1

Introduction toPredictive Learning

Electrical and Computer Engineering

LECTURE SET 9Nonstandard Learning Approaches

2

OUTLINE• Motivation for non-standard approaches

- Learning with sparse high-dimensional data- Formalizing application requirements- Philosophical motivation

• New Learning Settings- Transduction- Universum Learning- Learning Using Privileged Information- Multi-Task Learning

• Summary

3

Sparse High-Dimensional Data • Recall standard inductive learning• High-dimensional, low sample size (HDLSS) data:

• Gene microarray analysis• Medical imaging (i.e., sMRI, fMRI)• Object and face recognition• Text categorization and retrieval• Web search

• Sample size smaller than dim: d ~10K–100K, n ~ 100• Standard learning methods don’t work for such data• How to improve generalization?• How humans learn?

4

How to improve generalization for HDLSS?

Conventional approaches use Standard inductive learning + a priori knowledge:• Preprocessing and feature selection (preceding learning)• Model parameterization (~ selection of good kernels)• Informative prior distributions (in statistical methods)

Non-standard learning formulations• Seek new generic formulations (not methods!) that better

reflect application requirements• A priori knowledge + additional data are used to derive

new problem formulations

5

Approach Known As Big Data

SRM Inductive Principle guaranteesConvergence to optimal/good generalization when• Training sample size grows large• Model complexity is optimally controlled Big Data Paradigm• The main focus/issue is how to design powerful learning

algorithms for large data sets DATA = SIGNAL + NOISE

• Standard inductive learning setting is usually assumed.

6

Formalizing Application Requirements• Classical statistics: parametric model is given (by experts)• Modern applications: complex iterative process

Non-standard (alternative) formulations may be better

APPLICATION NEEDS

LossFunction

Input, output,other variables

Training/test data

AdmissibleModels

FORMAL PROBLEM STATEMENT

LEARNING THEORY

7

Philosophical Motivation• Philosophical view 1 (Realism):

Learning ~ search for the truth (estimation of true dependency from available data)

System identification

~ Inductive Learning

where a priori knowledge is about the true model

8

Philosophical Motivation (cont’d)• Philosophical view (Instrumentalism):

Learning ~ search for the instrumental knowledge (estimation of useful dependency from available data)

VC-theoretical approach ~ focus on learning formulation

9

VC-theoretical approach• Focus on the learning setting rather than

on learning methods• Learning formulation depends on:

(1) available data(2) application requirements(3) a priori knowledge (assumptions)

• Factors (1)-(3) combined using Vapnik’s Keep-It-Direct (KID) Principle yield a learning formulation

10

Contrast these two approaches• Conventional (statistics, data mining):

a priori knowledge typically reflects properties of a true (good) model, i.e.a priori knowledge ~ parameterization

• Why a priori knowledge is about the true model?• VC-theoretic approach:

a priori knowledge ~ how to incorporate available data into problem formulationoften a priori knowledge ~ available data samples of different type new learning settings

),( wf x

11

Examples of Nonstandard Settings • Standard Inductive setting, e.g., digits 5 vs. 8

Finite training set Predictive model derived using only training dataPrediction for all possible test inputs

• Possible modifications- Transduction: Predict only for given test points - Universum Learning: available labeled data ~ examples of digits 5 and 8 and unlabeled examples ~ other digits- Learning using Privileged Information:training data provided by t different persons. Group label is known only for training data, but not available for test data.- Multi-Task Learning:training data ~ t groups (from different persons)test data ~ t groups (group label is known)

ii y,x

12

SVM-style Framework for New Settings • Conceptual Reasons:

Additional info/data new type of SRM structure

• Technical Reasons: New knowledge encoded as additional constraints on complexity (in SVM-like setting)

• Practical Reasons: - new settings directly comparable to standard SVM- standard SVM is a special case of a new setting- optimization s/w for new settings may require minor modification of (existing) standard SVM s/w

13

OUTLINE• Motivation for non-standard approaches• New Learning Settings

- Transduction- Universum Learning- Learning Using Privileged Information- Multi-Task Learning

Note: all settings assume binary classification• Summary

14

Transduction (Vapnik, 1982, 1995)• How to incorporate unlabeled test data into the

learning process? Assume binary classification

• Estimating function at given pointsGiven: labeled training data and unlabeled test points

Estimate: class labels at these test points

Goal of learning: minimization of risk on the test set:

where

*jx mj ,...,1 ii y,x ni ,...,1

),....( **1

*myyy

)/(,1)( **

1

*jj

y

m

j

ydPyyLm

R xy

),(),....,( **1

* mff xxy

15

Transduction vs Induction

a priori knowledge assumptions

estimated function

training data

predicted output

induction deduction

transduction

16

Transduction based on size of margin• Binary classification, linear parameterization,

joint set of (training + working) samplesNote: working sample = unlabeled test point

• Simplest case: single unlabeled (working) sample

• Goal of learning(1) explain well available data (~ joint set)(2) achieve max falsifiability (~ large margin)

~ Classify test (working + training) samples by the largest possible margin hyperplane (see below)

17

Margin-based Local Learning• Special case of transduction: single working point

• How to handle many unlabeled samples?

18

Transduction based on size of margin• Transductive SVM learning has two objectives:

(TL1) separate labeled data using a large-margin hyperplane ~ as in standard SVM(TL2) separate working data using a large-margin hyperplane

19

Loss function for unlabeled samples• Non-convex loss function:

• Transductive SVM constructs a large-margin hyperplane for labeled samples AND forces this hyperplane to stay away from unlabeled samples

20

Optimization formulation for SVM transduction• Given: joint set of (training + working) samples• Denote slack variables for training, for working • Minimize

subject to

where Solution (~ decision boundary)

• Unbalanced situation (small training/ large test) all unlabeled samples assigned to one class

• Additional constraint:

i*j

m

jj

n

ii CCbR

1

**

1

)(21),( www

mjnibyby

ji

jij

iii

,...,1,,...,1,0,1])[(1])[(

*

**

xwxw

mjbsigny jj ,...,1),(* xw** )()( bf xwx

m

ji

n

ii bm

yn 11

])[(11 xw

21

Optimization formulation (cont’d)• Hyperparameters control the trade-off

between explanation and falsifiability• Soft-margin inductive SVM is a special case of

soft-margin transduction with zero slacks• Dual + kernel version of SVM transduction• Transductive SVM optimization is not convex

(~ non-convexity of the loss for unlabeled data) – **elaborate/explain** different opt. heuristics ~ different solutions

• Exact solution (via exhaustive search) possible for small number of test samples (m) – but this solution is NOT very useful (~ inductive SVM).

*CandC

0* j

22

Many applications for transduction• Text categorization: classify word documents

into a number of predetermined categories• Email classification: Spam vs non-spam• Web page classification• Image database classification• All these applications:

- high-dimensional data- small labeled training set (human-labeled)- large unlabeled test set

23

Example application

• Prediction of molecular bioactivity for drug discovery

• Training data~1,909; test~634 samples• Input space ~ 139,351-dimensional• Prediction accuracy:SVM induction ~74.5%; transduction ~ 82.3%Ref: J. Weston et al, KDD cup 2001 data analysis: prediction

of molecular bioactivity for drug design – binding to thrombin, Bioinformatics 2003

24

Semi-Supervised Learning (SSL)• SSL assumes availability of labeled + unlabeled

data (similar to transduction)• SSL has a goal of estimating an inductive model

for predicting new (test) samples – different from transduction

• In machine learning, SSL and transduction are often used interchangeably, i.e. transduction can be used to estimate an SSL model.

• SSL methods usually combine supervised and unsupervised learning methods into one algorithm

25

SSL and Cluster Assumption• Cluster Assumption: real-life application data often

has clusters, due to (unknown) correlations between input variables. Discovering these clusters using unlabeled data helps supervised learning

• Example: document classification and info retrieval- individual words ~ input features (for classification)- uneven co-occurrence of words implies clustering of the documents in the input space- unlabeled documents can be used to identify this cluster structure, so that just a few labeled examples are sufficient for constructing a good decision rule

26

Toy Example for Text Classification• Data Set: 5 documents that need to be classified into 2

topics ~ “Economy’ and ‘Entertainment’• Each document is defined by 6 features (words)

- two labeled documents (shown in color)- need to classify three unlabeled documents

• Apply clustering 2 clusters (x1,x3) and (x2,x4,x5)

Budget deficit

Music Seafood Sex Interest Rates

Movies

L1 1L2 1 1 1U3 1 1U4 1 U5 1 1

27

Self-Learning Method (example of SSL)

Given initial labeled data set L and unlabeled set U• Repeat:

(1) estimate a classifier using L(2) classify randomly chosen unlabeled sample using the decision rule estimated in Step (1)(3) move this new labeled sample to L

Iterate steps (1)-(3) until all unlabeled samples are classified

28

Illustration (using 1-nearest neighbor classifier)

Hyperbolas data:- 10 labeled and 100 unlabeled samples (green)

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9Iteration 1

Unlabeled SamplesClass +1Class -1

29

Illustration (after 50 iterations)

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9Iteration 50

Unlabeled SamplesClass +1Class -1

30

Illustration (after 100 iterations)

All samples are labeled now:

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9Iteration 100

Class +1Class -1

31

Comparison: SSL vs T-SVM• Comparison 1 for low-dimensional data:

- Hyperbolas data set (10 labeled, 100 unlabeled)- 10 random realizations of training data

• Comparison 2 for high-dimensional data: - Digits 5 vs 8 (100 labeled, 1,866 unlabeled)- 10 random realizations of training/validation data

Note: validation data set for SVM model selection

• Methods used - Self-learning algorithm (using 1-NN classification)- Nonlinear T-SVM (needs parameter tuning)

32

Comparison 1: SSL vs T-SVM and SVMMethods used

- Self-learning algorithm (using 1-NN classification)- Nonlinear T-SVM (Poly kernel d=3)

• Self-learning method is better than SVM or T-SVM- Why?

HYPERBOLAS DATA SET

SETTING σ=0.025, No. of training and validation samples=10 (5 per class). No. of test samples= 100 (50 per class).

Polynomial SVM(d=3)

T-SVM(d=3) Self-Learning

Prediction Error ( %) 11.8(2.78) 10.9(2.60) 4.0(3.53)

Typical Parameters C= 700-7000 C*/C= 10-6 N/A

33

Comparison 2: SSL vs T-SVM and SVMMethods used

- Self-learning algorithm (using 1-NN classification)- Nonlinear T-SVM (RBF kernel)

• SVM or T-SVM is better than self-learning method - Why?

MNIST DATA SETSETTING: No. of training and validation samples=100. No. of test samples= 1866.

RBF SVM RBF T-SVM Self-LearningTest Error ( %) 5.79(1.80) 4.33(1.34) 7.68(1.26)Typical selected parameter values

C≈ 0.3- 7gamma≈5-3-5-2

C*/C≈0.01-1 N/A

34

Explanation of T-SVM for digits data setHistogram of projections of labeled + unlabeled data:

- for standard SVM (RBF kernel) ~ test error 5.94%- for T-SVM (RBF kernel) ~ test error 2.84%

Histogram for RBF SVM (with optimally tuned parameters):

-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 20

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

35

Explanation of T-SVM (cont’d)Histogram for T-SVM (with optimally tuned parameters)Note: (1) test samples are pushed outside the margin borders

(2) most labeled samples project away from margin

-3 -2 -1 0 1 2 30

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

36

Universum Learning (Vapnik 1998, 2006)

• Motivation: what is a priori knowledge?- info about the space of admissible models- info about admissible data samples

• Labeled training samples (as in inductive setting) + unlabeled samples from the Universum

• Universum samples encode info about the region of input space (where application data lives):- from a different distribution than training/test data- U-samples ~ neither positive nor negative class

• Examples of the Universum data• Large improvement for small sample size n

37

Cultural Interpretation of the Universum

neither Hillary nor Obama but looks like both

• Absurd examples, jokes, some art forms

38

Cultural Interpretation of the Universum

Marc Chagall:FACES

39

Cultural Interpretation of the Universum

Marcel Duchamp (1919)Mona Lisa with Mustache

• Some art formssurrealism, dadaism

40

More on Marcel Duchamp

Rrose Sélavy (Marcel Duchamp), 1921, Photo by Man Ray.

41

Main Idea of Universum Learning

Fig. courtesy of J. Weston (NEC Labs)

• Handwritten digit recognition: digit 5 vs 8

42

Learning with the Universum• Inductive setting for binary classification

Given: labeled training data and unlabeled Universum samples Goal of learning: minimization of prediction risk (as in standard inductive setting)

• Two goals of the Universum Learning:(UL1) separate/explain labeled training data using large-margin hyperplane (as in standard SVM)(UL2) maximize the number of contradictions on the Universum, i.e. Universum samples inside the margin

Goal (UL2) is achieved by using special loss function for Universum samples

*jx mj ,...,1

ii y,x ni ,...,1

43

Inference through contradictions

44

Universum SVM Formulation (U-SVM)• Given: labeled training + unlabeled Universum samples• Denote slack variables for training, for Universum • Minimize where

subject to for labeled data

for the Universum

where the Universum samples use -insensitive loss

• Convex optimization• Hyper-parameters control the trade-off btwn

minimization of errors and maximizing the # contradictions• When =0, standard soft-margin SVM

i*j

m

jj

n

ii CCbR

1

**

1

)(21),( www

iii by 1])[( xw nii ,...,1,0, *)( ii b xw mjj ,...,1,0*

0, * CC

0, * CC

*C

45

-insensitive loss for Universum samples

x

y

*2

1

46

Application Study (Vapnik, 2006)• Binary classification of handwritten digits 5 and 8• For this binary classification problem, the following

Universum sets had been used:U1: randomly selected other digits (0,1,2,3,4,6,7,9)

U2: randomly mixing pixels from images 5 and 8

U3: average of randomly selected examples of 5 and 8

Training set size tried: 250, 500, … 3,000 samples

Universum set size: 5,000 samples

• Prediction error: improved over standard SVM, i.e. for 500 training samples: 1.4% vs 2% (SVM)

47

Universum U3 via random averaging (RA)

Average

Class 1

Class -1 Hyper-plane

48

Random Averaging for digits 5 and 8• Two randomly selected examples

• Universum sample:

49

Application Study: gender of human faces• Binary classification setting• Difficult problem:

dimensionality ~ large (10K - 20K)labeled sample size ~ small (~ 10 - 20)

• Humans perform very well for this task• Issues:

- possible improvement (vs standard SVM)- how to choose Universum?- model parameter tuning

50

Male Faces: examples

51

Female Faces: examples

52

Universum Faces:neither male nor female

53

Empirical Study (Bai and Cherkassky 2008)• Gender classification of human faces (male/ female)• Data: 260 pictures of 52 individuals (20 females and 32 males,

5 pictures for each individual) from Univ. of Essex

• Data representation and pre-processing: image size 46x59 – converted into gray-scale image, following standard image processing (histogram equalization)

• Training data: 5 female and 8 male photos• Test data: remaining 39 photos (of other people)• Experimental procedure: randomly select 13 training samples

(and 39 test samples). Estimate and compare inductive SVM classifier with SVM classifier using N Universum samples (where N=100, 500, 1000).- Report results for 4 partitions (of training and test data)

54

Empirical Study (cont’d)• Universum generation:

U1 Average: of male and female samples randomly selected from the training set

U2 Empirical Distribution: estimate pixel-wise distribution of the training data. Generate a new picture from this distribution

U3 Animal faces:

5555

Universum generation: examples• U1 Averaging:

• U2 Empirical Distribution:

56

Results of gender classification• Classification accuracy: improves vs standard SVM by ~

2% with U1, and ~ 1% with U2 Universum

• Universum by averaging gives better results for this problem, when number of Universum samples N = 500 or 1,000 (shown above for 4 partitionings of the data)

5757

Results of gender classification• Universum ~ Animal Faces:

• Degrades classification accuracy by 2-5% (vs standard SVM)

• Animal faces are not relevant to this problem

58

Random Averaging Universum• How to come up with good Universum ?

- usually application – specificBut• RA Universum is generated from training data• Under what conditions RA U-SVM is effective? ~

better than standard SVM• Solution approach:

Analyze histogram of projections of training samples onto normal direction vector w of SVM model (for high-dim data)

59

Histogram of projections and RA Universum

• Typical histogram (for high-dimensional data)Case 1:

-1.5 -1 -0.5 0 0.5 1 1.510

0

101

102

103

60

Histogram of projections and RA Universum

• Typical histogram (for high-dimensional data)Case 2:

-1.5 -1 -0.5 0 0.5 1 1.50

2

4

6

8

10

12

14

16

61

Histogram of projections and RA Universum

• Typical histogram (for high-dimensional data)Case 3:

-3 -2 -1 0 1 2 30

50

100

150

200

250

62

Example: Handwritten digits 5 vs 8 • MNIST data set: each digit ~ 28x28=784 pixels• Training / validation set ~ 1,000 samples each

Test set ~ 1,866 samplesFor U-SVM ~ 1,000 Universum samples (via RA)

• Comparison of test error rate for SVM and U-SVM:

SVM U-SVM MNIST (RBF Kernel) 1.37% (0.22%) 1.20% (0.19%) MNIST (Linear Kernel) 4.58% (0.34%) 4.62% (0.37%)

63

Histogram of projections for MNIST data

for RBF SVM for Linear SVM

-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.50

50

100

150

200

250MNIST Data: Distribution of RBF kernel SVM output

-6 -4 -2 0 2 4 60

20

40

60

80

100

120MNIST Data: Disbution of the linear SVM output

64

Analyzing ‘other digits’ UniversumThe same set-up but using digits ‘1’ or ‘3’ UniversaDigit 1 Universum Digit 3 Universum

Test error: RBF SVM~ 1.47%, Digit1~1.31%, Digit3~1.01%

-3 -2 -1 0 1 2 30

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

-3 -2 -1 0 1 2 30

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

65

Discussion• Technical aspects: why the Universum data

improves generalization when d>>n? -SVM is solved in a low-dimensional subspace implicitly defined by the Universum data

• How to specify good Universum? - no formal procedure exists- conditions for the effectiveness of a particular Universum data set: see http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=5936738

• Philosophical aspects: - Relation to human learning (cultural Universum)?- New type of inference? Impact on psychology?

66

Learning Using Privileged Info (LUPI)Given: training data (x, x*, y) where x* is privileged info only for training dataEstimate: predictive model y = f(x)

This additional info x* is common in many apps:Handwritten digit recognition: ~ person’s label

Brain imaging (fMRI): human subject label (in a group of subjects performing the same cognitive task)

Medical diagnosis: ~ medical history after initial diagnosis

Time series prediction: ~ future values of the training data

67

Learning Using Privileged Info (LUPI)Given: training data (x, x*, y) where x* is privileged info only for training dataEstimate: predictive model y = f(x)Main Idea:standard learning x LUPI x + x*

+

68

LUPI (Vapnik, 2006)

• Goal of learning for both settings: find a function f(x) providing small error for test inputs

• Standard setting: training set ( , ), 1, ,i iy i nx

• LUPI setting: training set where is additional privileged info

• Privileged info: in the form of additional constraints on errors (slack variables)

*( , , ), 1, ,i i iy i nx x *ix

69

Many Application Settings for LUPI• Medical diagnosis: privileged info ~ expensive

test results, patient history after diagnosis, etc.• Time series prediction: privileged info ~ future

values of the training time series• Additional group information (~ structured data)• Feature selection etc.

SVM+ is a formal mechanismfor LUPI training

70

LUPI and SVM+ (Vapnik 2006)SVM+ is a new learning technology for LUPI• Improved prediction accuracy (proven

both theoretically and empirically)• Similar to learning with a human teacher:

‘privileged info’ ~ feedback from an experienced teacher(can be very beneficial)

• SVM+ is the most promising new technology for ‘difficult’ problems

• Computationally complex: no public domain software currently available

71

LUPI Formalization and Challenges• LUPI training data in the form

• Privileged Info - assumed to have informative value (for prediction)- is not available for predictive model

• Interpretation of privileged info- additional feedback from an expert teacher, different from ‘correct answers’ or y-labels- this feedback provided in a different space x*

• Challenge: how to handle/combine learning in two different feature spaces?

niyiii ,...2,1),,( xx

)( ix

)(ˆ xfy

72

SVM+ Technology for LUPI• LUPI training data in the form

• SVM+ Learning takes place in two spaces:- decision space Z where the decision rule is estimated- correction space Z* where the privileged info is encoded in the form of additional constraints on errors (slack variables) in the decision space

• SVM+ approach implements two (kernel) mappings- into decision space- into correction space

niyiii ,...2,1),,( xx

)( ix

)(ˆ zfy

Zxx :)( Zxx :)(

73

Illustration of SVM+ (privileged info = group label)

Decision function

Decision space

Correcting space Correcting functions

Group1Group2Class 1Class -1

Correcting space

mapping

2

2

1

1

r slack variable for group r

mapping

74

Two Objectives of SVM+ Learning• LUPI training data in the form

SVM+ Learning achieves two goals:(SVM+ 1)

- separate labeled training data in decision space Z using large-margin hyperplane (as in standard SVM)

(SVM+ 2) incorporate privileged info in the form of additional constraints on errors (slack variables) in the decision space. These constraints are modeled as (linear) functions in the correction space Z*

niyiii ,...2,1),,( xx

)( ix

)(ˆ zfy

75

SVM+ Formulation aka

r

tt Ti

ri

t

r

t

rrrddbwww

C 11,..,,,..,,

),(2

),(21 min

11

wwww

subject to:

trTiby rriii ,...,1,,1)),(( zw

trTi r

ri ,...,1,,0

trTid rrrri

ri ,...,1,,),( wz

]))(,[()]([ bsignfsigny Z xwxriz

r Zri

)(xz

Zizi )(xzix

Decision Space

Correcting Space

rSVM

76

SVM+ Formulation• Map the input vectors simultaneously into:

- Decision space (standard SVM classifier)- Correcting space (where correcting functions model slack variables for different groups)

• Decision space/function ~ the same for all groups• Correcting functions ~ different for each group

(but correcting space may be the same)• SVM+ optimization formulation incorporates:

- the capacity of decision function- capacity of correcting functions for group r- relative importance (weight) of these two capacities

ww, rr ww ,

77

Application Study (Liang and Cherkassky, 2007)

• SVM+ technology is fairly recent.Just a few known empirical comparison studies

• fMRI data analysis problem (CMU data set)- six subjects presented with {picture or sentence}- fMRI data is recorded over 16 time intervals- need to learn binary classifier fMRI image class

• Data preprocessing (Wang et al, 2004)- extract 7 input features (Regions of Interest) from high–dim. fMRI image input vector has 16x7=112 components

• Comparison betweenstandard SVM (data from all 6 subjects is pooled together)SVM+ method (6 groups where each group has 80 samples)

78

fMRI Application Study (cont’d)

• Experimental protocol: (randomly) split the data- 60% ~ training- 20% ~ validation (for tuning parameters)- 20% ~ test (for estimating prediction error)

• Details of methods used:- linear SVM classifier (single parameter C)- classifier (3 parameters: linear kernel for decision space, RBF kernel for correcting space, and parameter )

• Comparison results (over 10 trials): - standard SVM ~ ave test error: 22.2 % (st. dev. 3.8%)- ~ ave test error: 20.2% (st.dev. 3.1%)

SVM

SVM

79

Multi-Task Learning (MTL)

• Application: Handwritten digit recognitionLabeled training data provided by t persons (t >1) Goal 1: estimate a single classifier that will generalize well for future samples generated by these persons ~ LUPIGoal 2: estimate t related classifiers (for each person) ~ MTL

• Application: Medical diagnosisLabeled training data provided by t groups of patients (t >1), say men and women (t = 2) Goal 1: estimate a classifier to predict/diagnose a disease using training data from t groups of patients ~ LUPIGoal 2: estimate t classifiers specialized for each group of patients ~ MTL

Note: for MTL the group info is known for training and test data

80

Multi-task Learning

Training Data

Predictive Model

Training Data

Predictive Model

Training Data

Predictive Model

Training Data

Predictive Model

Task Relatedness Modeling

(a)

(b)

Task 1

Task 2

Task t

(a) Single task learning

(b) Multi-task learning (MTL)

81

Contrast Inductive Learning, MTL and LWSD

Training data for Group 1

Training data for Group t

…..

Inductive Learning

Multi-TaskLearning

Learning with Structured Data

f

dataGroup info

f1

ft

f

test data for Group 1

test data for Group t

…..…..

test data (no Group info)

test data (no Group info)

8282

Different Ways of Using Group Information

SVM

SVM+

SVM

SVM

svm+MTL

f(x)

f(x)

f1(x)

f2(x)

f1(x)

f2(x)

sSVM:

SVM+:

mSVM:

svm+MTL:

83

Problem setting for MTL• Assume binary classification problem• Training data ~ a union of t related groups(tasks)

Each group r has i.i.d. samples

generated from a joint distribution • Test data: task label for test samples is known

• Goal of learning: estimate t models such that the sum of expected losses for all tasks is

minimized:

rni ,...,2,1),( ri

ri yx

),( yPr x

rntr ,...,2,1

),( yPr x

},...,,{ 21 tfff

t

rrr ydPwfyLwR

1

),()),(,(()( xx

8484

SVM+ for Multi-Task Learning

• New learning formulation: SVM+MTL• Define decision function for each group as

• Common decision function models the relatedness among groups

• Correcting functions fine-tune the model for each group (task)

.

trdbf rrzzr r,...,1,)),(()),(()( wxwxx

bz )),(( wx

8585

t

r Ti

ri

t

rrrddbwww

rrt

C11,..,,,..,,

),(2

),(21 min

11

wwww subject to:

trTidby rrir

riri

ri ,...,1,,1)),(),(( zwzw

trTiri ,...,1,,0

rizr Z

ri )(xz

Zizi )(xzix Correcting Space

SVM+MTL FormulationDecision Space

trdbsignf rrzzr r,...,1,))),(()),((()( wxwxx

8686

Empirical Validation• Different ways of using group info

different learning settings:- which one yields better generalization?- how performance is affected by sample size?

• Empirical comparisons:- synthetic data set

8787

Different Ways of Using Group Information

SVM

SVM+

SVM

SVM

svm+MTL

f(x)

f(x)

f1(x)

f2(x)

f1(x)

f2(x)

sSVM:

SVM+:

mSVM:

svm+MTL:

8888

Synthetic Data

• Generate x with each component • The coefficient vectors of three tasks are specified as

• For each task and each data vector, • Details of methods used:

- linear SVM classifier (single parameter C)- SVM+, SVM+MTL classifier (3 parameters: linear kernel for decision space, RBF kernel for correcting space, and parameter γ) - Independent validation set for model selection

20R 20,...,1,)1,1(~ i�uniformxi

]0,0,0,0,0,1,0,0,0,0,0,1,1,1,1,1,1,1,1,1[

]0,0,0,0,0,0,0,0,1,0,1,0,1,1,1,1,1,1,1,1[

]0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1[

3

2

1

)5.0( xisigny

8989

Experimental Results • Comparison results (over 10 trials):

ave test accuracy(%) ±std. dev.

#per group sSVM SVM+ mSVM svm+MTLn=100 88.11±0.65 88.31±0.84 91.18±1.26 91.47±1.03

n=50 85.74±1.36 86.49±1.69 85.09±1.40 87.39±2.29

n=15 80.10±3.42 80.84±3.16 70.73±2.69 79.24±2.81

Note1: relative performance depends on sample size ‘correct’ problem setting depends on sample size

Note2: SVM+ always better than SVMSVM+MTL always better than mSVM

90

OUTLINE• Motivation for non-standard approaches• New Learning Settings• Summary

- Advantages/limitations of non-standard settings- Vapnik’s Imperative

91

Advantages/Limitations of nonstandard settings• Advantages

- makes common sense- follows methodological framework (VC-theory)- yields better generalization (but not always)- new directions for research

• Limitations- need to formalize application requirements need to understand application domain- generally more complex learning formulations- more difficult model selection- few known empirical comparisons (to date)

92

Vapnik’s Imperative• Vapnik’s Imperative:

asking the right question (with finite data) most direct learning problem formulation

• ~ Controlling the goals of inference, in order to produce a better-posed problem

• Three types of such restrictions (Vapnik, 2006)- regularization (constraining function smoothness)- SRM ~ constraints on VC-dimension of models- constraints on the mode of inference

• Philosophical interpretation (of restrictions):Find a model that explains well available data and can be easily falsified

93

Course Summary: Knowledge Discovery and Philosophical Connections

• Aristotle: All men by nature desire knowledge

• Einstein: Man has an intense desire for assured knowledge

• Assured Knowledge- religion- belief in reason- belief in science (i.e., genetic causes)

94

Scientific Knowledge

• Scientific Knowledge (last 3-4 centuries):- objective- recurrent events (repeatable by others)- quantifiable (described by math models)

• Knowledge ~ causal, deterministic, logical humans cannot reason well about

- noisy/random data- multivariate (high-dimensional) data

95

The Nature of Knowledge

• Idealism: knowledge is pure thought• Realism: knowledge comes from experience• Western History of knowledge

- Ancient Greece ~ idealism- Classical science ~ realism

• But it is not possible to obtain assured knowledge from empirical data (Hume):Whatever in knowledge is of empirical nature is never certain

96

Digital Age• Most knowledge in the form of data from

sensors (not human sense perceptions)• How to get assured knowledge from data?• Naïve realism: data knowledge ?Wired Magazine, 16/07: We can stop looking for

(scientific) models. We can analyze the data without hypotheses about what it might show. We can throw the numbers into the biggest computing clusters the world has ever seen and let statistical algorithms find patterns where science cannot

97

(Over) Promise of ScienceArchimedes: Give me a place to stand, and a

lever long enough, and I will move the worldLaplace: Present events are connected with

preceding ones by a tie based upon the evident principle that a thing cannot occur without a cause that produces it.

Digital Age: more data more knowledgemore connectivity new/better knowledge

98

Classical Statistics

How to obtain assured knowledge from data?• Classical approach:

(1) specify a parametric model(2) estimate its parameters from datamore data better (more accurate) model

• Parametric form is based on first-principle knowledge and therefore correct

99

Philosophical Views on Science

• Karl Popper: Science starts from problems, and not from observations

• Werner Heisenberg :What we observe is not nature itself, but nature exposed to our method of questioning

• Albert Einstein:- God is subtle but he is not malicious.- Reality is merely an illusion, albeit a very persistent one.

100

Predictive Data Modeling Philosophy

• Single ‘correct’ causal model cannot be estimated from data.

• Good predictive models:- not unique- depend on the problem setting ( ~ our method of questioning)

• Predictive Learning methodology is the most important aspect of learning from data.

101

Predictive Modeling: Technical Aspects

• The same methodology applies to financial, bio/medical, weather, marketing etc

• Interpretation of predictive models:- big problem (not addressed in this course)- reflects the need for assured knowledge- requires understanding of the predictive methodology and its assumptions, rather than just interpretable parameterization

102

Non-Technical Aspects• Predictive methodology + philosophy can be

applied to other areas • Examples: writing assignments etc.• Lessons from philosophy and history:

- Amount of information grows exponentially, but the number of ‘big ideas’ remains quite small- Critical thinking ability becomes even more important in our digital age - Freedom of information does not guarantee freedom of thought

103

Predictive Learning and Humanities

• Specialization/ Separation btwn science and humanities – only recently ~ 200 years ago

• Predictive modeling and humanities/culture: - both advocate constraints/ restrictions

• Cultural constraints:~ ethical or religious principles

104

References• Vapnik, V. Estimation of Dependencies Based on Empirical Data. Empirical

Inference Science: Afterword of 2006, Springer, 2006 • Cherkassky, V. and F. Mulier, Learning from Data, second edition, Wiley, 2007• Cherkassky, V., Predictive Learning, available at http://vctextbook.com/ 2013 • Cherkassky, V., S. Dhar, and W. Dai, Practical Conditions for Effectiveness of the

Universum Learning, IEEE Trans. on Neural Networks, 22, 8, 1241-1255, 2011• Chapelle, O., Schölkopf, B., and A. Zien, Eds., Semi-Supervised Learning, MIT

Press, 2006• Weston, J., Collobert, R., Sinz, F., Bottou, L. and V. Vapnik, Inference with the

Universum, Proc. ICML, 2006• Wang, X., Hutchingson, R. and Mitchell, T., Training fMRI classifier to

discriminate cognitive states across multiple subjects, Proc. NIPS, 2004• Liang, L. and Cherkassky, V., Learning using structured data: application to fMRI

data analysis, Proc. IJCNN, 2007• Ando, R. and Zhang, T., A Framework for learning predictive structures from

multiple tasks and unlabeled data, JMLR, 2005• Evgeniou, T. and Pontil, M., Regularized multi-task learning, Proc of 10-th Conf.

on Knowledge Discovery and Data Mining, 2004