Transcript Statistical Learning

11/16: After Sanity Test
Post-mortem
Project presentations in the last 2-3 classes
Start of Statistical Learning
Sanity Test..
• Max: 52.5 Min: 6 Avg: 24.6 Stdev: 14.8
– Including those sitting-in: Avg: 24.42; Stdev: 13.8
»
»
»
»
»
»
»
»
70+: 0
60-70: 0
50-60: 2
40-50: 0
30-40: 3
20-30: 4
10-20: 5
0-10: 2
• Students with low scores have to re-do the test at
home (with access to notes, web etc.). An eventual
score of less than 45 will be viewed as failing on the
content
P(hi ) is called the hypothesis prior
Nothing special about “learning” – just vanilla probabilistic inference
Where is the
hypothesis prior?
i.i.d. assumption
How did this prediction come about?
Which hypothesis did we use?
The analogy with diagnosis
Medical diagnosis
Full Bayesian learning
•
•
•
•
•
•
Given symptoms of a patient, predict
whether she will have other symptoms
(such as death…)
Can try predicting directly from
symptoms (is what we did before the
advent of medicine)
But we normally assume that diseases
cause symptoms.. Thus we want to first
figure out the disease and then predict
other symptoms
Diseases have prior probabilities (in
fact, the “ignored prior” fallacy is the
main reason for internet induced
hypochondria)
Given the symptoms, we compute the
posterior on the diseases, and then use
that to predict other symptoms
•
•
•
•
Given training data, predict test
data
Can try predicting test data directly
from training data (e.g. k-NN)
But we normally assume that
hypothesis explain data. Thus we
want to first figure out the
hypotheses causing the data and
then using them predict test data
Hypotheses have prior probabilities
(as to how likely they are—
independent of the data being seen
right now).
Given the data, we compute the
posterior on the hypotheses, and
then use that to predict test data
How many
Equivalently minimize
- log P(d|hi) – log P(hi)
Additional bits required to specify d
Bits required to specify hi
Why should P(hi ) be low for complex hypotheses?
--connection to MDL principle
--because “statisticians” distrust priors
(and want the data to speak for itself)
When will ML hit roadblock? Small data
Should AI also distrust priors? Priors can encode background knowledge..
(There is even evidence that human brain uses priors)
http://web.mit.edu/cocosci/Papers/significance.pdf
A technical head ache with priors: What if the
posterior keeps changing parametrically?
Conjugate priors..
11/18
Density Estimation
• The general task of learning a probability model, given data that are
assumed to be generated from that model
• Given data D whose instances are made-up of attributes x that are
distributed according to P*(x), we want to learn an estimate P’ to
P* such that the distance between P* and P’ is minimized
– Distance between distributions is typically measured using KL
Divergence
– D(P*||P’) = EP*[log P*(x)/P’(x)]
• But alas, we don’t know P* 
= EP* log P*(x) - EP* log P’(x)
• The first term is constant and can be ignored in comparing two estimates P’
and P’’
• But how do we get the second term? Since the data instances are drawn
from P*, a P’ that maximizes their log likelihood is the best.
• If data are drawn i.i.d, then their joint likelihood is a product and log
terms over all data instances can be summed..
Bias-Variance Tradeoff
(learning with bias vs. regularization)
• So, we want to get the distribution P’ that maximizes EP* log P’(x)
– Learning is just optimization!
• But how do we select the candidate space of distributions (hypotheses)?
• [Bias problem] If the class of distributions we consider is too
small/inflexible (“highly biased”), then the best we get may still be too far
from P*
• [Variance problem] If the class of distributions considered is too
large/expressive, then small random fluctuations in the choice of data can
radically change the properties of the model, thus exhibiting high variance
on the test data
• Standard solutions:
1.
2.
3.
Limit attention to a “reasonable class of distributions” (e.g. Naïve Bayes)
Allow a large class of distributions, but “penalize” the more complex ones
[Also called “regularization”].
Combination of both..
A class of distributions can be defined by restricting the class of graphical models
(e.g. only naïve bayes models) or CPDs (only noisy-or or conditional Gaussians)
allowed in the hypothesis space
Generative vs. Discriminative Learning
• Often, we are really more interested in predicting only
a subset of the attributes given the rest.
– E.g. we have data attributes split into subsets X and Y, and
we are interested in predicting Y given the values of X
• You can do this by either by
– learning the joint distribution P(X, Y) [Generative learning]
– or learning just the conditional distribution P(Y|X)
[Discriminative learning]
• Often a given classification problem can be handled
either generatively or discriminatively
– E.g. Naïve Bayes and Logistic Regression
• Which is better?
Generative vs. Discriminative
P(y)P(x|y) =
P(y,x)
= P(x)P(y|x)
Generative Learning
Discriminative Learning
• More general (after all if you have
P(Y, X) you can predict Y given X
as well as do other inferences
• More to the point (if what you
want is P(Y|X), why bother with
P(Y,X) which is after all P(Y|X)
*P(X) and thus models the
dependencies between X’s also?
• Since we don’t need to model
dependencies among X, we don’t
need to make any independence
assumptions among them. So, we
can merrily use highly correlated
features..
– You can predict jokes as well as
make them up (or predict spam
mails as well as generate them)
• In trying to learn P(Y,X), we are
often forced to make many
independence assumptions both
in Y and X—and these may be
wrong..
– Interestingly, this type of high bias
can help generative techniques
when there is too little data
– Interestingly, this freedom can hurt
discriminative learners when there
is too little data (as over fitting is
easy)
Bayes networks are not well suited for discriminative learning; Markov Networks are
--thus Conditional Random Fields are basically MNs doing discriminative learning
--Logistic regression can be seen as a simple CRF
Taxonomy of (statistical) Learning
Tasks
Model constraints
Observability of data
• Type of network being learned
• Complete data
– Bayes Network vs. Markov
network
• Topology given; CPTs to be
learned
• Only relevant attributes are
given; need to learn topology
as well as CPTs
– Tricky part for MLE is that
increasing the connectivity of a
network cannot reduce
likelihood
• We don’t know what the
relevant attributes are
– Each data instance gives the values
of each of the attributes
• Incomplete data
– Some of the data instances might
be missing the values for some of
the attributes
• Hidden attributes (variables)
– None of the data instances have
values for some of the attributes
(which often correspond to
“intermediate” concepts which
help improve the sparsity of
network. E.g. “syndromes” which
connect symptoms to diseases; or
class variables in mixture models
Sample complexity linearly varies with # parameters to be learned,
and #parameters vary exponentially with # edges in the graphical model
11/23
Steps in ML based learning
1.
Write down an expression for the likelihood of the data as a function of
the parameter(s)
Assume i.i.d. distribution
2.
3.
Write down the derivative of the log likelihood with respect to each
parameter
Find the parameter values such that the derivatives are zero
There are two ways this step can become complex
Individual (partial) derivatives lead to non-linear functions (depends on
the type of distribution the parameters are controlling; binomial is a very
easy case)
Individual (partial) derivatives will involve more than one parameter (thus
leading to simultaneous equations)
 In general, we will need to use continuous function optimization techniques

One idea is to use gradient descent to find the point where the derivative goes to zero. But
for gradient descent to find global optimum, we need to know for sure that the function we
are optimizing has a single optimum (this is why convex functions are important. If the
likelihood is a convex function, then gradient descent will be guaranteed to find the global
minimum).
Note that for us, data
are 2-attribute tuples
[Flavor, Wrapper]
No entanglement of parameters for complete data
for Bayes nets with known topology and tabular CPTs
Specifically, each partial derivative will involve only one parameter
i.e., each partial derivative contains only one parameter
so you are solving single variable equations rather than simultaneous
equations.
doesn’t hold for markov nets ; doesn’t also hold for Bayes nets where CPDs induce
direct parameter dependencies
Celebrating ease of learning for bayes
nets with complete data!
• So we just noted that if we know the topology
of the Bayes net, and we have complete data
then the parameters are un-entangled, and
can be learned separately from just data
counts.
• Questions: How big a deal is this?
– Can we have complete data?
– Can we have known topology?
Learning the parameters of a Gaussian
Case Study: Learning Bayes Net
models for Relational database tables
•
Consider a relational table in RDBMS with n
attributes
–
•
•
•
Say an employee table giving the age,
position, salary etc of each employee
Suppose we want to learn the generative
model underlying it
Suppose we were able to hypothesize the
topology
–
•
We might be able to do so if (a) we know the
domain or (b) we know some of the causal
dependencies in the data
If the relational table is “complete” –i.e..,
every tuple gives the value for every
attribute, (which is the standard RDBMS
model), then learning the parameters of this
network is easy!
Now, suppose the table is slightly “dirty”—in
that there are tuples that have some missing
values for some of the attributes
–
•
If only a small percent of the tuples are
incomplete, then we can
–
–
•
Say, some of the employee tuples are missing
age information, others are missing salary
information etc.
1. Learn the model using the complete tuples
2. predict the null values in the dirty tuples
using the learned model
But, if a non-trivial percent of the tuples are
incomplete, then, we might want to continue
for step 2 above by
–
3. Now that we have “completed” all the
incomplete tuples, we have fully complete
data. Learn the model with this Completed
data; and see if it is any better
•
•
A model is better if it provides a higher
likelihood for the observed data
But why stop here? Continue and use the
new model to re-predict the missing values,
and iterate
–
This is the basic idea of EM (Expectation
Maximization) algorithm
What if the best generative model contains
attribute that are not mentioned in the table?
•
•
In the previous relational table scenario, we
assumed that some of the tuples are missing some
of the attribute values.
What if all tuples are missing some attribute values?
–
•
Can we still use EM?
–
–
E.g. Educational level of the employee can be an
attribute that is missing in the current table.
This is like having an attribute column whose value
is not known for any of the tuples
–
–
•
•
But why would we do it?
–
Why would we do it?
Can we still use EM?
–
–
•
Surprisingly, it turns out yes. In the earlier scenario, we
used the complete tuples for setting up the initial model,
but then used it to complete the data, and loop
There is no reason why we should initialize using
complete data. We can initialize the model (parameters)
randomly, and still do the EM looping!
Given a complete relational table, such as the employee
one, why would we start hypothesizing hidden
attributes?
Because the right hypothesis on the hidden attribute can
significantly reduce the number of parameters
For example, the educational level of the employee
might cluster employees into “PhD” folks (who
presumably have high salaries, interesting positions, and
mature ages), and “non-PhD” folks (who presumably
have low salaries, green-behind-the-ears ages, and
assembly programming kind of jobs), and in each cluster
the distributions of the attribute values are different (as
described above)
So,
–
–
Hypothesizing hidden attributes reduces the parameters
to be estimated, but makes their estimation hard
Not hypothesizing them allows us to deal with complete
data, but might require exponentially many parameters
to be learned (from the same data—making the
parameters, while easy to estimate, pretty worthless in
terms of accuracy.
Missing data (but no hidden variables)
Involves Bayes Net inference; can get by with approximate inference
Which is a more representative Obama
Picture?
11/25
Hidden Variable: Your party affiliation (liberal vs. conservative)
Study by Eugene Caruso, To be published in PNAS
Where P(Xj|Ci) can be
any distribution…
0. Initialize the parameters randomly
Loop
Inference
Candy Example
Start with 1000 samples
Initialize parameters as
Why does EM Work?
The “size of the step” is
determined adaptively
by where the max of the
lowerbound is..
--In contrast, gradient descent
requires a stepsize parameter
--Newton Raphson requires
second derivative..
Log of Sums don’t have easy
closed form optima; use Jensen’s
inequality and focus on Sum of logs
which will be a lower bound
Structure (Topology) Learning
• Search over different network topologies
• Question: How do we decide which topology is
better?
– Idea 1: Check if the independence relations posited by
the topology actually hold
– Idea 2: Consider which topology agrees with the data
more (i.e., provides higher likelihood)
• But need to be careful--increasing edges in a network cannot
reduce likelihood
– Idea 3: Need to penalize complexity of the network (either using
prior on network topologies, or using syntactic complexity
measures)
11/30
Problems with ML
and Bayesian Learning..
• ML based learning is
unable to take the size of
the data into account (1/3
is the same as 1M/3M)
• We however tend to start
with a prior, and are less
willing to change the prior
unless shown enough
evidence
– Bayesian learning can
handle this..
If a thumbtack came
up heads once when
you tossed it 3 times,
what is the probability
that it will come up
heads the next time?
Now, a coin came up
heads once when
you tossed it heads
three times.
What do you think
is the probability that
it will come up heads
next time?
How about if it came up
heads 1million times
in 3 million trials?
Bayesian Learning (for coin toss..)
• Let q be the probability that the coin comes
heads
– Each different value of q is a different
hypothesis
– So P(h) –the hypothesis prior– can be specified
by specifying P(q)
• Starting with prior on q, we just need to
compute the posterior
• Challenge: Find a distribution over continuous
space that
– can be represented compactly
– and updated efficiently when we get new data
• Example: Uniform; but what if we have a more
information?
• Beta distributions
– Think of a and b as the number of heads and
tails you have seen prior to the start of this
experiment
– Update:
“Conjugate Prior”
• A prior distribution family Pc is considered a
conjugate prior for a likelihood function family Pl
if starting with a hypothesis prior Pc1 from Pc and
seeing data with likelihood Pl from Pl the
posterior of the hypothesis prior will also be in Pc
– Beta distributions are conjugate priors for bernouli
(Binomial) likelihood distributions
– Dirichlet distributions are conjugate priors for
Multinomial likelihood distributions
– Normal-Wishart distributions are conjugate priors for
Gaussian likelihood distributions
Bayesian Prediction
• So suppose we started with Beta[a,b] as the prior
– Probability of heads will be a/(a+b)
• Need to integrate P(heads|q )P(q) dq over 0 to 1
• Now after seeing Dh heads and Dt tails, the
posterior will be Beta[a+Dh, b+Dt ]
– Probability of heads now will be
• (a+Dh )/(a+Dh + b+Dt )
• So, to the ML estimate, you just add a + b virtual
samples…
– Is what you did with Laplace Smoothing…
Laplace smoothing is a backdoor way of making ML predictions
be in line with full Bayesian learning…
Multi-parameter case
(Assume Parameter Independence)
Priors and Background Knowledge
• Hypothesis priors can be seen as providing
background knowledge
• Background knowledge is also helpful in
“logical learning”
– Sao Paulo airport example
Undirected Probabilistic Graphical
Models
(Markov Nets)
(Slides from Sam Roweis Lecture)
Connection to MCMC:
MCMC requires sampling a node given its markov blanket
Need to use P(x|MB(x)). For Bayes nets MB(x) contains more
nodes than are mentioned in the local distribution CPT(x)
 For Markov nets,
Because neighbor relation is symmetric
nodes xi and xj are both neighbors of each other..
Markov Networks
• Undirected graphical models
Smoking
Cancer
Asthma

Cough
Potential functions defined over cliques
1
P( x)    c ( xc )
Z c
Z    c ( xc )
x
c
Smoking Cancer
Ф(S,C)
False
False
4.5
False
True
4.5
True
False
2.7
True
True
4.5
Markov Networks
• Undirected graphical models
Smoking
Cancer
Asthma

Cough
Log-linear model:
1


P( x)  exp   wi f i ( x) 
Z
 i

Weight of Feature i
Feature i
 1 if  Smoking  Cancer
f1 (Smoking, Cancer )  
 0 otherwise
w1  1.5
Markov Nets vs. Bayes Nets
Property
Markov Nets
Bayes Nets
Form
Prod. potentials
Prod. potentials
Potentials
Arbitrary
Cond. probabilities
Cycles
Allowed
Forbidden
Partition func. Z = ? global
Indep. check
Z = 1 local
Graph separation D-separation
Indep. props. Some
Some
Inference
Convert to Markov
MCMC, BP, etc.
Inference in Markov Networks
• Goal: Compute marginals & conditionals of
P( X ) 
1


exp   wi f i ( X ) 
Z
 i



Z   exp   wi f i ( X ) 
X
 i

• Exact inference is #P-complete
• Conditioning on Markov blanket is easy:
w f ( x) 


P( x | MB( x )) 
exp   w f ( x  0)   exp   w f ( x  1) 
exp
i
i
i i
• Gibbs sampling exploits this
i i
i
i i
Partition function cancels out
MCMC: Gibbs Sampling
state ← random truth assignment
for i ← 1 to num-samples do
for each variable x
sample x according to P(x|neighbors(x))
state ← state with new value of x
P(F) ← fraction of states in which F is true
Other Inference Methods
•
•
•
•
Many variations of MCMC
Belief propagation (sum-product)
Variational approximation
Exact methods
Learning Markov Networks
• Learning parameters (weights)
– Generatively
– Discriminatively
• Learning structure (features)
• In this tutorial: Assume complete data
(If not: EM versions of algorithms)
Entanglement in log likelihood…
a
b
c
P( X ) 
1


exp   wi f i ( X ) 
Z
 i



Z   exp   wi f i ( X ) 
X
 i

Generative Weight Learning
• Maximize likelihood or posterior probability
• Numerical optimization (gradient or 2nd order)
• No local maxima

log Pw ( x)  ni ( x)  Ew ni ( x)
wi
No. of times feature i is true in data
Expected no. times feature i is true according to model
• Requires inference at each step (slow!)
Discriminative Weight Learning
• Maximize conditional likelihood of query (y)
given evidence (x)

log Pw ( y | x)  ni ( x, y )  Ew ni ( x, y )
wi
No. of true groundings of clause i in data
Expected no. true groundings according to model
• Approximate expected counts by counts in
MAP state of y given x
Structure Learning
• How to learn the structure of a Markov
network?
– … not too different from learning structure for a
Bayes network: discrete search through space of
possible graphs, trying to maximize data
probability….