Transcript BN-01

‫‪Introduction to Bayesian Networks‬‬
‫‪Instructor: Dan Geiger‬‬
‫על מה המהומה ? נישואים מאושרים בין תורת ההסתברות ותורת הגרפים‪.‬‬
‫הילדים המוצלחים‪ :‬אלגוריתמים לגילוי תקלות‪ ,‬קודים לגילוי שגיאות‪,‬‬
‫מודלים למערכות מורכבות‪ .‬שימושים במגוון רחב של תחומים‪.‬‬
‫קורס ממבט אישי למדי על נושא מחקריי לאורך שנים רבות‪.‬‬
‫‪Web page: www.cs.technion.ac.il/~dang/courseBN‬‬
‫‪http://webcourse.cs.technion.ac.il/236372‬‬
‫‪Email: [email protected]‬‬
‫‪Phone: 829 4339‬‬
‫‪Office: Taub 616.‬‬
‫‪.‬‬
What is it all about ?
•How to use graphs to represent probability distributions
over thousands of random variables ?
•How to encode conditional independence in directed and
undirected graphs ?
•How to use such representations for efficient computations
of the probability of events of interest ?
•How to learn such models from data ?
2
Course Information
Meetings:
 Lecture: Mondays 8:30 –10:30
 Tutorial: Mondays 12:30 – 13:30
Grade:
 50% in 5 question sets. These questions sets are obligatory. Each contains mostly
theoretical problems. Submit in pairs before due time (three weeks).
 40% Bochan on January 7. (Maybe 20%-20% with a lecture).
 10% Attending at least 12 lectures and recitation classes for a **passing grade**.
(In very special circumstances 2% per missing item).




Prerequisites:
Data structure 1 (cs234218)
Algorithms 1 (cs234247)
Probability (any course)
Information and handouts:

http://webcourse.cs.technion.ac.il/236372

http://www.cs.technion.ac.il/~dang/courseBN/ (Only lecture slides)
3
Relations to Some Other Courses
.‫אמור לי מי חבריך ואומר לך מי אתה‬
 Introduction
to Artificial Intelligence (cs236501)
 Introduction to Machine Learning (cs236756)
 Introduction to Neural Networks (cs236950)
 Algorithms in Computational Biology (cs236522)
 Error correcting codes
 Data mining
4

Mathematical Foundations (4 weeks including students’ lectures, based on Pearl’s
Chapter 3 + papers).
1.
Properties of Conditional Independence (Soundness and completeness of marginal
independence, graphoid axioms and their interpretation as “irrelevance”,
incompleteness of conditional independence, no disjunctive axioms possible.)
2.
Properties of graph separation (Paz and Pearl 85, Theorem 3), soundness and
completeness of saturated independence statements. Undirected Graphs as I-maps
of probability distributions. Markov-Blankets, Pairwise independence basis.
Representation theorems (Pearl and Paz, from each basis to I-maps). Markov
networks, HC representation theorem, Completeness theorem. Markov chains
3.
Bayesian Networks, d-separation, Soundness, Completeness.
4.
Chordal Graphs as the intersection of BN and Markov networks. Equivalence of their
4 definitions.

Combinatorial Optimiziation of Exact Inference in Graphical models (3 weeks including
students lectures).
1.
HMMs
2.
Exact inference and their combinatorial optimization.
3.
Clique tree algorithm. Conditioning.
4.
Tree-width. Feedback Vertex Set.

Learning (5 weeks including students lectures).
1.
Introduction to Bayesian statistics
2.
Learning Bayesian networks
3.
Chow and Liu’s algorithm; the TAN model.
4.
Structural EM
5.
Searching for Bayesian networks

Applications (2 weeks including student lectures).
6
Homeworks
• HMW #1. Read Chapter 3.1 & 3.2.1. Answer Questions 3.1,
3.2, Prove Eq 3.5b, and fully expand/fix the proof of Theorem 2.
Submit in pairs no later than 5/11/12 (Two weeks from now).
•HMW #2. Read fully Chapter 3. Answer additional 5 questions
of choice at the end. Submit in pairs no later than 19/11/12.
•HMW #3. Read Chapter 2 in Pearl’s book and answer 6
questions of choice at the end. Submit in pairs no later than
3/12/12.
•HMW #4. Submit in pairs 24/12/12
• HMW #5. Submit in pairs 14/1/12
Pearl’s book contains all the notations that I happen not to define
in these slides – consult it often – it is also a very unique and
interesting classic text book.
7
The Traditional View of Probability
in Text Books
• Probability theory provides the impression that we need to
literally represent a joint distribution explicitly as P(x1,…,xn)
on all propositions and their combinations. It is consistent and
exhaustive.
• This representation stands in sharp contrast to human
reasoning: It requires exponential computations to compute
marginal probabilities like P(x1) or conditionals like P(x1|x2).
•Humans judge pairwise conditionals swiftly while
conjunctions are judged hesitantly. Numerical calculations do
not reflect simple reasoning tasks.
8
The Traditional View of Probability
in Text Books
Given ?
Computed ?
Estimated or Computed ?
P(e | h)
Given ?
.‫ כחום גבוה‬e ‫ כמחלה ויראלית נדירה ועל‬h ‫חישבו על‬
9
The Qualitative Notion of
Dependence
•Marginal independence is defined numerically as P(x,y)=P(x) P(y).
The truth of this equation is hard to judge by humans, while judging
whether X and Y are dependent is often easy.
•“Burglary within a day” and “nuclear war within five years”
• Likewise, people tend to judge
x
Y
Z
10
•The notions of relevance and dependence are far more basic
than the numerical values. In a resonating system it should be
asserted once and not be sensitive to numerical changes.
•Acquisition of new facts may destroy existing dependencies
as well as creating new once.
• Learning child’s age Z destroys the dependency between
height X and reading ability Y.
•Learning symptoms Z of a patient creates dependencies
between the diseases X and Y that could account for it.
Probability theory provides in principle such a device via
P(X | Y, K) = P(X |K)
But can we model the dynamics of dependency changes based
on logic, without reference to numerical quantities ?
11
Definition of Marginal Independence
Definition: IP(X,Y) iff for all xDX and yDy
Pr(X=x, Y=y) = Pr(X=x) Pr(Y=y)
Comments:
 Each Variable X has a domain DX with value (or state) x in DX.
 We often abbreviate via P(x, y) = P(x) P(y).
 When Y is the emptyset, we get Pr(X=x) = Pr(X=x).
 Alternative notations to IP(X,Y) such as: I(X,Y) or XY
 Next few slides on properties of marginal independence are based
on “Axioms and algorithms for inferences involving probabilistic
independence.”
12
Properties of Marginal Independence
Trivial Independence: Ip(X,)
Symmetry: Ip(X,Y)  Ip(Y,X)
Decomposition: Ip(X,YW)  Ip(X,Y)
Mixing: Ip(X,Y) and Ip(XY,W)  Ip(X,YW)
Proof (Soundness).
Trivial independence and Symmetry follow from the definition.
Decomposition: Given P(x,y,w) = P(x) P(y,w), simply sum over w
on both sides of the given equation.
Mixing: Given: P(x,y) = P(x) P(y) and P(x,y,w) = P(x,y) P(w).
Hence, P(x,y,w) = P(x) P(y) P(w) = P(x) P(y,w).
13
Properties of Marginal Independence
Are there more such independent properties of independence ?
No. There are none !
Horn axioms are of the form 1 & … & n   where each
statement  stands for an independence statement. We use the symbol
 for a set of independence statements.
Namely:  is derivable from  via these properties if and only if  is
entailed by  (i.e.,  holds in all probability distributions that satisfy
).
Put differently: For every set  and a statement  not derivable from
, there exists a probability distribution P that satisfies  and not .
14
Properties of Marginal Independence
Can we use these properties to infer a new independence
statements  from a set of given independence statements  in
polynomial time ?
YES. The “membership algorithm” and completeness proof in
Recitation class (Paper P2).
Comment. The question “does  entail ” could in principle be
undecidable, drops to being decidable via a complete set of
axioms, and then drops to polynomial with this claim.
15
Properties of Marginal Independence
we check consistency of a set + independence plus a
set - of negated independence statements ?
Can
The membership algorithm in previous slide applies only
for G that includes one negated statement – simply use the
algorithm to check that it is not entailed from + .
But another property of independence called “Amstrong
Relation” guarantees that consistency is indeed verified by
checking separately (in isolation) that each statement in is not entailed from + .
16
Definitions of Conditional Independence
Ip(X,Z,Y) if and only if whenever P(Y=y,Z=z) >0 (3.1)
Pr(X=x | Y=y , Z=z) = Pr(X=x |Z=z)
17
Properties of Conditional Independence
Same Properties of conditional independence:
Symmetry: I(X,Z,Y)  I(Y,Z,X)
Decomposition: I(X,Z,YW)  I(X,Z,Y)
Mixing: I(X,Z,Y) and I(XY,Z,W)  I(X,Z,YW)
BAD NEWS. Are there more properties of independence ?
Yes, infinitely many independent Horn axioms. No answer to
the membership problem, nor to the consistency problem.
18
Important Properties of Conditional
Independence
Recall there are some more notations.
19
Graphical Interpretation
20
Use graphs and not pure logic
•Variables represented by nodes and dependencies by edges.
• Common in our language: “threads of thoughts”, “lines of
reasoning”, “connected ideas”, “far-fetched arguments”.
•Still, capturing the essence of dependence is not an easy task.
When modeling causation, association, and relevance, it is hard to
distinguish between direct and indirect neighbors.
•If we just connect “dependent variables” we will get cliques.
21
Undirected Graphs can represent
Independence
Let G be an undirected graph (V,E).
Define IG(X,Z,Y) for disjoint sets of nodes X,Y, and Z
if and only if all paths between a node in X and a node
in Y pass via a node in Z.
In the text book another notation used is <X|Z|Y>G.
22
Other semantics.
The color of each pixel depends on its neighbor.
M = { IG(M1,{F1,F2},M2), IG(F1,{M1,M2},F2) + symmetry }
23
Dependency Models – abstraction of
Probability distributions
A dependency model M over a finite set of elements
U is a rule that assigns truth values to the predicate
IM(X,Z,Y) where X,Y, Z are (disjoint) subsets of U.
24
Definitions:
1. G=(U,E) is an I-map of a model M over U if
IG(X,Z,Y)  IM(X,Z,Y)
for all disjoint subsets X,Y, Z of U.
2. G is a D-map of M if IM(X,Z,Y)  IG(X,Z,Y)
for all disjoint subsets X,Y, Z of U.
3. G is a perfect map of M if IG(X,Z,Y)  IM(X,Z,Y)
for all disjoint subsets X,Y, Z of U.
4. M is graph-isomorph if there exists a graph G such
that G is a perfect map of M.
25
Representation of independencies with Undirected graphs
can not always be perfect
Strong Union: IG(X,Z,Y)  IG(X,ZW,Y)
This property holds for graph separation but not for
conditional independence in probability.
If G is an I-map of P it can represent IP(X,Z,Y) but can
not represent the negation IP(X,ZW,Y).
If G is a D-map of P it can represent IP(X,ZW,Y) but
can not represent IP(X,Z,Y).
27