Upload
drea
View
37
Download
0
Embed Size (px)
DESCRIPTION
NIPS 2003 Workshop on Information Theory and Learning: The Bottleneck and Distortion Approach Organizers: Thomas Gedeon Naftali Tishby http://www.cs.huji.ac.il/~tishby/NIPS-Workshop. Saturday, December 13 Morning session - PowerPoint PPT Presentation
Citation preview
NIPS 2003 Workshop onInformation Theory and Learning:
The Bottleneck and Distortion Approach
Organizers:
Thomas Gedeon Naftali Tishby
http://www.cs.huji.ac.il/~tishby/NIPS-Workshop
Saturday, December 13Morning session
7:30-8:20 N. Tishby Introduction – A new look at Shannon’s Information Theory 8:20-9:00 T. Gedeon Mathematical structure of Information Distortion methods 9:00-9:30 D. Miller Deterministic annealing9:30-10:00 N. Slonim Multivariate Information Bottleneck10:00-10:30 B. Mumey Optimal Mutual Information Quantization is NP-complete
Afternoon session16:00-16:20 G. Elidan The Information Bottleneck EM algorithm16:20-16:40 A. Globerson Sufficient Dimensionality Reduction with Side Information16:40-17:00 A. Parker Phase Transitions in the Information Distortion 17:00-17:20 A. Dimitrov Information Distortion as a model of sensory processing break17:40-18:00 J. Sinkkonen IB-type clustering for continuous finite data18:00-18:20 G. Chechik GIB: Gaussian Information Bottleneck18:20-18:40 Y. Crammer IB and data-representation – Bregman to the rescue 18:40-19:00 B. Westover Achievable rates for Pattern Recognition
Outline:Outline: The fundamental dilemmaThe fundamental dilemma: : Simplicity/Complexity Simplicity/Complexity versusversus Accuracy Accuracy
Lessons from Lessons from Statistical PhysicsStatistical Physics for Machine Learning for Machine Learning Is there a Is there a “right level” “right level” of description?of description?
A Variational A Variational PrinciplePrinciple Shannon’s source and channel Shannon’s source and channel codingcoding – – dual problemsdual problems [Mutual] [Mutual] InformationInformation as the as the Fundamental QuantityFundamental Quantity Information BottleneckInformation Bottleneck (IB) (IB) andand Efficient Efficient RepresentationsRepresentations Finite Data IssuesFinite Data Issues
Algorithms, Applications and ExtensionsAlgorithms, Applications and Extensions Words, documents and meaning…Words, documents and meaning… Understanding neural codes …Understanding neural codes … Quantization of side-information Quantization of side-information Relevant linear dimension reduction: Gaussian IB Relevant linear dimension reduction: Gaussian IB Using “irrelevant” side informationUsing “irrelevant” side information Multivariate-IB and Graphical ModelsMultivariate-IB and Graphical Models Sufficient Dimensionality Reduction (SDR)Sufficient Dimensionality Reduction (SDR)
The importance of being principled…The importance of being principled…
Learning Theory:The fundamental dilemma…
XX
YY
y=f(x)y=f(x)
Good modelsGood models should enable should enable Prediction Prediction of new data…of new data…
Tradeoff between Tradeoff between accuracy and accuracy and simplicitysimplicity
Or in unsupervised learning …
XX
YY
20 25 30 35 40 4510
20
30
40
50
60
70
Cluster Cluster models?models?
Lessons from Statistical Physics Physical systems with many degrees of freedom:Physical systems with many degrees of freedom: What is the right level of description?What is the right level of description?
“ “Microscopic Variables” – Microscopic Variables” – dynamical variablesdynamical variables ( (positions, positions,
momenta,…momenta,…)) “ “Variational Functions” – Variational Functions” – Thermodynamic Thermodynamic potentialspotentials
(Energy, Entropy, Free-(Energy, Entropy, Free-energy,…)energy,…)
Competition Competition between between order order ((energyenergy)) and and disorder disorder (entropy)(entropy) ……
Emergent “Emergent “relevant” relevant” descriptiondescription
Order Order ParametersParameters
(magnetization,…)(magnetization,…)
Statistical Models in Learning and AI
Statistical models of complex systems:Statistical models of complex systems: What is the right level of description?What is the right level of description?
“ “Microscopic Variables” – Microscopic Variables” – probability probability distributionsdistributions ( (over all over all observed and hidden variables)observed and hidden variables) “ “Variational Functions” – Variational Functions” – Log-likelihood, Log-likelihood, Entropy,… ?Entropy,… ? (Multi-(Multi-Information, “Free-energy”,… )Information, “Free-energy”,… ) Competition between Competition between accuracy accuracy and and complexitycomplexity Emerged “Emerged “relevant” relevant”
descriptiondescriptionFeaturesFeatures
(clusters, sufficient statistics,(clusters, sufficient statistics,…)…)
The Fundamental Dilemma (of science):
Model Complexity vs Prediction Accuracy
Complexity
Accu
racy
Possible Models/representations
Efficiency Efficiency
Surv
ivab
ilitSu
rviv
abilit
yy
Limited dataLimited data
Bounded Bounded ComputationComputation
Can we quantify it…?
When there is a (relevant) prediction or distortion measure
Accuracy small average distortion (good predictions)
Complexity short description (high compression)
A general tradeoff between distortion and compression:
Shannon’s Information Theory
Shannon’s Information Theory
SentSentmessagemessagess
sourcsourcee
receivreceiverer
ReceivedReceivedmessagesmessages
channechannell
symbolssymbols
01101010011100110101001110……
encodeencoderr
decodedecoderr
SourceSourcecodingcoding
ChannelChannelcodingcoding
SourceSourcedecodingdecoding
ChannelChanneldecodingdecoding
Error Error CorrectionCorrectionChannel CapacityChannel Capacity
CompressioCompressionnSource EntropySource Entropy
DecompressiDecompressionon
Rate vs DistortionRate vs Distortion Capacity vs EfficiencyCapacity vs Efficiency
We need to index the max number of non-overlapping green blobs inside the blue blob:
(mutual information!)
X X̂)x|x̂(p
)ˆ|(2 XXnH
)(2 XnH
)ˆ,()ˆ|()( 22/2 XXnIXXnHXnH
Compression and Mutual Compression and Mutual InformationInformation
Axioms for “(multi) information”(w I Nemenman)
M1: For two nodes, X and Y, the shared information, I[X;Y], is a continuous function of p(x,y).
M2: when one of the variables defines the other uniquely (1-1 mapping from X to Y) and both p(x)=p(y)=1/K , then I[X;Y] is a monotonic function of K.
M3: If X is partitioned into two subsets, X1 and X2, then the amount of shared information is additive in the following sense: I[X;Y]=I[(X1,X2);Y]+p(x1)I[X;Y|X1]+p(x2)I[X;Y|X2]
M4: Shared information is symmetric: I[X;Y]=I[Y;X]. M5:
Theorem:
1 2 1 1 2 1 2 1[ ; ;...; ] [ ; ;...; ] [( , ,..., ); ]N N N NI X X X I X X X I X X X X
1 2
1 21 2 1 2
, ,..., 1 2
( , ,..., )[ ; ;...; ] ( , ,..., ) log( ) ( ) ( )
N
NN N
x x x N
p x x xI X X X p x x xp x p x p x
The Dual Variational problems of IT:The Dual Variational problems of IT:RateRate vs vs Distortion Distortion and and CapacityCapacity vs vs
EfficiencyEfficiency Q: What is the simplest representation – fewer bits(/sec) (Rate) for a given expected distortion (accuracy)?
A: (RDT, Shannon 1960) solve:
Q: What is the maximum number of bits(/sec) that can be sent reliably (prediction rate) for a given efficiency of the channel (power, cost)?
A: (Capacity-power tradeoff, Shannon 1956) solve:
)ˆ,(min)( XXIDRDd
),ˆ(max)( YXIECEe
Rate-DistortionRate-Distortion theory theory The tradeoff between expected representation size (Rate) and the expected distortion is expressed by the rate-distortion function:
By introducing a Lagrange multiplier, , we have a variational principle:
with:
)ˆ,(min)( XXIDRDd
)ˆ,()ˆ,()ˆ,()]|ˆ([
xxpxxdXXIxxpL
)ˆ,()|ˆ()()ˆ,(ˆ,
)ˆ,(xxdxxpxpxxd
xxxxp
The Rate-Distortion function, R(D), is the optimal rate for a given distortion and is a convex function.
R(D)
D
Achievable region
impossible
DDR )(
-bits/-bits/accuracyaccuracy
00
The Capacity – Efficiency TradeoffThe Capacity – Efficiency Tradeoff
AchievableAchievable RegionRegion
E (Efficiency (power))E (Efficiency (power))
C
(E)
C(E)
Ca
pacit
y at
Ca
pacit
y at
EEbits/ergbits/erg
Double matching of source to channel- eliminates the need for
coding (?!)21( ) log
2R D
D
2
1( ) log 12
EC E
R(D)R(D)
DD
C(E)C(E)
EE
In the case of a Gaussian channel In the case of a Gaussian channel (w. power constraint)(w. power constraint) and and square distortion, double matching means:square distortion, double matching means:
221 ED
And No coding!And No coding!
““There isThere is a curious and provocative duality between the a curious and provocative duality between the properties of a source with a distortion measure and those of a properties of a source with a distortion measure and those of a channel … if we consider channels in which there is a “cost” channel … if we consider channels in which there is a “cost” associated with the different input letters…associated with the different input letters…
The solution to this problem amounts, mathematically, to The solution to this problem amounts, mathematically, to maximizingmaximizing a mutual information under linear inequality a mutual information under linear inequality constraint… which leads to a capacity-cost function C(E) for the constraint… which leads to a capacity-cost function C(E) for the channel…channel…
In a somewhat dual way, evaluating the rate-distortion function In a somewhat dual way, evaluating the rate-distortion function R(D) for source amounts, mathematically, to R(D) for source amounts, mathematically, to minimizingminimizing a a mutual Information … again with a linear inequality constraint.mutual Information … again with a linear inequality constraint.
This duality can be pursued further and is related to the duality This duality can be pursued further and is related to the duality between past and future and the notions of control and between past and future and the notions of control and knowledge. Thus we may have knowledge of the past but knowledge. Thus we may have knowledge of the past but cannot control it; we may control the future but have no cannot control it; we may control the future but have no knowledge of it.”knowledge of it.”
C. E. Shannon (1959)C. E. Shannon (1959)
Bottlenecks and Neural Nets Auto association: forcing compact representations
provide a “relevant code” of w.r.t.
X Y
X̂
Input Output
Sample 1 Sample 2
Past Future
X̂ X Y
Bottlenecks in Nature…(w E Schniedman, Rob de Ruyter, W Bialek, N Brenner)
The idea: find a compressed variable that enables short encoding of ( small
)while preserving as much as possible the
information on the relevant signal ( )
X X̂)x|x̂(p Y)x̂|y(p
)x̂(p)X̂,X(I )Y,X̂(I
)Y,X(I
X̂)X̂,X(I
)Y,X̂(I
X
Y
A A Variational PrincipleVariational PrincipleWe want a short representation of X that keeps the information about another variable, Y, if possible.
X
YX YXI
ˆ
),(
)ˆ,( XXI
),ˆ( YXI
),ˆ()ˆ,()|ˆ( YXIXXIxxpL
Self Consistent EquationsSelf Consistent Equations Marginal:
Markov condition:
Bayes’ rule:
x
xpxxpxp )()|ˆ()ˆ(
x
xxpxypxyp )ˆ|()|()ˆ|(
)|ˆ()ˆ()()ˆ|( xxpxpxpxxp
0)|ˆ()]|ˆ([
xxpxxpL
)ˆ,(exp
),()ˆ()|ˆ( xxD
xZxpxxp KL
The emerged effective distortioneffective distortion measure:
y
KLKL
xypxypxyp
xypxypDxxD
)ˆ|()|(log)|(
)ˆ|(|)|(ˆ,
• Regular if is absolutely continuous w.r.t.• Small if predicts y as well as x:
)ˆ|( xyp )|( xypx̂
yx
yx
xyp
xxp
xyp
)ˆ|(
)|ˆ(
)|(
ˆ
I-Projections on a set of distributions (Csiszar 75,84)
• The I-projection of a distribution q(x) on a convex set of distributions L:
qpDLqIPR KLLp |~minarg),( ~
),( pfL
)(xq
)(~ xp)(* xp
The The Blahut-Arimoto Blahut-Arimoto AlgorithmAlgorithm
)ˆ(xp )|ˆ( xxp
dXXIxZxxpxpxxpxp
)ˆ,(min),(logminmin)|ˆ(),ˆ()|ˆ()ˆ(
)|ˆ()()ˆ(
)ˆ,(exp),()ˆ()|ˆ(1
xxpxpxp
xxdxZxpxxp
tx
t
t
tt
The Information BottleneckInformation Bottleneck Algorithm
)ˆ,()ˆ,(min
),(logminminmin
)|ˆ(),ˆ(),ˆ|(
)|ˆ()ˆ()ˆ|(
xxDXXI
xZ
KLxxpxpxyp
xxpxpxyp
xtt
tx
t
KLt
t
tt
xxpxypxyp
xxpxpxp
xxDxZxpxxp
)ˆ|()|()ˆ|(
)|ˆ()()ˆ(
)ˆ,(exp),()ˆ()|ˆ(1
“free energy”
The IB emergent effective distortion measure:
)ˆ|(|)|(ˆ, xypxypDxxD KLKL
)ˆ(xp )|ˆ( xxp
)ˆ|( xypThe IBalgorithm
Key TheoremsKey TheoremsIB Coding Theorem (Wyner 1975, Bachrach, Navot, Tishby 2003):The IB variational problem provides a tight convex lower bound on the expected size of the achievable representations of X, that maintain at least of mutual information on the variable Y. The bound can be achieved by: .
Formally equivalent to “Channel coding with Side-Information” (introduced in a very different context by Wyner, 75).
The same bound is true for the dual channel coding problem and the optimal representation constitutes a “perfectly matched source-channel” and requires “no further coding”.
IB Algorithm Convergence (Tishby 99, Friedman, Slonim,Tishby 01):
The Generalized Arimoto-Blauht IB algorithm converges to a local minima of the IB functional, (including its generalized multivariate case).
),ˆ( YXI,X̂
2|||ˆ| XX
Source (p(x))Channel: (q(y|x))
Entropy: Capacity:
Rate-Distortion function:Capacity-Expense function:
Constrained Reliability-Rate functionConstrained Reliability-Rate function
Information Bottleneck (source)Information Bottleneck (channel)
( )
( ) max ( , )pe p E
C E I p q
max ( , )pC I p q
ˆ( , )
ˆ( , ( , )) max ( ; )X
Y qI p q I
L I p x y I Y X
min ( , )qH I p q
( ) min ( , )qd D
R D I p q
ˆ
ˆˆ( , )
ˆ( , ) max min ( | )p
q p KLI p q R d D
F R D D p p
ˆˆ( , )
ˆ[ , ( , )] min ( ; )Y
X pI p q I
L I p x y I X X
ˆˆ( , ) ;
ˆ( , ) max min ( | )p
p q KLI p q R e E
F R E D q q
Summary
-The IB method turns this unified coding principle into algorithms for extracting “informative” relevant structures from empirical joint probability distributions P(X1,X2)…
-The Multivariate IB extends this framework to extract “informative” structures from multivariate distributions, P(X1,…,Xn), via Min Max I(X1,…,Xn), and Graphical models- IB can be further extended for Sufficient Dimensionality Reduction a method for finding informative continuous low dimensional representations via Min Max I(X1,X2)
- Shannon’s Information Theory suggests a compelling framework for quantifying the “Complexity-Accuracy” dilemma, by unifying source and channel coding as a Min Max of mutual information.
Many thanks to:
Fernando PereiraWilliam BialekNir Friedman
Noam SlonimGal Chechik
Amir GlobersonRan Bachrach
Amir NavotIlya Nemenman