1 How to be a Bayesian without believing Yoav Freund Joint work with Rob Schapire and Yishay Mansour.

1

How to be a Bayesian without believing

Yoav Freund

Joint work with Rob Schapire and Yishay Mansour

2

Motivation

• Statistician: “Are you a Bayesian or a Frequentist?”

• Yoav: “I don’t know, you tell me…”• I need a better answer….

3

Toy example

• Computer receives telephone call• Measures Pitch of voice• Decides gender of caller

HumanVoice

Male

Female

4

Generative modeling

Voice Pitch

Pro

babi

lity

mean1

var1

mean2

var2

5

Discriminative approach

Voice Pitch

No.

of

mis

take

s

6

Discriminative Bayesian approach

Voice Pitch

Pro

babi

lity

Conditional probability:

€

P g = m | x( ) =1

1+ e− x−θ( )

Prior

€

P0 θ( ) =1

Ze− θ−a( )

2

Posterior

7

Suggested approach

Voice Pitch

No.

of

mis

take

s

Unsure

Definitelyfemale

Definitelymale

8

Formal Frameworks

For stating theorems regarding the dependence of the generalization error on the size of the training set.

9

The PAC set-up

1. Learner chooses classifier set C c C, c: X {-1,+1}

and requests m training examples

2. Nature chooses a target classifier c from C and a distribution P over X

3. Nature generates training set(x1,y1), (x2,y2), … ,(xm,ym)

4. Learner generates h: X {-1,+1} Goal: P(h(x) c(x)) < c,P

10

The agnostic set-up

1. Learner chooses classifier set C c C, c: X {-1,+1}

and requests m training examples

2. Nature chooses distribution D over X {-1,+1}

3. Nature generates training set according to D

(x1,y1), (x2,y2), … ,(xm,ym)

4. Learner generates h: X {-1,+1} Goal: PD(h(x) y) < PD(c*(x) y) +

D Where c* = argminc C(PD(c(x) y))

Vapnik’s pattern-recognition problem

11

Self-bounding learning

1. Learner selects concept class C2. Nature generates training set

T=(x1,y1), (x2,y2), … ,(xm,ym)IID according to a distribution D over X {-1,+1}

3. Learner generates h: X {-1,+1} and a bound T such that with high probability over the random choice of the training set T

PD(h(x) y) < PD(c*(x) y) + T

Freund 97

12

Learning a region predictor

1. Learner selects concept class C2. Nature generates training set

(x1,y1), (x2,y2), … ,(xm,ym)IID according to a distribution D over X {-1,+1}

3. Learner generates h: X { {-1}, {+1}, {-1,+1} , {} }

such that with high probability PD(y h(x)) < PD(c*(x) y) + 1

and PD(h(x)={-1,+1} ) < 2

Vovk 2000

13

Intuitions

The rough idea

14

A motivating example

-

-

-+

+

+

++

+

++

++

-

-

-

-

-

-

-

-

-

--

-

-

-

-

-

--

-

-

-

-

-

-

--

-

-

-

-

-

--

-

--

-

-

-

-

-

-

+

++

+

++

+

+

++

+

++

+ +

++

+

+

+

+

+

++

+

++

+

+

++

+

++

+--

-

-- -

--

---

--

?

?

?

15

Distribution of errors

0 1/2True error

Empirical error

0 1/2Worst case

Contendersfor best.-> Predict with majority vote

Non-contenders -> ignore!

0 1/2Typical case

16

Main result

Finite concept class

17

Notation

€

x,y( ) ~ D; y ∈ −1,+1{ }Data distribution:

€

h( ) ˙ = P x,y( )~D h(x) ≠ y( )Generalization error:

€

T = x1,y1( ), x2 ,y2( ),..., xm ,ym( ); T ~ DmTraining set:

€

ˆ ε (h) ˙ = 1

m1 h(x) ≠ y[ ]

x,y( )∈T

∑ ˙ = P x,y( )~T h(x) ≠ y[ ]

Training error:

18

The algorithm

€

η > 0, Δ > 0Parameters

€

w(h) ˙ = e−η ˆ ε h( )Hypothesis weight:

€

ˆ l η x( ) ˙ = 1

ηln

w(h)h:h ( x)=+1

∑

w(h)h:h ( x)=−1

∑

⎛

⎝

⎜ ⎜ ⎜

⎞

⎠

⎟ ⎟ ⎟

EEmpirical mpirical LLog og RRatioatio::

€

ˆ p η ,Δ x( ) =

+1 if ˆ l x( ) > Δ

−1,+1{ } if ˆ l x( ) ≤ Δ

−1 if ˆ l x( ) < −Δ

⎧

⎨ ⎪

⎩ ⎪

Prediction rule:

19

Suggested tuning

€

η=ln 8 H( )m1 2−θ

€

Δ=2ln 2 δ( )

m+

ln 8 H( )

8m1 2+θ

€

P(abstain) = P x,y( )~Dˆ p (x) = −1,+1{ }( ) = 5ε h*

( ) +Oln 1 δ( ) + ln H( )

m1/2−θ

⎛

⎝ ⎜ ⎜

⎞

⎠ ⎟ ⎟

€

2) for m = Ω ln 1 δ( ) ln H( )( )1/θ ⎛

⎝ ⎜

⎞ ⎠ ⎟

Yields:

€

1) P mistake( ) = P x,y( )~D y ∉ ˆ p (x)( ) = 2ε h*( ) +O

ln m( )m1/2−θ

⎛

⎝ ⎜

⎞

⎠ ⎟

20

Main properties

1. The ELR is very stable. Probability of large deviations is independent of size of concept class.

2. Expected value of ELR is close to the True Log Ratio (using true hypothesis errors instead of estimates.)

3. TLR is a good proxy of the best concept in the class.

21

McDiarmid’s theorem

€

f : X m →R

€

∀x1,K ,xm; ′ x i ∈ X

€

f x1,K ,xm( ) − f x1,K ,xi−1, ′ x i ,xi+1,K ,xm( ) ≤ ci

If

And

€

X1,K ,Xm are independent random variables

Then

€

P f X1,K ,Xm( ) − E f X1,K ,Xm( )[ ] ≥ ε( ) ≤ 2exp −2ε 2

ci2

i=1

m

∑

⎛

⎝

⎜ ⎜

⎞

⎠

⎟ ⎟

22

Empirical log ratio is stable

€

ˆ ′ ε h( ) = training error with one example changed

€

ˆ ′ ε h( ) − ˆ ε h( ) ≤ 1 m

€

ˆ R η′ K( ) ˙ = 1

ηln e

−η ˆ ′ ε h ⎛ ⎝ ⎜

⎞ ⎠ ⎟

h∈K

∑ ⎛

⎝ ⎜ ⎜

⎞

⎠ ⎟ ⎟

€

ˆ l η x( ) = ˆ R η h | h x( ) = +1{ }( ) − ˆ R η h | h x( ) = −1{ }( )

€

ˆ R η K( ) ˙ = 1

ηln e−η ˆ ε h( )

h∈K

∑ ⎛

⎝ ⎜

⎞

⎠ ⎟

€

∀K ⊆H

23

Bounded variation proof

€

ˆ R η′ K( ) − ˆ R η K( ) =1

ηln e−η ˆ ′ ε h( )

h∈K

∑ e−η ˆ ε h( )

h∈K

∑ ⎛

⎝ ⎜

⎞

⎠ ⎟

€

≤1

ηln max

h∈Ke−η ˆ ′ ε h( ) e−η ˆ ε h( )

[ ]( )

€

=maxh∈K

ˆ ′ ε h( ) − ˆ ε h( )[ ] ≤ 1 m

24

Infinite concept classes

Geometry of the concept class

25

Infinite concept classes

• Stated bounds are vacuous.• How to approximate a infinite class

with a finite class?• Unlabeled examples give useful

information.

26

A metric space of classifiers

gf

Classifier space Example Space

d

d(f,g) = P( f(x) = g(x) )

Neighboring models make similar predictions

27

-covers

Classifier space

Classifier class

€

=1/10

€

=1/20

No. of neighborsincreases like

€

1 ε

No. of neighborsincreases like

€

1 ε 2

28

Computational issues

• How to compute the -cover?• We can use unlabeled examples to

generate cover.• Estimate prediction by ignoring

concepts with high error.

29

Application: comparing perfect features

• 45,000 features• Training Examples:

102 negative 2-10 positive 104 unlabeled

• >1 features have zero training error.• Which feature(s) should we use?• How to combine them?

30

A typical perfect feature

Feature value

No.

of

imag

es

Negativeexamples

Positive examples

Unlabeldexamples

31

Pseudo-Bayes for single threhold

• Set of possible thresholds is uncountably infinite

• Using an -cover over thresholds • Equivalent to using the distribution of

unlabeled examples as the prior distribution over the set of thresholds.

32

What it will do

Feature value

Negativeexamples

Prior weights Error factor

-10

+1

33

Relation to large margins

Neighborhood of good classifiers

SVM and Adaboost search for a linear discriminator with a large margin

34

Relation to Bagging

• Bagging: Generate classifiers from random subsets of

training set. Predict according to the majority vote among

classifiers.(Another possibility: flip label of a small random

subset of the training set)

• Can be seen as a randomized estimate of the log ratio.

35

Bias/Variance for classification

• Bias: error of predicting with the sign of the True Log Ratio (infinite training set).

• Variance: additional error from predicting with the sign of the Empirical Log Ratio which is based on a finite training sample.

36

New directions

How a measure of confidence can help in practice

37

Face Detection

• Paul Viola and Mike Jones developed a face detector that can work in real time (15 frames per second).

QuickTime™ and aYUV420 codec decompressorare needed to see this picture.

38

Using confidence to save time

The detector combines 6000 simple features using Adaboost.

In most boxes, only 8-9 features are calculated.

Feature 1Allboxes

Definitely not a face

Might be a face

Feature 2

39

Selective sampling

Unlabeled dataPartially trained

classifier Sample of unconfident examples

Labeledexamples

40

Co-training

Images thatMight contain

faces

Color info

Shape info

Partially trainedColor basedClassifier

Partially trainedShape based

Classifier

Confident Predictions

Confident Predictions

41

Summary

• Bayesian averaging is justifiable even without Bayesian assumptions.

• Infinite concept classes: use -covers• Efficient implementations: Thresholds, SVM,

boosting, bagging… still largely open.• Calibration (Recent work of Vovk)• A good measure of confidence is very

important in practice.• >2 classes (predicting with a subset)