occam’s Razor at Work

M=0 M=1 M=2 M=3

40 40 40 40
Model Evidence
20 20 20 20 1

0.8
0 0 0 0

0.6

P(Y|M)
−20 −20 −20 −20
0 5 10 0 5 10 0 5 10 0 5 10
0.4
M=4 M=5 M=6 M=7
0.2
40 40 40 40

0
20 20 20 20 0 1 2 3 4 5 6 7
M

0 0 0 0

−20 −20 −20 −20
0 5 10 0 5 10 0 5 10 0 5 10

demo: polybayes

On Choosing Priors

• Objective Priors: noninformative priors that attempt to capture ignorance and
have good frequentist properties.

• Subjective Priors: priors should capture our beliefs as well as possible. They
are subjective but not arbitrary.

• Hierarchical Priors: multiple levels of priors:

p(θ) = dα p(θ|α)p(α)

= dα p(θ|α) dβ p(α|β)p(β) (etc...)

• Empirical Priors: learn some of the parameters of the prior from the data
(“Empirical Bayes”)

Objective Priors

Non-informative priors:

Consider a Gaussian with mean µ and variance σ 2.

The parameter µ informs about the location of the data.
If we pick p(µ) = p(µ − a) ∀a then predictions are location invariant

p(x|x ) = p(x − a|x − a)

But p(µ) = p(µ − a) ∀a implies p(µ) = Unif(−∞, ∞) which is improper.

Similarly, σ informs about the scale of the data, so we can pick p(σ) ∝ 1/σ

Problems: It is hard (impossible) to generalize to all parameters of a complicated
model. Risk of incoherent inferences (e.g. ExEy [Y |X] = Ey [Y ]), paradoxes, and
improper posteriors.

Objective Priors

Reference Priors:

Captures the following notion of noninformativeness. Given a model p(x|θ) we wish
to ﬁnd the prior on θ such that an experiment involving observing x is expected to
provide the most information about θ.

That is, most of the information about θ will come from the experiment rather than
the prior. The information about θ is:

I(θ|x) = − p(θ) log p(θ)dθ − (− p(θ, x) log p(θ|x)dθ dx)

This can be generalized to experiments with n obserations (giving diﬀerent answers)

Problems: Hard to compute in general (e.g. MCMC schemes), prior depends on
the size of data to be observed.

Objective Priors

Jeﬀreys Priors:

Motivated by invariance arguments: the principle for choosing priors should not
depend on the parameterization.

dθ
p(φ) = p(θ)
dφ
p(θ) ∝ h(θ)1/2
∂2
h(θ) = − p(x|θ) 2 log p(x|θ) dx (Fisher information)
∂θ

Problems: It is hard (impossible) to generalize to all parameters of a complicated
model. Risk of incoherent inferences (e.g. ExEy [Y |X] = Ey [Y ]), paradoxes, and
improper posteriors.

Subjective Priors

Priors should capture out beliefs as well as possible.

Otherwise we are not coherent.

End of story!

How do we know our beliefs?
• Think about the problems domain (no black box view of machine learning)
• Generate data from the prior. Does it match expectations?
Even very vague beliefs can be useful.

Hierarchical Priors

p(θ) = dα p(θ|α)p(α)

= dα p(θ|α) dβ p(α|β)p(β)

= dα p(θ|α) dβ p(α|β) dγ p(β|γ)p(γ) (etc...)

In models with many parameters, priors over parameters have hyperparameters.
These in turn can also have priors with hyper-hyperparameters, etc.

A α a,b
ut
B β ν1 ν2 ... νS
Σ0, µ0 xt-1 xt
C γ Λ1 Λ2 ... ΛS α

R a, b sn π
yt Ψ yn
t=1... T(i)
D δ
xn
i=1... n n=1...N

Empirical “Priors”

Consider a hierarchical model with parameters θ and hyperparameters α

p(D|α) = p(D|θ)p(θ|α) dθ

Estimate hyperparameters from the data

α = argmaxp(D|α)
ˆ (level II ML)
α

Prediction:
p(x|D, α) =
ˆ p(x|θ)p(θ|D, α) dθ
ˆ

Advantages: Robust—overcomes some limitations of mis-speciﬁcation of the prior.

Problem: Double counting of evidence / overﬁtting.

Exponential Family and Conjugate Priors

p(x|θ) in the exponential family if it can be written as:

p(x|θ) = f (x)g(θ) exp{φ(θ) s(x)}

φ vector of natural parameters
s(x) vector of suﬃcient statistics
f and g positive functions of x and θ, respectively.

The conjugate prior for this is

p(θ) = h(η, ν) g(θ)η exp{φ(θ) ν}

where η and ν are hyperparameters and h is the normalizing function.

The posterior for N data points is also conjugate (by deﬁnition), with
hyperparameters η + N and ν + n s(xn). This is computationally convenient.

p(θ|x1, . . . , xN ) = h η + N, ν + s(xn) g(θ)η+N exp φ(θ) (ν + s(xn))
n n

Computing Marginal Likelihoods can be
Computationally Intractable

Observed data y, hidden variables x, parameters θ, model class m.

p(y|m) = p(y|θ, m) p(θ|m) dθ

• This can be a very high dimensional integral.

• The presence of latent variables results in additional dimensions that need to
be marginalized out.

p(y|m) = p(y, x|θ, m) p(θ|m) dx dθ

• The likelihood term can be complicated.

Approximation Methods for Posteriors and Marginal Likelihoods

• Laplace approximation

• Bayesian Information Criterion (BIC)

• Variational approximations

• Expectation Propagation (EP)

• Markov chain Monte Carlo methods (MCMC)

• Exact Sampling

• ...

Note: there are other deterministic approximations; we won’t review them all.

Bayesian Information Criterion (BIC)

BIC can be obtained from the Laplace approximation:

ˆ ˆ d 1
ln p(y|m) ≈ ln p(θ|m) + ln p(y|θ, m) + ln 2π − ln |A|
2 2

by taking the large sample limit (n → ∞) where n is the number of data points:

ˆ m) − d ln n
ln p(y|m) ≈ ln p(y|θ,
2

Properties:
• Quick and easy to compute
• It does not depend on the prior
• We can use the ML estimate of θ instead of the MAP estimate
• It is equivalent to the MDL criterion
• Assumes that as n → ∞ , all the parameters are well-determined (i.e. the model
is identiﬁable; otherwise, d should be the number of well-determined parameters)
• Danger: counting parameters can be deceiving! (c.f. sinusoid, inﬁnite models)

Lower Bounding the Marginal Likelihood
Variational Bayesian Learning

Let the latent variables be x, observed data y and the parameters θ.
We can lower bound the marginal likelihood (Jensen’s inequality):

ln p(y|m) = ln p(y, x, θ|m) dx dθ

p(y, x, θ|m)
= ln q(x, θ) dx dθ
q(x, θ)
p(y, x, θ|m)
≥ q(x, θ) ln dx dθ.
q(x, θ)

Use a simpler, factorised approximation to q(x, θ) ≈ qx(x)qθ (θ):

p(y, x, θ|m)
ln p(y|m) ≥ qx(x)qθ (θ) ln dx dθ
qx(x)qθ (θ)
= Fm(qx(x), qθ (θ), y).

Variational Bayesian Learning . . .

Maximizing this lower bound, Fm, leads to EM-like iterative updates:

(t+1) (t)
qx (x) ∝ exp ln p(x,y|θ, m) qθ (θ) dθ E-like step

(t+1) (t+1)
qθ (θ) ∝ p(θ|m) exp ln p(x,y|θ, m) qx (x) dx M-like step

Maximizing Fm is equivalent to minimizing KL-divergence between the approximate
posterior, qθ (θ) qx(x) and the true posterior, p(θ, x|y, m):

qx(x) qθ (θ)
ln p(y|m) − Fm(qx(x), qθ (θ), y) = qx(x) qθ (θ) ln dx dθ = KL(q p)
p(θ, x|y, m)

In the limit as n → ∞, for identiﬁable models, the variational lower bound
approaches the BIC criterion.

The Variational Bayesian EM algorithm

EM for MAP estimation Variational Bayesian EM
Goal: maximize p(θ|y, m) w.r.t. θ Goal: lower bound p(y|m)
E Step: compute VB-E Step: compute

(t+1)
qx (x) = p(x|y, θ (t)) (t+1)
qx ¯ (t)
(x) = p(x|y, φ )

M Step: VB-M Step:
(t+1) (t+1) (t+1) (t+1)
θ =argmax qx (x) ln p(x, y, θ) dx qθ (θ) ∝ exp qx (x) ln p(x, y, θ) dx
θ

Properties:
• Reduces to the EM algorithm if qθ (θ) = δ(θ − θ ∗).
• Fm increases monotonically, and incorporates the model complexity penalty.
• Analytical parameter distributions (but not constrained to be Gaussian).
• VB-E step has same complexity as corresponding E step.
• We can use the junction tree, belief propagation, Kalman ﬁlter, etc, algorithms
in the VB-E step of VB-EM, but using expected natural parameters, φ. ¯

Variational Bayesian EM

The Variational Bayesian EM algorithm has been used to approximate Bayesian
learning in a wide range of models such as:
• probabilistic PCA and factor analysis
• mixtures of Gaussians and mixtures of factor analysers
• hidden Markov models
• state-space models (linear dynamical systems)
• independent components analysis (ICA)
• discrete graphical models...
The main advantage is that it can be used to automatically do model selection
and does not suﬀer from overﬁtting to the same extent as ML methods do.

Also it is about as computationally demanding as the usual EM algorithm.

See: www.variational-bayes.org

mixture of Gaussians demo: run simple

Expectation Propagation (EP)
Data (iid) D = {x(1) . . . , x(N )}, model p(x|θ), with parameter prior p(θ).
N
1
The parameter posterior is: p(θ|D) = p(θ) p(x(i)|θ)
p(D) i=1
N N
We can write this as product of factors over θ: p(θ) p(x(i)|θ) = fi(θ)
i=1 i=0
def def
where f0(θ) = p(θ) and fi(θ) = p(x(i)|θ) and we will ignore the constants.

N
def ˜
We wish to approximate this by a product of simpler terms: q(θ) = fi(θ)
i=0
N N
min KL fi(θ) ˜
fi(θ) (intractable)
q(θ)
i=0 i=0
˜
min KL fi(θ) fi(θ) (simple, non-iterative, inaccurate)
˜
fi (θ)

min KL fi(θ) ˜ ˜
fj (θ) fi(θ) ˜
fj (θ) (simple, iterative, accurate) ← EP
˜
fi (θ)
j=i j=i

Expectation Propagation
Input f0(θ) . . . fN (θ)
˜ ˜ ˜
Initialize f0(θ) = f0(θ), fi(θ) = 1 for i > 0, q(θ) = i fi(θ)
repeat
for i = 0 . . . N do
q(θ) ˜
Deletion: qi(θ) ← = fj (θ)
˜
fi(θ) j=i
˜
Projection: finew (θ) ← arg min KL(fi(θ)qi(θ) f (θ)qi(θ))
f (θ)
˜
Inclusion: q(θ) ← finew (θ) qi(θ)
end for
until convergence

The EP algorithm. Some variations are possible: here we assumed that f0 is in
the exponential family, and we updated sequentially over i. The names for the steps
(deletion, projection, inclusion) are not the same as in (Minka, 2001)
• Minimizes the opposite KL to variational methods
• ˜
fi(θ) in exponential family → projection step is moment matching
• Loopy belief propagation and assumed density ﬁltering are special cases
• No convergence guarantee (although convergent forms can be developed)

An Overview of Sampling Methods

Monte Carlo Methods:
• Simple Monte Carlo
• Rejection Sampling
• Importance Sampling
• etc.
Markov Chain Monte Carlo Methods:
• Gibbs Sampling
• Metropolis Algorithm
• Hybrid Monte Carlo
• etc.
Exact Sampling Methods

Markov chain Monte Carlo (MCMC) methods

Assume we are interested in drawing samples from some desired distribution p∗(θ),
e.g. p∗(θ) = p(θ|D, m).

We define a Markov chain:
θ0 → θ1 → θ2 → θ3 → θ4 → θ5 . . .

where θ0 ∼ p0(θ), θ1 ∼ p1(θ), etc, with the property that:

pt(θ ) = pt−1(θ) T (θ → θ ) dθ

where T (θ → θ ) is the Markov chain transition probability from θ to θ .

We say that p∗(θ) is an invariant (or stationary) distribution of the Markov chain
defined by T iff:
p∗(θ ) = p∗(θ) T (θ → θ ) dθ

Markov chain Monte Carlo (MCMC) methods

We have a Markov chain θ0 → θ1 → θ2 → θ3 → . . . where θ0 ∼ p0(θ), θ1 ∼ p1(θ),
etc, with the property that:

pt(θ ) = pt−1(θ) T (θ → θ ) dθ

where T (θ → θ ) is the Markov chain transition probability from θ to θ .
A useful condition that implies invariance of p∗(θ) is detailed balance:

p∗(θ )T (θ → θ) = p∗(θ)T (θ → θ )

MCMC methods deﬁne ergodic Markov chains, which converge to a unique
stationary distribution (also called an equilibrium distribution) regardless of the
initial conditions p0(θ):
lim pt(θ) = p∗(θ)
t→∞

Procedure: deﬁne an MCMC method with equilibrium distribution p(θ|D, m), run
method and collect samples. There are also sampling methods for p(D|m).

On-screen viewing permitted. Printing not permitted. http://www.cambridge.org/0521642981
See http://www.inference.phy.cam.ac.uk/mackay/itila/ for links.

Exact Sampling
a.k.a. perfect simulation, coupling from the past
250 250

• Coupling: running multiple Markov chains (MCs) using the same
random seeds. E.g. imagine starting a Markov chain at each possible
200 200
value of the state (θ).
• Coalescence: if two coupled MCs end up at the same state at time
t, then they will forever follow the same path.
150 150 • Monotonicity: Rather than running an MC starting from every
state, ﬁnd a partial ordering of the states preserved by the coupled
transitions, and track the highest and lowest elements of the partial
100 100 ordering. When these coalesce, MCs started from all initial states
would have coalesced.
• Running from the past: Start at t = −K in the past, if highest and
50 50 lowest elements of the MC have coalesced by time t = 0 then all
MCs started at t = −∞ would have coalesced, therefore the chain
must be at equilibrium, therefore θ0 ∼ p∗(θ).
0 0

0 15 20 0 5 10 15 20 0
Bottom Line This procedure, when it produces a sample, will produce
5 10 15 20

(from MacKay 2003) one from the exact distribution p∗(θ).

BREAK

Ideally, but I’m sure we’re running really late, so we probably only have 1 hour left...

Further Topics

• Feature Selection and ARD

• Bayesian Discriminative Learning (BPM vs SVM)

• From Parametric to Nonparametric Methods
– Gaussian Processes
– Dirichlet Process Mixtures
– Other Non-parametric Bayesian Methods

• Bayesian Decision Theory and Active Learning

• Bayesian Semi-supervised Learning

Feature Selection

Example: classification

input x = (x1, . . . , xD ) ∈ RD
output y ∈ {+1, −1}

2D possible subsets of relevant input features.

One approach, consider all models m ∈ {0, 1}D and find

m = argmax p(D|m)
ˆ
m

Problems: intractable, overfitting, we should really average

Feature Selection

• Why are we doing feature selection?

• What does it cost us to keep all the features?

• Usual answer (overﬁtting) does not apply to fully Bayesian methods, since they
don’t involve any ﬁtting.

• We should only do feature selection if there is a cost associated with measuring
features or predicting with many features.

Note: Radford Neal won the NIPS feature selection competition using Bayesian
methods that used 100% of the features.

Feature Selection: Automatic Relevance Determination
Bayesian neural network

Data: D = {(x(i), y (i))}N = (X, y)
i=1
Parameters (weights): θ = {{wij }, {vk }}

prior p(θ|α)
posterior p(θ|α, D) ∝ p(y|X, θ)p(θ|α)
evidence p(y|X, α) = p(y|X, θ)p(θ|α) dθ
prediction p(y |D, x , α) = p(y |x , θ)p(θ|D, α) dθ

Automatic Relevance Determination (ARD):

Let the weights from feature xd have variance α−1: p(wdj |αd) = N (0, α−1)

αd → ∞ variance → 0 weights → 0 (irrelevant)
Let’s think about this: αd ∞ ﬁnite variance weight can vary (relevant)

ˆ
ARD: optimize α = argmaxp(y|X, α).
α

During optimization some αd will go to ∞, so the model will discover irrelevant
inputs.

Bayesian Discriminative Modeling

Terminology for classification with inputs x and classes y:
• Generative Model: models prior p(y) and class-conditional density p(x|y)
• Discriminative Model: directly models the conditional distribution p(y|x) or
the class boundary e.g. {x : p(y = +1|x) = 0.5}
Myth: Bayesian Methods = Generative Models
For example, it is possible to define Bayesian kernel classifiers (e.g. Bayes point
machines, and Gaussian processes) analogous to support vector machines (SVMs).
3 3 3
BPM BPM BPM
2 2 2

1 1 1

0 0 0

−1 −1 −1

−2 −2 −2

−3
SVM −3
SVM −3
SVM
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

(figure adapted from Minka, 2001)

Parametric vs Nonparametric Models

Terminology (roughly):
• Parametric Models have a finite fixed number of parameters θ, regardless of
the size of the data set. Given θ, the predictions are independent of the data D:

p(x, θ|D) = p(x|θ) p(θ|D)

The parameters are a finite summary of the data. We can also call this
model-based learning (e.g. mixture of k Gaussians)

• Non-parametric Models allow the number of “parameters” to grow with the
data set size, or alternatively we can think of the predictions as depending on the
data, and possible a usually small number of parameters α

p(x|D, α)

We can also call this memory-based learning (e.g. kernel density estimation)

From Parametric to Nonparametric

Consider data set D = {x(1), . . . , x(n)}, new data point x, parameters θ, and
hyperparameters α

By integrating out parameters one can seemingly turn a parametric method into a
non-parametric method:

p(x, D, θ|α) = p(x|θ)p(D|θ)p(θ|α) (parametric)
p(x|D, α) ∝ p(x|θ)p(D|θ)p(θ|α) dθ (non-parametric)

hyperparameters hyperparameters

parameters

data ... ...

A key question: are a fixed finite number of sufficient statistics of the data needed
to make predictions?

Nonparametric Bayesian Methods (Infinite Models)

We ought not to limit the complexity of our model a priori (e.g. number of hidden
states, number of basis functions, number of mixture components, etc) since we
don’t believe that the real data was actually generated from a statistical model with
a small number of parameters.

Therefore, regardless of how much training data we have, we should consider models
with as many parameters as we can handle computationally.

Here there is no model order selection task:
• No need to compare marginal likelihoods to select model order (which is often
difficult).
• No need to use Occam’s razor to limit the number of parameters in the model.
In fact, we may even want to consider doing inference in models with an infinite
number of parameters...

Gaussian Processes for Regression

Two ways of understanding Gaussian processes (GPs)...
• Starting from multivariate Gaussians

• Starting from linear regression

...from multivariate Gaussians to GPs...

univariate Gaussian density for t

2 −1/2 t2
p(t) = (2πσ ) exp − 2
2σ

multivariate Gaussian density for t = (t1, t2, t3, . . . tN )

1
p(t) = |2πΣ|−1/2 exp − t Σ−1t
2

Σ is an N × N covariance matrix.

Imagine that Σij depends on i and j and we plot samples of t as if they were
functions...

gpdemogen and gpdemo

...from linear regression to GPs...

• Linear regression with inputs xi and outputs ti: ti = wd xid + i
d

• Linear regression with basis functions (“kernel trick”): ti = wd φd(xi) + i
d

• Bayesian linear regression with basis functions:

wd ∼ N (0, βd) [ independent of w , ∀ = d], i ∼ N (0, σ 2)

• Integrating out the weights, wd, we find:
def
E[ti] = 0, E[titj ] = Cij = βd φd(xi) φd(xj ) + δij σ 2
d

def
This is a Gaussian process with covariance function C(xi, xj ) = Cij .

This Gaussian process has a finite number of basis functions. Many useful GP
covariance functions correspond to infinitely many basis functions.

Gaussian Process Regression

A Gaussian Process (GP) places a prior directly on the space of functions such that
at any ﬁnite selection of points x(1), . . . , x(N ) the corresponding function values
t(1), . . . , t(N ) have a multivariate Gaussian distribution.

Samples from the Prior Samples from the Posterior

2 2
output, t(x)

output, t(x)
0 0

−2 −2

−2 0 2 −2 0 2
input, x input, x
The covariance between two function values t(i) and t(j) under the prior is given
by the covariance function C(x(i), x(j)), which typically decays monotonically with
x(i) − x(j) , encoding smoothness.

GPs are “Bayesian Kernel Regression Machines”

Using Gaussian Processes for Classiﬁcation
Binary classiﬁcation problem: Given a data set D = {(x(n), y (n))}N , where
n=1
(n)
y ∈ {−1, +1}, infer class label probabilities at new points.
1.5

y = +1
1 y = −1

5
0.5
f
f 0
0

−5
1 1
−0.5
0.5 0.5
0 0
−1 y −0.5 −0.5 x
−1 −0.5 0 0.5 1
x −1 −1

There are many ways to relate function values f (n) to class probabilities:
1


 1+exp(−yf ) sigmoid (logistic)
Φ(yf ) cumulative normal (probit)

p(y|f ) =

 H(yf ) threshold
+ (1 − 2 )H(yf ) robust threshold


Dirichlet Process Mixtures (Inﬁnite Mixtures)

Consider using a ﬁnite mixture of K components to model a data set
D = {x(1), . . . , x(n)}

K
p(x(i)|θ) = πj pj (x(i)|θ j )
j=1
K
= P (s(i) = j|π) pj (x(i)|θ j , s(i) = j)
j=1

Distribution of indicators s = (s(1), . . . , s(n)) given π is multinomial
K n
n def
P (s(1), . . . , s(n)|π) = πj j , nj = δ(s(i), j) .
j=1 i=1

Assume mixing proportions π have a given symmetric conjugate Dirichlet prior
K
Γ(α) α/K−1
p(π|α) = πj
Γ(α/K)K j=1

Dirichlet Process Mixtures (Inﬁnite Mixtures) - II

Distribution of indicators s = (s(1), . . . , s(n)) given π is multinomial
K n
n def
P (s(1), . . . , s(n)|π) = πj j , nj = δ(s(i), j) .
j=1 i=1

Mixing proportions π have a symmetric conjugate Dirichlet prior
K
Γ(α) α/K−1
p(π|α) = πj
Γ(α/K)K j=1

Integrating out the mixing proportions, π, we obtain
K
Γ(α) Γ(nj + α/K)
P (s(1), . . . , s(n)|α) = dπ P (s|π)P (π|α) =
Γ(n + α) j=1 Γ(α/K)

Dirichlet Process Mixtures (Infinite Mixtures) - III
K
Γ(α) Γ(nj + α/K)
Starting from P (s|α) =
Γ(n + α) j=1 Γ(α/K)
Conditional Probabilities: Finite K
n−i,j + α/K
P (s(i) = j|s−i, α) =
n−1+α
def
where s−i denotes all indices except i, and n−i,j = =i δ(s( ), j)

DP: more populous classes are more more likely to be joined

Conditional Probabilities: Infinite K
Taking the limit as K → ∞ yields the conditionals
 n
−i,j
 n−1+α j represented
P (s(i) = j|s−i, α) =
α

n−1+α all j not represented

Left over mass, α, ⇒ countably infinite number of indicator settings.
Gibbs sampling from posterior of indicators is often easy!

Other Non-parametric Bayesian Models

• Infinite Hidden Markov models

• Hierarchical Dirichlet Processes

• Dirichlet Diffusion Trees

• Infinite mixtures of Gaussian Processes

• ...

Bayesian Decision Theory

Bayesian decision theory deals with the problem of making optimal decisions—that
is, decisions or actions that minimize an expected loss.
• Let’s say we have a choice of taking one of k possible actions a1 . . . ak .
• Assume that the world can be in one of m diﬀerent states s1, . . . , sm.
• If we take action ai and the world is in state sj we incur a loss ij
• Given all the observed data D and prior background knowledge B , our beliefs
about the state of the world are summarized by p(s|D, B).
• The optimal action is the one which is expected to minimize loss (or maximize
utility):
m
a∗ = argmin ij p(sj |D, B)
ai
j=1

Bayesian sequential decision theory (statistics)
Optimal control theory (engineering)
Reinforcement learning (computer science / psychology)

Bayesian Active Learning

Active Learning is a special case of Bayesian Decision Theory

An example:
• Consider an active classification problem with a labeled data set
D = {(x(1), y (1)), . . . , (x(n), y (n))} and a pool of unlabeled data points
Du = {x(n+1), . . . , x(n+m)}.
• Assume there is a cost associated with finding the true label of a point.
• The action is picking which unlabeled point to find the true label for. The
remaining points will be labeled according to most probable predicted label after
including this true label.
• The state of the world is the true labels of all the points.
• The beliefs are p(y (n+1), . . . , y (n+m)|D , Du, B)
• The loss is the misclassification loss on the remaining points and the loss due to
the cost of labeling the point.

Bayesian Semi-supervised Learning

In semi-supervised learning you have a labeled data set
D = {(x(1), y (1)), . . . , (x(n), y (n))} and an unlabeled data set
Du = {x(n+1), . . . , x(n+m)}. Often m n.

Goal: to learn a model p(y|x) (e.g. a classifier, y ∈ {±1})

Question: how should knowledge about p(x) from Du affect the classifier p(y|x)?

Answer: it all depends on your priors!

Two Bayesian approaches:
• Generative: Express your beliefs about the generative process p(y) and p(x|y)—
this induces a relationship between p(x) and p(y|x).

• Discriminative: Directly express some prior that relates parameters of p(y|x) to
the parameters in p(x). One simple example is the notion that the decision
boundary should prefer to go through regions of low density.

Reconciling Bayesian and Frequentist Views

Frequentist theory tends to focus on minimax performance of methods – i.e.
what is the worst case performance if the environment is adversarial. Frequentist
methods often optimize some penalized cost function.

Bayesian methods focus on expected loss under the posterior. Bayesian methods,
in theory, do not make use of optimization, except at the point at which decisions
are to be made.

There are some reasons why frequentist procedures are useful to Bayesians:
• Communication: If Bayesian A wants to convince Bayesians B, C, and D of the validity of some
inference (or even non-Bayesians) then she must determine that not only does this inference
follows from prior pA but also would have followed from pB , pC and pD , etc. For this reason it’s
useful sometimes to ﬁnd a prior which has good frequentist (worst-case) properties, even though
acting on the prior would not be coherent with our beliefs.
• Robustness: Priors with good frequentist properties can be more robust to mis-speciﬁcations of
the prior. Two ways of dealing with robustness issues are to make sure that the prior is vague
enough, and to make use of a loss function to penalize costly errors.

also, recently, PAC-Bayesian frequentist bounds on Bayesian procedures.

Limitations and Criticisms of Bayesian Methods

• It is subjective

• It is hard to come up with a prior, the assumptions are usually wrong.

• The closed world assumption: need to consider all possible hypotheses for the
data before observing the data

• It is computationally demanding

Summary

• Introduce Foundations • Approximation Tools
– Some canonical problems: classiﬁcation, – Laplace’s Approximation
regression, density estimation, coin toss – Bayesian Information Criterion (BIC)
– Representing beliefs and the Cox axioms – Variational Approximations
– The Dutch Book Theorem – Expectation Propagation
– Asymptotic Certainty and Consensus – MCMC
– Occam’s Razor and Marginal Likelihoods – Exact Sampling
– Choosing Priors • Advanced Topics
∗ Objective Priors: – Feature Selection and ARD
Noninformative, Jeﬀreys, Reference – Bayesian Discriminative Learning (BPM vs SVM)
∗ Subjective Priors – From Parametric to Nonparametric Methods
∗ Hierarchical Priors ∗ Gaussian Processes
∗ Empirical Priors ∗ Dirichlet Process Mixtures
∗ Conjugate Priors ∗ Other Non-parametric Bayesian Methods
• The Intractability Problem – Bayesian Decision Theory and Active Learning
– Bayesian Semi-supervised Learning
– Reconciling Bayesian and Frequentist Views
– Limitations and Criticisms of Bayesian Methods
– Discussion

Conclusions

• Bayesian methods provide a coherent framework for doing inference under
uncertainty and for learning from data

• The ideas are simple, although execution can be hard

• There are still many open research directions

http://www.gatsby.ucl.ac.uk/∼zoubin
(for more resources, also to contact me
if interested in a PhD or postdoc)

Thanks for your patience!

icml2004 tutorial on bayesian methods for machine learning

More Related Content

What's hot (18)

Viewers also liked (7)

Similar to icml2004 tutorial on bayesian methods for machine learning (20)

More from zukun (20)

Recently uploaded (20)

icml2004 tutorial on bayesian methods for machine learning