Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
MACHINE PERCEPTION
CZECH TECHNICAL
UNIVERSITY
R
E
S
E
A
R
C
H
R
E
P
O
R
T
I
S
S
N
1
2
1
3

2
3
6
5
Statistical Pattern Recognition
Toolbox for Matlab
User’s guide
Vojtˇech Franc and V´aclav Hlav´ aˇc
¦xfrancv,hlavac¦@cmp.felk.cvut.cz
CTU–CMP–2004–08
June 24, 2004
Available at
ftp://cmp.felk.cvut.cz/pub/cmp/articles/FrancTR200408.pdf
The authors were supported by the European Union under project
IST200135454 ECVision, the project IST200132184 ActIPret and
by the Czech Grant Agency under project GACR 102/03/0440 and
by the Austrian Ministry of Education under project CONEX GZ
45.535.
Research Reports of CMP, Czech Technical University in Prague, No. 8, 2004
Published by
Center for Machine Perception, Department of Cybernetics
Faculty of Electrical Engineering, Czech Technical University
Technick´a 2, 166 27 Prague 6, Czech Republic
fax +420 2 2435 7385, phone +420 2 2435 7637, www: http://cmp.felk.cvut.cz
Statistical Pattern Recognition Toolbox for Matlab
Vojtˇech Franc and V´aclav Hlav´ aˇc
June 24, 2004
Contents
1 Introduction 3
1.1 What is the Statistical Pattern Recognition Toolbox and how it has been
developed? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 The STPRtool purpose and its potential users . . . . . . . . . . . . . . 3
1.3 A digest of the implemented methods . . . . . . . . . . . . . . . . . . . 4
1.4 Relevant Matlab toolboxes of others . . . . . . . . . . . . . . . . . . . . 5
1.5 STPRtool document organization . . . . . . . . . . . . . . . . . . . . . 5
1.6 How to contribute to STPRtool? . . . . . . . . . . . . . . . . . . . . . 5
1.7 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Linear Discriminant Function 8
2.1 Linear classiﬁer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Linear separation of ﬁnite sets of vectors . . . . . . . . . . . . . . . . . 10
2.2.1 Perceptron algorithm . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Kozinec’s algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 Multiclass linear classiﬁer . . . . . . . . . . . . . . . . . . . . . 14
2.3 Fisher Linear Discriminant . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Generalized Anderson’s task . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.1 Original Anderson’s algorithm . . . . . . . . . . . . . . . . . . . 19
2.4.2 General algorithm framework . . . . . . . . . . . . . . . . . . . 21
2.4.3 εsolution using the Kozinec’s algorithm . . . . . . . . . . . . . 21
2.4.4 Generalized gradient optimization . . . . . . . . . . . . . . . . . 22
3 Feature extraction 25
3.1 Principal Component Analysis . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Linear Discriminant Analysis . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Kernel Principal Component Analysis . . . . . . . . . . . . . . . . . . . 31
3.4 Greedy kernel PCA algorithm . . . . . . . . . . . . . . . . . . . . . . . 32
3.5 Generalized Discriminant Analysis . . . . . . . . . . . . . . . . . . . . . 33
3.6 Combining feature extraction and linear classiﬁer . . . . . . . . . . . . 34
1
4 Density estimation and clustering 41
4.1 Gaussian distribution and Gaussian mixture model . . . . . . . . . . . 41
4.2 MaximumLikelihood estimation of GMM . . . . . . . . . . . . . . . . 43
4.2.1 Complete data . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2.2 Incomplete data . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3 Minimax estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.4 Probabilistic output of classiﬁer . . . . . . . . . . . . . . . . . . . . . . 48
4.5 Kmeans clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5 Support vector and other kernel machines 57
5.1 Support vector classiﬁers . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.2 Binary Support Vector Machines . . . . . . . . . . . . . . . . . . . . . 58
5.3 OneAgainstAll decomposition for SVM . . . . . . . . . . . . . . . . . 63
5.4 OneAgainstOne decomposition for SVM . . . . . . . . . . . . . . . . . 64
5.5 Multiclass BSVM formulation . . . . . . . . . . . . . . . . . . . . . . . 66
5.6 Kernel Fisher Discriminant . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.7 Preimage problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.8 Reduced set method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6 Miscellaneous 76
6.1 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.2 Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.3 Bayesian classiﬁer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.4 Quadratic classiﬁer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.5 Knearest neighbors classiﬁer . . . . . . . . . . . . . . . . . . . . . . . 82
7 Examples of applications 85
7.1 Optical Character Recognition system . . . . . . . . . . . . . . . . . . 85
7.2 Image denoising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
8 Conclusions and future planes 94
2
Chapter 1
Introduction
1.1 What is the Statistical Pattern Recognition Tool
box and how it has been developed?
The Statistical Pattern Recognition Toolbox (abbreviated STPRtool) is a collection of
pattern recognition (PR) methods implemented in Matlab.
The core of the STPRtool is comprised of statistical PR algorithms described in
the monograph Schlesinger, M.I., Hlav´ aˇc, V: Ten Lectures on Statistical and Structural
Pattern Recognition, Kluwer Academic Publishers, 2002 [26]. The STPRtool has been
developed since 1999. The ﬁrst version was a result of a diploma project of Vojtˇech
Franc [8]. Its author became a PhD student in CMP and has been extending the
STPRtool since. Currently, the STPRtool contains much wider collection of algorithms.
The STPRtool is available at
http://cmp.felk.cvut.cz/cmp/cmp software.html.
We tried to create concise and complete documentation of the STPRtool. This
document and associated help ﬁles in .html should give all information and parameters
the user needs when she/he likes to call appropriate method (typically a Matlab m
ﬁle). The intention was not to write another textbook. If the user does not know
the method then she/he will likely have to read the description of the method in a
referenced original paper or in a textbook.
1.2 The STPRtool purpose and its potential users
• The principal purpose of STPRtool is to provide basic statistical pattern recog
nition methods to (a) researchers, (b) teachers, and (c) students. The STPRtool
is likely to help both in research and teaching. It can be found useful by practi
tioners who design PR applications too.
3
• The STPRtool oﬀers a collection of the basic and the stateofthe art statistical
PR methods.
• The STPRtool contains demonstration programs, working examples and visual
ization tools which help to understand the PR methods. These teaching aids are
intended for students of the PR courses and their teachers who can easily create
examples and incorporate them into their lectures.
• The STPRtool is also intended to serve a designer of a new PR method by
providing tools for evaluation and comparison of PR methods. Many standard
and stateofthe art methods have been implemented in STPRtool.
• The STPRtool is open to contributions by others. If you develop a method
which you consider useful for the community then contact us. We are willing to
incorporate the method into the toolbox. We intend to moderate this process to
keep the toolbox consistent.
1.3 A digest of the implemented methods
This section should give the reader a quick overview of the methods implemented in
STPRtool.
• Analysis of linear discriminant function: Perceptron algorithm and multi
class modiﬁcation. Kozinec’s algorithm. Fisher Linear Discriminant. A collection
of known algorithms solving the Generalized Anderson’s Task.
• Feature extraction: Linear Discriminant Analysis. Principal Component Anal
ysis (PCA). Kernel PCA. Greedy Kernel PCA. Generalized Discriminant Analy
sis.
• Probability distribution estimation and clustering: Gaussian Mixture
Models. ExpectationMaximization algorithm. Minimax probability estimation.
Kmeans clustering.
• Support Vector and other Kernel Machines: Sequential Minimal Opti
mizer (SMO). Matlab Optimization toolbox based algorithms. Interface to the
SV M
light
software. Decomposition approaches to train the Multiclass SVM clas
siﬁers. Multiclass BSVM formulation trained by Kozinec’s algorithm, Mitchell
DemyanovMolozenov algorithm and Nearest Point Algorithm. Kernel Fisher
Discriminant.
4
1.4 Relevant Matlab toolboxes of others
Some researchers in statistical PR provide code of their methods. The Internet oﬀers
many software resources for Matlab. However, there are not many compact and well
documented packages comprised of PR methods. We found two useful packages which
are brieﬂy introduced below.
NETLAB is focused to the Neural Networks applied to PR problems. The classical
PR methods are included as well. This Matlab toolbox is closely related to the
popular PR textbook [3]. The documentation published in book [21] contains de
tailed description of the implemented methods and many working examples. The
toolbox includes: Probabilistic Principal Component Analysis, Generative Topo
graphic Mapping, methods based on the Bayesian inference, Gaussian processes,
Radial Basis Functions (RBF) networks and many others.
The toolbox can be downloaded from
http://www.ncrg.aston.ac.uk/netlab/.
PRTools is a Matlab Toolbox for Pattern Recognition [7]. Its implementation is
based on the object oriented programming principles supported by the Matlab
language. The toolbox contains a wide range of PR methods including: analysis
of linear and nonlinear classiﬁers, feature extraction, feature selection, methods
for combining classiﬁers and methods for clustering.
The toolbox can be downloaded from
http://www.ph.tn.tudelft.nl/prtools/.
1.5 STPRtool document organization
The document is divided into chapters describing the implemented methods. The last
chapter is devoted to more complex examples. Most chapters begin with deﬁnition of
the models to be analyzed (e.g., linear classiﬁer, probability densities, etc.) and the list
of the implemented methods. Individual sections have the following ﬁxed structure: (i)
deﬁnition of the problem to be solved, (ii) brief description and list of the implemented
methods, (iii) reference to the literature where the problem is treaded in details and
(iv) an example demonstrating how to solve the problem using the STPRtool.
1.6 How to contribute to STPRtool?
If you have an implementation which is consistent with the STPRtool spirit and you
want to put it into public domain through the STPRtool channel then send an email
to xfrancv@cmp.felk.cvut.cz.
5
Please prepare your code according to the pattern you ﬁnd in STPRtool implemen
tations. Provide piece of documentation too which will allow us to incorporate it to
STPRtool easily.
Of course, your method will be bound to your name in STPRtool.
1.7 Acknowledgements
First, we like to thank to Prof. M.I. Schlesinger from the Ukrainian Academy of
Sciences, Kiev, Ukraine who showed us the beauty of pattern recognition and gave us
many valuable advices.
Second, we like to thank for discussions to our colleagues in the Center for Machine
Perception, the Czech Technical University in Prague, our visitors, other STPRtool
users and students who used the toolbox in their laboratory exercises.
Third, thanks go to several Czech, EU and industrial grants which indirectly con
tributed to STPRtool development.
Forth, the STPRtool was rewritten and its documentation radically improved in
May and June 2004 with the ﬁnancial help of ECVision European Research Network
on Cognitive Computer Vision Systems (IST200135454) lead by D. Vernon.
6
Notation
Uppercase bold letters denote matrices, for instance A. Vectors are implicitly column
vectors. Vectors are denoted by lowercase bold italic letters, for instance x. The
concatenation of column vectors w ∈ R
n
, z ∈ R
m
is denoted as v = [w; z] where the
column vector v is (n + m)dimensional.
¸x x
) Dot product between vectors x and x
.
Σ Covariance matrix.
µ Mean vector (mathematical expectation).
S Scatter matrix.
A ⊆ R
n
ndimensional input space (space of observations).
¸ Set of c hidden states – class labels ¸ = ¦1, . . . , c¦.
T Feature space.
φ: A → T Mapping from input space A to the feature space T.
k: A A →R Kernel function (Mercer kernel).
T
XY
Labeled training set (more precisely multiset) T
XY
=
¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦.
T
X
Unlabeled training set T
X
= ¦x
1
, . . . , x
l
¦.
q: A → ¸ Classiﬁcation rule.
f
y
: A →R Discriminant function associated with the class y ∈ ¸.
f: A →R Discriminant function of the binary classiﬁer.
[ [ Cardinality of a set.
  Euclidean norm.
det(()) Matrix determinant.
δ(i, j) Kronecker delta δ(i, j) = 1 for i = j and δ(i, j) = 0
otherwise.
_
Logical or.
7
Chapter 2
Linear Discriminant Function
The linear classiﬁcation rule q: A ⊆ R
n
→ ¸ = ¦1, 2, . . . , c¦ is composed of a set of
discriminant functions
f
y
(x) = ¸w
y
x) + b
y
, ∀y ∈ ¸ ,
which are linear with respect to both the input vector x ∈ R
n
and their parameter
vectors w ∈ R
n
. The scalars b
y
, ∀y ∈ ¸ introduce bias to the discriminant functions.
The input vector x ∈ R
n
is assigned to the class y ∈ ¸ its corresponding discriminant
function f
y
attains maximal value
y = argmax
y
∈Y
f
y
(x) = argmax
y
∈Y
(¸w
y
x) + b
y
) . (2.1)
In the particular binary case ¸ = ¦1, 2¦, the linear classiﬁer is represented by a
single discriminant function
f(x) = ¸w x) + b ,
given by the parameter vector w ∈ R
n
and bias b ∈ R. The input vector x ∈ R
n
is
assigned to class y ∈ ¦1, 2¦ as follows
q(x) =
_
1 if f(x) = ¸w x) + b ≥ 0 ,
2 if f(x) = ¸w x) + b < 0 .
(2.2)
The datatype used to describe the linear classiﬁer and the implementation of the
classiﬁer is described in Section 2.1. Following sections describe supervised learning
methods which determine the parameters of the linear classiﬁer based on available
knowledge. The implemented learning methods can be distinguished according to the
type of the training data:
I. The problem is described by a ﬁnite set T = ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ containing pairs
of observations x
i
∈ R
n
and corresponding class labels y
i
∈ ¸. The methods
using this type of training data is described in Section 2.2 and Section 2.3.
8
II. The parameters of class conditional distributions are (completely or partially)
known. The Anderson’s task and its generalization are representatives of such
methods (see Section 2.4).
The summary of implemented methods is given in Table 2.1.
References: The linear discriminant function is treated throughout for instance in
books [26, 6].
Table 2.1: Implemented methods: Linear discriminant function
linclass Linear classiﬁer (linear machine).
andrerr Classiﬁcation error of the Generalized Anderson’s
task.
androrig Original method to solve the AndersonBahadur’s
task.
eanders Epsilonsolution of the Generalized Anderson’s task.
ganders Solves the Generalized Anderson’s task.
ggradander Gradient method to solve the Generalized Anderson’s
task.
ekozinec Kozinec’s algorithm for epsoptimal separating hyper
plane.
mperceptr Perceptron algorithm to train multiclass linear ma
chine.
perceptron Perceptron algorithm to train binary linear classiﬁer.
fld Fisher Linear Discriminant.
fldqp Fisher Linear Discriminant using Quadratic Program
ming.
demo linclass Demo on the algorithms learning linear classiﬁers.
demo anderson Demo on Generalized Anderson’s task.
2.1 Linear classiﬁer
The STPRtool uses a speciﬁc structure array to describe the binary (see Table 2.2)
and the multiclass (see Table 2.3) linear classiﬁer. The implementation of the linear
classiﬁer itself provides function linclass.
Example: Linear classiﬁer
The example shows training of the Fisher Linear Discriminant which is the classical
example of the linear classiﬁer. The Riply’s data set riply trn.mat is used for training.
The resulting linear classiﬁer is then evaluated on the testing data riply tst.mat.
9
Table 2.2: Datatype used to describe binary linear classiﬁer.
Binary linear classiﬁer (structure array):
.W [n 1] The normal vector w ∈ R
n
of the separating hyper
plane f(x) = ¸w x) + b.
.b [1 1] Bias b ∈ R of the hyperplane.
.fun = ’linclass’ Identiﬁes function associated with this data type.
Table 2.3: Datatype used to describe multiclass linear classiﬁer.
Multiclass linear classiﬁer (structure array):
.W [n c] Matrix of parameter vectors w
y
∈ R
n
, y = 1, . . . , c of
the linear discriminant functions f
y
(x) = ¸w
y
x) +b.
.b [c 1] Parameters b
y
∈ R, y = 1, . . . , c.
.fun = ’linclass’ Identiﬁes function associated with this data type.
trn = load(’riply_trn’); % load training data
tst = load(’riply_tst’); % load testing data
model = fld( trn ); % train FLD classifier
ypred = linclass( tst.X, model ); % classify testing data
cerror( ypred, tst.y ) % evaluate testing error
ans =
0.1080
2.2 Linear separation of ﬁnite sets of vectors
The input training data T = ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ consists of pairs of observations
x
i
∈ R
n
and corresponding class labels y
i
∈ ¸ = ¦1, . . . , c¦. The implemented methods
solve the task of (i) training separating hyperplane for binary case c = 2, (ii) training
εoptimal hyperplane and (iii) training the multiclass linear classiﬁer. The deﬁnitions
of these tasks are given below.
The problem of training the binary (c = 2) linear classiﬁer (2.2) with zero training
error is equivalent to ﬁnding the hyperplane H = ¦x: ¸w x) +b = 0¦ which separates
the training vectors of the ﬁrst y = 1 and the second y = 2 class. The problem is
formally deﬁned as solving the set of linear inequalities
¸w x
i
) + b ≥ 0 , y
i
= 1 ,
¸w x
i
) + b < 0 , y
i
= 2 .
(2.3)
with respect to the vector w ∈ R
n
and the scalar b ∈ R. If the inequalities (2.3) have a
solution then the training data T is linearly separable. The Perceptron (Section 2.2.1)
10
and Kozinec’s (Section 2.2.2) algorithm are useful to train the binary linear classiﬁers
from the linearly separable data.
If the training data is linearly separable then the optimal separating hyperplane
can be deﬁned the solution of the following task
(w
∗
, b
∗
) = argmax
w,b
m(w, b)
= argmax
w,b
min
_
min
i∈I
1
¸w x
i
) + b
w
, min
i∈I
2
−
¸w x
i
) + b
w
_
,
(2.4)
where 1
1
= ¦i: y
i
= 1¦ and 1
2
= ¦i: y
i
= 2¦ are sets of indices. The optimal separating
hyperplane H
∗
= ¦x ∈ R
n
: ¸w
∗
x) + b
∗
= 0¦ separates training data T with maxi
mal margin m(w
∗
, b
∗
). This task is equivalent to training the linear Support Vector
Machines (see Section 5). The optimal separating hyperplane cannot be found exactly
except special cases. Therefore the numerical algorithms seeking the approximate so
lution are applied instead. The εoptimal solution is deﬁned as the vector w and scalar
b such that the inequality
m(w
∗
, b
∗
) −m(w, b) ≤ ε , (2.5)
holds. The parameter ε ≥ 0 deﬁnes closeness to the optimal solution in terms of the
margin. The εoptimal solution can be found by Kozinec’s algorithm (Section 2.2.2).
The problem of training the multiclass linear classiﬁer c > 2 with zero training
error is formally stated as the problem of solving the set of linear inequalities
¸w
y
i
x
i
) + b
y
i
> ¸w
y
x
i
) + b
y
, i = 1, . . . , l , y
i
,= y , (2.6)
with respect to the vectors w
y
∈ R
n
, y ∈ ¸ and scalars b
y
∈ R, y ∈ ¸. The task (2.6)
for linearly separable training data T can be solved by the modiﬁed Perceptron (Sec
tion 2.2.3) or Kozinec’s algorithm. An interactive demo on the algorithms separating
the ﬁnite sets of vectors by a hyperplane is implemented in demo linclass.
2.2.1 Perceptron algorithm
The input is a set T = ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ of binary labeled y
i
∈ ¦1, 2¦ training
vectors x
i
∈ R
n
. The problem of training the separating hyperplane (2.3) can be
formally rewritten to a simpler form
¸v z
i
) > 0 , i = 1, . . . , l , (2.7)
where the vector v ∈ R
n+1
is constructed as
v = [w; b] , (2.8)
11
and transformed training data Z = ¦z
1
, . . . , z
l
¦ are deﬁned as
z
i
=
_
[x
i
; 1] if y
i
= 1 ,
−[x
i
; 1] if y
i
= 2 .
(2.9)
The problem of solving (2.7) with respect to the unknown vector v ∈ R
n+1
is equivalent
to the original task (2.3). The parameters (w, b) of the linear classiﬁer are obtained
from the found vector v by inverting the transformation (2.8).
The Perceptron algorithm is an iterative procedure which builds a series of vectors
v
(0)
, v
(1)
, . . . , v
(t)
until the set of inequalities (2.7) is satisﬁed. The initial vector v
(0)
can
be set arbitrarily (usually v = 0). The Novikoﬀ theorem ensures that the Perceptron
algorithm stops after ﬁnite number of iterations t if the training data are linearly
separable. Perceptron algorithm is implemented in function perceptron.
References: Analysis of the Perceptron algorithm including the Novikoﬀ’s proof of
convergence can be found in Chapter 5 of the book [26]. The Perceptron from the
neural network perspective is described in the book [3].
Example: Training binary linear classiﬁer with Perceptron
The example shows the application of the Perceptron algorithm to ﬁnd the binary
linear classiﬁer for synthetically generated 2dimensional training data. The generated
training data contain 50 labeled vectors which are linearly separable with margin 1.
The found classiﬁer is visualized as a separating hyperplane (line in this case) H =
¦x ∈ R
2
: ¸w x) + b = 0¦. See Figure 2.1.
data = genlsdata( 2, 50, 1); % generate training data
model = perceptron( data ); % call perceptron
figure;
ppatterns( data ); % plot training data
pline( model ); % plot separating hyperplane
2.2.2 Kozinec’s algorithm
The input is a set T = ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ of binary labeled y
i
∈ ¦1, 2¦ training
vectors x
i
∈ R
n
. The Kozinec’s algorithm builds a series of vectors w
(0)
1
, w
(1)
1
, . . . , w
(t)
1
and w
(0)
2
, w
(1)
2
, . . . , w
(t)
2
which converge to the vector w
∗
1
and w
∗
2
respectively. The
vectors w
∗
1
and w
∗
2
are the solution of the following task
w
∗
1
, w
∗
2
= argmin
w
1
∈X
1
,w
2
∈X
2
w
1
−w
2
 ,
where A
1
stands for the convex hull of the training vectors of the ﬁrst class A
1
=
¦x
i
: y
i
= 1¦ and A
2
for the convex hull of the second class likewise. The vector w
∗
=
12
−55 −50 −45 −40 −35 −30
−55
−50
−45
−40
−35
−30
Figure 2.1: Linear classiﬁer trained by the Perceptron algorithm
w
∗
1
−w
∗
2
and the bias b
∗
=
1
2
(w
∗
2

2
−w
∗
1

2
) determine the optimal hyperplane (2.4)
separating the training data with the maximal margin.
The Kozinec’s algorithm is proven to converge to the vectors w
∗
1
and w
∗
2
in inﬁnite
number of iterations t = ∞. If the εoptimal optimality stopping condition is used
then the Kozinec’s algorithm converges in ﬁnite number of iterations. The Kozinec’s
algorithm can be also used to solve a simpler problem of ﬁnding the separating hyper
plane (2.3). Therefore the following two stopping conditions are implemented:
I. The separating hyperplane (2.3) is sought for (ε < 0). The Kozinec’s algorithm is
proven to converge in a ﬁnite number of iterations if the separating hyperplane
exists.
II. The εoptimal hyperplane (2.5) is sought for (ε ≥ 0). Notice that setting ε = 0
force the algorithm to seek the optimal hyperplane which is generally ensured to
be found in an inﬁnite number of iterations t = ∞.
The Kozinec’s algorithm which trains the binary linear classiﬁer is implemented in the
function ekozinec.
References: Analysis of the Kozinec’s algorithm including the proof of convergence
can be found in Chapter 5 of the book [26].
Example: Training εoptimal hyperplane by the Kozinec’s algorithm
The example shows application of the Kozinec’s algorithm to ﬁnd the (ε = 0.01)
optimal hyperplane for synthetically generated 2dimensional training data. The gen
erated training data contains 50 labeled vectors which are linearly separable with
13
margin 1. The found classiﬁer is visualized (Figure 2.2) as a separating hyperplane
H = ¦x ∈ R
2
: ¸w x) +b = 0¦ (straight line in this case). It could be veriﬁed that the
found hyperplane has margin model.margin which satisﬁes the εoptimality condition.
data = genlsdata(2,50,1); % generate data
% run ekozinec
model = ekozinec(data, struct(’eps’,0.01));
figure;
ppatterns(data); % plot data
pline(model); % plot found hyperplane
−10 −8 −6 −4 −2 0 2 4 6 8 10
−10
−8
−6
−4
−2
0
2
4
6
8
Figure 2.2: εoptimal hyperplane trained with the Kozinec’s algorithm.
2.2.3 Multiclass linear classiﬁer
The input training data T = ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ consists of pairs of observations
x
i
∈ R
n
and corresponding class labels y
i
∈ ¸ = ¦1, . . . , c¦. The task is to design
the linear classiﬁer (2.1) which classiﬁes the training data T without error. The prob
lem of training (2.1) is equivalent to solving the set of linear inequalities (2.6). The
inequalities (2.6) can be formally transformed to a simpler set
¸v z
y
i
) > 0 , i = 1, . . . , l , y = ¸ ¸ ¦y
i
¦ , (2.10)
where the vector v ∈ R
(n+1)c
contains the originally optimized parameters
v = [w
1
; b
1
; w
2
; b
2
; . . . ; w
c
; b
c
] .
14
The set of vectors z
y
i
∈ R
(n+1)c
, i = 1, . . . , l, y ∈ ¸ ¸ ¦y
i
¦ is created from the training
set T such that
z
y
i
(j) =
⎧
⎨
⎩
[x
i
; 1] , for j = y
i
,
−[x
i
; 1] , for j = y ,
0 , otherwise.
(2.11)
where z
y
i
(j) stands for the jth slot between coordinates (n+1)(j −1)+1 and (n+1)j.
The described transformation is known as the Kesler’s construction.
The transformed task (2.10) can be solved by the Perceptron algorithm. The simple
form of the Perceptron updating rule allows to implement the transformation implicitly
without mapping the training data T into (n + 1)cdimensional space. The described
method is implemented in function mperceptron.
References: The Kesler’s construction is described in books [26, 6].
Example: Training multiclass linear classiﬁer by the Perceptron
The example shows application of the Perceptron rule to train the multiclass linear
classiﬁer for the synthetically generated 2dimensional training data pentagon.mat. A
user’s own data can be generated interactively using function createdata(’finite’,10)
(number 10 stands for number of classes). The found classiﬁer is visualized as its sep
arating surface which is known to be piecewise linear and the class regions should be
convex (see Figure 2.3).
data = load(’pentagon’); % load training data
model = mperceptron(data); % run training algorithm
figure;
ppatterns(data); % plot training data
pboundary(model); % plot decision boundary
2.3 Fisher Linear Discriminant
The input is a set T = ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ of binary labeled y
i
∈ ¦1, 2¦ training
vectors x
i
∈ R
n
. Let 1
y
= ¦i: y
i
= y¦, y ∈ ¦1, 2¦ be sets of indices of training
vectors belonging to the ﬁrst y = 1 and the second y = 2 class, respectively. The class
separability in a direction w ∈ R
n
is deﬁned as
F(w) =
¸w S
B
w)
¸w S
W
w)
, (2.12)
where S
B
is the betweenclass scatter matrix
S
B
= (µ
1
−µ
2
)(µ
1
−µ
2
)
T
, µ
y
=
1
[1
y
[
i∈Iy
x
i
, y ∈ ¦1, 2¦ ,
15
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Figure 2.3: Multiclass linear classiﬁer trained by the Perceptron algorithm.
and S
W
is the within class scatter matrix deﬁned as
S
W
= S
1
+S
2
, S
y
=
i∈Yy
(x
i
−µ
y
)(x
i
−µ
y
)
T
, y ∈ ¦1, 2¦ .
In the case of the Fisher Linear Discriminant (FLD), the parameter vector w of the
linear discriminant function f(x) = ¸w x) + b is determined to maximize the class
separability criterion (2.12)
w = argmax
w
F(w
) = argmax
w
¸w
S
B
w
)
¸w
S
W
w
)
. (2.13)
The bias b of the linear rule must be determined based on another principle. The
STPRtool implementation, for instance, computes such b that the equality
¸w µ
1
) + b = −(¸w µ
2
) + b) ,
holds. The STPRtool contains two implementations of FLD:
I. The classical solution of the problem (2.13) using the matrix inversion
w = S
W
−1
(µ
1
−µ
2
) .
This case is implemented in the function fld.
16
II. The problem (2.13) can be reformulated as the quadratic programming (QP) task
w = argmin
w
¸w
S
W
w
) ,
subject to
¸w
(µ
1
−µ
2
)) = 2 .
The QP solver quadprog of the Matlab Optimization Toolbox is used in the
toolbox implementation. This method can be useful when the matrix inversion
is hard to compute. The method is implemented in the function fldqp.
References: The Fisher Linear Discriminant is described in Chapter 3 of the book [6]
or in the book [3].
Example: Fisher Linear Discriminant
The binary linear classiﬁer based on the Fisher Linear Discriminant is trained on
the Riply’s data set riply trn.mat. The QP formulation fldqp of FLD is used.
However, the function fldqp can be replaced by the standard approach implemented
in the function fld which would yield the same solution. The found linear classiﬁer is
visualized (see Figure 2.4) and evaluated on the testing data riply tst.mat.
trn = load(’riply_trn’); % load training data
model = fldqp(trn); % compute FLD
figure;
ppatterns(trn); pline(model); % plot data and solution
tst = load(’riply_tst’); % load testing data
ypred = linclass(tst.X,model); % classify testing data
cerror(ypred,tst.y) % compute testing error
ans =
0.1080
2.4 Generalized Anderson’s task
The classiﬁed object is described by the vector of observations x ∈ A ⊆ R
n
and a
binary hidden state y ∈ ¦1, 2¦. The class conditional distributions p
XY
(x[y), y ∈ ¦1, 2¦
are known to be multivariate Gaussian distributions. The parameters (µ
1
, Σ
1
) and
(µ
2
, Σ
2
) of these class distributions are unknown. However, it is known that the
parameters (µ
1
, Σ
1
) belong to a certain ﬁnite set of parameters ¦(µ
i
, Σ
i
): i ∈ 1
1
¦.
Similarly (µ
2
, Σ
2
) belong to a ﬁnite set ¦(µ
i
, Σ
i
): i ∈ 1
2
¦. Let q: A ⊆ R
n
→ ¦1, 2¦
17
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Figure 2.4: Linear classiﬁer based on the Fisher Linear Discriminant.
be a binary linear classiﬁer (2.2) with discriminant function f(x) = ¸w x) + b. The
probability of misclassiﬁcation is deﬁned as
Err(w, b) = max
i∈I
1
∪I
2
ε(w, b, µ
i
, Σ
i
) ,
where ε(w, b, µ
i
, Σ
i
) is probability that the Gaussian random vector x with mean
vector µ
i
and the covariance matrix Σ
i
satisﬁes q(x) = 1 for i ∈ 1
2
or q(x) = 2 for
i ∈ 1
1
. In other words, it is the probability that the vector x will be misclassiﬁed by
the linear rule q.
The Generalized Anderson’s task (GAT) is to ﬁnd the parameters (w
∗
, b
∗
) of the
linear classiﬁer
q(x) =
_
1 , for f(x) = ¸w
∗
x) + b
∗
≥ 0 ,
2 , for f(x) = ¸w
∗
x) + b
∗
< 0 ,
such that the error Err(w
∗
, b
∗
) is minimal
(w
∗
, b
∗
) = argmin
w,b
Err(w, b) = argmin
w,b
max
i∈I
1
∪I
2
ε(w, b, µ
i
, Σ
i
) . (2.14)
The original Anderson’s task is a special case of (2.14) when [1
1
[ = 1 and [1
2
[ = 1.
The probability ε(w, b, µ
i
, Σ
i
) is proportional to the reciprocal of the Mahalanobis
distance r
i
between the (µ
i
, Σ
i
) and the nearest vector of the separating hyperplane
H = ¦x ∈ R
n
: ¸w x) + b = 0¦, i.e.,
r
i
= min
x∈H
¸(µ
i
−x) (Σ
i
)
−1
(µ
i
−x)) =
¸w µ
i
) + b
_
¸w Σ
i
w)
.
18
The exact relation between the probability ε(w, b, µ
i
, Σ
i
) and the corresponding Ma
halanobis distance r
i
is given by the integral
ε(w, b, µ
i
, Σ
i
) =
_
∞
r
i
1
√
2π
e
−
1
2
t
2
dt . (2.15)
The optimization problem (2.14) can be equivalently rewritten as
(w
∗
, b
∗
) = argmax
w,b
F(w, b) = argmax
w,b
min
i∈I
1
∪I
2
¸w µ
i
) + b
_
¸w Σ
i
w)
.
which is more suitable for optimization. The objective function F(w, b) is proven to
be convex in the region where the probability of misclassiﬁcation Err(w, b) is less than
0.5. However, the objective function F(w, b) is not diﬀerentiable.
The STPRtool contains implementations of the algorithm solving the original An
derson’s task as well as implementations of three diﬀerent approaches to solve the
Generalized Anderson’s task which are described bellow. An interactive demo on the
algorithms solving the Generalized Anderson’s task is implemented in demo anderson.
References: The original Anderson’s task was published in [1]. A detailed description
of the Generalized Anderson’s task and all the methods implemented in the STPRtool
is given in book [26].
2.4.1 Original Anderson’s algorithm
The algorithm is implemented in the function androrig. This algorithm solves the
original Anderson’s task deﬁned for two Gaussians only [1
1
[ = [1
2
[ = 1. In this case,
the optimal solution vector w ∈ R
n
is obtained by solving the following problem
w = ((1 −λ)Σ
1
+ λΣ
2
)
−1
(µ
1
−µ
2
) ,
1 −λ
λ
=
¸
¸w Σ
2
w)
¸w Σ
1
w)
,
where the scalar 0 < λ < 1 is unknown. The problem is solved by an iterative algorithm
which stops while the condition
¸
¸
¸
¸
γ −
1 −λ
λ
¸
¸
¸
¸
≤ ε , γ =
¸
¸w Σ
2
w)
¸w Σ
1
w)
,
is satisﬁed. The parameter ε deﬁnes the closeness to the optimal solution.
Example: Solving the original Anderson’s task
19
The original Anderson’s task is solved for the Riply’s data set riply trn.mat. The
parameters (µ
1
, Σ
1
) and (µ
2
, Σ
2
) of the Gaussian distribution of the ﬁrst and the
second class are obtained by the MaximumLikelihood estimation. The found linear
classiﬁer is visualized and the Gaussian distributions (see Figure 2.5). The classiﬁer is
validated on the testing set riply tst.mat.
trn = load(’riply_trn’); % load training data
distrib = mlcgmm(data); % estimate Gaussians
model = androrig(distrib); % solve Anderson’s task
figure;
pandr( model, distrib ); % plot solution
tst = load(’riply_tst’); % load testing data
ypred = linclass( tst.X, model ); % classify testing data
cerror( ypred, tst.y ) % evaluate error
ans =
0.1150
−0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
gauss 1
0.929
gauss 2
1.37
Figure 2.5: Solution of the original Anderson’s task.
20
2.4.2 General algorithm framework
The algorithm is implemented in the function ganders. This algorithm repeats the
following steps:
I. An improving direction (∆w, ∆b) in the parameter space is computed such that
moving along this direction increases the objective function F. The quadprog
function of the Matlab optimization toolbox is used to solve this subtask.
II. The optimal movement k along the direction (∆w, ∆b). The optimal movement k
is determined by 1Dsearch based on the cutting interval algorithm according to
the Fibonacci series. The number of interval cuttings of is an optional parameter
of the algorithm.
III. The new parameters are set
w
(t+1)
: = w
(t)
+ k∆w , b
(t+1)
: = b
(t)
+ k∆b .
The algorithm iterates until the minimal change in the objective function is less than
the prescribed ε, i.e., the condition
F(w
(t+1)
, b
(t+1)
) −F(w
(t)
, b
(t)
) > ε ,
is violated.
Example: Solving the Generalized Anderson’s task
The Generalized Anderson’s task is solved for the data mars.mat. Both the ﬁrst
and second class are described by three 2Ddimensional Gaussian distributions. The
found linear classiﬁer as well as the input Gaussian distributions are visualized (see
Figure 2.6).
distrib = load(’mars’); % load input Gaussians
model = ganders(distrib); % solve the GAT
figure;
pandr(model,distrib); % plot the solution
2.4.3 εsolution using the Kozinec’s algorithm
This algorithm is implemented in the function eanders. The εsolution is deﬁned as a
linear classiﬁer with parameters (w, b) such that
F(w, b) < ε
21
−5 −4 −3 −2 −1 0 1 2 3 4 5
−4
−3
−2
−1
0
1
2
3
4
5
gauss 1
0.0461
gauss 2
0.0483
gauss 3
0.0503
gauss 4
0.0421
gauss 5
0.047
gauss 6
0.0503
Figure 2.6: Solution of the Generalized Anderson’s task.
is satisﬁed. The prescribed ε ∈ (0, 0.5) is the desired upper bound on the probability
of misclassiﬁcation. The problem is transformed to the task of separation of two sets
of ellipsoids which is solved the Kozinec’s algorithm. The algorithm converges in ﬁnite
number of iterations if the given εsolution exists. Otherwise it can get stuck in an
endless loop.
Example: εsolution of the Generalized Anderson’s task
The εsolution of the generalized Anderson’s task is solved for the data mars.mat.
The desired maximal possible probability of misclassiﬁcation is set to 0.06 (it is 6%).
Both the ﬁrst and second class are described by three 2Ddimensional Gaussian distri
butions. The found linear classiﬁer and the input Gaussian distributions are visualized
(see Figure 2.7).
distrib = load(’mars’); % load Gaussians
options = struct(’err’,0.06’); % maximal desired error
model = eanders(distrib,options); % training
figure; pandr(model,distrib); % visualization
2.4.4 Generalized gradient optimization
This algorithm is implemented in the function ggradandr. The objective function
F(w, b) is convex but not diﬀerentiable thus the standard gradient optimization cannot
22
−5 −4 −3 −2 −1 0 1 2 3 4 5
−4
−3
−2
−1
0
1
2
3
4
5
gauss 1
0.0473
gauss 2
0.0496
gauss 3
0.0516
gauss 4
0.0432
gauss 5
0.0483
gauss 6
0.0516
Figure 2.7: εsolution of the Generalized Anderson’s task with maximal 0.06 probability
of misclassiﬁcation.
be used. However, the generalized gradient optimization method can be applied. The
algorithm repeats the following two steps:
I. The generalized gradient (∆w, ∆b) is computed.
II. The optimized parameters are updated
w
(t+1)
: = w
(t)
+ ∆w , b
(t+1)
= b
(t)
+ ∆b .
The algorithm iterates until the minimal change in the objective function is less than
the prescribed ε, i.e., the condition
F(w
(t+1)
, b
(t+1)
) −F(w
(t)
, b
(t)
) > ε ,
is violated. The maximal number of iterations can be used as the stopping condition
as well.
Example: Solving the Generalized Anderson’s task using the generalized
gradient optimization
The generalized Anderson’s task is solved for the data mars.mat. Both the ﬁrst
and second class are described by three 2Ddimensional Gaussian distributions. The
maximal number of iterations is limited to tmax=10000. The found linear classiﬁer as
well as the input Gaussian distributions are visualized (see Figure 2.8).
23
distrib = load(’mars’); % load Gaussians
options = struct(’tmax’,10000); % max # of iterations
model = ggradandr(distrib,options); % training
figure;
pandr( model, distrib ); % visualization
−5 −4 −3 −2 −1 0 1 2 3 4 5
−4
−3
−2
−1
0
1
2
3
4
5
gauss 1
0.0475
gauss 2
0.0498
gauss 3
0.0518
gauss 4
0.0434
gauss 5
0.0485
gauss 6
0.0518
Figure 2.8: Solution of the Generalized Anderson’s task after 10000 iterations of the
generalized gradient optimization algorithm.
24
Chapter 3
Feature extraction
The feature extraction step consists of mapping the input vector of observations x ∈ R
n
onto a new feature description z ∈ R
m
which is more suitable for given task. The linear
projection
z = W
T
x +b , (3.1)
is commonly used. The projection is given by the matrix W [n m] and bias vector
b ∈ R
m
. The Principal Component Analysis (PCA) (Section 3.1) is a representative of
the unsupervised learning method which yields the linear projection (3.1). The Linear
Discriminant Analysis (LDA) (Section 3.2) is a representative of the supervised learning
method yielding the linear projection. The linear data projection is implemented in
function linproj (See examples in Section 3.1 and Section 3.2). The data type used
to describe the linear data projection is deﬁned in Table 3.
The quadratic data mapping projects the input vector x ∈ R
n
onto the feature
space R
m
where m = n(n + 3)/2. The quadratic mapping φ: R
n
→R
m
is deﬁned as
φ(x) = [ x
1
, x
2
, . . . , x
n
,
x
1
x
1
, x
1
x
2
, . . . , x
1
x
n
,
x
2
x
2
, . . . , x
2
x
n
,
.
.
.
x
n
x
n
]
T
,
(3.2)
where x
i
, i = 1, . . . , n are entries of the vector x ∈ R
n
. The quadratic data mapping
is implemented in function qmap.
The kernel functions allow for nonlinear extensions of the linear feature extrac
tion methods. In this case, the input vectors x ∈ R
n
are mapped into a new higher
dimensional feature space T in which the linear methods are applied. This yields a
nonlinear (kernel) projection of data which has generally deﬁned as
z = A
T
k(x) +b , (3.3)
25
where A [l m] is a parameter matrix and b ∈ R
m
a parameter vector. The vector
k(x) = [k(x, x
1
), . . . , k(x, x
l
)]
T
is a vector of kernel functions centered in the training
data or their subset. The x ∈ R
n
stands for the input vector and z ∈ R
m
is the
vector of extracted features. The extracted features are projections of φ(x) onto m
vectors (or functions) from the feature space T. The Kernel PCA (Section 3.3) and
the Generalized Discriminant Analysis (GDA) (Section 3.5) are representatives of the
feature extraction methods yielding the kernel data projection (3.3). The kernel data
projection is implemented in the function kernelproj (See examples in Section 3.3
and Section 3.5). The data type used to describe the kernel data projection is deﬁned
in Table 3. The summary of implemented methods for feature extraction is given in
Table 3.1
Table 3.1: Implemented methods for feature extraction
linproj Linear data projection.
kerproj Kernel data projection.
qmap Quadratic data mapping.
lin2quad Merges linear rule and quadratic mapping.
lin2svm Merges linear rule and kernel projection.
lda Linear Discriminant Analysis.
pca Principal Component Analysis.
pcarec Computes reconstructed vector after PCA projection.
gda Generalized Discriminant Analysis.
greedykpca Greedy Kernel Principal Component Analysis.
kpca Kernel Principal Component Analysis.
kpcarec Reconstructs image after kernel PCA.
demo pcacomp Demo on image compression using PCA.
Table 3.2: Datatype used to describe linear data projection.
Linear data projection (structure array):
.W [n m] Projection matrix W = [w
1
, . . . , w
m
].
.b [m1] Bias vector b.
.fun = ’linproj’ Identiﬁes function associated with this data type.
3.1 Principal Component Analysis
Let T
X
= ¦x
1
, . . . , x
l
¦ be a set of training vectors from the ndimensional input space
R
n
. The set of vectors T
Z
= ¦z
1
, . . . , z
l
¦ is a lower dimensional representation of the
26
Table 3.3: Datatype used to describe kernel data projection.
Kernel data projection (structure array):
.Alpha [l m] Parameter matrix A [l m].
.b [m1] Bias vector b.
.sv.X [n l] Training vectors ¦x
1
, . . . , x
l
¦.
.options.ker [string] Kernel identiﬁer.
.options.arg [1 p] Kernel argument(s).
.fun = ’kernelproj’ Identiﬁes function associated with this data type.
input training vectors T
X
in the mdimensional space R
m
. The vectors T
Z
are obtained
by the linear orthonormal projection
z = W
T
x +b , (3.4)
where the matrix W [n m] and the vector b [m 1] are parameters of the projec
tion. The reconstructed vectors T
˜
X
= ¦ ˜ x
1
, . . . , ˜ x
l
¦ are computed by the linear back
projection
˜ x = W(z −b) , (3.5)
obtained by inverting (3.4). The mean square reconstruction error
ε
MS
(W, b) =
1
l
l
i=1
x
i
− ˜ x
i

2
, (3.6)
is a function of the parameters of the linear projections (3.4) and (3.5). The Principal
Component Analysis (PCA) is the linear orthonormal projection (3.4) which allows
for the minimal mean square reconstruction error (3.6) of the training data T
X
. The
parameters (W, b) of the linear projection are the solution of the optimization task
(W, b) = argmin
W
,b
ε
MS
(W
, b
) , (3.7)
subject to
¸w
i
w
j
) = δ(i, j) , ∀i, j ,
where w
i
, i = 1, . . . , m are column vectors of the matrix W = [w
1
, . . . , w
m
] and
δ(i, j) is the Kronecker delta function. The solution of the task (3.7) is the matrix
W = [w
1
, . . . , w
m
] containing the m eigenvectors of the sample covariance matrix
which have the largest eigen values. The vector b equals to W
T
µ, where µ is the
sample mean of the training data. The PCA is implemented in the function pca. A
demo showing the use of PCA for image compression is implemented in the script
demo pcacomp.
27
References: The PCA is treated throughout for instance in books [13, 3, 6].
Example: Principal Component Analysis
The input data are synthetically generated from the 2dimensional Gaussian dis
tribution. The PCA is applied to train the 1dimensional subspace to approximate
the input data with the minimal reconstruction error. The training data, the recon
structed data and the 1dimensional subspace given by the ﬁrst principal component
are visualized (see Figure 3.1).
X = gsamp([5;5],[1 0.8;0.8 1],100); % generate data
model = pca(X,1); % train PCA
Z = linproj(X,model); % lower dim. proj.
XR = pcarec(X,model); % reconstr. data
figure; hold on; axis equal; % visualization
h1 = ppatterns(X,’kx’);
h2 = ppatterns(XR,’bo’);
[dummy,mn] = min(Z);
[dummy,mx] = max(Z);
h3 = plot([XR(1,mn) XR(1,mx)],[XR(2,mn) XR(2,mx)],’r’);
legend([h1 h2 h3], ...
’Input vectors’,’Reconstructed’, ’PCA subspace’);
2 3 4 5 6 7 8
2.5
3
3.5
4
4.5
5
5.5
6
6.5
7
7.5
Input vectors
Reconstructed
PCA subspace
Figure 3.1: Example on the Principal Component Analysis.
28
3.2 Linear Discriminant Analysis
Let T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦, x
i
∈ R
n
, y ∈ ¸ = ¦1, 2, . . . , c¦ be labeled set
of training vectors. The withinclass S
W
and betweenclass S
B
scatter matrices are
described as
S
W
=
y∈Y
S
y
, S
y
=
i∈Iy
(x
i
−µ
i
)(x
i
−µ
i
)
T
,
S
B
=
y∈Y
[1
y
[(µ
y
−µ)(µ
y
−µ)
T
,
(3.8)
where the total mean vector µ and the class mean vectors µ
y
, y ∈ ¸ are deﬁned as
µ =
1
l
l
i=1
x
i
, µ
y
=
1
[1
y
[
i∈Iy
x
i
, y ∈ ¸ .
The goal of the Linear Discriminant Analysis (LDA) is to train the linear data projec
tion
z = W
T
x ,
such that the class separability criterion
F(W) =
det(
˜
S
B
)
det(
˜
S
W
)
=
det(S
B
)
det(S
W
)
,
is maximized. The
˜
S
B
is the betweenclass and
˜
S
W
is the withinclass scatter matrix
of projected data which is deﬁned similarly to (3.8).
References: The LDA is treated for instance in books [3, 6].
Example 1: Linear Discriminant Analysis
The LDA is applied to extract 2 features from the Iris data set iris.mat which
consists of the labeled 4dimensional data. The data after the feature extraction step
are visualized in Figure 3.2.
orig_data = load(’iris’); % load input data
model = lda(orig_data,2); % train LDA
ext_data = linproj(orig_data,model); % feature extraction
figure; ppatterns(ext_data); % plot ext. data
Example 2: PCA versus LDA
This example shows why the LDA is superior to the PCA when features good for
classiﬁcation are to be extracted. The synthetical data are generated from the Gaussian
mixture model. The LDA and PCA are trained on the generated data. The LDA and
29
−1.5 −1 −0.5 0 0.5 1 1.5
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
Figure 3.2: Example shows 2dimensional data extracted from the originally 4
dimensional Iris data set using Linear Discriminant Analysis.
PCA directions are displayed as connection of the vectors projected to the trained LDA
and PCA subspaces and reprojected back to the input space (see Figure 3.3a). The
extracted data using the LDA and PCA model are displayed with the Gaussians ﬁtted
by the MaximumLikelihood (see Figure 3.3b).
% Generate data
distrib.Mean = [[5;4] [4;5]]; % mean vectors
distrib.Cov(:,:,1) = [1 0.9; 0.9 1]; % 1st covariance
distrib.Cov(:,:,2) = [1 0.9; 0.9 1]; % 2nd covariance
distrib.Prior = [0.5 0.5]; % Gaussian weights
data = gmmsamp(distrib,250); % sample data
lda_model = lda(data,1); % train LDA
lda_rec = pcarec(data.X,lda_model);
lda_data = linproj(data,lda_model);
pca_model = pca(data.X,1); % train PCA
pca_rec = pcarec(data.X,pca_model);
pca_data = linproj( data,pca_model);
figure; hold on; axis equal; % visualization
ppatterns(data);
30
h1 = plot(lda_rec(1,:),lda_rec(2,:),’r’);
h2 = plot(pca_rec(1,:),pca_rec(2,:),’b’);
legend([h1 h2],’LDA direction’,’PCA direction’);
figure; hold on;
subplot(2,1,1); title(’LDA’); ppatterns(lda_data);
pgauss(mlcgmm(lda_data));
subplot(2,1,2); title(’PCA’); ppatterns(pca_data);
pgauss(mlcgmm(pca_data));
3.3 Kernel Principal Component Analysis
The Kernel Principal Component Analysis (Kernel PCA) is the nonlinear extension of
the ordinary linear PCA. The input training vectors T
X
= ¦x
1
, . . . , x
l
¦, x
i
∈ A ⊆ R
n
are mapped by φ: A → T to a high dimensional feature space T. The linear PCA
is applied on the mapped data T
Φ
= ¦φ(x
1
), . . . , φ(x
l
)¦. The computation of the
principal components and the projection on these components can be expressed in
terms of dot products thus the kernel functions k: A A →R can be employed. The
kernel PCA trains the kernel data projection
z = A
T
k(x) +b , (3.9)
such that the reconstruction error
ε
KMS
(A, b) =
1
l
l
i=1
φ(x
i
) −
˜
φ(x
i
)
2
, (3.10)
is minimized. The reconstructed vector
˜
φ(x) is given as a linear combination of the
mapped data T
Φ
˜
φ(x) =
l
i=1
β
i
φ(x
i
) , β = A(z −b) . (3.11)
In contrast to the linear PCA, the explicit projection from the feature space T to the
input space A usually do not exist. The problem is to ﬁnd the vector x ∈ A its image
φ(x) ∈ T well approximates the reconstructed vector
˜
φ(x) ∈ T. This procedure is
implemented in the function kpcarec for the case of the Radial Basis Function (RBF)
kernel. The procedure consists of the following step:
I. Project input vector x
in
∈ R
n
onto its lower dimensional representation z ∈ R
m
using (3.9).
31
II. Computes vector x
out
∈ R
n
which is good preimage of the reconstructed vector
˜
φ(x
in
), such that
x
out
= argmin
x
φ(x) −
˜
φ(x
in
)
2
.
References: The kernel PCA is described at length in book [29, 30].
Example: Kernel Principal Component Analysis
The example shows using the kernel PCA for data denoising. The input data are
synthetically generated 2dimensional vectors which lie on a circle and are corrupted by
the Gaussian noise. The kernel PCA model with RBF kernel is trained. The function
kpcarec is used to ﬁnd preimages of the reconstructed data (3.11). The result is
visualized in Figure 3.4.
X = gencircledata([1;1],5,250,1); % generate circle data
options.ker = ’rbf’; % use RBF kernel
options.arg = 4; % kernel argument
options.new_dim = 2; % output dimension
model = kpca(X,options); % compute kernel PCA
XR = kpcarec(X,model); % compute reconstruced data
figure; % Visualization
h1 = ppatterns(X);
h2 = ppatterns(XR, ’+r’);
legend([h1 h2],’Input vectors’,’Reconstructed’);
3.4 Greedy kernel PCA algorithm
The Greedy Kernel Principal Analysis (Greedy kernel PCA) is an eﬃcient algorithm
to compute the ordinary kernel PCA. Let T
X
= ¦x
1
, . . . , x
l
¦, x
i
∈ R
n
be the set of
input training vectors. The goal is to train the kernel data projection
z = A
T
k
S
(x) +b , (3.12)
where A [d m] is the parameter matrix, b [m 1] is the parameter vector and
k
S
= [k(x, s
1
), . . . , k(x, s
l
)]
T
are the kernel functions centered in the vectors T
S
=
¦s
1
, . . . , s
d
¦. The vector set T
S
is a subset of training data T
X
. In contrast to the
ordinary kernel PCA, the subset T
S
does not contain all the training vectors T
X
thus
the complexity of the projection (3.12) is reduced compared to (3.9). The objective of
the Greedy kernel PCA is to minimize the reconstruction error (3.11) while the size d
of the subset T
S
is kept small. The Greedy kernel PCA algorithm is implemented in
the function greedykpca.
32
References: The implemented Greedy kernel PCA algorithm is described in [11].
Example: Greedy Kernel Principal Component Analysis
The example shows using kernel PCA for data denoising. The input data are
synthetically generated 2dimensional vectors which lie on a circle and are corrupted
with the Gaussian noise. The Greedy kernel PCA model with RBF kernel is trained.
The number of vectors deﬁning the kernel projection (3.12) is limited to 25. In contrast
the ordinary kernel PCA (see experiment in Section 3.3) requires all the 250 training
vectors. The function kpcarec is used to ﬁnd preimages of the reconstructed data
obtained by the kernel PCA projection. The training and reconstructed vectors are
visualized in Figure 3.5. The vector of the selected subset T
X
are visualized as well.
X = gencircledata([1;1],5,250,1); % generate training data
options.ker = ’rbf’; % use RBF kernel
options.arg = 4; % kernel argument
options.new_dim = 2; % output dimension
options.m = 25; % size of sel. subset
model = greedykpca(X,options); % run greedy algorithm
XR = kpcarec(X,model); % reconstructed vectors
figure; % visualization
h1 = ppatterns(X);
h2 = ppatterns(XR,’+r’);
h3 = ppatterns(model.sv.X,’ob’,12);
legend([h1 h2 h3], ...
’Training set’,’Reconstructed set’,’Selected subset’);
3.5 Generalized Discriminant Analysis
The Generalized Discriminant Analysis (GDA) is the nonlinear extension of the ordi
nary Linear Discriminant Analysis (LDA). The input training data T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦,
x
i
∈ A ⊆ R
n
, y ∈ ¸ = ¦1, . . . , c¦ are mapped by φ: A → T to a high dimen
sional feature space T. The ordinary LDA is applied on the mapped data T
ΦY
=
¦(φ(x
1
), y
1
), . . . , (φ(x
l
), y
l
)¦. The computation of the projection vectors and the pro
jection on these vectors can be expressed in terms of dot products thus the kernel
functions k: A A → R can be employed. The resulting kernel data projection is
deﬁned as
z = A
T
k(x) +b ,
33
where A [l m] is the matrix and b [m 1] the vector trained by the GDA. The
parameters (A, b) are trained to increase the betweenclass scatter and decrease the
withinclass scatter of the extracted data T
ZY
= ¦(z
1
, y
1
), . . . , (z
l
, y
l
)¦, z
i
∈ R
m
. The
GDA is implemented in the function gda.
References: The GDA was published in the paper [2].
Example: Generalized Discriminant Analysis
The GDA is trained on the Iris data set iris.mat which contains 3 classes of
4dimensional data. The RBF kernel with kernel width σ = 1 is used. Notice, that
setting σ too low leads to the perfect separability of the training data T
XY
but the kernel
projection is heavily overﬁtted. The extracted data ¸
ZY
are visualized in Figure 3.6.
orig_data = load(’iris’); % load data
options.ker = ’rbf’; % use RBF kernel
options.arg = 1; % kernel arg.
options.new_dim = 2; % output dimension
model = gda(orig_data,options); % train GDA
ext_data = kernelproj(orig_data,model); % feature ext.
figure; ppatterns(ext_data); % visualization
3.6 Combining feature extraction and linear classi
ﬁer
Let T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ be a training set of observations x
i
∈ A ⊆ R
n
and
corresponding hidden states y
i
∈ ¸ = ¦1, . . . , c¦. Let φ: A →R
m
be a mapping
φ(x) = [φ
1
(x), . . . , φ
m
(x)]
T
deﬁning a nonlinear feature extraction step (e.g., quadratic data mapping or kernel
projection). The feature extraction step applied to the input data T
XY
yields extracted
data T
ΦY
= ¦(φ(x
1
), y
1
), . . . , (φ(x
l
), y
l
)¦. Let f(x) = ¸w φ(x)) + b be a linear
discriminant function found by an arbitrary training method on the extracted data
T
ΦY
. The function f can be seen as a nonlinear function of the input vectors x ∈ A,
as
f(x) = ¸w φ(x)) + b =
m
i=1
w
i
φ
i
(x) + b .
In the case of the quadratic data mapping (3.2) used as the feature extraction step,
the discriminant function can be written in the following form
f(x) = ¸x Ax) +¸x b) + c . (3.13)
34
The discriminant function (3.13) deﬁnes the quadratic classiﬁer described in Section 6.4.
Therefore the quadratic classiﬁer can be trained by (i) applying the quadratic data
mapping qmap, (ii) training a linear rule on the mapped data (e.g., perceptron, fld)
and (iii) combining the quadratic mapping and the linear rule. The combining is
implement in function lin2quad.
In the case of the kernel projection (3.3) applied as the feature extraction step, the
resulting discriminant function has the following form
f(x) = ¸α k(x)) + b . (3.14)
The discriminant function (3.14) deﬁnes the kernel (SVM type) classiﬁer described
in Section 5. This classiﬁer can be trained by (i) applying the kernel projection
kernelproj, (ii) training a linear rule on the extracted data and (iii) combing the kernel
projection and the linear rule. The combining is implemented in function lin2svm.
Example: Quadratic classiﬁer trained the Perceptron
This example shows how to train the quadratic classiﬁer using the Perceptron al
gorithm. The Perceptron algorithm produces the linear rule only. The input training
data are linearly nonseparable but become separable after the quadratic data map
ping is applied. Function lin2quad is used to compute the parameters of the quadratic
classiﬁer explicitly based on the found linear rule in the feature space and the mapping.
Figure 3.7 shows the decision boundary of the quadratic classiﬁer.
% load training data living in the input space
input_data = load(’vltava’);
% map data to the feature space
map_data = qmap(input_data);
% train linear rule in the feature sapce
lin_model = perceptron(map_data);
% compute parameters of the quadratic classifier
quad_model = lin2quad(lin_model);
% visualize the quadratic classifier
figure; ppatterns(input_data);
pboundary(quad_model);
Example: Kernel classiﬁer from the linear rule
This example shows how to train the kernel classiﬁer using arbitrary training algo
rithm for linear rule. The Fisher Linear Discriminant was used in this example. The
35
greedy kernel PCA algorithm was applied as the feature extraction step. The Riply’s
data set was used for training and evaluation. Figure 3.8 shows the found kernel clas
siﬁer, the training data and vectors used in the kernel expansion of the discriminant
function.
% load input data
trn = load(’riply_trn’);
% train kernel PCA by greedy algorithm
options = struct(’ker’,’rbf’,’arg’,1,’new_dim’,10);
kpca_model = greedykpca(trn.X,options);
% project data
map_trn = kernelproj(trn,kpca_model);
% train linear classifier
lin_model = fld(map_trn);
% combine linear rule and kernel PCA
kfd_model = lin2svm(kpca_model,lin_model);
% visualization
figure;
ppatterns(trn); pboundary(kfd_model);
ppatterns(kfd_model.sv.X,’ok’,13);
% evaluation on testing data
tst = load(’riply_tst’);
ypred = svmclass(tst.X,kfd_model);
cerror(ypred,tst.y)
ans =
0.0970
36
0 1 2 3 4 5 6 7 8
1
2
3
4
5
6
7
LDA direction
PCA direction
(a)
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
LDA
−4 −3 −2 −1 0 1 2 3 4 5
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
PCA
(b)
Figure 3.3: Comparison between PCA and LDA. Figure (a) shows input data and
found LDA (red) and PCA (blue) directions onto which the data are projected. Figure
(b) shows the class conditional Gaussians estimated from data projected data onto
LDA and PCA directions.
37
−8 −6 −4 −2 0 2 4 6 8 10
−8
−6
−4
−2
0
2
4
6
8
10
Input vectors
Reconstructed
Figure 3.4: Example on the Kernel Principal Component Analysis.
−8 −6 −4 −2 0 2 4 6 8 10
−6
−4
−2
0
2
4
6
8
Training set
Reconstructed set
Selected set
Figure 3.5: Example on the Greedy Kernel Principal Component Analysis.
38
−0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
Figure 3.6: Example shows 2dimensional data extracted from the originally 4
dimensional Iris data set using the Generalized Discriminant Analysis. In contrast
to LDA (see Figure 3.2 for comparison), the class clusters are more compact but at the
expense of a higher risk of overﬁtting.
−0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
Figure 3.7: Example shows quadratic classiﬁer found by the Perceptron algorithm on
the data mapped to the feature by the quadratic mapping.
39
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Figure 3.8: Example shows the kernel classiﬁer found by the Fisher Linear Discriminant
algorithm on the data mapped to the feature by the kernel PCA. The greedy kernel
PCA was used therefore only a subset of training data was used to determine the kernel
projection.
40
Chapter 4
Density estimation and clustering
The STPRtool represents a probability distribution function P
X
(x), x ∈ R
n
as a
structure array which must contain ﬁeld .fun (see Table 4.2). The ﬁeld .fun speciﬁes
function which is used to evaluate P
X
(x) for given set of realization ¦x
1
, . . . , x
l
¦ of a
random vector. Let the structure model represent a given distribution function. The
Matlab command
y = feval(X,model.fun)
is used to evaluate y
i
= P
X
(x
i
). The matrix X [n l] contains column vectors
¦x
1
, . . . , x
l
¦. The vector y [1 l] contains values of the distribution function. In
particular, the Gaussian distribution and the Gaussian mixture models are described
in Section 4.1. The summary of implemented methods for dealing with probability
density estimation and clustering is given in Table 4.1
4.1 Gaussian distribution and Gaussian mixture model
The value of the multivariate Gaussian probability distribution function in the vector
x ∈ R
n
is given by
P
X
(x) =
1
(2π)
n
2
_
det(Σ)
exp(−
1
2
(x −µ)
T
Σ
−1
(x −µ)) , (4.1)
where µ [n 1] is the mean vector and Σ [n n] is the covariance matrix. The
STPRtool uses speciﬁc datatype to describe a set of Gaussians which can be labeled,
i.e., each pair (µ
i
, Σ
i
) is assigned to the class y
i
∈ ¸. The covariance matrix Σ is
always represented as the square matrix [n n], however, the shape of the matrix can
be indicated in the optimal ﬁeld .cov type (see Table 4.3). The datatype is described
in Table 4.4. The evaluation of the Gaussian distribution is implemented in the function
41
Table 4.1: Implemented methods: Density estimation and clustering
gmmsamp Generates sample from Gaussian mixture model.
gsamp Generates sample from Gaussian distribution.
kmeans Kmeans clustering algorithm.
pdfgauss Evaluates for multivariate Gaussian distribution.
pdfgauss Evaluates Gaussian mixture model.
sigmoid Evaluates sigmoid function.
emgmm ExpectationMaximization Algorithm for Gaussian
mixture model.
melgmm Maximizes Expectation of LogLikelihood for Gaus
sian mixture.
mlcgmm Maximal Likelihood estimation of Gaussian mixture
model.
mlsigmoid Fitting a sigmoid function using ML estimation.
mmgauss Minimax estimation of Gaussian distribution.
rsde Reduced Set Density Estimator.
demo emgmm Demo on ExpectationMaximization (EM) algorithm.
demo mmgauss Demo on minimax estimation for Gaussian.
demo svmpout Fitting a posteriori probability to SVM output.
Table 4.2: Datatype used to represent probability distribution function.
Probability Distribution Function (structure array):
.fun [string] The name of a function which evaluates probability
distribution y = feval( X, model.fun ).
pdfgauss. The random sampling from the Gaussian distribution is implemented in the
function gsamp.
The Gaussian Mixture Model (GMM) is weighted sum of the Gaussian distributions
P
X
(x) =
y∈Y
P
Y
(y)P
XY
(x[y) , (4.2)
where P
Y
(y), y ∈ ¸ = ¦1, . . . , c¦ are weights and P
XY
(x[y) is the Gaussian distribu
tion (4.1) given by parameters (µ
y
, Σ
y
). The datatype used to represent the GMM
is described in Table 4.5. The evaluation of the GMM is implemented in the function
pdfgmm. The random sampling from the GMM is implemented in the function gmmsamp.
Example: Sampling from the Gaussian distribution
The example shows how to represent the Gaussian distribution and how to sample
data it. The sampling from univariate and bivariate Gaussians is shown. The distribu
tion function of the univariate Gaussian is visualized and compared to the histogram
42
Table 4.3: Shape of covariance matrix.
Parameter cov type:
’full’ Full covariance matrix.
’diag’ Diagonal covariance matrix.
’spherical’ Spherical covariance matrix.
Table 4.4: Datatype used to represent the labeled set of Gaussians.
Labeled set of Gaussians (structure array):
.Mean [n k] Mean vectors ¦µ
1
, . . . , µ
k
¦.
.Cov [n n k] Covariance matrices ¦Σ
1
, . . . , Σ
k
¦.
.y [1 k] Labels ¦y
1
, . . . , y
k
¦ assigned to Gaussians.
.cov type [string] Optional ﬁeld which indicates shape of the covariance
matrix (see Table 4.3).
.fun = ’pdfgauss’ Identiﬁes function associated with this data type.
of the sample data (see Figure 4.1a). The isoline of the distribution function of the
bivariate Gaussian is visualized together with the sample data (see Figure 4.1b).
% univariate case
model = struct(’Mean’,1,’Cov’,2); % Gaussian parameters
figure; hold on;
% plot pdf. of the Gaussisan
plot([4:0.1:5],pdfgauss([4:0.1:5],model),’r’);
[Y,X] = hist(gsamp(model,500),10); % compute histogram
bar(X,Y/500); % plot histogram
% bivariate case
model.Mean = [1;1]; % Mean vector
model.Cov = [1 0.6; 0.6 1]; % Covariance matrix
figure; hold on;
ppatterns(gsamp(model,250)); % plot sampled data
pgauss(model); % plot shape
4.2 MaximumLikelihood estimation of GMM
Let the probability distribution
P
XY Θ
(x, y[θ) = P
XY Θ
(x[y, θ)P
Y Θ
(y[θ) ,
43
Table 4.5: Datatype used to represent the Gaussian Mixture Model.
Gaussian Mixture Model (structure array):
.Mean [n c] Mean vectors ¦µ
1
, . . . , µ
c
¦.
.Cov [n n c] Covariance matrices ¦Σ
1
, . . . , Σ
c
¦.
.Prior [c 1] Weights of Gaussians ¦P
K
(1), . . . , P
K
(c)¦.
.fun = ’pdfgmm’ Identiﬁes function associated with this data type.
−4 −3 −2 −1 0 1 2 3 4 5 6
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
−2 −1 0 1 2 3 4 5
−3
−2
−1
0
1
2
3
4
5
gauss 1
0.127
(a) (b)
Figure 4.1: Sampling from the univariate (a) and bivariate (b) Gaussian distribution.
be known up to parameters θ ∈ Θ. The vector of observations x ∈ R
n
and the hidden
state y ∈ ¸ = ¦1, . . . , c¦ are assumed to be realizations of random variables which
are independent and identically distributed according to the P
XY Θ
(x, y[θ). The con
ditional probability distribution P
XY Θ
(x[y, θ) is assumed to be Gaussian distribution.
The θ involves parameters (µ
y
, Σ
y
), y ∈ ¸ of Gaussian components and the values
of the discrete distribution P
Y Θ
(y[θ), y ∈ ¸. The marginal probability P
XΘ
(x[θ) is
the Gaussian mixture model (4.2). The MaximumLikelihood estimation of the pa
rameters θ of the GMM for the case of complete and incomplete data is described in
Section 4.2.1 and Section 4.2.2, respectively.
4.2.1 Complete data
The input is a set T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ which contains both vectors of ob
servations x
i
∈ R
n
and corresponding hidden states y
i
∈ ¸. The pairs (x
i
, y
i
) are
assumed to samples from random variables which are independent and identically dis
tributed (i.i.d.) according to P
XY Θ
(x, y[θ). The logarithm of probability P(T
XY
[θ)
is referred to as the loglikelihood L(θ[T
XY
) of θ with respect to T
XY
. Thanks to the
44
i.i.d. assumption the loglikelihood can be factorized as
L(θ[T
XY
) = log P(T
XY
[θ) =
l
i=1
log P
XY Θ
(x
i
, y
i
[θ) .
The problem of MaximumLikelihood (ML) estimation from the complete data T
XY
is
deﬁned as
θ
∗
= argmax
θ∈Θ
L(θ[T
XY
) = argmax
θ∈Θ
l
i=1
log P
XY Θ
(x
i
, y
i
[θ) . (4.3)
The ML estimate of the parameters θ of the GMM from the complete data set T
XY
has an analytical solution. The computation of the ML estimate is implemented in the
function mlcgmm.
Example: MaximumLikelihood estimate from complete data
The parameters of the Gaussian mixture model is estimated from the complete
Riply’s data set riply trn. The data are binary labeled thus the GMM has two Gaus
sians components. The estimated GMM is visualized as: (i) shape of components of the
GMM (see Figure 4.2a) and (ii) contours of the distribution function (see Figure 4.2b)
.
data = load(’riply_trn’); % load labeled (complete) data
model = mlcgmm(data); % ML estimate of GMM
% visualization
figure; hold on;
ppatterns(data); pgauss(model);
figure; hold on;
ppatterns(data);
pgmm(model,struct(’visual’,’contour’));
4.2.2 Incomplete data
The input is a set T
X
= ¦x
1
, . . . , x
l
¦ which contains only vectors of observations
x
i
∈ R
n
without knowledge about the corresponding hidden states y
i
∈ ¸. The
logarithm of probability P(T
X
[θ) is referred to as the loglikelihood L(θ[T
X
) of θ with
respect to T
X
. Thanks to the i.i.d. assumption the loglikelihood can be factorized as
L(θ[T
X
) = log P(T
X
[θ) =
l
i=1
y∈Y
P
XY Θ
(x
i
[y
i
, θ)P
Y Θ
(y
i
[θ) .
45
The problem of MaximumLikelihood (ML) estimation from the incomplete data T
X
is
deﬁned as
θ
∗
= argmax
θ∈Θ
L(θ[T
X
) = argmax
θ∈Θ
l
i=1
y∈Y
P
XY Θ
(x
i
[y
i
, θ)P
Y Θ
(y
i
[θ) . (4.4)
The ML estimate of the parameters θ of the GMM from the incomplete data set
T
XY
does not have an analytical solution. The numerical optimization has to be used
instead.
The ExpectationMaximization (EM) algorithm is an iterative procedure for ML
estimation from the incomplete data. The EM algorithm builds a sequence of parameter
estimates
θ
(0)
, θ
(1)
, . . . , θ
(t)
,
such that the loglikelihood L(θ
(t)
[T
X
) monotonically increases, i.e.,
L(θ
(0)
[T
X
) < L(θ
(1)
[T
X
) < . . . < L(θ
(t)
[T
X
)
until a stationary point L(θ
(t−1)
[T
X
) = L(θ
(t)
[T
X
) is achieved. The EM algorithm is a
local optimization technique thus the estimated parameters depend on the initial guess
θ
(0)
. The random guess or Kmeans algorithm (see Section 4.5) are commonly used to
select the initial θ
(0)
. The EM algorithm for the GMM is implemented in the function
emgmm. An interactive demo on the EM algorithm for the GMM is implemented in
demo emgmm.
References: The EM algorithm (including convergence proofs) is described in book [26].
A nice description can be found in book [3]. The monograph [18] is entirely devoted
to the EM algorithm. The ﬁrst papers describing the EM are [25, 5].
Example: MaximumLikelihood estimate from incomplete data
The estimation of the parameters of the GMM is demonstrated on a synthetic
data. The ground truth model is the GMM with two univariate Gaussian components.
The ground truth model is sampled to obtain a training (incomplete) data. The EM
algorithm is applied to estimate the GMM parameters. The random initialization is
used by default. The ground truth model, sampled data and the estimate model are
visualized in Figure 4.3a. The plot of the loglikelihood function L(θ
(t)
[T
X
) with respect
to the number of iterations t is visualized in Figure 4.3b.
% ground truth Gaussian mixture model
true_model.Mean = [2 2];
true_model.Cov = [1 0.5];
true_model.Prior = [0.4 0.6];
sample = gmmsamp(true_model, 250);
46
% ML estimation by EM
options.ncomp = 2; % number of Gaussian components
options.verb = 1; % display progress info
estimated_model = emgmm(sample.X,options);
% visualization
figure;
ppatterns(sample.X);
h1 = pgmm(true_model,struct(’color’,’r’));
h2 = pgmm(estimated_model,struct(’color’,’b’));
legend([h1(1) h2(1)],’Ground truth’, ’ML estimation’);
figure; hold on;
xlabel(’iterations’); ylabel(’loglikelihood’);
plot( estimated_model.logL );
4.3 Minimax estimation
The input is a set T
X
= ¦x
1
, . . . , x
l
¦ contains vectors x
i
∈ R
n
which are realizations of
random variable distributed according to P
XΘ
(x[θ). The vectors T
X
are assumed to be
a realizations with high value of the distribution function P
XΘ
(x[θ). The P
XΘ
(x[θ)
is known up to the parameter θ ∈ Θ. The minimax estimation is deﬁned as
θ
∗
= argmax
θ∈Θ
min
x∈T
X
log P
XΘ
(x[θ) . (4.5)
If a procedure for the MaximumLikelihood estimate (global solution) of the distri
bution P
XΘ
(x[θ) exists then a numerical iterative algorithm which converges to the
optimal estimate θ
∗
can be constructed. The algorithm builds a series of estimates
θ
(0)
, θ
(1)
, . . . , θ
(t)
which converges to the optimal estimate θ
∗
. The algorithm converges
in a ﬁnite number of iterations to such an estimate θ that the condition
min
x∈T
X
log P
XΘ
(x[θ
∗
) − min
x∈T
X
log P
XΘ
(x[θ) < ε , (4.6)
holds for ε > 0 which deﬁnes closeness to the optimal solution θ
∗
. The inequality (4.6)
can be evaluated eﬃciently thus it is a natural choice for the stopping condition of
the algorithm. The STPRtool contains an implementation of the minimax algorithm
for the Gaussian distribution (4.1). The algorithm is implemented in the function
mmgauss. An interactive demo on the minimax estimation for the Gaussian distribution
is implemented in the function demo mmgauss.
References: The minimax algorithm for probability estimation is described and ana
lyzed in Chapter 8 of the book [26].
47
Example: Minimax estimation of the Gaussian distribution
The bivariate Gaussian distribution is described by three samples which are known
to have high value of the distribution function. The minimax algorithm is applied
to estimate the parameters of the Gaussian. In this particular case, the minimax
problem is equivalent to searching for the minimal enclosing ellipsoid. The found
solution is visualized as the isoline of the distribution function at the point P
XΘ
(x[θ)
(see Figure 4.4).
X = [[0;0] [1;0] [0;1]]; % points with high p.d.f.
model = mmgauss(X); % run minimax algorithm
% visualization
figure; ppatterns(X,’xr’,13);
pgauss(model, struct(’p’,exp(model.lower_bound)));
4.4 Probabilistic output of classiﬁer
In general, the discriminant function of an arbitrary classiﬁer does not have meaning of
probability (e.g., SVM, Perceptron). However, the probabilistic output of the classiﬁer
can help in postprocessing, for instance, in combining more classiﬁers together. Fitting
a sigmoid function to the classiﬁer output is a way how to solve this problem.
Let T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ is the training set composed of the vectors x
i
∈
A ⊆ R
n
and the corresponding binary hidden states y
i
∈ ¸ = ¦1, 2¦. The training
set T
XY
is assumed to be identically and independently distributed from underlying
distribution. Let f: A ⊆ R
n
→R be a discriminant function trained from the data T
XY
by an arbitrary learning method. The aim is to estimate parameters of a posteriori
distribution P
Y FΘ
(y[f(x), θ) of the hidden state y given the value of the discriminant
function f(x). The distribution is modeled by a sigmoid function
P
Y FΘ
(1[f(x), θ) =
1
1 + exp(a
1
f(x) + a
2
)
,
which is determined by parameters θ = [a
1
, a
2
]
T
. The loglikelihood function is deﬁned
as
L(θ[T
XY
) =
l
i=1
log P
Y FΘ
(y
i
[f(x
i
), θ) .
The parameters θ = [a
1
, a
2
]
T
can be estimated by the maximumlikelihood method
θ = argmax
θ
L(θ
[T
XY
) = argmax
θ
l
i=1
log P
Y FΘ
(y
i
[f(x
i
), θ
) . (4.7)
48
The STPRtool provides an implementation mlsigmoid which uses the Matlab Opti
mization toolbox function fminunc to solve the task (4.7). The function produces the
parameters θ = [a
1
, a
2
]
T
which are stored in the datatype describing the sigmoid (see
Table 4.6). The evaluation of the sigmoid is implemented in function sigmoid.
Table 4.6: Datatype used to describe sigmoid function.
Sigmoid function (structure array):
.A [2 1] Parameters θ = [a
1
, a
2
]
T
of the sigmoid function.
.fun = ’sigmoid’ Identiﬁes function associated with this data type.
References: The method of ﬁtting the sigmoid function is described in [22].
Example: Probabilistic output for SVM
This example is implemented in the script demo svmpout. The example shows how
to train the probabilistic output for the SVM classiﬁer. The Riply’s data set riply trn
is used for training the SVM and the ML estimation of the sigmoid. The ML estimated
of Gaussian mixture model (GMM) of the SVM output is used for comparison. The
GMM allows to compute the a posteriori probability for comparison to the sigmoid
estimate. The example is implemented in the script demo svmpout. Figure 4.5a shows
found probabilistic models of a posteriori probability of the SVM output. Figure 4.5b
shows the decision boundary of SVM classiﬁer for illustration.
% load training data
data = load(’riply_trn’);
% train SVM
options = struct(’ker’,’rbf’,’arg’,1,’C’,10);
svm_model = smo(data,options);
% compute SVM output
[dummy,svm_output.X] = svmclass(data.X,svm_model);
svm_output.y = data.y;
% fit sigmoid function to svm output
sigmoid_model = mlsigmoid(svm_output);
% ML estimation of GMM model of SVM output
gmm_model = mlcgmm(svm_output);
% visulization
figure; hold on;
49
xlabel(’svm output f(x)’); ylabel(’p(y=1f(x))’);
% sigmoid model
fx = linspace(min(svm_output.X), max(svm_output.X), 200);
sigmoid_apost = sigmoid(fx,sigmoid_model);
hsigmoid = plot(fx,sigmoid_apost,’k’);
ppatterns(svm_output);
% GMM model
pcond = pdfgauss( fx, gmm_model);
gmm_apost = (pcond(1,:)*gmm_model.Prior(1))./...
(pcond(1,:)*gmm_model.Prior(1)+(pcond(2,:)*gmm_model.Prior(2)));
hgmm = plot(fx,gmm_apost,’g’);
hcomp = pgauss(gmm_model);
legend([hsigmoid,hgmm,hcomp],’P(y=1f(x)) MLSigmoid’,...
’P(y=1f(x)) MLGMM’,’P(f(x)y=1) MLGMM’,’P(f(x)y=2) MLGMM’);
% SVM decision boundary
figure; ppatterns(data); psvm(svm_model);
4.5 Kmeans clustering
The input is a set T
X
= ¦x
1
, . . . , x
l
¦ which contains vectors x
i
∈ R
n
. Let θ =
¦µ
1
, . . . , µ
c
¦, µ
i
∈ R
n
be a set of unknown cluster centers. Let the objective function
F(θ) be deﬁned as
F(θ) =
1
l
l
i=1
min
y=1,...,c
µ
y
−x
i

2
.
The small value of F(θ) indicates that the centers θ well describe the input set T
X
.
The task is to ﬁnd such centers θ which describe the set T
X
best, i.e., the task is to
solve
θ
∗
= argmin
θ
F(θ) = argmin
θ
1
l
l
i=1
min
y=1,...,c
µ
y
−x
i

2
.
The Kmeans
1
is an iterative procedure which iteratively builds a series θ
(0)
, θ
(1)
, . . . , θ
(t)
.
The Kmeans algorithm converges in a ﬁnite number of steps to the local optimum of
the objective function F(θ). The found solution depends on the initial guess θ
(0)
which
1
The K in the name Kmeans stands for the number of centers which is here denoted by c.
50
is usually selected randomly. The Kmeans algorithm is implemented in the function
kmeans.
References: The Kmeans algorithm is described in the books [26, 6, 3].
Example: Kmeans clustering
The Kmeans algorithm is used to ﬁnd 4 clusters in the Riply’s training set. The
solution is visualized as the cluster centers. The vectors classiﬁed to the corresponding
centers are distinguished by color (see Figure 4.6a). The plot of the objective function
F(θ
(t)
) with respect to the number of iterations is shown in Figure 4.6b.
data = load(’riply_trn’); % load data
[model,data.y] = kmeans(data.X, 4 ); % run kmeans
% visualization
figure; ppatterns(data);
ppatterns(model.X,’sk’,14);
pboundary( model );
figure; hold on;
xlabel(’t’); ylabel(’F’);
plot(model.MsErr);
51
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
gauss 1
0.981
gauss 2
1.45
(a)
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
(b)
Figure 4.2: Example of the MaximumLikelihood of the Gaussian Mixture Model from
the incomplete data. Figure (a) shows components of the GMM and Figure (b) shows
contour plot of the distribution function.
52
−5 −4 −3 −2 −1 0 1 2 3 4 5
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Ground truth
ML estimation
(a)
0 5 10 15 20 25
−460
−458
−456
−454
−452
−450
−448
−446
−444
−442
iterations
l
o
g
−
l
i
k
e
l
i
h
o
o
d
(b)
Figure 4.3: Example of using the EM algorithm for ML estimation of the Gaussian
mixture model. Figure (a) shows ground truth and the estimated GMM and Figure(b)
shows the loglikelihood with respect to the number of iterations of the EM.
53
−0.4 −0.2 0 0.2 0.4 0.6 0.8 1
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
gauss 1
0.304
Figure 4.4: Example of the Minimax estimation of the Gaussian distribution.
54
−4 −3 −2 −1 0 1 2 3 4 5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
svm output f(x)
p
(
y
=
1

f
(
x
)
)
P(y=1f(x)) ML−Sigmoid
P(y=1f(x)) ML−GMM
P(f(x)y=1) ML−GMM
P(f(x)y=2) ML−GMM
(a)
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
(b)
Figure 4.5: Example of ﬁtting a posteriori probability to the SVM output. Figure (a)
shows the sigmoid model and GMM model both ﬁtted by ML estimated. Figure (b)
shows the corresponding SVM classiﬁer.
55
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
(a)
0 2 4 6 8 10 12
0.04
0.06
0.08
0.1
0.12
0.14
0.16
t
F
(b)
Figure 4.6: Example of using the Kmeans algorithm to cluster the Riply’s data set.
Figure (a) shows found clusters and Figure (b) contains the plot of the objective func
tion F(θ
(t)
) with respect to the number of iterations.
56
Chapter 5
Support vector and other kernel
machines
5.1 Support vector classiﬁers
The binary support vector classiﬁer uses the discriminant function f: A ⊆ R
n
→R of
the following form
f(x) = ¸α k
S
(x)) + b . (5.1)
The k
S
(x) = [k(x, s
1
), . . . , k(x, s
d
)]
T
is the vector of evaluations of kernel functions
centered at the support vectors o = ¦s
1
, . . . , s
d
¦, s
i
∈ R
n
which are usually subset of
the training data. The α ∈ R
l
is a weight vector and b ∈ R is a bias. The binary
classiﬁcation rule q: A → ¸ = ¦1, 2¦ is deﬁned as
q(x) =
_
1 for f(x) ≥ 0 ,
2 for f(x) < 0 .
(5.2)
The datatype used by the STPRtool to represent the binary SVM classiﬁer is described
in Table 5.2.
The multiclass generalization involves a set of discriminant functions f
y
: A ⊆ R
n
→
R, y ∈ ¸ = ¦1, 2, . . . , c¦ deﬁned as
f
y
(x) = ¸α
y
k
S
(x)) + b
y
, y ∈ ¸ .
Let the matrix A = [α
1
, . . . , α
c
] be composed of all weight vectors and b = [b
1
, . . . , b
c
]
T
be a vector of all biases. The multiclass classiﬁcation rule q: A → ¸ = ¦1, 2, . . . , c¦ is
deﬁned as
q(x) = argmax
y∈Y
f
y
(x) . (5.3)
The datatype used by the STPRtool to represent the multiclass SVM classiﬁer is
described in Table 5.3. Both the binary and the multiclass SVM classiﬁers are imple
mented in the function svmclass.
57
The majority voting strategy is other commonly used method to implement the
multiclass SVM classiﬁer. Let q
j
: A ⊆ R
n
→ ¦y
1
j
, y
2
j
¦, j = 1, . . . , g be a set of g binary
SVM rules (5.2). The jth rule q
j
(x) classiﬁes the inputs x into class y
1
j
∈ ¸ or y
2
j
∈ ¸.
Let v(x) be a vector [c 1] of votes for classes ¸ when the input x is to be classiﬁed.
The vector v(x) = [v
1
(x), . . . , v
c
(x)]
T
is computed as:
Set v
y
= 0, ∀y ∈ ¸.
For j = 1, . . . , g do
v
y
(x) = v
y
(x) + 1 where y = q
j
(x).
end.
The majority votes based multiclass classiﬁer assigns the input x into such class y ∈ ¸
having the majority of votes
y = argmax
y
=1,...,g
v
y
(x) . (5.4)
The datatype used to represent the majority voting strategy for the multiclass SVM
classiﬁer is described in Table 5.4. The strategy (5.4) is implemented in the function
mvsvmclass. The implemented methods to deal with the SVM and related kernel
classiﬁers are enlisted in Table 5.1
References: Books describing Support Vector Machines are for example [31, 4, 30]
5.2 Binary Support Vector Machines
The input is a set T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ of training vectors x
i
∈ R
n
and cor
responding hidden states y
i
∈ ¸ = ¦1, 2¦. The linear SVM aims to train the linear
discriminant function f(x) = ¸w x) + b of the binary classiﬁer
q(x) =
_
1 for f(x) ≥ 0 ,
2 for f(x) < 0 .
The training of the optimal parameters (w
∗
, b
∗
) is transformed to the following quadratic
programming task
(w
∗
, b
∗
) = argmin
w,b
1
2
w
2
+ C
l
i=1
ξ
p
i
, (5.5)
¸w x
i
) + b ≥ +1 −ξ
i
, i ∈ 1
1
,
¸w x
i
) + b ≤ −1 + ξ
i
, i ∈ 1
2
,
ξ
i
≥ 0 , i ∈ 1
1
∪ 1
2
.
58
Table 5.1: Implemented methods: Support vector and other kernel machines.
svmclass Support Vector Machines Classiﬁer.
mvsvmclass Majority voting multiclass SVM classiﬁer.
evalsvm Trains and evaluates Support Vector Machines classi
ﬁer.
bsvm2 Multiclass BSVM with L2soft margin.
oaasvm Multiclass SVM using OneAgainstAll decomposi
tion.
oaosvm Multiclass SVM using OneAgainstOne decomposi
tion.
smo Sequential Minimal Optimization for binary SVM
with L1soft margin.
svmlight Interface to SV M
light
software.
svmquadprog SVM trained by Matlab Optimization Toolbox.
diagker Returns diagonal of kernel matrix of given data.
kdist Computes distance between vectors in kernel space.
kernel Evaluates kernel function.
kfd Kernel Fisher Discriminant.
kperceptr Kernel Perceptron.
minball Minimal enclosing ball in kernel feature space.
rsrbf Reduced Set Method for RBF kernel expansion.
rbfpreimg RBF preimage by Schoelkopf’s ﬁxedpoint algorithm.
rbfpreimg2 RBF preimage problem by Gradient optimization.
rbfpreimg3 RBF preimage problem by KwokTsang’s algorithm.
demo svm Demo on Support Vector Machines.
The 1
1
= ¦i: y
i
= 1¦ and 1
2
= ¦i: y
i
= 2¦ are sets of indices. The C > 0 stands for the
regularization constant. The ξ
i
≥ 0, i ∈ 1
1
∪ 1
2
are slack variables used to relax the
inequalities for the case of nonseparable data. The linear p = 1 and quadratic p = 2
term for the slack variables ξ
i
is used. In the case of the parameter p = 1, the L1soft
margin SVM is trained and for p = 2 it is denoted as the L2soft margin SVM.
The training of the nonlinear (kernel) SVM with L1soft margin corresponds to
solving the following QP task
β
∗
= argmax
β
¸β 1) −
1
2
¸β Hβ) , (5.6)
subject to
¸β γ) = 1 ,
β ≥ 0 ,
β ≤ 1C .
59
Table 5.2: Datatype used to describe binary SVM classiﬁer.
Binary SVM classiﬁer (structure array):
.Alpha [d 1] Weight vector α.
.b [1 1] Bias b.
.sv.X [n d] Support vectors o = ¦s
1
, . . . , s
d
¦.
.options.ker [string] Kernel identiﬁer.
.options.arg [1 p] Kernel argument(s).
.fun = ’svmclass’ Identiﬁes function associated with this data type.
Table 5.3: Datatype used to describe multiclass SVM classiﬁer.
Multiclass SVM classiﬁer (structure array):
.Alpha [d c] Weight vectors A = [α
1
, . . . , α
c
].
.b [c 1] Vector of biased b = [b
1
, . . . , b
c
]
T
.
.sv.X [n d] Support vectors o = ¦s
1
, . . . , s
d
¦.
.options.ker [string] Kernel identiﬁer.
.options.arg [1 p] Kernel argument(s).
.fun = ’svmclass’ Identiﬁes function associated with this data type.
The 0 [l 1] is vector of zeros and the 1 [l 1] is vector of ones. The vector γ [l 1]
and the matrix H [l l] are constructed as follows
γ
i
=
_
1 if y
i
= 1 ,
−1 if y
i
= 2 ,
H
i,j
=
_
k(x
i
, x
j
) if y
i
= y
j
,
−k(x
i
, x
j
) if y
i
,= y
j
,
where k: A A → R is selected kernel function. The discriminant function (5.1) is
determined by the weight vector α, bias b and the set of support vectors o. The
set of support vectors is denoted as o = ¦x
i
: i ∈ 1
SV
¦, where the set of indices is
1
SV
= ¦i: β
i
> 0¦. The weight vectors are compute as
α
i
=
_
β
i
if y
i
= 1 ,
−β
i
if y
i
= 2 ,
for i ∈ 1
SV
. (5.7)
The bias b can be computed from the KarushKuhnTucker (KKT) conditions as the
following constrains
f(x
i
) = ¸α k
S
(x
i
)) + b = +1 for i ∈ ¦j: y
j
= 1, 0 < β
j
< C¦ ,
f(x
i
) = ¸α k
S
(x
i
)) + b = −1 for i ∈ ¦j: y
j
= 2, 0 < β
j
< C¦ ,
must hold for the optimal solution. The bias b is computed as an average over all the
constrains such that
b =
1
[1
b
[
i∈I
b
γ
i
−¸α k
S
(x
i
)) .
60
Table 5.4: Datatype used to describe the majority voting strategy for multiclass SVM
classiﬁer.
Majority voting SVM classiﬁer (structure array):
.Alpha [d g] Weight vectors A = [α
1
, . . . , α
g
].
.b [g 1] Vector of biased b = [b
1
, . . . , b
g
]
T
.
.bin y [2 g] Targets for the binary rules. The vector bin y(1,:)
contains [y
1
1
, y
1
2
, . . . , y
1
g
] and the bin y(2,:) contains
[y
2
1
, y
2
2
, . . . , y
2
g
].
.sv.X [n d] Support vectors o = ¦s
1
, . . . , s
d
¦.
.options.ker [string] Kernel identiﬁer.
.options.arg [1 p] Kernel argument(s).
.fun = ’mvsvmclass’ Identiﬁes function associated with this data type.
where 1
b
= ¦i: 0 < β
i
< C¦ are the indices of the vectors on the boundary.
The training of the nonlinear (kernel) SVM with L2soft margin corresponds to
solving the following QP task
β
∗
= argmax
β
¸β 1) −
1
2
¸β (H+
1
2C
E)β) , (5.8)
subject to
¸β γ) = 1 ,
β ≥ 0 ,
where E is the identity matrix. The set of support vectors is set to o = ¦x
i
: i ∈ 1
SV
¦
where the set of indices is 1
SV
= ¦i: β
i
> 0¦. The weight vector α is given by (5.7).
The bias b can be computed from the KKT conditions as the following constrains
f(x
i
) = ¸α k
S
(x
i
)) + b = +1 −
α
i
2C
for i ∈ ¦i: y
j
= 1, β
j
> 0¦ ,
f(x
i
) = ¸α k
S
(x
i
)) + b = −1 +
α
i
2C
for i ∈ ¦i: y
j
= 2, β
j
> 0¦ ,
must hold at the optimal solution. The bias b is computed as an average over all the
constrains thus
b =
1
[1
SV
[
i∈I
SV
γ
i
(1 −
α
i
2C
) −¸α k
S
(x
i
)) .
The binary SVM solvers implemented in the STPRtool are enlisted in Table 5.5.
An interactive demo on the binary SVM is implemented in demo svm.
References: The Sequential Minimal Optimizer (SMO) is described in the books [23,
27, 30]. The SV M
light
is described for instance in the book [27].
Example: Training binary SVM.
61
Table 5.5: Implemented binary SVM solvers.
Function description
smo Implementation of the Sequential Minimal Optimizer (SMO)
used to train the binary SVM with L1soft margin. It can
handle moderate size problems.
svmlight Matlab interface to SV M
light
software (Version: 5.00) which
trains the binary SVM with L1soft margin. The executable
ﬁle svm learn for the Linux OS must be installed. It can be
downloaded from http://svmlight.joachims.org/. The
SV M
light
is very powerful optimizer which can handle large
problems.
svmquadprog SVM solver based on the QP solver quadprog of the Mat
lab optimization toolbox. It trains both L1soft and L2soft
margin binary SVM. The function quadprog is a part of the
Matlab Optimization toolbox. It useful for small problems
only.
The binary SVM is trained on the Riply’s training set riply trn. The SMO solver
is used in the example but other solvers (svmlight, svmquadprog) can by used as well.
The trained SVM classiﬁer is evaluated on the testing data riply tst. The decision
boundary is visualized in Figure 5.1.
trn = load(’riply_trn’); % load training data
options.ker = ’rbf’; % use RBF kernel
options.arg = 1; % kernel argument
options.C = 10; % regularization constant
% train SVM classifier
model = smo(trn,options);
% model = svmlight(trn,options);
% model = svmquadprog(trn,options);
% visualization
figure;
ppatterns(trn); psvm(model);
tst = load(’riply_tst’); % load testing data
ypred = svmclass(tst.X,model); % classify data
cerror(ypred,tst.y) % compute error
62
ans =
0.0920
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Figure 5.1: Example showing the binary SVM classiﬁer trained on the Riply’s data.
5.3 OneAgainstAll decomposition for SVM
The input is a set T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ of training vectors x
i
∈ A ⊆ R
n
and corresponding hidden states y
i
∈ ¸ = ¦1, 2, . . . , c¦. The goal is to train the
multiclass rule q: A ⊆ R
n
→ ¸ deﬁned by (5.3). The problem can be solved by
the OneAgainstAll (OAA) decomposition. The OAA decomposition transforms the
multiclass problem into a series of c binary subtasks that can be trained by the binary
SVM. Let the training set T
y
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ contain the modiﬁed hidden
states deﬁned as
y
i
=
_
1 for y = y
i
,
2 for y ,= y
i
.
The discriminant functions
f
y
(x) = ¸α
y
k
S
(x)) + b
y
, y ∈ ¸ ,
are trained by the binary SVM solver from the set T
y
XY
, y ∈ ¸. The OAA decomposition
is implemented in the function oaasvm. The function allows to specify any binary SVM
63
solver (see Section 5.2) used to solve the binary problems. The result is the multiclass
SVM classiﬁer (5.3) represented by the datatype described in Table 5.3.
References: The OAA decomposition to train the multiclass SVM and comparison
to other methods is described for instance in the paper [12].
Example: Multiclass SVM using the OAA decomposition
The example shows the training of the multiclass SVM using the OAA decomposi
tion. The SMO binary solver is used to train the binary SVM subtasks. The training
data and the decision boundary is visualized in Figure 5.2.
data = load(’pentagon’); % load data
options.solver = ’smo’; % use SMO solver
options.ker = ’rbf’; % use RBF kernel
options.arg = 1; % kernel argument
options.C = 10; % regularization constant
model = oaasvm(data,options ); % training
% visualization
figure;
ppatterns(data);
ppatterns(model.sv.X,’ko’,12);
pboundary(model);
5.4 OneAgainstOne decomposition for SVM
The input is a set T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ of training vectors x
i
∈ A ⊆ R
n
and
corresponding hidden states y
i
∈ ¸ = ¦1, 2, . . . , c¦. The goal is to train the multiclass
rule q: A ⊆ R
n
→ ¸ based on the majority voting strategy deﬁned by (5.4). The
oneagainstone (OAO) decomposition strategy can be used train such a classiﬁer. The
OAO decomposition transforms the multiclass problem into a series of g = c(c −1)/2
binary subtasks that can be trained by the binary SVM. Let the training set T
j
XY
=
¦(x
1
, y
1
), . . . , (x
l
j
, y
l
j
)¦ contain the training vectors x
i
∈ 1
j
= ¦i: y
i
= y
1
_
y
i
= y
2
¦
and the modiﬁed the hidden states deﬁned as
y
i
=
_
1 for y
1
j
= y
i
,
2 for y
2
j
= y
i
,
i ∈ 1
j
.
The training set T
j
XY
, j = 1, . . . , g is constructed for all g = c(c −1)/2 combinations of
classes y
1
j
∈ ¸ and y
2
j
∈ ¸ ¸¦y
1
j
¦. The binary SVM rules q
j
, j = 1, . . . , g are trained on
the data T
j
XY
. The OAO decomposition for multiclass SVM is implemented in function
64
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Figure 5.2: Example showing the multiclass SVM classiﬁer build by oneagainstall
decomposition.
oaosvm. The function oaosvm allows to specify an arbitrary binary SVM solver to train
the subtasks.
References: The OAO decomposition to train the multiclass SVM and comparison
to other methods is described for instance in the paper [12].
Example: Multiclass SVM using the OAO decomposition
The example shows the training of the multiclass SVM using the OAO decomposi
tion. The SMO binary solver is used to train the binary SVM subtasks. The training
data and the decision boundary is visualized in Figure 5.3.
data = load(’pentagon’); % load data
options.solver = ’smo’; % use SMO solver
options.ker = ’rbf’; % use RBF kernel
options.arg = 1; % kernel argument
options.C = 10; % regularization constant
model = oaosvm(data,options); % training
% visualization
figure;
ppatterns(data);
ppatterns(model.sv.X,’ko’,12);
pboundary(model);
65
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Figure 5.3: Example showing the multiclass SVM classiﬁer build by oneagainstone
decomposition.
5.5 Multiclass BSVM formulation
The input is a set T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ of training vectors x
i
∈ A ⊆ R
n
and
corresponding hidden states y
i
∈ ¸ = ¦1, 2, . . . , c¦. The goal is to train the multiclass
rule q: A ⊆ R
n
→ ¸ deﬁned by (5.3). In the linear case, the parameters of the multi
class rule can be found by solving the multiclass BSVM formulation which is deﬁned
as
(W
∗
, b
∗
, ξ
∗
) = argmin
W,b
1
2
y∈Y
(w
y

2
+ b
2
y
) + C
i∈I
y∈Y\{y}
(ξ
i
)
p
. ¸¸ .
F(W,b,ξ)
, (5.9)
subject to
¸w
y
i
x
i
) + b
y
i
−(¸w
y
x
i
) + b
y
) ≥ 1 −ξ
y
i
, i ∈ 1 , y ∈ ¸ ¸ ¦y
i
¦ ,
ξ
y
i
≥ 0 , i ∈ 1 , y ∈ ¸ ¸ ¦y
i
¦ ,
where W = [w
1
, . . . , w
c
] is a matrix of normal vectors, b = [b
1
, . . . , b
c
]
T
is vector of
biases, ξ is a vector of slack variables and 1 = [1, . . . , l] is set of indices. The L1soft
margin is used for p = 1 and L2soft margin for p = 2. The letter B in BSVM stands
for the bias term added to the objective (5.9).
In the case of the L2soft margin p = 2, the optimization problem (5.9) can be
66
expressed in its dual form
A
∗
= argmax
A
i∈I
y∈Y\{y
i
}
α
y
i
−
1
2
i∈I
y∈Y\{y
i
}
j∈I
u∈Y\{y
j
}
α
y
i
α
u
j
h(y, u, i, j) , (5.10)
subject to
α
y
i
≥ 0 , i ∈ 1 , y ∈ ¸ ¸ ¦y
i
¦ ,
where A = [α
1
, . . . , α
c
] are optimized weight vectors. The function h: ¸¸11 →R
is deﬁned as
h(y, u, i, j) = (¸x
i
x
j
)+1)(δ(y
i
, y
j
)+δ(y, u)−δ(y
i
, u)−δ(y
j
, y))+
δ(y, u)δ(i, j)
2C
. (5.11)
The criterion (5.10) corresponds to the singleclass SVM criterion which is simple to
optimize. The multiclass rule (5.3) is composed of the set of discriminant functions
f
y
: A →R which are computed as
f
y
(x) =
i∈I
¸x
i
x)
u∈Y\{y
i
}
α
u
i
(δ(u, y
i
) −δ(u, y)) + b
y
, y ∈ ¸ , (5.12)
where the bias b
y
, y ∈ ¸ is given by
b
y
=
i∈I
y∈Y\{y
i
}
α
u
i
(δ(y, y
i
) −δ(y, u)) , y ∈ ¸ .
The nonlinear (kernel) classiﬁer is obtained by substituting the selected kernel k: A
A →R for the dot products ¸xx
) to (5.11) and (5.12). The training of the multiclass
BSVM classiﬁer with L2soft margin is implemented in the function bsvm2. The opti
mization task (5.10) can be solved by most the optimizers. The optimizers implemented
in the function bsvm2 are enlisted bellow:
MitchellDemyanovMalozemov (solver = ’mdm’).
Kozinec’s algorithm (solver = ’kozinec’).
Nearest Point Algorithm (solver = ’npa’).
The above mentioned algorithms are of iterative nature and converge to the optimal
solution of the criterion (5.9). The upper bound F
UB
and the lower bound F
LB
on the
optimal value F(W
∗
, b
∗
, ξ
∗
) of the criterion (5.9) can be computed. The bounds are
used in the algorithms to deﬁne the following two stopping conditions:
Relative tolerance:
F
UB
−F
LB
F
LB
+ 1
≤ ε
rel
.
67
Absolute tolerance: F
UB
≤ ε
abs
.
References: The multiclass BSVM formulation is described in paper [12]. The trans
formation of the multiclass BSVM criterion to the singleclass criterion (5.10) imple
mented in the STPRtool is published in paper [9]. The MitchellDemyanovMalozemov
and Nearest Point Algorithm are described in paper [14]. The Kozinec’s algorithm
based SVM solver is described in paper [10].
Example: Multiclass BSVM with L2soft margin.
The example shows the training of the multiclass BSVM using the NPA solver. The
other solvers can be used instead by setting the variable options.solver (other options
are ’kozinec’, ’mdm’). The training data and the decision boundary is visualized in
Figure 5.4.
data = load(’pentagon’); % load data
options.solver = ’npa’; % use NPA solver
options.ker = ’rbf’; % use RBF kernel
options.arg = 1; % kernel argument
options.C = 10; % regularization constant
model = bsvm2(data,options); % training
% visualization
figure;
ppatterns(data);
ppatterns(model.sv.X,’ko’,12);
pboundary(model);
5.6 Kernel Fisher Discriminant
The input is a set T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ of training vectors x
i
∈ A ⊆ R
n
and
corresponding values of binary state y
i
∈ ¸ = ¦1, 2¦. The Kernel Fisher Discriminant
is the nonlinear extension of the linear FLD (see Section 2.3). The input training
vectors are assumed to be mapped into a new feature space T by a mapping function
φ: A → T. The ordinary FLD is applied on the mapped data training data T
ΦY
=
¦(φ(x
1
), y
1
), . . . , (φ(x
l
), y
l
)¦. The mapping is performed eﬃciently by means of kernel
functions k: A A → R being the dot products of the mapped vectors k(x, x
) =
¸φ(x) φ(x
)). The aim is to ﬁnd a direction ψ =
l
i=1
α
i
φ(x
i
) in the feature space
T given by weights α = [α
1
, . . . , α
l
]
T
such that the criterion
F(α) =
¸α Mα)
¸α (N+ µE)α)
=
¸α m)
2
¸α (N+ µE)α)
, (5.13)
68
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Figure 5.4: Example showing the multiclass BSVM classiﬁer with L2soft margin.
is maximized. The criterion (5.13) is the feature space counterpart of the crite
rion (2.13) deﬁned in the input space A. The vector m [l 1], matrices N [l l]
M [l l] are constructed as
m
y
=
1
[1
y
[
K1
y
, y ∈ ¸ , m = m
1
−m
2
,
N = KK
T
−
y∈Y
[1
y
[m
y
m
T
y
, M = (m
1
−m
2
)(m
1
−m
2
)
T
,
where the vector 1
y
has entries i ∈ 1
y
equal to and remaining entries zeros. The kernel
matrix K [l l] contains kernel evaluations K
i,j
= k(x
i
, x
j
). The diagonal matrix
µE in the denominator of the criterion (5.13) serves as the regularization term. The
discriminant function f(x) of the binary classiﬁer (5.2) is given by the found vector α
as
f(x) = ¸ψ φ(x)) + b = ¸α k(x)) + b , (5.14)
where the k(x) = [k(x
1
, x), . . . , k(x
l
, x)]
T
is vector of evaluations of the kernel. The
discriminant function (5.14) is known up to the bias b which is determined by the linear
SVM with L1soft margin is applied on the projected data T
ZY
= ¦(z
1
, y
1
), . . . , (z
l
, y
l
)¦.
The projections ¦z
1
, . . . , z
l
¦ are computed as
z
i
= ¸α k(x
i
)) .
The KFD training algorithm is implemented in the function kfd.
69
References: The KFD formulation is described in paper [19]. The detailed analysis
can be found in book [30].
Example: Kernel Fisher Discriminant
The binary classiﬁer is trained by the KFD on the Riply’s training set riply trn.mat.
The found classiﬁer is evaluated on the testing data riply tst.mat. The training data
and the found kernel classiﬁer is visualized in Figure 5.5. The classiﬁer is visualized as
the SVM classiﬁer, i.e., the isolines with f(x) = +1 and f(x) = −1 are visualized and
the support vectors (whole training set) are marked as well.
trn = load(’riply_trn’); % load training data
options.ker = ’rbf’; % use RBF kernel
options.arg = 1; % kernel argument
options.C = 10; % regularization of linear SVM
options.mu = 0.001; % regularization of KFD criterion
model = kfd(trn, options) % KFD training
% visualization
figure;
ppatterns(trn);
psvm(model);
% evaluation on testing data
tst = load(’riply_tst’);
ypred = svmclass( tst.X, model );
cerror( ypred, tst.y )
ans =
0.0960
5.7 Preimage problem
Let T
X
= ¦x
1
, . . . , x
l
¦ be a set of vector from the input space A ⊆ R
n
. The kernel
methods works with the data nonlinearly mapped φ: A → T into a new feature space
T. The mapping is performed implicitly by using the kernel functions k: A A → R
which are dot products of the input data mapped into the feature space T such that
k(x
a
, x
b
) = ¸φ(x
a
) φ(x
b
)). Let ψ ∈ T be given by the following kernel expansion
ψ =
l
i=1
α
i
φ(x
i
) , (5.15)
70
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Figure 5.5: Example shows the binary classiﬁer trained based on the Kernel Fisher
Discriminant.
where α = [α
1
, . . . , α
l
]
T
is a weight vector of real coeﬃcients. The preimage problem
aims to ﬁnd an input vector x ∈ A, its feature space image φ(x) ∈ T would well
approximate the expansion point Ψ ∈ T given by (5.15). In other words, the pre
image problem solves mapping from the feature space T back to the input space A.
The preimage problem can be stated as the following optimization task
x = argmin
x
φ(x) −ψ
2
= argmin
x
φ(x
) −
l
i=1
α
i
φ(x
i
)
2
= argmin
x
k(x
, x
) −2
l
i=1
α
i
k(x
, x
i
) +
l
i=1
l
j=1
α
i
α
j
k(x
i
, x
j
) .
(5.16)
Let the Radial Basis Function (RBF) k(x
a
, x
b
) = exp(−0.5x
a
−x
b

2
/σ
2
) be assumed.
The preimage problem (5.16) for the particular case of the RBF kernel leads to the
following task
x = argmax
x
l
i=1
α
i
exp(−0.5x
−x
i

2
/σ
2
) . (5.17)
The STPRtool provides three optimization algorithms to solve the RBF preimage
problem (5.17) which are enlisted in Table 5.6. However, all the methods are guaranteed
to ﬁnd only a local optimum of the task (5.17).
71
Table 5.6: Methods to solve the RBF preimage problem implemented in the STPRtool.
rbfpreimg An iterative ﬁxed point algorithm proposed in [28].
rbfpreimg2 A standard gradient optimization method using the
Matlab optimization toolbox for 1D search along the
gradient.
rbfpreimg3 A method exploiting the correspondence between dis
tances in the input space and the feature space pro
posed in [16].
References: The preimage problem and the implemented algorithms for its solution
are described in [28, 16].
Example: Preimage problem
The example shows how to visualize a preimage of the data mean in the feature
induced by the RBF kernel. The result is visualized in Figure 5.6.
% load input data
data = load(’riply_trn’);
% define the mean in the feature space
num_data = size(data,2);
expansion.Alpha = ones(num_data,1)/num_data;
expansion.sv.X = data.X;
expansion.options.ker = ’rbf’;
expansion.options.arg = 1;
% compute preimage problem
x = rbfpreimg(expansion);
% visualization
figure;
ppatterns(data.X);
ppatterns(x,’or’,13);
72
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Figure 5.6: Example shows training data and the preimage of their mean in the feature
space induced by the RBF kernel.
5.8 Reduced set method
The reduced set method is useful to simplify the kernel classiﬁer trained by the SVM
or other kernel machines. Let the function f: A ⊆ R
n
→R be given as
f(x) =
l
i=1
α
i
k(x
i
, x) + b = ¸ψ φ(x)) + b , (5.18)
where T
X
= ¦x
1
, . . . , x
l
¦ are vectors from R
n
, k: A A → R is a kernel function,
α = [α
1
, . . . , α
l
]
T
is a weight vector of real coeﬃcients and b ∈ R is a scalar. Let
φ: A → T be a mapping from the input space A to the feature space T associated
with kernel function such that k(x
a
, x
b
) = ¸φ(x
a
) φ(x
b
)). The function f(x) can be
expressed in the feature space T as the linear function f(x) = ¸ψ φ(x)) + b. The
reduced set method aims to ﬁnd a new function
˜
f(x) =
m
i=1
β
i
k(s
i
, x) + b = ¸φ(x)
m
i=1
β
i
φ(s
i
)
. ¸¸ .
˜
ψ
) + b , (5.19)
such that the expansion (5.19) is shorter, i.e., m < l, and well approximates the original
one (5.18). The weight vector β = [β
1
, . . . , β
m
]
T
and the vectors T
S
= ¦s
1
, . . . , s
m
¦,
s
i
∈ R
n
determine the reduced kernel expansion (5.19). The problem of ﬁnding the
73
reduced kernel expansion (5.19) can be stated as the optimization task
(β, T
S
) = argmin
β
,T
S

˜
ψ −ψ
2
= argmin
β
,T
S
_
_
_
_
_
m
i=1
β
i
φ(s
i
) −
l
i=1
α
i
φ(x
i
)
_
_
_
_
_
2
. (5.20)
Let the Radial Basis Function (RBF) k(x
a
, x
b
) = exp(−0.5x
a
−x
b

2
/σ
2
) be assumed.
In this case, the optimization task (5.20) can be solved by an iterative greedy algo
rithm. The algorithm converts the task (5.20) into a series of m preimage tasks (see
Section 5.7). The reduced set method for the RBF kernel is implemented in function
rsrbf.
References: The implemented reduced set method is described in [28].
Example: Reduced set method for SVM classiﬁer
The example shows how to use the reduced set method to simplify the SVM clas
siﬁer. The SVM classiﬁer is trained from the Riply’s training set riply trn/mat. The
trained kernel expansion (discriminant function) is composed of 94 support vectors.
The reduced expansion with 10 support vectors is found. Decision boundaries of the
original and the reduced SVM are visualized in Figure 5.7. The original and reduced
SVM are evaluated on the testing data riply tst.mat.
% load training data
trn = load(’riply_trn’);
% train SVM classifier
model = smo(trn,struct(’ker’,’rbf’,’arg’,1,’C’,10));
% compute the reduced rule with 10 support vectors
red_model = rsrbf(model,struct(’nsv’,10));
% visualize decision boundaries
figure; ppatterns(trn);
h1 = pboundary(model,struct(’line_style’,’r’));
h2 = pboundary(red_model,struct(’line_style’,’b’));
legend([h1(1) h2(1)],’Original SVM’,’Reduced SVM’);
% evaluate SVM classifiers
tst = load(’riply_tst’);
ypred = svmclass(tst.X,model);
err = cerror(ypred,tst.y);
red_ypred = svmclass(tst.X,model);
74
red_err = cerror(red_ypred,tst.y);
fprintf([’Original SVM: nsv = %d, err = %.4f\n’ ...
’Reduced SVM: nsv = %d, err = %.4f\n’], ...
model.nsv, err, red_model.nsv, red_err);
Original SVM: nsv = 94, err = 0.0960
Reduced SVM: nsv = 10, err = 0.0960
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Original SVM
Reduced SVM
Figure 5.7: Example shows the decision boundaries of the SVM classiﬁer trained by
SMO and the approximated SVM classiﬁer found by the reduced set method. The
original decision rule involves 94 support vectors while the reduced one only 10 support
vectors.
75
Chapter 6
Miscellaneous
6.1 Data sets
The STPRtool provides several datasets and functions to generate synthetic data. The
datasets stored in the Matlab data ﬁle are enlisted in Table 6.1. The functions which
generate synthetic data and scripts for converting external datasets to the Matlab are
given in Table 6.2.
Table 6.1: Datasets stored in /data/ directory of the STPRtool.
/riply trn.mat Training part of the Riply’s data set [24].
/riply tst.mat Testing part of the Riply’s data set [24].
/iris.mat Well known Fisher’s data set.
/anderson task/*.mat Input data for the demo on the Generalized Ander
son’s task demo anderson.
/binary separable/*.mat Input data for the demo on training linear classiﬁers
from ﬁnite vector sets demo linclass.
/gmm samles/*.mat Input data for the demo on the Expectation
Maximization algorithm for the Gaussian mixture
model demo emgmm.
/multi separable/*.mat Linearly separable multiclass data in 2D.
/ocr data/*.mat Examples of handwritten numerals. A description is
given in Section 7.1.
/svm samples/*.mat Input data for the demo on the Support Vector Ma
chines demo svm.
76
Table 6.2: Data generation.
createdata GUI which allows to create interactively labeled vec
tors sets and labeled sets of Gaussian distributions. It
operates in 2dimensional space only.
genlsdata Generates linearly separable binary data with pre
scribed margin.
gmmsamp Generates data from the Gaussian mixture model.
gsamp Generates data from the multivariate Gaussian distri
bution.
gencircledata Generates data on circle corrupted by the Gaussian
noise.
usps2mat Converts U.S. Postal Service (USPS) database of
handwritten numerals to the Matlab ﬁle.
6.2 Visualization
The STPRtool provides tools for visualization of common models and data. The list
of functions for visualization is given in Table 6.3. Each function contains an example
of using.
Table 6.3: Functions for visualization.
pandr Visualizes solution of the Generalized Anderson’s
task.
pboundary Plots decision boundary of given classiﬁer in 2D.
pgauss Visualizes set of bivariate Gaussians.
pgmm Visualizes the bivariate Gaussian mixture model.
pkernelproj Plots isolines of kernel projection.
plane3 Plots plane in 3D.
pline Plots line in 2D.
ppatterns Visualizes patterns as points in the feature space. It
works for 1D, 2D and 3D.
psvm Plots decision boundary of binary SVM classiﬁer in
2D.
showim Displays images entered as column vectors.
77
6.3 Bayesian classiﬁer
The object under study is assumed to be described by a vector of observations x ∈ A
and a hidden state y ∈ ¸. The x and y are realizations of random variables with joint
probability distribution P
XY
(x, y). A decision rule q: A → T takes a decision d ∈ T
based on the observation x ∈ A. Let W: T¸ →R be a loss function which penalizes
the decision q(x) ∈ T when the true hidden state is y ∈ ¸. Let A ⊆ R
n
and the sets
¸ and T be ﬁnite. The Bayesian risk R(q) is an expectation of the value of the loss
function W when the decision rule q is applied, i.e.,
R(q) =
_
X
y∈Y
P
XY
(x, y)W(q(x), y) dx . (6.1)
The optimal rule q
∗
which minimizes the Bayesian risk (6.1) is referred to as the
Bayesian rule
q
∗
(x) = argmin
y
y∈Y
P
XY
(x, y)W(q(x), y) , ∀x ∈ A .
The STPRtool implements the Bayesian rule for two particular cases:
Minimization of misclassiﬁcation: The set of decisions T coincides to the set of
hidden states ¸ = ¦1, . . . , c¦. The 0/1loss function
W
0/1
(q(x), y) =
_
0 for q(x) = y ,
1 for q(x) ,= y .
(6.2)
is used. The Bayesian risk (6.1) with the 0/1loss function corresponds to the
expectation of misclassiﬁcation. The rule q: A → ¸ which minimizes the expec
tation of misclassiﬁcation is deﬁned as
q(x) = argmax
y∈Y
P
Y X
(y[x) ,
= argmax
y∈Y
P
XY
(x[y)P
Y
(y) .
(6.3)
Classiﬁcation with rejectoption: The set of decisions T is assumed to be T =
¸ ∪ ¦dont know¦. The loss function is deﬁned as
W
ε
(q(x), y) =
⎧
⎨
⎩
0 for q(x) = y ,
1 for q(x) ,= y ,
ε for q(x) = dont know ,
(6.4)
where ε is penalty for the decision dont know. The rule q: A → ¸ which mini
mizes the Bayesian risk with the loss function (6.4) is deﬁned as
q(x) =
⎧
⎨
⎩
argmax
y∈Y
P
XY
(x[y)P
Y
(y) if 1 −max
y∈Y
P
Y X
(y[x) < ε ,
dont know if 1 −max
y∈Y
P
Y X
(y[x) ≥ ε .
(6.5)
78
To apply the optimal classiﬁcation rules one has to know the classconditional
distributions P
XY
and a priory distribution P
Y
(or their estimates). The datatype
used by the STPRtool is described in Table 6.4. The Bayesian classiﬁers (6.3) and (6.5)
are implemented in function bayescls.
Table 6.4: Datatype used to describe the Bayesian classiﬁer.
Bayesian classiﬁer (structure array):
.Pclass [cell 1 c] Class conditional distributions P
XY
. The distribution
P
XY
(x[y) is given by Pclassy which uses the data
type described in Table 4.2.
.Prior [1 c] Discrete distribution P
Y
(y), y ∈ ¸.
.fun = ’bayescls’ Identiﬁes function associated with this data type.
References: The Bayesian decision making with applications to PR is treated in
details in book [26].
Example: Bayesian classiﬁer
The example shows how to design the plugin Bayesian classiﬁer for the Riply’s
training set riply trn.mat. The classconditional distributions P
XY
are modeled by
the Gaussian mixture models (GMM) with two components. The GMM are estimated
by the EM algorithm emgmm. The a priori probabilities P
Y
(y), y ∈ ¦1, 2¦ are estimated
by the relative occurrences (it corresponds to the ML estimate). The designed Bayesian
classiﬁer is evaluated on the testing data riply tst.mat.
% load input training data
trn = load(’riply_trn’);
inx1 = find(trn.y==1);
inx2 = find(trn.y==2);
% Estimation of classconditional distributions by EM
bayes_model.Pclass{1} = emgmm(trn.X(:,inx1),struct(’ncomp’,2));
bayes_model.Pclass{2} = emgmm(trn.X(:,inx2),struct(’ncomp’,2));
% Estimation of priors
n1 = length(inx1); n2 = length(inx2);
bayes_model.Prior = [n1 n2]/(n1+n2);
% Evaluation on testing data
tst = load(’riply_tst’);
ypred = bayescls(tst.X,bayes_model);
cerror(ypred,tst.y)
79
ans =
0.0900
The following example shows how to visualize the decision boundary of the found
Bayesian classiﬁer minimizing the misclassiﬁcation (solid black line). The Bayesian
classiﬁer splits the feature space into two regions corresponding to the ﬁrst class and
the second class. The decision boundary of the rejectoptions rule (dashed line) is
displayed for comparison. The rejectoption rule splits the feature space into three
regions corresponding to the ﬁrst class, the second class and the region in between
corresponding to the dont know decision. The visualization is given in Figure 6.1.
% Visualization
figure; hold on; ppatterns(trn);
bayes_model.fun = ’bayescls’;
pboundary(bayes_model);
% Penalization for don’t know decision
reject_model = bayes_model;
reject_model.eps = 0.1;
% Vislualization of rejetoption rule
pboundary(reject_model,struct(’line_style’,’k’));
6.4 Quadratic classiﬁer
The quadratic classiﬁcation rule q: A ⊆ R
n
→ ¸ = ¦1, 2, . . . , c¦ is composed of a set
of discriminant functions
f
y
(x) = ¸x A
y
x) +¸b
y
x) + c
y
, ∀y ∈ ¸ ,
which are quadratic with respect to the input vector x ∈ R
n
. The quadratic discrimi
nant function f
y
is determined by a matrix A
y
[n n], a vector b
y
[n 1] and a scalar
c
y
[1 1]. The input vector x ∈ R
n
is assigned to the class y ∈ ¸ its corresponding
discriminant function f
y
attains maximal value
y = argmax
y
∈Y
f
y
(x) = argmax
y
∈Y
(¸x A
y
x) +¸b
y
x) + c
y
) . (6.6)
In the particular binary case ¸ = ¦1, 2¦, the quadratic classiﬁer is represented by
a single discriminant function
f(x) = ¸x Ax) +¸b x) + c ,
80
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Figure 6.1: Figure shows the decision boundary of the Bayesian classiﬁer (solid line)
and the decision boundary of the rejectoption rule with ε = 0.1 (dashed line). The
classconditional probabilities are modeled by the Gaussian mixture models with two
components estimated by the EM algorithm.
given by a matrix A [n n], a vector b [n 1] and a scalar c [1 1]. The input vector
x ∈ R
n
is assigned to class y ∈ ¦1, 2¦ as follows
q(x) =
_
1 if f(x) = ¸x Ax) +¸b x) + c ≥ 0 ,
2 if f(x) = ¸x Ax) +¸b x) + c < 0 .
(6.7)
The STPRtool uses a speciﬁc structure array to describe the binary (see Table 6.5)
and the multiclass (see Table 6.6) quadratic classiﬁer. The implementation of the
quadratic classiﬁer itself provides function quadclass.
Table 6.5: Datatype used to describe binary quadratic classiﬁer.
Binary linear quadratic (structure array):
.A [n n] Parameter matrix A.
.b [n 1] Parameter vector b.
.c [1 1] Parameter scalar c.
.fun = ’quadclass’ Identiﬁes function associated with this data type.
Example: Quadratic classiﬁer
The example shows how to train the quadratic classiﬁer using the Fisher Linear Dis
criminant and the quadratic data mapping. The classiﬁer is trained from the Riply’s
81
Table 6.6: Datatype used to describe multiclass quadratic classiﬁer.
Multiclass quadratic classiﬁer (structure array):
.A [n n c] Parameter matrices A
y
, y ∈ ¸.
.b [n c] Parameter vectors b
y
, y ∈ ¸.
.c [1 c] Parameter scalars c
y
, y ∈ ¸.
.fun = ’quadclass’ Identiﬁes function associated with this data type.
training set riply trn.mat. The found quadratic classiﬁer is evaluated on the test
ing data riply trn.mat. The boundary of the quadratic classiﬁer is visualized (see
Figure 6.2).
trn = load(’riply_trn’); % load input data
quad_data = qmap(trn); % quadratic mapping
lin_model = fld(quad_data); % train FLD
% make quadratic classifier
quad_model = lin2quad(lin_model);
% visualization
figure;
ppatterns(trn); pboundary(quad_model);
% evaluation
tst = load(’riply_tst’);
ypred = quadclass(tst.X,quad_model);
cerror(ypred,tst.y)
ans =
0.0940
6.5 Knearest neighbors classiﬁer
Let T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ be a set of prototype vectors x
i
∈ A ⊆ R
n
and
corresponding hidden states y
i
∈ ¸ = ¦1, . . . , c¦. Let R
n
(x) = ¦x
: x −x
 ≤ r
2
¦ be
a ball centered in the vector x in which lie K prototype vectors x
i
, i ∈ ¦1, . . . , l¦, i.e.,
[¦x
i
: x
i
∈ R
n
(x)¦[ = K. The Knearest neighbor (KNN) classiﬁcation rule q: A → ¸
is deﬁned as
q(x) = argmax
y∈Y
v(x, y) , (6.8)
82
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Figure 6.2: Figure shows the decision boundary of the quadratic classiﬁer.
where v(x, y) is number of prototype vectors x
i
with hidden state y
i
= y which lie in
the ball x
i
∈ R
n
(x). The classiﬁcation rule (6.8) is computationally demanding if the
set of prototype vectors is large. The datatype used by the STPRtool to represent the
Knearest neighbor classiﬁer is described in Table 6.7.
Table 6.7: Datatype used to describe the Knearest neighbor classiﬁer.
KNN classiﬁer (structure array):
.X [n l] Prototype vectors ¦x
1
, . . . , x
l
¦.
.y [1 l] Labels ¦y
1
, . . . , y
l
¦.
.fun = ’knnclass’ Identiﬁes function associated with this data type.
References: The Knearest neighbors classiﬁer is described in most books on PR, for
instance, in book [6].
Example: Knearest neighbor classiﬁer
The example shows how to create the (K = 8)NN classiﬁcation rule from the
Riply’s training data riply trn.mat. The classiﬁer is evaluated on the testing data
riply tst.mat. The decision boundary is visualized in Figure 6.3.
% load training data and setup 8NN rule
trn = load(’riply_trn’);
model = knnrule(trn,8);
83
% visualize decision boundary and training data
figure; ppatterns(trn); pboundary(model);
% evaluate classifier
tst = load(’riply_tst’);
ypred = knnclass(tst.X,model);
cerror(ypred,tst.y)
ans =
0.1160
−1.5 −1 −0.5 0 0.5 1
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Figure 6.3: Figure shows the decision boundary of the (K = 8)nearest neighbor
classiﬁer.
84
Chapter 7
Examples of applications
We intend to demonstrate how methods implemented in the STPRtool can be used to
solve a more complex applications. The applications selected are the optical charac
ter recognition (OCR) system based on the multiclass SVM (Section 7.1) and image
denoising system using the kernel PCA (Section 7.2).
7.1 Optical Character Recognition system
The use of the STPRtool is demonstrated on a simple Optical Character Recognition
(OCR) system for handwritten numerals from 0 to 9. The STPRtool provides GUI
which allows to use standard computer mouse as an input device to draw images of
numerals. The GUI is intended just for demonstration purposes and the scanner or
other device would be used in practice instead. The multiclass SVM are used as the
base classiﬁer of the numerals. A numeral to be classiﬁed is represented as a vector
x containing pixel values taken from the numeral image. The training stage of the
OCR system involves (i) collection of training examples of numerals and (ii) training
the multiclass SVM. To see the trained OCR system run the demo script demo ocr.
The GUI for drawing numerals is implemented in a function mpaper. The function
mpaper creates a form composed of 50 cells to which the numerals can be drawn (see
Figure 7.1a). The drawn numerals are normalized to size 16 16 pixels (diﬀerent
size can be speciﬁed). The function mpaper is designed to invoke a speciﬁed function
whenever the middle mouse button is pressed. In particular, the functions save chars
and ocr fun are called from the mpaper to process drawn numerals. The function
save chars saves drawn numerals in the stage of collecting training examples and the
function ocr fun is used to recognize the numerals if the trained OCR is applied.
The set of training examples has to be collected before the OCR is trained. The
function collect chars invokes the form for drawing numerals and allows to save the
numerals to a speciﬁed ﬁle. The STPRtool contains examples of numerals collected
85
from 6 diﬀerent persons. Each person drew 50 examples of each of numerals from
’0’ to ’9’. The numerals are stored in the directory /data/ocr numerals/. The ﬁle
name format name xx.mat (the character can be omitted from the name) is used.
The string name is name of the person who drew the particular numeral and xx stands
for a label assigned to the numeral. The labels from y = 1 to y = 9 corresponds to the
numerals from ’1’ to ’9’ and the label y = 10 is used for numeral ’0’. For example,
to collect examples of numeral ’1’ from person honza cech
collect_chars(’honza_cech1’)
was called. Figure 7.1a shows the form ﬁlled with the examples of the numeral ’1’.
The Figure 7.1b shows numerals after normalization which are saved to the speci
ﬁed ﬁle honza cech.mat. The images of normalized numerals are stored as vectors
to a matrix X of size [256 50]. The vector dimension n = 256 corresponds to
16 16 pixels of each image numeral and l = 50 is the maximal number of numerals
drawn to a single form. If the ﬁle name format name xx.mat is met then the function
mergesets(Directory,DataFile) can be used to merge all the ﬁle to a single one.
The resulting ﬁle contains a matrix X of all examples and a vector y containing labels
are assigned to examples according the number xx. The string Directory speciﬁes the
directory where the ﬁles name xx.mat are stored and the string FileName is a ﬁle name
to which the resulting training set is stored.
Having the labeled training set T
XY
= ¦(x
1
, y
1
), . . . , (x
l
, y
l
)¦ created the multiclass
SVM classiﬁer can be trained. The methods to train the multiclass SVM are described
in Chapter 5. In particular, the OneAgainstOne (OAO) decomposition was used as
it requires the shortest training time. However, if a speed of classiﬁcation is the major
objective the multiclass BSVM formulation is often a better option as it produce less
number of support vectors. The SVM has two free parameters which have to be tuned:
the regularization constant C and the kernel function k(x
a
, x
b
) with its argument.
The RBF kernel k(x
a
, x
b
) = exp(−0.5x
a
− x
b

2
/σ
2
) determined by the parameter
σ was used in the experiment. Therefore, the free SVM parameters are C and σ. A
common method to tune the free parameters is to use the crossvalidation to select the
best parameters from a preselected set Θ = ¦(C
1
, σ
1
), . . . , (C
k
, σ
k
)¦. The STPRtool
contains a function evalsvm which evaluates the set of SVM parameters Θ based on the
crossvalidation estimation of the classiﬁcation error. The function evalsvm returns the
SVM model its parameters (C
∗
, σ
∗
) turned out to be have the lowest crossvalidation
error over the given set Θ. The parameter tuning is implemented in the script tune ocr.
The parameters C = ∞ (data are separable) and σ = 5 were found to be the best.
After the parameters (C
∗
, σ
∗
) are tuned the SVM classiﬁer is trained using all
training data available. The training procedure for OCR classiﬁer is implemented in
the function train ocr. The directory /demo/ocr/models/ contains the SVM model
trained by OAO decomposition strategy (function oaosvm) and model trained based
on the multiclass BSVM formulation (function bsvm2).
86
The particular model used by OCR is speciﬁed in the function ocr fun which
implicitly uses the ﬁle ocrmodel.mat stored in the directory /demo/ocr/. To run the
resulting OCR type demo ocr or call
mpaper(struct(’fun’,’ocr_fun’))
If one intends to modify the classiﬁer or the way how the classiﬁcation results are
displayed then modify the function ocr fun. A snapshot of running OCR is shown in
Figure 7.2.
7.2 Image denoising
The aim is to train a model of a given class of images in order to restore images
corrupted by noise. The U.S. Postal Service (USPS) database of images of hand
written numerals [17] was used as the class of images to model. The additive Gaussian
noise was assumed as the source of image degradation. The kernel PCA was used to
model the image class. This method was proposed in [20] and further developed in [15].
The STPRtool contains a script which converts the USPS database to the Matlab
data ﬁle. It is implemented in the function usps2mat. The corruption of the USPS
images by the Gaussian noise is implemented in script make noisy data. The script
make noisy data produces the Matlab ﬁle usps noisy.mat its contains is summarized
in Table 7.1.
The aim is to train a model of images of all numerals at ones. Thus the unlabeled
training data set T
X
= ¦x
1
, . . . , x
l
¦ is used. The model should describe the nonlinear
subspace A
img
⊆ R
n
where the images live. It is assumed that the nonlinear subspace
A
img
forms a linear subspace T
img
of a new feature space T. A link between the
input space A and the feature space T is given by a mapping φ: A → T which is
implicitly determined by a selected kernel function k: A A →R. The kernel function
corresponds to the dot product in the feature space, i.e., k(x
a
, x
b
) = ¸φ(x
a
) φ(x
b
)).
The kernel PCA can be used to ﬁnd the basis vectors of the subspace T
img
from the
input training data T
X
. Having the kernel PCA model of the subspace T
img
it can be
used to map the images x ∈ A into the subspace A
img
. A noisy image ˆ x is assumed to
lie outside the subspace A
img
. The denoising procedure is based on mapping the noisy
image ˆ x into the subspace A
img
. The idea of using the kernel PCA for image denoising
is demonstrated in Figure 7.3. The example which produced the ﬁgure is implemented
in demo kpcadenois.
The subspace T
img
is modeled by the kernel PCA trained from the set T
X
mapped
into the feature space T by a function φ: A → T. Let T
Φ
= ¦φ(x
1
), . . . , φ(x
l
)¦ denote
the feature space representation of the training images T
X
. The kernel PCA computes
the lower dimensional representation T
Z
= ¦z
1
, . . . , z
l
¦, z
i
∈ R
m
of the mapped images
87
Table 7.1: The content of the ﬁle usps noisy.mat produced by the script
make noisy data.
Structure array trn containing the training data:
gnd X [256 7291] Training images of numerals represented as column
vectors. The input are 7291 grayscale images of size
16 16.
X [256 7291] Training images corrupted by the Gaussian noise with
signaltonoise ration SNR = 1 (it can be changed).
y [1 7291] Labels y ∈ ¸ = ¦1, . . . , 10¦ of training images.
Structure array tst containing the testing data:
gnd X [256 2007] Testing images of numerals represented as column vec
tors. The input are 2007 grayscale images of size
16 16.
X [256 2007] Training images corrupted by the Gaussian noise with
signaltonoise ration SNR = 1 (it can be changed).
y [1 2007] Labels y ∈ ¸ = ¦1, . . . , 10¦ of training images.
T
Φ
to minimize the mean square reconstruction error
ε
KMS
=
l
i=1
φ(x
i
) −
˜
φ(x
i
)
2
,
where
˜
φ(x) =
l
i=1
β
i
φ(x) , β = A(z −b) , z = A
T
k(x) +b . (7.1)
The vector k(x) = [k(x
1
, x), . . . , k(x
l
, x)]
T
contains kernel functions evaluated at the
training data T
X
. The matrix A [l m] and the vector b [m 1] are trained by the
kernel PCA algorithm. Given a projection
˜
φ(x) onto the kernel PCA subspace an
image x
∈ R
n
of the input space can be recovered by solving the preimage problem
x
= argmin
x
φ(x) −
˜
φ(x)
2
. (7.2)
The Radial Basis Function (RBF) kernel k(x
a
, x
b
) = exp(−0.5x
a
− x
b

2
/σ
2
) was
used here as the preimage problem (7.2) can be well solved for this case. The whole
procedure of reconstructing the corrupted image ˆ x ∈ R
n
involves (i) using (7.1) to
compute z which determines
˜
φ(x) and (ii) solving the preimage problem (7.2) which
yields the resulting x
. This procedure is implemented in function kpcarec.
In contrast to the linear PCA, the kernel PCA allows to model the nonlinearity
which is very likely in the case of images. However, the nonlinearity hidden in
88
the parameter σ of the used RBF kernel has to be tuned properly to prevent over
ﬁtting. Another parameter is the dimension m of the subspace kernel PCA subspace
to be trained. The best parameters (σ
∗
, m
∗
) was selected out of a prescribed set
Θ = ¦(σ
1
, m
1
), . . . , (σ
k
, m
k
)¦ such that the crossvalidation estimate of the mean square
error
ε
MS
(σ, m) =
_
X
P
X
(x)
_
x −x

2
_
dx ,
was minimal. The x stands for the ground truth image and x
denotes the image
restored from noisy one ˆ x using (7.1) and (7.2). The images are assumed to be gen
erated from distribution P
X
(x). Figure 7.4 shows the idea of tuning the parameters
of the kernel PCA. The linear PCA was used to solve the same task for comparison.
The same idea was used to tune the output dimension of the linear PCA. Having the
parameters tuned the resulting kernel PCA and linear PCA are trained on the whole
training set. The whole training procedure of the kernel PCA is implemented in the
script train kpca denois. The training for linear PCA is implemented in the script
train lin denois. The greedy kernel PCA Algorithm 3.4 is used for training the
model. It produces sparse model and can deal with large data sets. The results of
tuning the parameters are displayed in Figure 7.5. The ﬁgure was produced by a script
show denois tuning.
The function kpcarec is used to denoise images using the kernel PCA model. In
the case of the linear PCA, the function pcarec was used. An example of denoising of
the testing data of USPS is implemented in script show denoising. Figure 7.6 shows
examples of ground truth images, noisy images, images restored by the linear PCA and
kernel PCA.
89
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Control: left_button − draw, middle_button − classify, right_button − erase.
(a)
(b)
Figure 7.1: Figure (a) shows handwritten training examples of numeral ’1’ and Figure
(b) shows corresponding normalized images of size 16 16 pixels.
90
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Control: left_button − draw, middle_button − classify, right_button − erase.
(a)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
1 2 3 4 5 6 7 8 9 0
4 0
5 3 1 9
6 2 8
(b)
Figure 7.2: A snapshot of running OCR system. Figure (a) shows handwritten nu
merals and Figure (b) shows recognized numerals.
91
−4 −3 −2 −1 0 1 2 3 4 5 6 7
−4
−3
−2
−1
0
1
2
3
4
5
6
7
Ground truth
Noisy examples
Corrupted
Reconstructed
Figure 7.3: Demonstration of idea of image denoising by kernel PCA.
Data
generator
model
model
Distortion
KPCA Preimage
problem
θ ∈ Θ
Error
e
r
r
o
r
θ
assessment
Figure 7.4: Schema of tuning parameters of kernel PCA model used for image restora
tion.
92
0 10 20 30 40 50 60 70 80 90 100
10
11
12
13
14
15
16
17
18
dim
ε
M
S
3.5 4 4.5 5 5.5 6 6.5 7 7.5 8
6
8
10
12
14
16
18
20
σ
ε
M
S
dim = 50
dim = 100
dim = 200
dim = 300
(a) (b)
Figure 7.5: Figure (a) shows plot of ε
MS
with respect to output dimension of lin
ear PCA. Figure(b) shows plot of ε
MS
with respect to output dimension and kernel
argument of the kernel PCA.
Ground truth
Numerals with added Gaussian noise
Linear PCA reconstruction
Kernel PCA reconstruction
Figure 7.6: The results of denosing the handwritten images of USPS database.
93
Chapter 8
Conclusions and future planes
The STPRtool as well as its documentation is being updating. Therefore we are grateful
for any comment, suggestion or other feedback which would help in the development
of the STPRtool.
There are many relevant methods and proposals which have not been implemented
yet. We welcome anyone who likes to contribute to the STPRtool in any way. Please
send us an email and your potential contribution.
The future planes involve the extensions of the toolbox by the following methods:
• Feature selection methods.
• AdaBoost and related learning methods.
• Eﬃcient optimizers to for large scale Support Vector Machines learning problems.
• Model selection methods for SVM.
• Probabilistic Principal Component Analysis (PPCA) and mixture of PPCA esti
mated by the ExpectationMaximization (EM) algorithm.
• Discrete probability models estimated by the EM algorithm.
94
Bibliography
[1] T.W. Anderson and R.R. Bahadur. Classiﬁcation into two multivariate normal
distributions with diﬀerrentia covariance matrices. Anals of Mathematical Statis
tics, 33:420–431, June 1962.
[2] G. Baudat and F. Anouar. Generalized discriminant analysis using a kernel ap
proach. Neural Computation, 12(10):2385–2404, 2000.
[3] C.M. Bishop. Neural Networks for Pattern Recognition. Clarendon Press, Oxford,
Great Britain, 3th edition, 1997.
[4] N. Cristianini and J. ShaweTaylor. Support Vector Machines. Cambridge Uni
versity Press, 2000.
[5] A.P. Dempster, N.M. Laird, and D.B. Rubin. Maximum likelihood from incom
plete data via the EM Algorithm. Journal of the Royal Statistical Society, 39:185–
197, 1977.
[6] R.O. Duda, P.E. Hart, and D.G. Stork. Pattern Classiﬁcation. John Wiley &
Sons, 2nd. edition, 2001.
[7] R.P.W. Duin. Prtools: A matlab toolbox for pattern recognition, 2000.
[8] V. Franc. Programov´e n´astroje pro rozpozn´ av´ an´ı (Pattern Recognition Program
ming Tools, In Czech). Master’s thesis,
ˇ
Cesk´e vysok´e uˇcen´ı technick´e, Fakulta
elektrotechnick´a, Katedra kybernetiky, February 2000.
[9] V. Franc and V. Hlav´aˇc. Multiclass support vector machine. In R. Kasturi,
D. Laurendeau, and Suen C., editors, 16th International Conference on Pattern
Recognition, volume 2, pages 236–239. IEEE Computer Society, 2002.
[10] V. Franc and V. Hlav´ aˇc. An iterative algorithm learning the maximal margin
classiﬁer. Pattern Recognition, 36(9):1985–1996, 2003.
95
[11] V. Franc and V. Hlav´ aˇc. Greedy algorithm for a training set reduction in the
kernel methods. In N. Petkov and M.A. Westenberg, editors, Computer Analysis
of Images and Patterns, pages 426–433, Berlin, Germany, 2003. Springer.
[12] C.W. Hsu and C.J. Lin. A comparison of methods for multiclass support vector
machins. IEEE Transactions on Neural Networks, 13(2), March 2002.
[13] I.T. Jollife. Principal Component Analysis. SpringerVerlag, New York, 1986.
[14] S.S. Keerthi, S.K. Shevade, C. Bhattacharya, and K.R.K. Murthy. A fast itera
tive nearest point algorithm for support vector machine classiﬁer design. IEEE
Transactions on Neural Networks, 11(1):124–136, January 2000.
[15] Kim Kwang, In, Franz Matthias, O., and Sch¨olkopf Bernhard. Kernel hebbian
algorithm for singlefram superresolution. In Leonardis Aleˇs and Bischof Horst,
editors, Statisical Learning in Computer Vision, ECCV Workshop. Springer, May
2004.
[16] J.T. Kwok and I.W. Tsang. The preimage problem in kernel methods. In Pro
ceedings of the Twentieth International Conference on Machine Learning (ICML
2003), pages 408–415, Washington, D.C., USA, August 2003.
[17] Y. LeCun, B. Boser, J.S. Denker, D. Henderson, R.E. Howard, W. Hubbard, and
L.J Jackel. Backpropagation applied to handwritten zip code recognition. Neural
Computation, 1:541–551, 1989.
[18] G. McLachlan and T. Krishnan. The EM Algorithm and Extensions. John Wiley
& Sons, New York, 1997.
[19] S. Mika, G. R¨ atsch, J. Weston, B. Sch¨olkpf, and K. M¨ uller. Fisher discriminant
analysis with kernel. In Y.H. Hu, J. Larsen, and S. Wilson, E. Douglas, editors,
Neural Networks for Signal Processing, pages 41–48. IEEE, 1999.
[20] S. Mika, B. Sch¨ olkopf, A. Smola, K.R. M¨ uller, M. Scholz, and G. R¨ atsch. Kernel
pca and denoising in feature spaces. In M.S. Kearns, S.A. Solla, and D.A. Cohn,
editors, Advances in Neural Information Processing Systems 11, pages 536 – 542,
Cambridge, MA, 1999. MIT Press.
[21] I.T. Nabney. NETLAB : algorithms for pattern recognition. Advances in pattern
recognition. Springer, London, 2002.
[22] J. Platt. Probabilities for sv machines. In A.J. Smola, P.J. Bartlett, B. Scholkopf,
and D. Schuurmans, editors, Advances in Large Margin Classiﬁers (Neural Infor
mation Processing Series). MIT Press, 2000.
96
[23] J.C. Platt. Sequential minimal optimizer: A fast algorithm for training support
vector machines. Technical Report MSRTR9814, Microsoft Research, Redmond,
1998.
[24] B.D. Riply. Neural networks and related methods for classiﬁcation (with discu
sion). J. Royal Statistical Soc. Series B, 56:409–456, 1994.
[25] M.I. Schlesinger. A connection between learning and selflearning in the pattern
recognition (in Russian). Kibernetika, 2:81–88, 1968.
[26] M.I. Schlesinger and V. Hlav´ aˇc. Ten lectures on statistical and structural pattern
recognition. Kluwer Academic Publishers, 2002.
[27] B. Sch¨olkopf, C. Burges, and A.J. Smola. Advances in Kernel Methods – Support
Vector Learning. MIT Press, Cambridge, 1999.
[28] B. Sch¨olkopf, P. Knirsch, and C. Smola, A. Burges. Fast approximation of support
vector kernel expansions, and an interpretation of clustering as approximation in
feature spaces. In P. Levi, M. Schanz, R.J. Ahler, and F. May, editors, Mus
tererkennung 199820. DAGM., pages 124–132, Berlin, Germany, 1998. Springer
Verlag.
[29] B. Scholkopf, A. Smola, and K.R. Muller. Nonlinear component analysis as a
kernel eigenvalue problem. Neural Computation, 10:1299–1319, 1998.
[30] B. Sch¨olkopf and A.J. Smola. Learning with Kernels. The MIT Press, MA, 2002.
[31] V. Vapnik. The nature of statistical learning theory. Springer Verlag, 1995.
97
Statistical Pattern Recognition Toolbox for Matlab
Vojtˇch Franc and V´clav Hlav´ˇ e a ac June 24, 2004
Contents
1 Introduction 1.1 What is the Statistical Pattern Recognition Toolbox and how it has been developed? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 The STPRtool purpose and its potential users . . . . . . . . . . . . . . 1.3 A digest of the implemented methods . . . . . . . . . . . . . . . . . . . 1.4 Relevant Matlab toolboxes of others . . . . . . . . . . . . . . . . . . . . 1.5 STPRtool document organization . . . . . . . . . . . . . . . . . . . . . 1.6 How to contribute to STPRtool? . . . . . . . . . . . . . . . . . . . . . 1.7 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Linear Discriminant Function 2.1 Linear classiﬁer . . . . . . . . . . . . . . . . . . 2.2 Linear separation of ﬁnite sets of vectors . . . . 2.2.1 Perceptron algorithm . . . . . . . . . . . 2.2.2 Kozinec’s algorithm . . . . . . . . . . . . 2.2.3 Multiclass linear classiﬁer . . . . . . . . 2.3 Fisher Linear Discriminant . . . . . . . . . . . . 2.4 Generalized Anderson’s task . . . . . . . . . . . 2.4.1 Original Anderson’s algorithm . . . . . . 2.4.2 General algorithm framework . . . . . . 2.4.3 εsolution using the Kozinec’s algorithm 2.4.4 Generalized gradient optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 3 3 4 5 5 5 6 8 9 10 11 12 14 15 17 19 21 21 22 25 26 29 31 32 33 34
3 Feature extraction 3.1 Principal Component Analysis . . . . . . . . . . . 3.2 Linear Discriminant Analysis . . . . . . . . . . . 3.3 Kernel Principal Component Analysis . . . . . . . 3.4 Greedy kernel PCA algorithm . . . . . . . . . . . 3.5 Generalized Discriminant Analysis . . . . . . . . . 3.6 Combining feature extraction and linear classiﬁer
1
. 5. . . . . . . .2 MaximumLikelihood estimation of GMM .5 Kmeans clustering . . . . . . . . . .2. . . .1 Optical Character Recognition system . . . . . . . . 5 Support vector and other kernel machines 5. . . . . . . . . . .2 Image denoising . . 4. . .3 Bayesian classiﬁer . . . . . 6. . . . . . . . . . . .7 Preimage problem . . . . . . . . . . . 41 41 43 44 45 47 48 50 57 57 58 63 64 66 68 70 73 76 76 77 78 80 82 85 85 87 94 7 Examples of applications 7. . . . . .2. . . . . . . . . . . . .8 Reduced set method . . . . . . . . . . . . . . . . . . .5 Knearest neighbors classiﬁer . . . . . . . . .3 OneAgainstAll decomposition for SVM . . . . . 4. . . . . . . . . . . . . . . . . . . . . . . . . . .1 Complete data . . . . . . . . . . . . . . . . . . . .4 Quadratic classiﬁer . . . . . . . . . . . .4 OneAgainstOne decomposition for SVM . . . . . . . . . . . . . . . . . . . . . 4. . . . . . . . . 6 Miscellaneous 6. . . . . . . . . . . .1 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6 Kernel Fisher Discriminant . . . . . . . . . . . . . . . . . .2 Visualization . . . . . . . . . . . . . . . . 4. . . . . . . . . . . . . . . . . 6. . . . . . . . . . . . . . . . . . . . . . . . . 6. . . . . . . . 4. . . . . . . . . . . . . . . . . . .2 Incomplete data . . . . . . . . . . . . . . . . . . . 5. . . . . . . . . . 5. . . . . . . . . . . . . . .2 Binary Support Vector Machines . . . . . . . . . . . . . . . .5 Multiclass BSVM formulation . . . . . .3 Minimax estimation . . . . . . . . . 5. . . . . . . .1 Support vector classiﬁers .4 Probabilistic output of classiﬁer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . . . . . . . . . . . . . 7. . . .1 Gaussian distribution and Gaussian mixture model 4. . . . . . . 8 Conclusions and future planes 2 . . . . . . . . . . . . . . . 5. . . . . . 6. . . . . . .4 Density estimation and clustering 4. . 5. . .
Chapter 1 Introduction
1.1 What is the Statistical Pattern Recognition Toolbox and how it has been developed?
The Statistical Pattern Recognition Toolbox (abbreviated STPRtool) is a collection of pattern recognition (PR) methods implemented in Matlab. The core of the STPRtool is comprised of statistical PR algorithms described in the monograph Schlesinger, M.I., Hlav´ˇ, V: Ten Lectures on Statistical and Structural ac Pattern Recognition, Kluwer Academic Publishers, 2002 [26]. The STPRtool has been e developed since 1999. The ﬁrst version was a result of a diploma project of Vojtˇch Franc [8]. Its author became a PhD student in CMP and has been extending the STPRtool since. Currently, the STPRtool contains much wider collection of algorithms. The STPRtool is available at http://cmp.felk.cvut.cz/cmp/cmp software.html. We tried to create concise and complete documentation of the STPRtool. This document and associated help ﬁles in .html should give all information and parameters the user needs when she/he likes to call appropriate method (typically a Matlab mﬁle). The intention was not to write another textbook. If the user does not know the method then she/he will likely have to read the description of the method in a referenced original paper or in a textbook.
1.2
The STPRtool purpose and its potential users
• The principal purpose of STPRtool is to provide basic statistical pattern recognition methods to (a) researchers, (b) teachers, and (c) students. The STPRtool is likely to help both in research and teaching. It can be found useful by practitioners who design PR applications too.
3
• The STPRtool oﬀers a collection of the basic and the stateofthe art statistical PR methods. • The STPRtool contains demonstration programs, working examples and visualization tools which help to understand the PR methods. These teaching aids are intended for students of the PR courses and their teachers who can easily create examples and incorporate them into their lectures. • The STPRtool is also intended to serve a designer of a new PR method by providing tools for evaluation and comparison of PR methods. Many standard and stateofthe art methods have been implemented in STPRtool. • The STPRtool is open to contributions by others. If you develop a method which you consider useful for the community then contact us. We are willing to incorporate the method into the toolbox. We intend to moderate this process to keep the toolbox consistent.
1.3
A digest of the implemented methods
This section should give the reader a quick overview of the methods implemented in STPRtool. • Analysis of linear discriminant function: Perceptron algorithm and multiclass modiﬁcation. Kozinec’s algorithm. Fisher Linear Discriminant. A collection of known algorithms solving the Generalized Anderson’s Task. • Feature extraction: Linear Discriminant Analysis. Principal Component Analysis (PCA). Kernel PCA. Greedy Kernel PCA. Generalized Discriminant Analysis. • Probability distribution estimation and clustering: Gaussian Mixture Models. ExpectationMaximization algorithm. Minimax probability estimation. Kmeans clustering. • Support Vector and other Kernel Machines: Sequential Minimal Optimizer (SMO). Matlab Optimization toolbox based algorithms. Interface to the SV M light software. Decomposition approaches to train the Multiclass SVM classiﬁers. Multiclass BSVM formulation trained by Kozinec’s algorithm, MitchellDemyanovMolozenov algorithm and Nearest Point Algorithm. Kernel Fisher Discriminant.
4
1.4
Relevant Matlab toolboxes of others
Some researchers in statistical PR provide code of their methods. The Internet oﬀers many software resources for Matlab. However, there are not many compact and well documented packages comprised of PR methods. We found two useful packages which are brieﬂy introduced below. NETLAB is focused to the Neural Networks applied to PR problems. The classical PR methods are included as well. This Matlab toolbox is closely related to the popular PR textbook [3]. The documentation published in book [21] contains detailed description of the implemented methods and many working examples. The toolbox includes: Probabilistic Principal Component Analysis, Generative Topographic Mapping, methods based on the Bayesian inference, Gaussian processes, Radial Basis Functions (RBF) networks and many others. The toolbox can be downloaded from http://www.ncrg.aston.ac.uk/netlab/. PRTools is a Matlab Toolbox for Pattern Recognition [7]. Its implementation is based on the object oriented programming principles supported by the Matlab language. The toolbox contains a wide range of PR methods including: analysis of linear and nonlinear classiﬁers, feature extraction, feature selection, methods for combining classiﬁers and methods for clustering. The toolbox can be downloaded from http://www.ph.tn.tudelft.nl/prtools/.
1.5
STPRtool document organization
The document is divided into chapters describing the implemented methods. The last chapter is devoted to more complex examples. Most chapters begin with deﬁnition of the models to be analyzed (e.g., linear classiﬁer, probability densities, etc.) and the list of the implemented methods. Individual sections have the following ﬁxed structure: (i) deﬁnition of the problem to be solved, (ii) brief description and list of the implemented methods, (iii) reference to the literature where the problem is treaded in details and (iv) an example demonstrating how to solve the problem using the STPRtool.
1.6
How to contribute to STPRtool?
If you have an implementation which is consistent with the STPRtool spirit and you want to put it into public domain through the STPRtool channel then send an email to xfrancv@cmp.felk.cvut.cz. 5
other STPRtool users and students who used the toolbox in their laboratory exercises. M.I. the Czech Technical University in Prague. 6 . Kiev. 1. Ukraine who showed us the beauty of pattern recognition and gave us many valuable advices.7 Acknowledgements First. EU and industrial grants which indirectly contributed to STPRtool development. Second. Third. Vernon. we like to thank for discussions to our colleagues in the Center for Machine Perception. Forth.Please prepare your code according to the pattern you ﬁnd in STPRtool implementations. Provide piece of documentation too which will allow us to incorporate it to STPRtool easily. Schlesinger from the Ukrainian Academy of Sciences. our visitors. your method will be bound to your name in STPRtool. we like to thank to Prof. Of course. thanks go to several Czech. the STPRtool was rewritten and its documentation radically improved in May and June 2004 with the ﬁnancial help of ECVision European Research Network on Cognitive Computer Vision Systems (IST200135454) lead by D.
Logical or. Cardinality of a set. Vectors are implicitly column vectors. Mean vector (mathematical expectation). Kernel function (Mercer kernel). j) Dot product between vectors x and x . z] where the column vector v is (n + m)dimensional. j) = 1 for i = j and δ(i. . Euclidean norm. . . . . Matrix determinant. Feature space. for instance x. Unlabeled training set TX = {x1 . for instance A. y1 ). Classiﬁcation rule. . The concatenation of column vectors w ∈ Rn . z ∈ Rm is denoted as v = [w. c}. xl }. . Discriminant function of the binary classiﬁer. . Vectors are denoted by lowercase bold italic letters.Notation Uppercase bold letters denote matrices. j) = 0 otherwise. Kronecker delta δ(i. x·x Σ µ S X ⊆ Rn Y F φ: X → F k: X × X → R TXY TX q: X → Y fy : X → R f: X → R · · det(()·) δ(i. Mapping from input space X to the feature space F . Labeled training set (more precisely multiset) TXY = {(x1 . ndimensional input space (space of observations). . Scatter matrix. . Discriminant function associated with the class y ∈ Y. 7 . Set of c hidden states – class labels Y = {1. . Covariance matrix. (xl . yl )}. .
∀y ∈ Y introduce bias to the discriminant functions.3. Following sections describe supervised learning methods which determine the parameters of the linear classiﬁer based on available knowledge. c} is composed of a set of discriminant functions fy (x) = w y · x + by . 2 if f (x) = w · x + b < 0 . . (xl . . y1 ). 2} as follows q(x) = 1 if f (x) = w · x + b ≥ 0 . . .Chapter 2 Linear Discriminant Function The linear classiﬁcation rule q: X ⊆ Rn → Y = {1. (2. The implemented learning methods can be distinguished according to the type of the training data: I.2 and Section 2. . The scalars by .2) The datatype used to describe the linear classiﬁer and the implementation of the classiﬁer is described in Section 2. 2.1. The input vector x ∈ Rn is assigned to class y ∈ {1. The methods using this type of training data is described in Section 2. ∀y ∈ Y . The problem is described by a ﬁnite set T = {(x1 . y ∈Y y ∈Y (2. The input vector x ∈ Rn is assigned to the class y ∈ Y its corresponding discriminant function fy attains maximal value y = argmax fy (x) = argmax ( wy · x + by ) . the linear classiﬁer is represented by a single discriminant function f (x) = w · x + b . . yl )} containing pairs of observations xi ∈ Rn and corresponding class labels yi ∈ Y. 8 .1) In the particular binary case Y = {1. . given by the parameter vector w ∈ Rn and bias b ∈ R. . which are linear with respect to both the input vector x ∈ Rn and their parameter vectors w ∈ Rn . 2}.
6].1: Implemented methods: Linear discriminant function linclass Linear classiﬁer (linear machine). The parameters of class conditional distributions are (completely or partially) known. 9 .mat is used for training. demo linclass Demo on the algorithms learning linear classiﬁers.3) linear classiﬁer.1. Table 2. mperceptr Perceptron algorithm to train multiclass linear machine. The implementation of the linear classiﬁer itself provides function linclass. fldqp Fisher Linear Discriminant using Quadratic Programming. andrerr Classiﬁcation error of the Generalized Anderson’s task.mat.1 Linear classiﬁer The STPRtool uses a speciﬁc structure array to describe the binary (see Table 2. ggradander Gradient method to solve the Generalized Anderson’s task. fld Fisher Linear Discriminant. The summary of implemented methods is given in Table 2.4). ganders Solves the Generalized Anderson’s task. Example: Linear classiﬁer The example shows training of the Fisher Linear Discriminant which is the classical example of the linear classiﬁer. perceptron Perceptron algorithm to train binary linear classiﬁer. demo anderson Demo on Generalized Anderson’s task.2) and the multiclass (see Table 2.II. 2. The Anderson’s task and its generalization are representatives of such methods (see Section 2. The resulting linear classiﬁer is then evaluated on the testing data riply tst. androrig Original method to solve the AndersonBahadur’s task. References: The linear discriminant function is treated throughout for instance in books [26. eanders Epsilonsolution of the Generalized Anderson’s task. ekozinec Kozinec’s algorithm for epsoptimal separating hyperplane. The Riply’s data set riply trn.
b [1 × 1] Bias b ∈ R of the hyperplane. trn = load(’riply_trn’).2. tst.1) 10 . . The Perceptron (Section 2. The problem is formally deﬁned as solving the set of linear inequalities w · xi + b ≥ 0 .fun = ’linclass’ Identiﬁes function associated with this data type. w · xi + b < 0 . .1080 % % % % % load training data load testing data train FLD classifier classify testing data evaluate testing error 2.3: Datatype used to describe multiclass linear classiﬁer. .2: Datatype used to describe binary linear classiﬁer. model ). The deﬁnitions of these tasks are given below. y = 1. Binary linear classiﬁer (structure array): . . . yi = 1 . y1 ). The implemented methods solve the task of (i) training separating hyperplane for binary case c = 2. yi = 2 . .y ) ans = 0. c. If the inequalities (2.W [n × c] Matrix of parameter vectors w y ∈ Rn . Multiclass linear classiﬁer (structure array): .X. (2. yl )} consists of pairs of observations xi ∈ Rn and corresponding class labels yi ∈ Y = {1. cerror( ypred. . c of the linear discriminant functions fy (x) = w y · x + b. .3) with respect to the vector w ∈ Rn and the scalar b ∈ R. c}. ypred = linclass( tst. . . . model = fld( trn ). . . The problem of training the binary (c = 2) linear classiﬁer (2.Table 2.3) have a solution then the training data T is linearly separable.fun = ’linclass’ Identiﬁes function associated with this data type. y = 1. . . .2 Linear separation of ﬁnite sets of vectors The input training data T = {(x1 . . Table 2. . tst = load(’riply_tst’).b [c × 1] Parameters by ∈ R. .2) with zero training error is equivalent to ﬁnding the hyperplane H = {x: w · x + b = 0} which separates the training vectors of the ﬁrst y = 1 and the second y = 2 class. . (xl . (ii) training εoptimal hyperplane and (iii) training the multiclass linear classiﬁer.W [n × 1] The normal vector w ∈ Rn of the separating hyperplane f (x) = w · x + b.
. The problem of training the multiclass linear classiﬁer c > 2 with zero training error is formally stated as the problem of solving the set of linear inequalities w yi · xi + byi > wy · xi + by . 2.8) i = 1. The optimal separating hyperplane H∗ = {x ∈ Rn : w∗ · x + b∗ = 0} separates training data T with maximal margin m(w ∗ . (2. The optimal separating hyperplane cannot be found exactly except special cases. Therefore the numerical algorithms seeking the approximate solution are applied instead. y ∈ Y.2. An interactive demo on the algorithms separating the ﬁnite sets of vectors by a hyperplane is implemented in demo linclass. y ∈ Y and scalars by ∈ R.2. .b w. If the training data is linearly separable then the optimal separating hyperplane can be deﬁned the solution of the following task (w∗ . b∗ ). yl )} of binary labeled yi ∈ {1. (xl . b∗ ) = argmax m(w.7) .5) holds. b) ≤ ε .6) with respect to the vectors wy ∈ Rn . b∗ ) − m(w. . (2.2. where the vector v ∈ Rn+1 is constructed as v = [w. . i = 1. .2).b = argmax min min i∈I1 w · xi + b w · xi + b . The task (2.2) algorithm are useful to train the binary linear classiﬁers from the linearly separable data. l . .3) can be formally rewritten to a simpler form v · zi > 0 . .1 Perceptron algorithm The input is a set T = {(x1 .and Kozinec’s (Section 2. The εoptimal solution can be found by Kozinec’s algorithm (Section 2.4) where I1 = {i: yi = 1} and I2 = {i: yi = 2} are sets of indices. The parameter ε ≥ 0 deﬁnes closeness to the optimal solution in terms of the margin. y1). The εoptimal solution is deﬁned as the vector w and scalar b such that the inequality m(w ∗ . The problem of training the separating hyperplane (2. b] . yi = y . l .2. This task is equivalent to training the linear Support Vector Machines (see Section 5). b) w. min − w w i∈I2 .6) for linearly separable training data T can be solved by the modiﬁed Perceptron (Section 2. 11 (2. . .3) or Kozinec’s algorithm. . 2} training vectors xi ∈ Rn . (2. . . (2.
1.7) with respect to the unknown vector v ∈ Rn+1 is equivalent to the original task (2. y1). (2. . 1). . z l } are deﬁned as zi = [xi . The initial vector v (0) can be set arbitrarily (usually v = 0). 1] if yi = 2 . −[xi . . . v (1) . Example: Training binary linear classiﬁer with Perceptron The example shows the application of the Perceptron algorithm to ﬁnd the binary linear classiﬁer for synthetically generated 2dimensional training data.w2 ∈X2 w1 − w2 . The Perceptron from the neural network perspective is described in the book [3]. The Novikoﬀ theorem ensures that the Perceptron algorithm stops after ﬁnite number of iterations t if the training data are linearly separable. .9) The problem of solving (2. ppatterns( data ). w∗ = 1 2 argmin w1 ∈X1 . The Perceptron algorithm is an iterative procedure which builds a series of vectors (0) v . w 2 . The found classiﬁer is visualized as a separating hyperplane (line in this case) H = {x ∈ R2 : w · x + b = 0}. The vector w∗ = 12 . . figure. where X1 stands for the convex hull of the training vectors of the ﬁrst class X1 = {xi : yi = 1} and X2 for the convex hull of the second class likewise. . yl )} of binary labeled yi ∈ {1. . w1 . .and transformed training data Z = {z 1 . . b) of the linear classiﬁer are obtained from the found vector v by inverting the transformation (2. The parameters (w. pline( model ).2 Kozinec’s algorithm The input is a set T = {(x1 . w2 which converge to the vector w ∗ and w ∗ respectively. data = genlsdata( 2. . . 2} training (0) (1) (t) vectors xi ∈ Rn . . 50. The Kozinec’s algorithm builds a series of vectors w 1 . . w 1 (0) (1) (t) and w 2 . See Figure 2. Perceptron algorithm is implemented in function perceptron.2.8). v (t) until the set of inequalities (2. . The generated training data contain 50 labeled vectors which are linearly separable with margin 1.7) is satisﬁed. % generate training data % call perceptron % plot training data % plot separating hyperplane 2. References: Analysis of the Perceptron algorithm including the Novikoﬀ’s proof of convergence can be found in Chapter 5 of the book [26]. 1] if yi = 1 .3). (xl . The 1 2 vectors w∗ and w ∗ are the solution of the following task 1 2 w∗. . . model = perceptron( data ). . . .
3).4) 1 2 2 1 2 separating the training data with the maximal margin. Example: Training εoptimal hyperplane by the Kozinec’s algorithm The example shows application of the Kozinec’s algorithm to ﬁnd the (ε = 0. The εoptimal hyperplane (2. References: Analysis of the Kozinec’s algorithm including the proof of convergence can be found in Chapter 5 of the book [26].01)optimal hyperplane for synthetically generated 2dimensional training data. The Kozinec’s algorithm can be also used to solve a simpler problem of ﬁnding the separating hyperplane (2. Therefore the following two stopping conditions are implemented: I.1: Linear classiﬁer trained by the Perceptron algorithm w∗ − w ∗ and the bias b∗ = 1 ( w ∗ 2 − w∗ 2 ) determine the optimal hyperplane (2.5) is sought for (ε ≥ 0).−30 −35 −40 −45 −50 −55 −55 −50 −45 −40 −35 −30 Figure 2. The Kozinec’s algorithm is proven to converge to the vectors w∗ and w ∗ in inﬁnite 1 2 number of iterations t = ∞. The Kozinec’s algorithm which trains the binary linear classiﬁer is implemented in the function ekozinec. The separating hyperplane (2. The generated training data contains 50 labeled vectors which are linearly separable with 13 .3) is sought for (ε < 0). II. If the εoptimal optimality stopping condition is used then the Kozinec’s algorithm converges in ﬁnite number of iterations. Notice that setting ε = 0 force the algorithm to seek the optimal hyperplane which is generally ensured to be found in an inﬁnite number of iterations t = ∞. The Kozinec’s algorithm is proven to converge in a ﬁnite number of iterations if the separating hyperplane exists.
yl )} consists of pairs of observations xi ∈ Rn and corresponding class labels yi ∈ Y = {1.margin which satisﬁes the εoptimality condition. . l . % plot data pline(model). . .10) where the vector v ∈ R(n+1)c contains the originally optimized parameters v = [w1 .2: εoptimal hyperplane trained with the Kozinec’s algorithm. . y1 ). w 2 . c}. bc ] . . (xl .6). ppatterns(data).2) as a separating hyperplane H = {x ∈ R2 : w · x + b = 0} (straight line in this case).0. . 2. . . b1 .1) which classiﬁes the training data T without error.1). The inequalities (2. i i = 1. % plot found hyperplane 8 6 4 2 0 −2 −4 −6 −8 −10 −10 −8 −6 −4 −2 0 2 4 6 8 10 Figure 2. . w c . 14 . struct(’eps’.2. .50.3 Multiclass linear classiﬁer The input training data T = {(x1 . figure. data = genlsdata(2. (2.01)). .1) is equivalent to solving the set of linear inequalities (2.6) can be formally transformed to a simpler set v · zy > 0 . . . % generate data % run ekozinec model = ekozinec(data. The found classiﬁer is visualized (Figure 2. The problem of training (2. . y = Y \ {yi } . b2 .margin 1. It could be veriﬁed that the found hyperplane has margin model. . The task is to design the linear classiﬁer (2. .
y ∈ {1.mat. i The described transformation is known as the Kesler’s construction.11) where z y (j) stands for the jth slot between coordinates (n + 1)(j − 1) + 1 and (n + 1)j. Example: Training multiclass linear classiﬁer by the Perceptron The example shows application of the Perceptron rule to train the multiclass linear classiﬁer for the synthetically generated 2dimensional training data pentagon.10) (number 10 stands for number of classes). . data = load(’pentagon’).10) can be solved by the Perceptron algorithm. µy = 15 . figure. Iy  i∈I y (2. l. . w · SW w 1 xi . . i = 1. z i (j) = ⎩ 0. % load training data % run training algorithm % plot training data % plot decision boundary 2. for j = y . y −[xi . otherwise. (2. Let Iy = {i: yi = y}. (xl . y ∈ {1. 6].3 Fisher Linear Discriminant The input is a set T = {(x1 . respectively. yl )} of binary labeled yi ∈ {1. 2} . y i set T such that ⎧ ⎨ [xi . The class separability in a direction w ∈ Rn is deﬁned as F (w) = w · SB w . 2} be sets of indices of training vectors belonging to the ﬁrst y = 1 and the second y = 2 class. 1] . . 2} training vectors xi ∈ Rn . A user’s own data can be generated interactively using function createdata(’finite’. ∈ Y \ {yi } is created from the training for j = yi . 1] . ppatterns(data). . The transformed task (2.3). model = mperceptron(data). . The described method is implemented in function mperceptron. pboundary(model). References: The Kesler’s construction is described in books [26. The simple form of the Perceptron updating rule allows to implement the transformation implicitly without mapping the training data T into (n + 1)cdimensional space. . y1).The set of vectors z y ∈ R(n+1)c . .12) where SB is the betweenclass scatter matrix SB = (µ1 − µ2 )(µ1 − µ2 )T . The found classiﬁer is visualized as its separating surface which is known to be piecewise linear and the class regions should be convex (see Figure 2.
The classical solution of the problem (2.8 0. computes such b that the equality w · µ1 + b = −( w · µ2 + b) .3: Multiclass linear classiﬁer trained by the Perceptron algorithm. The STPRtool implementation. This case is implemented in the function fld.1 0.6 −0.2 −0.4 −0. y ∈ {1.6 0.2 0 0. The STPRtool contains two implementations of FLD: I.8 1 Figure 2.2 0 −0. Sy = i∈Yy (xi − µy )(xi − µy )T . holds.13) The bias b of the linear rule must be determined based on another principle.8 −0. 2} . the parameter vector w of the linear discriminant function f (x) = w · x + b is determined to maximize the class separability criterion (2. w · SW w (2. and SW is the within class scatter matrix deﬁned as SW = S1 + S2 .4 −0.4 0.8 −1 −1 −0. 16 .6 0.13) using the matrix inversion w = SW −1 (µ1 − µ2 ) .12) w = argmax F (w ) = argmax w w w · SB w .4 0.6 −0. for instance. In the case of the Fisher Linear Discriminant (FLD).2 0.
The QP formulation fldqp of FLD is used. 2}.y) ans = 0. it is known that the parameters (µ1 . Σi ): i ∈ I1 }. the function fldqp can be replaced by the standard approach implemented in the function fld which would yield the same solution. figure. The found linear classiﬁer is visualized (see Figure 2. Σ2 ) of these class distributions are unknown. The class conditional distributions pXY (xy). y ∈ {1. trn = load(’riply_trn’). However.4) and evaluated on the testing data riply tst. Σ1 ) and (µ2 . The problem (2. However. cerror(ypred.model).II. tst = load(’riply_tst’). Let q: X ⊆ Rn → {1.1080 % load training data % compute FLD % % % % plot data and solution load testing data classify testing data compute testing error 2. The method is implemented in the function fldqp.tst. Similarly (µ2 . ppatterns(trn). Σi ): i ∈ I2 }. w subject to w · (µ1 − µ2 ) = 2 . Σ1 ) belong to a certain ﬁnite set of parameters {(µi . This method can be useful when the matrix inversion is hard to compute. References: The Fisher Linear Discriminant is described in Chapter 3 of the book [6] or in the book [3].13) can be reformulated as the quadratic programming (QP) task w = argmin w · SW w . model = fldqp(trn).X.mat. 2} 17 . ypred = linclass(tst. The parameters (µ1 . The QP solver quadprog of the Matlab Optimization Toolbox is used in the toolbox implementation.mat. pline(model). Σ2 ) belong to a ﬁnite set {(µi . 2} are known to be multivariate Gaussian distributions. Example: Fisher Linear Discriminant The binary linear classiﬁer based on the Fisher Linear Discriminant is trained on the Riply’s data set riply trn.4 Generalized Anderson’s task The classiﬁed object is described by the vector of observations x ∈ X ⊆ Rn and a binary hidden state y ∈ {1.
b∗ ) = argmin Err(w.5 1 Figure 2.4: Linear classiﬁer based on the Fisher Linear Discriminant. w.5 −1 −0. such that the error Err(w∗ . The probability of misclassiﬁcation is deﬁned as Err(w.14) when I1  = 1 and I2  = 1.5 0 0. b. for f (x) = w∗ · x + b∗ < 0 . q(x) = 2 . The probability ε(w. µi . µi .2 −1. Σi ) is proportional to the reciprocal of the Mahalanobis distance r i between the (µi . 18 . µi .2) with discriminant function f (x) = w · x + b. for f (x) = w∗ · x + b∗ ≥ 0 .4 0. b. it is the probability that the vector x will be misclassiﬁed by the linear rule q.6 0.2 1 0. In other words.e. µi . i. Σi ) . b) = max ε(w. i∈I1 ∪I2 where ε(w. b∗ ) of the linear classiﬁer 1 . b∗ ) is minimal (w ∗ . be a binary linear classiﬁer (2.8 0. Σi ) is probability that the Gaussian random vector x with mean vector µi and the covariance matrix Σi satisﬁes q(x) = 1 for i ∈ I2 or q(x) = 2 for i ∈ I1 .1. Σi ) . r i = min (µi − x) · (Σi )−1 (µi − x) = x∈H w · µi + b w · Σi w . The Generalized Anderson’s task (GAT) is to ﬁnd the parameters (w ∗ .2 0 −0.b w. b. b) = argmin max ε(w.14) The original Anderson’s task is a special case of (2.b i∈I1 ∪I2 (2.. Σi ) and the nearest vector of the separating hyperplane H = {x ∈ Rn : w · x + b = 0}. b.
4. A detailed description of the Generalized Anderson’s task and all the methods implemented in the STPRtool is given in book [26]. The objective function F (w. b) is less than 0.1 Original Anderson’s algorithm The algorithm is implemented in the function androrig. b) = argmax min w.5. b. µi .b w. 2. Σi ) and the corresponding Mahalanobis distance r i is given by the integral ε(w. w · Σ1 w where the scalar 0 < λ < 1 is unknown. b. 1−λ = λ w · Σ2 w .14) can be equivalently rewritten as (w ∗ . References: The original Anderson’s task was published in [1]. Σi ) = ∞ ri 1 2 1 √ e− 2 t dt . the optimal solution vector w ∈ Rn is obtained by solving the following problem w = ((1 − λ)Σ1 + λΣ2 )−1 (µ1 − µ2 ) . the objective function F (w. µi .b i∈I1 ∪I2 w · µi + b . The STPRtool contains implementations of the algorithm solving the original Anderson’s task as well as implementations of three diﬀerent approaches to solve the Generalized Anderson’s task which are described bellow.15) The optimization problem (2. The problem is solved by an iterative algorithm which stops while the condition γ− 1−λ ≤ε. λ γ= w · Σ2 w . 2π (2. The parameter ε deﬁnes the closeness to the optimal solution.The exact relation between the probability ε(w. However. b) is proven to be convex in the region where the probability of misclassiﬁcation Err(w. w · Σ1 w is satisﬁed. In this case. An interactive demo on the algorithms solving the Generalized Anderson’s task is implemented in demo anderson. w · Σi w which is more suitable for optimization. b) is not diﬀerentiable. b∗ ) = argmax F (w. Example: Solving the original Anderson’s task 19 . This algorithm solves the original Anderson’s task deﬁned for two Gaussians only I1  = I2  = 1.
pandr( model. % cerror( ypred.4 −0. Σ2 ) of the Gaussian distribution of the ﬁrst and the second class are obtained by the MaximumLikelihood estimation.6 0.7 gauss 2 0. % model = androrig(distrib).5 0.8 −0.6 −0. trn = load(’riply_trn’).37 0.4 gauss 1 0. distrib ).929 0.mat.8 0. % figure.2 0 0.mat. The classiﬁer is validated on the testing set riply tst. The parameters (µ1 .1150 load training data estimate Gaussians solve Anderson’s task plot solution load testing data classify testing data evaluate error 0. tst. % ypred = linclass( tst. % distrib = mlcgmm(data). model ).3 0.5).9 1. The found linear classiﬁer is visualized and the Gaussian distributions (see Figure 2.6 Figure 2.4 0.2 0. Σ1 ) and (µ2 .X.y ) % ans = 0. 20 .The original Anderson’s task is solved for the Riply’s data set riply trn.5: Solution of the original Anderson’s task.2 0.1 −0. % tst = load(’riply_tst’).
% load input Gaussians % solve the GAT % plot the solution 2. The quadprog function of the Matlab optimization toolbox is used to solve this subtask. Both the ﬁrst and second class are described by three 2Ddimensional Gaussian distributions. is violated. the condition F (w(t+1) . This algorithm repeats the following steps: I. figure.mat. distrib = load(’mars’). b(t+1) ) − F (w (t) . II. An improving direction (∆w. III..6). model = ganders(distrib). b(t+1) : = b(t) + k∆b . pandr(model. b) such that F (w. The εsolution is deﬁned as a linear classiﬁer with parameters (w.e. The number of interval cuttings of is an optional parameter of the algorithm. i. b) < ε 21 .distrib). The new parameters are set w(t+1) : = w (t) + k∆w . The optimal movement k along the direction (∆w.2. Example: Solving the Generalized Anderson’s task The Generalized Anderson’s task is solved for the data mars. The found linear classiﬁer as well as the input Gaussian distributions are visualized (see Figure 2.4.4.3 εsolution using the Kozinec’s algorithm This algorithm is implemented in the function eanders. The optimal movement k is determined by 1Dsearch based on the cutting interval algorithm according to the Fibonacci series. b(t) ) > ε .2 General algorithm framework The algorithm is implemented in the function ganders. ∆b) in the parameter space is computed such that moving along this direction increases the objective function F . The algorithm iterates until the minimal change in the objective function is less than the prescribed ε. ∆b).
distrib options model = figure. The found linear classiﬁer and the input Gaussian distributions are visualized (see Figure 2. Otherwise it can get stuck in an endless loop. The problem is transformed to the task of separation of two sets of ellipsoids which is solved the Kozinec’s algorithm. The desired maximal possible probability of misclassiﬁcation is set to 0. pandr(model.06 (it is 6%).distrib). b) is convex but not diﬀerentiable thus the standard gradient optimization cannot 22 .options).0461 gauss 2 gauss 3 2 gauss 1 0.047 0.0483 4 3 0.mat. The objective function F (w.5 0. The prescribed ε ∈ (0.4.6: Solution of the Generalized Anderson’s task. = struct(’err’. eanders(distrib.06’). Both the ﬁrst and second class are described by three 2Ddimensional Gaussian distributions. is satisﬁed. % % % % load Gaussians maximal desired error training visualization 2.4 Generalized gradient optimization This algorithm is implemented in the function ggradandr. = load(’mars’).7).0421 gauss 6 gauss 4 0.0.0503 1 0 gauss 5 0.5) is the desired upper bound on the probability of misclassiﬁcation. The algorithm converges in ﬁnite number of iterations if the given εsolution exists.0503 −1 −2 −3 −4 −5 −4 −3 −2 −1 0 1 2 3 4 5 Figure 2. Example: εsolution of the Generalized Anderson’s task The εsolution of the generalized Anderson’s task is solved for the data mars. 0.
be used. The algorithm repeats the following two steps: I.e. i. Example: Solving the Generalized Anderson’s task using the generalized gradient optimization The generalized Anderson’s task is solved for the data mars..0432 −4 −5 −4 −3 −2 −1 0 1 2 3 4 5 Figure 2.7: εsolution of the Generalized Anderson’s task with maximal 0. the generalized gradient optimization method can be applied. However.5 4 3 gauss 2 0. b(t) ) > ε . II. The algorithm iterates until the minimal change in the objective function is less than the prescribed ε.0483 gauss 5 0 −1 gauss 6 −2 gauss 4 0.0516 −3 0.0516 gauss 3 2 1 0. The generalized gradient (∆w. The found linear classiﬁer as well as the input Gaussian distributions are visualized (see Figure 2.8). is violated. b(t+1) ) − F (w (t) . the condition F (w(t+1) . The maximal number of iterations can be used as the stopping condition as well. The optimized parameters are updated w(t+1) : = w (t) + ∆w . Both the ﬁrst and second class are described by three 2Ddimensional Gaussian distributions.06 probability of misclassiﬁcation.0473 0.0496 gauss 1 0. The maximal number of iterations is limited to tmax=10000. 23 . b(t+1) = b(t) + ∆b .mat. ∆b) is computed.
% load Gaussians % max # of iterations % training % visualization 5 4 0. options = struct(’tmax’. distrib ). model = ggradandr(distrib.distrib = load(’mars’).10000). pandr( model.0434 −4 −5 −4 −3 −2 −1 0 1 2 3 4 5 Figure 2.0498 0.8: Solution of the Generalized Anderson’s task after 10000 iterations of the generalized gradient optimization algorithm. 24 .0475 1 0 gauss 5 0. figure.0518 gauss 2 gauss 3 3 2 gauss 1 0.0518 −1 0.0485 gauss 6 gauss 4 −2 −3 0.options).
1). . .1) is a representative of the unsupervised learning method which yields the linear projection (3.2) . the input vectors x ∈ Rn are mapped into a new higher dimensional feature space F in which the linear methods are applied. (3. The linear data projection is implemented in function linproj (See examples in Section 3. i = 1. The kernel functions allow for nonlinear extensions of the linear feature extraction methods. The quadratic data mapping is implemented in function qmap. . x2 x2 .3) (3. The quadratic data mapping projects the input vector x ∈ Rn onto the feature space Rm where m = n(n + 3)/2. x2 xn . x1 x1 . . This yields a nonlinear (kernel) projection of data which has generally deﬁned as z = AT k(x) + b . . . The linear projection z = WT x + b . The data type used to describe the linear data projection is deﬁned in Table 3. x1 xn . . . . .2). The Linear Discriminant Analysis (LDA) (Section 3. x1 x2 . where xi . The Principal Component Analysis (PCA) (Section 3. The quadratic mapping φ: Rn → Rm is deﬁned as φ(x) = [ x1 . The projection is given by the matrix W [n × m] and bias vector b ∈ Rm .Chapter 3 Feature extraction The feature extraction step consists of mapping the input vector of observations x ∈ Rn onto a new feature description z ∈ Rm which is more suitable for given task. . xn xn ]T . . In this case. . . . x2 . .1 and Section 3. . 25 (3. . n are entries of the vector x ∈ Rn . .1) is commonly used. xn .2) is a representative of the supervised learning method yielding the linear projection.
lin2quad Merges linear rule and quadratic mapping. . lin2svm Merges linear rule and kernel projection. xl } be a set of training vectors from the ndimensional input space Rn . qmap Quadratic data mapping. .1 Table 3. kpcarec Reconstructs image after kernel PCA.2: Datatype used to describe linear data projection. lda Linear Discriminant Analysis.5). . The x ∈ Rn stands for the input vector and z ∈ Rm is the vector of extracted features. The summary of implemented methods for feature extraction is given in Table 3. . . x1 ). xl )]T is a vector of kernel functions centered in the training data or their subset.5) are representatives of the feature extraction methods yielding the kernel data projection (3. k(x.3). greedykpca Greedy Kernel Principal Component Analysis. The vector k(x) = [k(x. Table 3. kerproj Kernel data projection. 3. wm ]. . The extracted features are projections of φ(x) onto m vectors (or functions) from the feature space F . . . z l } is a lower dimensional representation of the 26 . The kernel data projection is implemented in the function kernelproj (See examples in Section 3. The Kernel PCA (Section 3.fun = ’linproj’ Identiﬁes function associated with this data type. . . .W [n × m] Projection matrix W = [w1 .b [m × 1] Bias vector b. .1 Principal Component Analysis Let TX = {x1 . pca Principal Component Analysis. . . The set of vectors TZ = {z 1 . Linear data projection (structure array): . .3) and the Generalized Discriminant Analysis (GDA) (Section 3. gda Generalized Discriminant Analysis.3 and Section 3. The data type used to describe the kernel data projection is deﬁned in Table 3. . .where A [l × m] is a parameter matrix and b ∈ Rm a parameter vector.1: Implemented methods for feature extraction linproj Linear data projection. . pcarec Computes reconstructed vector after PCA projection. demo pcacomp Demo on image compression using PCA. kpca Kernel Principal Component Analysis.
arg [1 × p] Kernel argument(s).b (3. . m are column vectors of the matrix W = [w1 . . .Table 3. Kernel data projection (structure array): . . b) = l l ˜ xi − xi i=1 2 .7) is the matrix W = [w 1 . The solution of the task (3. . The parameters (W. where w i . . The mean square reconstruction error 1 εM S (W. . 27 . .6) of the training data TX . . The vector b equals to WT µ. input training vectors TX in the mdimensional space Rm .6) is a function of the parameters of the linear projections (3. . (3. j . . i = 1.X [n × l] Training vectors {x1 . j) . . . .fun = ’kernelproj’ Identiﬁes function associated with this data type. .ker [string] Kernel identiﬁer.4).options. ∀i.options. . . .4) and (3. b) = argmin εM S (W . . The vectors TZ are obtained by the linear orthonormal projection z = WT x + b . xl }. . A demo showing the use of PCA for image compression is implemented in the script demo pcacomp. . The PCA is implemented in the function pca. .7) subject to wi · wj = δ(i. b) of the linear projection are the solution of the optimization task (W. The Principal Component Analysis (PCA) is the linear orthonormal projection (3. xl } are computed by the linear back ˜ projection ˜ x = W(z − b) . .3: Datatype used to describe kernel data projection. j) is the Kronecker delta function.sv. The reconstructed vectors TX = {x1 .Alpha [l × m] Parameter matrix A [l × m].4) where the matrix W [n × m] and the vector b [m × 1] are parameters of the projec˜ ˜ tion. W . . w m ] containing the m eigenvectors of the sample covariance matrix which have the largest eigen values. (3.b [m × 1] Bias vector b.5).5) obtained by inverting (3. (3. where µ is the sample mean of the training data. wm ] and δ(i. b ) . .4) which allows for the minimal mean square reconstruction error (3.
5 3 2. hold on.5]. Z = linproj(X. XR = pcarec(X. ’PCA subspace’). [dummy. ’Input vectors’. The PCA is applied to train the 1dimensional subspace to approximate the input data with the minimal reconstruction error. model = pca(X. Example: Principal Component Analysis The input data are synthetically generated from the 2dimensional Gaussian distribution. The training data. axis equal. reconstr.1: Example on the Principal Component Analysis.model).5 2 3 4 5 6 7 Input vectors Reconstructed PCA subspace 8 Figure 3. 7.1).’r’).mx] = max(Z). legend([h1 h2 h3].mx)].’Reconstructed’. . the reconstructed data and the 1dimensional subspace given by the ﬁrst principal component are visualized (see Figure 3. X = gsamp([5.5 5 4.5 7 6.References: The PCA is treated throughout for instance in books [13.. h3 = plot([XR(1.0.mn) XR(2. proj. data figure.mn) XR(1.’kx’).5 4 3. [dummy. 28 .’bo’).model). 6].5 6 5.[XR(2.1).8 1].100). % % % % generate data train PCA lower dim. h2 = ppatterns(XR..[1 0.mx)].8.mn] = min(Z). % visualization h1 = ppatterns(X. 3.
. (3.8) SB = y∈Y Iy (µy − µ)(µy − µ)T . The LDA and 29 . Sy = i∈Iy (xi − µi )(xi − µi )T . Iy  i∈I y The goal of the Linear Discriminant Analysis (LDA) is to train the linear data projection z = WT x . c} be labeled set of training vectors. y ∈ Y = {1. References: The LDA is treated for instance in books [3. .2 Linear Discriminant Analysis Let TXY = {(x1 .mat which consists of the labeled 4dimensional data. . y1). data Example 2: PCA versus LDA This example shows why the LDA is superior to the PCA when features good for classiﬁcation are to be extracted. .8). = ˜ det(SW ) det(SW ) ˜ ˜ is maximized. ppatterns(ext_data).2. yl )}. % % % % load input data train LDA feature extraction plot ext. 6]. The synthetical data are generated from the Gaussian mixture model. The SB is the betweenclass and SW is the withinclass scatter matrix of projected data which is deﬁned similarly to (3. The LDA and PCA are trained on the generated data. Example 1: Linear Discriminant Analysis The LDA is applied to extract 2 features from the Iris data set iris. model = lda(orig_data. figure. xi ∈ Rn .3. orig_data = load(’iris’).model). . y ∈ Y . . ext_data = linproj(orig_data. . . (xl . such that the class separability criterion F (W) = ˜ det(SB ) det(SB ) . The withinclass SW and betweenclass SB scatter matrices are described as SW = y∈Y Sy . The data after the feature extraction step are visualized in Figure 3. 2. where the total mean vector µ and the class mean vectors µy .2). i=1 µy = 1 xi . y ∈ Y are deﬁned as 1 µ= l l xi .
4] [4.2) = [1 0. lda_model = lda(data.1 0 −0.4 0.X.1) = [1 0.lda_model).5 Figure 3.pca_model).1 −0. figure.3b).:.1).Prior = [0. pca_data = linproj( data.5 0. axis equal. ppatterns(data).9 1].5].4 −0.3a). 0.:. distrib.5]]. pca_rec = pcarec(data.2 −0.Cov(:.2: Example shows 2dimensional data extracted from the originally 4dimensional Iris data set using Linear Discriminant Analysis.3 −0. distrib.5 1 1.pca_model). hold on.X. 30 % % % % % mean vectors 1st covariance 2nd covariance Gaussian weights sample data % train LDA % train PCA % visualization . lda_rec = pcarec(data.5 0 0. % Generate data distrib.1). The extracted data using the LDA and PCA model are displayed with the Gaussians ﬁtted by the MaximumLikelihood (see Figure 3.9 1]. PCA directions are displayed as connection of the vectors projected to the trained LDA and PCA subspaces and reprojected back to the input space (see Figure 3.5 −1 −0. distrib. data = gmmsamp(distrib.lda_model).9.Mean = [[5.X. 0.0.2 0.5 0.250).9.Cov(:. pca_model = pca(data.3 0. lda_data = linproj(data.5 −1.
3 Kernel Principal Component Analysis The Kernel Principal Component Analysis (Kernel PCA) is the nonlinear extension of the ordinary linear PCA.lda_rec(2. such that the reconstruction error 1 εKM S (A. figure. The reconstructed vector φ(x) is given as a linear combination of the mapped data TΦ l ˜ φ(x) = i=1 βi φ(xi ) . The computation of the principal components and the projection on these components can be expressed in terms of dot products thus the kernel functions k: X × X → R can be employed.:).1. . . .1. 3. φ(xl )}. pgauss(mlcgmm(lda_data)). xi ∈ X ⊆ Rn are mapped by φ: X → F to a high dimensional feature space F . b) = l l (3. The problem is to ﬁnd the vector x ∈ X its image ˜ φ(x) ∈ F well approximates the reconstructed vector φ(x) ∈ F . subplot(2.’r’).2). 31 .:).’b’).9).10) ˜ is minimized. . The input training vectors TX = {x1 . xl }. . ppatterns(lda_data).pca_rec(2. hold on.11) In contrast to the linear PCA. pgauss(mlcgmm(pca_data)). (3. β = A(z − b) . Project input vector xin ∈ Rn onto its lower dimensional representation z ∈ Rm using (3.h1 = plot(lda_rec(1.:). This procedure is implemented in the function kpcarec for the case of the Radial Basis Function (RBF) kernel. (3. subplot(2. ppatterns(pca_data). The procedure consists of the following step: I. . .1).’LDA direction’. The kernel PCA trains the kernel data projection z = AT k(x) + b .:). .9) ˜ φ(xi ) − φ(xi ) i=1 2 . The linear PCA is applied on the mapped data TΦ = {φ(x1 ).’PCA direction’). title(’PCA’). legend([h1 h2]. h2 = plot(pca_rec(1. title(’LDA’). the explicit projection from the feature space F to the input space X usually do not exist.
. The kernel PCA model with RBF kernel is trained.11) while the size d of the subset TS is kept small. sl )]T are the kernel functions centered in the vectors TS = {s1 .II.new_dim = 2. The result is visualized in Figure 3. k(x.’Input vectors’. The vector set TS is a subset of training data TX .arg = 4. Example: Kernel Principal Component Analysis The example shows using the kernel PCA for data denoising. . 30].4. In contrast to the ordinary kernel PCA. ’+r’). The objective of the Greedy kernel PCA is to minimize the reconstruction error (3.model). h2 = ppatterns(XR.ker = ’rbf’. The input data are synthetically generated 2dimensional vectors which lie on a circle and are corrupted by the Gaussian noise.250. % Visualization h1 = ppatterns(X). sd }. (3. .9). % compute kernel PCA XR = kpcarec(X. .1). 3.4 Greedy kernel PCA algorithm The Greedy Kernel Principal Analysis (Greedy kernel PCA) is an eﬃcient algorithm to compute the ordinary kernel PCA. % use RBF kernel options.11). such that ˜ xout = argmin φ(x) − φ(xin ) 2 . .5.1]. 32 . legend([h1 h2]. x References: The kernel PCA is described at length in book [29.12) where A [d × m] is the parameter matrix. The goal is to train the kernel data projection z = AT kS (x) + b . xl }.’Reconstructed’). X = gencircledata([1. The function kpcarec is used to ﬁnd preimages of the reconstructed data (3. . % generate circle data options. % kernel argument options. s1 ). Let TX = {x1 . . .options). . . . The Greedy kernel PCA algorithm is implemented in the function greedykpca. b [m × 1] is the parameter vector and kS = [k(x. Computes vector xout ∈ Rn which is good preimage of the reconstructed vector ˜ φ(xin ).12) is reduced compared to (3. % output dimension model = kpca(X. xi ∈ Rn be the set of input training vectors. . % compute reconstruced data figure. the subset TS does not contain all the training vectors TX thus the complexity of the projection (3.
.1). . y1 ).model). The computation of the projection vectors and the projection on these vectors can be expressed in terms of dot products thus the kernel functions k: X × X → R can be employed. % reconstructed vectors figure. In contrast the ordinary kernel PCA (see experiment in Section 3. subset run greedy algorithm XR = kpcarec(X. . X = gencircledata([1. . The function kpcarec is used to ﬁnd preimages of the reconstructed data obtained by the kernel PCA projection. . The number of vectors deﬁning the kernel projection (3. .ker = ’rbf’.5 Generalized Discriminant Analysis The Generalized Discriminant Analysis (GDA) is the nonlinear extension of the ordinary Linear Discriminant Analysis (LDA). . y1). c} are mapped by φ: X → F to a high dimensional feature space F .5. options. model = greedykpca(X. h3 = ppatterns(model.’Selected subset’). The input data are synthetically generated 2dimensional vectors which lie on a circle and are corrupted with the Gaussian noise. yl )}. 33 . The Greedy kernel PCA model with RBF kernel is trained.options). % % % % % use RBF kernel kernel argument output dimension size of sel.12) is limited to 25. % visualization h1 = ppatterns(X). (φ(xl ). legend([h1 h2 h3]. y ∈ Y = {1. (xl . % generate training data options. options. . Example: Greedy Kernel Principal Component Analysis The example shows using kernel PCA for data denoising. 3.new_dim = 2.m = 25.References: The implemented Greedy kernel PCA algorithm is described in [11]..250. The resulting kernel data projection is deﬁned as z = AT k(x) + b . . h2 = ppatterns(XR. xi ∈ X ⊆ Rn .’Reconstructed set’.1].’ob’. . The input training data TXY = {(x1 .X. . yl )}. .12). The vector of the selected subset TX are visualized as well.sv. ’Training set’..3) requires all the 250 training vectors. .’+r’). options.5. The training and reconstructed vectors are visualized in Figure 3.arg = 4. The ordinary LDA is applied on the mapped data TΦY = {(φ(x1 ).
. . .model). .options). y1 ). Let φ: X → Rm be a mapping φ(x) = [φ1 (x). . . (xl . yl )} be a training set of observations xi ∈ X ⊆ Rn and corresponding hidden states yi ∈ Y = {1.6. Let f (x) = w · φ(x) + b be a linear discriminant function found by an arbitrary training method on the extracted data TΦY . .where A [l × m] is the matrix and b [m × 1] the vector trained by the GDA. (z l . y1 ). Notice. The feature extraction step applied to the input data TXY yields extracted data TΦY = {(φ(x1 ). The parameters (A. model = gda(orig_data.g. . . that setting σ too low leads to the perfect separability of the training data TXY but the kernel projection is heavily overﬁtted. options.. visualization 3. . The GDA is implemented in the function gda.ker = ’rbf’. as m f (x) = w · φ(x) + b = i=1 wiφi (x) + b . The function f can be seen as a nonlinear function of the input vectors x ∈ X . . ext_data = kernelproj(orig_data. .new_dim = 2. .6 Combining feature extraction and linear classiﬁer Let TXY = {(x1 . yl )}. orig_data = load(’iris’). The RBF kernel with kernel width σ = 1 is used. . . y1 ). quadratic data mapping or kernel projection). b) are trained to increase the betweenclass scatter and decrease the withinclass scatter of the extracted data TZY = {(z 1 . options.mat which contains 3 classes of 4dimensional data. yl )}.2) used as the feature extraction step.arg = 1. φm (x)]T deﬁning a nonlinear feature extraction step (e.13) . . c}. options. z i ∈ Rm . . . (φ(xl ). ppatterns(ext_data). figure. Example: Generalized Discriminant Analysis The GDA is trained on the Iris data set iris. the discriminant function can be written in the following form f (x) = x · Ax + x · b + c . . % % % % % % % load data use RBF kernel kernel arg. References: The GDA was published in the paper [2]. . output dimension train GDA feature ext. 34 (3. In the case of the quadratic data mapping (3. The extracted data YZY are visualized in Figure 3.
Figure 3. Function lin2quad is used to compute the parameters of the quadratic classiﬁer explicitly based on the found linear rule in the feature space and the mapping.7 shows the decision boundary of the quadratic classiﬁer. % load training data living in the input space input_data = load(’vltava’). The input training data are linearly nonseparable but become separable after the quadratic data mapping is applied. pboundary(quad_model).14) The discriminant function (3. % map data to the feature space map_data = qmap(input_data). The Fisher Linear Discriminant was used in this example. ppatterns(input_data). The Perceptron algorithm produces the linear rule only. The combining is implemented in function lin2svm. the resulting discriminant function has the following form f (x) = α · k(x) + b . Example: Kernel classiﬁer from the linear rule This example shows how to train the kernel classiﬁer using arbitrary training algorithm for linear rule.3) applied as the feature extraction step.4. % train linear rule in the feature sapce lin_model = perceptron(map_data). perceptron. Example: Quadratic classiﬁer trained the Perceptron This example shows how to train the quadratic classiﬁer using the Perceptron algorithm. (3.13) deﬁnes the quadratic classiﬁer described in Section 6. % visualize the quadratic classifier figure. Therefore the quadratic classiﬁer can be trained by (i) applying the quadratic data mapping qmap. (ii) training a linear rule on the extracted data and (iii) combing the kernel projection and the linear rule. The 35 . (ii) training a linear rule on the mapped data (e. The combining is implement in function lin2quad.. This classiﬁer can be trained by (i) applying the kernel projection kernelproj. % compute parameters of the quadratic classifier quad_model = lin2quad(lin_model).14) deﬁnes the kernel (SVM type) classiﬁer described in Section 5.g. In the case of the kernel projection (3. fld) and (iii) combining the quadratic mapping and the linear rule.The discriminant function (3.
’arg’.X.tst. % evaluation on testing data tst = load(’riply_tst’).sv. the training data and vectors used in the kernel expansion of the discriminant function. % train linear classifier lin_model = fld(map_trn).X.10). pboundary(kfd_model). % visualization figure. Figure 3. ypred = svmclass(tst. The Riply’s data set was used for training and evaluation. % train kernel PCA by greedy algorithm options = struct(’ker’.’ok’.8 shows the found kernel classiﬁer.lin_model).1.X.options).’new_dim’.0970 36 . kpca_model = greedykpca(trn. % combine linear rule and kernel PCA kfd_model = lin2svm(kpca_model. cerror(ypred. % load input data trn = load(’riply_trn’).13).kfd_model).kpca_model).greedy kernel PCA algorithm was applied as the feature extraction step. ppatterns(kfd_model.’rbf’. ppatterns(trn).y) ans = 0. % project data map_trn = kernelproj(trn.
7
LDA direction PCA direction
6
5
4
3
2
1
0
1
2
3
4
5
6
7
8
(a)
LDA 1.4 1.2 1 0.8 0.6 0.4 0.2 0 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
PCA 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 −4 −3 −2 −1 0 1 2 3 4 5
(b) Figure 3.3: Comparison between PCA and LDA. Figure (a) shows input data and found LDA (red) and PCA (blue) directions onto which the data are projected. Figure (b) shows the class conditional Gaussians estimated from data projected data onto LDA and PCA directions.
37
10 Input vectors Reconstructed 8
6
4
2
0
−2
−4
−6
−8 −8
−6
−4
−2
0
2
4
6
8
10
Figure 3.4: Example on the Kernel Principal Component Analysis.
8 Training set Reconstructed set Selected set 6
4
2
0
−2
−4
−6 −8
−6
−4
−2
0
2
4
6
8
10
Figure 3.5: Example on the Greedy Kernel Principal Component Analysis.
38
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
−0.2 −0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
Figure 3.6: Example shows 2dimensional data extracted from the originally 4dimensional Iris data set using the Generalized Discriminant Analysis. In contrast to LDA (see Figure 3.2 for comparison), the class clusters are more compact but at the expense of a higher risk of overﬁtting.
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6 −0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Figure 3.7: Example shows quadratic classiﬁer found by the Perceptron algorithm on the data mapped to the feature by the quadratic mapping.
39
40 .2 0 −0.2 1 0.5 −1 −0.2 −1.8 0. The greedy kernel PCA was used therefore only a subset of training data was used to determine the kernel projection.1.5 0 0.5 1 Figure 3.8: Example shows the kernel classiﬁer found by the Fisher Linear Discriminant algorithm on the data mapped to the feature by the kernel PCA.6 0.4 0.
fun (see Table 4. . . each pair (µi .3).4. The summary of implemented methods for dealing with probability density estimation and clustering is given in Table 4. . Σi) is assigned to the class yi ∈ Y. . The ﬁeld .e. 2 det(Σ) (4. i. Let the structure model represent a given distribution function.1. The datatype is described in Table 4.. however. the shape of the matrix can be indicated in the optimal ﬁeld .1 4. The covariance matrix Σ is always represented as the square matrix [n × n]. x ∈ Rn as a structure array which must contain ﬁeld . .model. The STPRtool uses speciﬁc datatype to describe a set of Gaussians which can be labeled.fun) is used to evaluate yi = PX (xi ).Chapter 4 Density estimation and clustering The STPRtool represents a probability distribution function PX (x). The Matlab command y = feval(X. xl }.2).1 Gaussian distribution and Gaussian mixture model The value of the multivariate Gaussian probability distribution function in the vector x ∈ Rn is given by PX (x) = 1 (2π) n 2 1 exp(− (x − µ)T Σ−1 (x − µ)) . The vector y [1 × l] contains values of the distribution function. . xl } of a random vector.fun speciﬁes function which is used to evaluate PX (x) for given set of realization {x1 .1) where µ [n × 1] is the mean vector and Σ [n × n] is the covariance matrix. The matrix X [n × l] contains column vectors {x1 . . The evaluation of the Gaussian distribution is implemented in the function 41 . In particular. .cov type (see Table 4. the Gaussian distribution and the Gaussian mixture models are described in Section 4.
rsde Reduced Set Density Estimator. The sampling from univariate and bivariate Gaussians is shown. y ∈ Y = {1. The Gaussian Mixture Model (GMM) is weighted sum of the Gaussian distributions PX (x) = y∈Y PY (y)PXY (xy) . mlsigmoid Fitting a sigmoid function using ML estimation. pdfgauss Evaluates for multivariate Gaussian distribution. c} are weights and PXY (xy) is the Gaussian distribution (4. pdfgauss. pdfgauss Evaluates Gaussian mixture model. Probability Distribution Function (structure array): . The distribution function of the univariate Gaussian is visualized and compared to the histogram 42 .1) given by parameters (µy . The random sampling from the Gaussian distribution is implemented in the function gsamp.1: Implemented methods: Density estimation and clustering gmmsamp Generates sample from Gaussian mixture model. emgmm ExpectationMaximization Algorithm for Gaussian mixture model. . Example: Sampling from the Gaussian distribution The example shows how to represent the Gaussian distribution and how to sample data it. gsamp Generates sample from Gaussian distribution. mmgauss Minimax estimation of Gaussian distribution. The random sampling from the GMM is implemented in the function gmmsamp.fun [string] The name of a function which evaluates probability distribution y = feval( X. (4. kmeans Kmeans clustering algorithm. Σy ).2: Datatype used to represent probability distribution function. demo svmpout Fitting a posteriori probability to SVM output. melgmm Maximizes Expectation of LogLikelihood for Gaussian mixture.2) where PY (y). demo emgmm Demo on ExpectationMaximization (EM) algorithm. .fun ). demo mmgauss Demo on minimax estimation for Gaussian. Table 4. model. mlcgmm Maximal Likelihood estimation of Gaussian mixture model. The datatype used to represent the GMM is described in Table 4. sigmoid Evaluates sigmoid function. The evaluation of the GMM is implemented in the function pdfgmm. .Table 4. .5.
3: Shape of covariance matrix. .1:5]. . % plot pdf.y [1 × k] Labels {y1 .pdfgauss([4:0. Σk }.model).Mean = [1. The isoline of the distribution function of the bivariate Gaussian is visualized together with the sample data (see Figure 4.Mean [n × k] Mean vectors {µ1 . Table 4.4: Datatype used to represent the labeled set of Gaussians. % Mean vector % Covariance matrix % plot sampled data % plot shape 4. .1:5]. . . yk } assigned to Gaussians. ’spherical’ Spherical covariance matrix. . pgauss(model). θ)PY Θ (yθ) . . .6. .’Cov’. µk }. . % compute histogram bar(X. % univariate case model = struct(’Mean’.2). ’diag’ Diagonal covariance matrix. Parameter cov type: ’full’ Full covariance matrix.Y/500). .Cov = [1 0. hold on. . .Table 4. 0.1].X] = hist(gsamp(model.10). .6 1].2 MaximumLikelihood estimation of GMM PXY Θ (x.250)). % plot histogram % bivariate case model. ppatterns(gsamp(model. 43 Let the probability distribution .3). . model.1a). of the sample data (see Figure 4. of the Gaussisan plot([4:0. Labeled set of Gaussians (structure array): . .500).1. yθ) = PXY Θ (xy. [Y.Cov [n × n × k] Covariance matrices {Σ1 .’r’). hold on.1b).fun = ’pdfgauss’ Identiﬁes function associated with this data type.cov type [string] Optional ﬁeld which indicates shape of the covariance matrix (see Table 4. % Gaussian parameters figure. figure.
35 5 4 0. . . . y1). .2. . The logarithm of probability P (TXY θ) is referred to as the loglikelihood L(θTXY ) of θ with respect to TXY .3 3 0.Mean [n × c] Mean vectors {µ1 .fun = ’pdfgmm’ Identiﬁes function associated with this data type. The conditional probability distribution PXY Θ (xy. . Σy ). yi) are assumed to samples from random variables which are independent and identically distributed (i. 0. µc }. The vector of observations x ∈ Rn and the hidden state y ∈ Y = {1. . PK (c)}.5: Datatype used to represent the Gaussian Mixture Model. θ) is assumed to be Gaussian distribution.) according to PXY Θ (x. c} are assumed to be realizations of random variables which are independent and identically distributed according to the PXY Θ (x. . . .2. (xl .1: Sampling from the univariate (a) and bivariate (b) Gaussian distribution. .1 and Section 4. y ∈ Y of Gaussian components and the values of the discrete distribution PY Θ (yθ). The MaximumLikelihood estimation of the parameters θ of the GMM for the case of complete and incomplete data is described in Section 4.2 1 0. .1 −1 0. . respectively. y ∈ Y. be known up to parameters θ ∈ Θ. The pairs (xi . .Table 4. .i.2. yθ). . Gaussian Mixture Model (structure array): .Cov [n × n × c] Covariance matrices {Σ1 . . .05 −2 0 −4 −3 −2 −1 0 1 2 3 4 5 6 −3 −2 −1 0 1 2 3 4 5 (a) (b) Figure 4. . The θ involves parameters (µy . yθ). . yl )} which contains both vectors of observations xi ∈ Rn and corresponding hidden states yi ∈ Y.15 gauss 1 0.25 2 0.127 0 0. Σc }. .d. The marginal probability PXΘ (xθ) is the Gaussian mixture model (4. Thanks to the 44 . .2).1 Complete data The input is a set TXY = {(x1 .2. . 4.Prior [c × 1] Weights of Gaussians {PK (1).
% load labeled (complete) data model = mlcgmm(data). pgauss(model). assumption the loglikelihood can be factorized as l L(θTX ) = log P (TX θ) = i=1 y∈Y PXY Θ (xi yi. Thanks to the i.i.2a) and (ii) contours of the distribution function (see Figure 4. hold on. Example: MaximumLikelihood estimate from complete data The parameters of the Gaussian mixture model is estimated from the complete Riply’s data set riply trn.2 Incomplete data The input is a set TX = {x1 . assumption the loglikelihood can be factorized as l L(θTXY ) = log P (TXY θ) = i=1 log PXY Θ (xi .2. .d. 45 .i.struct(’visual’.2b) . yiθ) . yiθ) . θ)PY Θ (yi θ) . The computation of the ML estimate is implemented in the function mlcgmm. The estimated GMM is visualized as: (i) shape of components of the GMM (see Figure 4. ppatterns(data). pgmm(model.3) The ML estimate of the parameters θ of the GMM from the complete data set TXY has an analytical solution. figure. . The data are binary labeled thus the GMM has two Gaussians components. .i.d. . data = load(’riply_trn’). The logarithm of probability P (TX θ) is referred to as the loglikelihood L(θTX ) of θ with respect to TX . The problem of MaximumLikelihood (ML) estimation from the complete data TXY is deﬁned as l θ = argmax L(θTXY ) = argmax θ∈Θ θ∈Θ i=1 ∗ log PXY Θ (xi . hold on. ppatterns(data). xl } which contains only vectors of observations xi ∈ Rn without knowledge about the corresponding hidden states yi ∈ Y. (4. % ML estimate of GMM % visualization figure. 4.’contour’)).
A nice description can be found in book [3]. The EM algorithm builds a sequence of parameter estimates θ (0) . The EM algorithm for the GMM is implemented in the function emgmm. such that the loglikelihood L(θ (t) TX ) monotonically increases. The numerical optimization has to be used instead.Cov = [1 0. . L(θ (0) TX ) < L(θ (1) TX ) < . References: The EM algorithm (including convergence proofs) is described in book [26]. .4 0. The EM algorithm is a local optimization technique thus the estimated parameters depend on the initial guess θ (0) .Mean = [2 2]. The random initialization is used by default. The ﬁrst papers describing the EM are [25. sampled data and the estimate model are visualized in Figure 4. (4. . The ground truth model is the GMM with two univariate Gaussian components.e. The EM algorithm is applied to estimate the GMM parameters. 46 .5) are commonly used to select the initial θ (0) . < L(θ (t) TX ) until a stationary point L(θ (t−1) TX ) = L(θ (t) TX ) is achieved. The random guess or Kmeans algorithm (see Section 4. The plot of the loglikelihood function L(θ (t) TX ) with respect to the number of iterations t is visualized in Figure 4. Example: MaximumLikelihood estimate from incomplete data The estimation of the parameters of the GMM is demonstrated on a synthetic data. % ground truth Gaussian mixture model true_model.5]. .4) The ML estimate of the parameters θ of the GMM from the incomplete data set TXY does not have an analytical solution. θ)PY Θ (yi θ) . 5].The problem of MaximumLikelihood (ML) estimation from the incomplete data TX is deﬁned as l θ = argmax L(θTX ) = argmax θ∈Θ θ∈Θ i=1 y∈Y ∗ PXY Θ (xi yi. true_model. true_model. θ (t) . . i. The monograph [18] is entirely devoted to the EM algorithm. θ (1) . sample = gmmsamp(true_model.3a.Prior = [0.3b. The ground truth model.6].. An interactive demo on the EM algorithm for the GMM is implemented in demo emgmm. The ground truth model is sampled to obtain a training (incomplete) data. . The ExpectationMaximization (EM) algorithm is an iterative procedure for ML estimation from the incomplete data. 250).
% number of Gaussian components options. % display progress info estimated_model = emgmm(sample. .6) holds for ε > 0 which deﬁnes closeness to the optimal solution θ ∗ .’b’)).logL ). h2 = pgmm(estimated_model. References: The minimax algorithm for probability estimation is described and analyzed in Chapter 8 of the book [26]. h1 = pgmm(true_model. % visualization figure.ncomp = 2. ylabel(’loglikelihood’).1). The minimax estimation is deﬁned as θ ∗ = argmax min log PXΘ (xθ) .3 Minimax estimation The input is a set TX = {x1 .6) can be evaluated eﬃciently thus it is a natural choice for the stopping condition of the algorithm. An interactive demo on the minimax estimation for the Gaussian distribution is implemented in the function demo mmgauss. The algorithm converges in a ﬁnite number of iterations to such an estimate θ that the condition x∈TX min log PXΘ (xθ ∗ ) − min log PXΘ (xθ) < ε . The vectors TX are assumed to be a realizations with high value of the distribution function PXΘ (xθ).X). . . .5) If a procedure for the MaximumLikelihood estimate (global solution) of the distribution PXΘ (xθ) exists then a numerical iterative algorithm which converges to the optimal estimate θ ∗ can be constructed. . ppatterns(sample. 4. xl } contains vectors xi ∈ Rn which are realizations of random variable distributed according to PXΘ (xθ).% ML estimation by EM options. The STPRtool contains an implementation of the minimax algorithm for the Gaussian distribution (4.’r’)).’Ground truth’. The algorithm builds a series of estimates θ (0) . θ (t) which converges to the optimal estimate θ ∗ . θ (1) .options). . . hold on. The inequality (4. x∈TX (4. θ∈Θ x∈TX (4. xlabel(’iterations’). The algorithm is implemented in the function mmgauss. 47 .X.struct(’color’. .struct(’color’. legend([h1(1) h2(1)]. plot( estimated_model. ’ML estimation’).verb = 1. The PXΘ (xθ) is known up to the parameter θ ∈ Θ. figure.
y1 ). yl )} is the training set composed of the vectors xi ∈ X ⊆ Rn and the corresponding binary hidden states yi ∈ Y = {1. (xl . The loglikelihood function is deﬁned as l L(θTXY ) = i=1 log PY F Θ (yif (xi ).0] [0. model = mmgauss(X). Let TXY = {(x1 . Perceptron). The aim is to estimate parameters of a posteriori distribution PY F Θ (yf (x). % run minimax algorithm % visualization figure. 4.0] [1.13). Let f : X ⊆ Rn → R be a discriminant function trained from the data TXY by an arbitrary learning method. for instance. the discriminant function of an arbitrary classiﬁer does not have meaning of probability (e. struct(’p’..exp(model. . X = [[0. θ) = 1 . The training set TXY is assumed to be identically and independently distributed from underlying distribution.lower_bound))).7) 48 . . the probabilistic output of the classiﬁer can help in postprocessing. (4.g.d.f. pgauss(model. The found solution is visualized as the isoline of the distribution function at the point PXΘ (xθ) (see Figure 4. a2 ]T can be estimated by the maximumlikelihood method l θ = argmax L(θ TXY ) = argmax θ θ i=1 log PY F Θ (yi f (xi ). The parameters θ = [a1 .1]]. θ) . in combining more classiﬁers together.’xr’.4 Probabilistic output of classiﬁer In general. However. . In this particular case. % points with high p. ppatterns(X.4). SVM. the minimax problem is equivalent to searching for the minimal enclosing ellipsoid. The distribution is modeled by a sigmoid function PY F Θ (1f (x). . Fitting a sigmoid function to the classiﬁer output is a way how to solve this problem. 1 + exp(a1 f (x) + a2 ) which is determined by parameters θ = [a1 . a2 ]T . θ ) .Example: Minimax estimation of the Gaussian distribution The bivariate Gaussian distribution is described by three samples which are known to have high value of the distribution function. θ) of the hidden state y given the value of the discriminant function f (x). 2}. The minimax algorithm is applied to estimate the parameters of the Gaussian.
References: The method of ﬁtting the sigmoid function is described in [22].’C’. The GMM allows to compute the a posteriori probability for comparison to the sigmoid estimate. Sigmoid function (structure array): .X] = svmclass(data.svm_output. Figure 4. svm_output. % compute SVM output [dummy.y = data.X.6: Datatype used to describe sigmoid function.6). % train SVM options = struct(’ker’.options).fun = ’sigmoid’ Identiﬁes function associated with this data type. svm_model = smo(data. Figure 4. Example: Probabilistic output for SVM This example is implemented in the script demo svmpout. % fit sigmoid function to svm output sigmoid_model = mlsigmoid(svm_output). % load training data data = load(’riply_trn’). The example is implemented in the script demo svmpout.’rbf’.10). % ML estimation of GMM model of SVM output gmm_model = mlcgmm(svm_output). The function produces the parameters θ = [a1 .5b shows the decision boundary of SVM classiﬁer for illustration. hold on. The Riply’s data set riply trn is used for training the SVM and the ML estimation of the sigmoid.The STPRtool provides an implementation mlsigmoid which uses the Matlab Optimization toolbox function fminunc to solve the task (4. a2 ]T which are stored in the datatype describing the sigmoid (see Table 4.7).1.A [2 × 1] Parameters θ = [a1 . The evaluation of the sigmoid is implemented in function sigmoid. The ML estimated of Gaussian mixture model (GMM) of the SVM output is used for comparison. . 49 . a2 ]T of the sigmoid function.5a shows found probabilistic models of a posteriori probability of the SVM output.y. Table 4. The example shows how to train the probabilistic output for the SVM classiﬁer.svm_model). % visulization figure.’arg’.
l i=1 y=1. Let the objective function F (θ) be deﬁned as l 1 F (θ) = min µy − xi 2 . hcomp = pgauss(gmm_model).X).5 Kmeans clustering The input is a set TX = {x1 ..gmm_apost. . µc }..sigmoid_model).Prior(2))).. % SVM decision boundary figure. ’P(y=1f(x)) MLGMM’. hgmm = plot(fx... . psvm(svm_model)./. ppatterns(svm_output).. (pcond(1.Prior(1)+(pcond(2. 200). i..c l i=1 θ θ The Kmeans1 is an iterative procedure which iteratively builds a series θ (0) .. θ (1) .. % GMM model pcond = pdfgauss( fx.Prior(1)).xlabel(’svm output f(x)’). θ = argmin F (θ) = argmin y=1.’g’). The Kmeans algorithm converges in a ﬁnite number of steps to the local optimum of the objective function F (θ).. . ppatterns(data). xl } which contains vectors xi ∈ Rn .:)*gmm_model.:)*gmm_model. .. 50 .. legend([hsigmoid. . Let θ = {µ1 . . the task is to solve l 1 ∗ min µy − xi 2 ..e. . gmm_apost = (pcond(1.’k’).sigmoid_apost. . hsigmoid = plot(fx. . θ (t) .’P(y=1f(x)) MLSigmoid’.’P(f(x)y=1) MLGMM’.:)*gmm_model. 4. max(svm_output.hgmm. The found solution depends on the initial guess θ (0) which 1 The K in the name Kmeans stands for the number of centers which is here denoted by c. sigmoid_apost = sigmoid(fx.. . . The task is to ﬁnd such centers θ which describe the set TX best.hcomp]. . µi ∈ Rn be a set of unknown cluster centers. gmm_model). ylabel(’p(y=1f(x))’).c The small value of F (θ) indicates that the centers θ well describe the input set TX .’P(f(x)y=2) MLGMM’). % sigmoid model fx = linspace(min(svm_output.X).
X.is usually selected randomly. hold on. pboundary( model ). 3].y] = kmeans(data. % load data [model. ppatterns(model. 51 . The vectors classiﬁed to the corresponding centers are distinguished by color (see Figure 4. xlabel(’t’). Example: Kmeans clustering The Kmeans algorithm is used to ﬁnd 4 clusters in the Riply’s training set. 6. 4 ).data.MsErr). data = load(’riply_trn’).14).6b.6a). ylabel(’F’). The plot of the objective function F (θ (t) ) with respect to the number of iterations is shown in Figure 4. The Kmeans algorithm is implemented in the function kmeans. plot(model. figure. % run kmeans % visualization figure.X. References: The Kmeans algorithm is described in the books [26.’sk’. ppatterns(data). The solution is visualized as the cluster centers.
5 0 0.45 0.6 0. 52 .8 0.5 −1 −0.5 1 (a) 1.2 −1.1.8 gauss 2 0.2: Example of the MaximumLikelihood of the Gaussian Mixture Model from the incomplete data.4 0.6 1.5 0 0.2 0 −0. Figure (a) shows components of the GMM and Figure (b) shows contour plot of the distribution function.5 1 (b) Figure 4.2 0 −0.5 −1 −0.4 gauss 1 0.2 −1.981 0.2 1 0.2 1 0.
15 0.25 0.05 0 −5 −4 −3 −2 −1 0 1 2 3 4 5 (a) −442 −444 −446 −448 log−likelihood −450 −452 −454 −456 −458 −460 0 5 10 iterations 15 20 25 (b) Figure 4.3 0.35 0.1 0.3: Example of using the EM algorithm for ML estimation of the Gaussian mixture model. Figure (a) shows ground truth and the estimated GMM and Figure(b) shows the loglikelihood with respect to the number of iterations of the EM. 53 .2 0.4 Ground truth ML estimation 0.0.
2 −0.4 −0.304 0.2 0 −0.4 −0.8 1 Figure 4.6 0.4: Example of the Minimax estimation of the Gaussian distribution.4 gauss 1 0.2 0.4 0.8 0.1 0.6 0.2 0 0. 54 .
1 0 −4 −3 −2 −1 0 1 svm output f(x) 2 3 4 5 (a) 1.5 −1 −0.4 0.7 0.2 1 0.2 −1.1 P(y=1f(x)) ML−Sigmoid P(y=1f(x)) ML−GMM P(f(x)y=1) ML−GMM P(f(x)y=2) ML−GMM 0.2 0 −0.5 1 (b) Figure 4. Figure (a) shows the sigmoid model and GMM model both ﬁtted by ML estimated.3 0. 55 .6 p(y=1f(x)) 0.2 0.6 0.9 0. Figure (b) shows the corresponding SVM classiﬁer.5 0.5 0 0.8 0.5: Example of ﬁtting a posteriori probability to the SVM output.4 0.8 0.
56 .6 0.2 0 −0.4 0.04 0 2 4 6 t 8 10 12 (b) Figure 4.12 0.1 F 0.1.06 0.6: Example of using the Kmeans algorithm to cluster the Riply’s data set.2 1 0.14 0.16 0. Figure (a) shows found clusters and Figure (b) contains the plot of the objective function F (θ (t) ) with respect to the number of iterations.8 0.08 0.2 −1.5 0 0.5 1 (a) 0.5 −1 −0.
2 for f (x) < 0 . The multiclass classiﬁcation rule q: X → Y = {1. . 57 . . (5. s1 ).Chapter 5 Support vector and other kernel machines 5.2.1 Support vector classiﬁers The binary support vector classiﬁer uses the discriminant function f : X ⊆ Rn → R of the following form f (x) = α · kS (x) + b . . .3) q(x) = argmax fy (x) . y ∈ Y = {1.3. c} is deﬁned as (5. . k(x. (5. . The multiclass generalization involves a set of discriminant functions fy : X ⊆ Rn → R. . y∈Y The datatype used by the STPRtool to represent the multiclass SVM classiﬁer is described in Table 5. bc ]T be a vector of all biases. . Let the matrix A = [α1 . . . The binary classiﬁcation rule q: X → Y = {1. si ∈ Rn which are usually subset of the training data. y∈Y. .1) The kS (x) = [k(x. . 2. sd )]T is the vector of evaluations of kernel functions centered at the support vectors S = {s1 . The α ∈ Rl is a weight vector and b ∈ R is a bias. . . sd }. Both the binary and the multiclass SVM classiﬁers are implemented in the function svmclass. 2. . c} deﬁned as fy (x) = αy · kS (x) + by .2) The datatype used by the STPRtool to represent the binary SVM classiﬁer is described in Table 5. 2} is deﬁned as q(x) = 1 for f (x) ≥ 0 . . . . . . . . αc ] be composed of all weight vectors and b = [b1 . . .
The linear SVM aims to train the linear discriminant function f (x) = w · x + b of the binary classiﬁer q(x) = 1 for f (x) ≥ 0 . The training of the optimal parameters (w ∗ ..g The datatype used to represent the majority voting strategy for the multiclass SVM classiﬁer is described in Table 5. b ) = argmin w + C ξip . . g do vy (x) = vy (x) + 1 where y = qj (x).4. . The majority votes based multiclass classiﬁer assigns the input x into such class y ∈ Y having the majority of votes y = argmax vy (x) . g be a set of g binary 1 2 SVM rules (5. yl )} of training vectors xi ∈ Rn and corresponding hidden states yi ∈ Y = {1. 30] 5. . . . . 4. (xl . y1 ). end.. b∗ ) is transformed to the following quadratic programming task l 1 ∗ ∗ 2 (w . .1 References: Books describing Support Vector Machines are for example [31. w · xi + b ≥ +1 − ξi .The majority voting strategy is other commonly used method to implement the 1 2 multiclass SVM classiﬁer. . . The implemented methods to deal with the SVM and related kernel classiﬁers are enlisted in Table 5. For j = 1. ξi ≥ 0 . . j = 1. Let v(x) be a vector [c × 1] of votes for classes Y when the input x is to be classiﬁed. 58 . ∀y ∈ Y. . i ∈ I1 ∪ I2 . The jth rule qj (x) classiﬁes the inputs x into class yj ∈ Y or yj ∈ Y.2 Binary Support Vector Machines The input is a set TXY = {(x1 .4) y =1.4) is implemented in the function mvsvmclass. (5. The vector v(x) = [v1 (x). . vc (x)]T is computed as: Set vy = 0. yj }.b i=1 i ∈ I1 . (5.. . Let qj : X ⊆ Rn → {yj .2). 2 for f (x) < 0 .. . 2}. w · xi + b ≤ −1 + ξi .5) 2 w. The strategy (5. . . i ∈ I2 .
svmlight Interface to SV M light software. oaasvm Multiclass SVM using OneAgainstAll decomposition. mvsvmclass Majority voting multiclass SVM classiﬁer. oaosvm Multiclass SVM using OneAgainstOne decomposition. rsrbf Reduced Set Method for RBF kernel expansion.6) subject to β·γ = 1. rbfpreimg RBF preimage by Schoelkopf’s ﬁxedpoint algorithm. 2 (5. The ξi ≥ 0. svmquadprog SVM trained by Matlab Optimization Toolbox. smo Sequential Minimal Optimization for binary SVM with L1soft margin. β ≥ 0. demo svm Demo on Support Vector Machines. diagker Returns diagonal of kernel matrix of given data. svmclass Support Vector Machines Classiﬁer. The C > 0 stands for the regularization constant. β ≤ 1C .1: Implemented methods: Support vector and other kernel machines.Table 5. The linear p = 1 and quadratic p = 2 term for the slack variables ξi is used. 59 . minball Minimal enclosing ball in kernel feature space. kperceptr Kernel Perceptron. In the case of the parameter p = 1. i ∈ I1 ∪ I2 are slack variables used to relax the inequalities for the case of nonseparable data. rbfpreimg2 RBF preimage problem by Gradient optimization. kfd Kernel Fisher Discriminant. The I1 = {i: yi = 1} and I2 = {i: yi = 2} are sets of indices. evalsvm Trains and evaluates Support Vector Machines classiﬁer. The training of the nonlinear (kernel) SVM with L1soft margin corresponds to solving the following QP task β ∗ = argmax β · 1 − β 1 β · Hβ . the L1soft margin SVM is trained and for p = 2 it is denoted as the L2soft margin SVM. bsvm2 Multiclass BSVM with L2soft margin. kdist Computes distance between vectors in kernel space. kernel Evaluates kernel function. rbfpreimg3 RBF preimage problem by KwokTsang’s algorithm.
ker [string] Kernel identiﬁer.1) is determined by the weight vector α. .Alpha [d × c] Weight vectors A = [α1 . (5. . The discriminant function (5.7) The bias b can be computed from the KarushKuhnTucker (KKT) conditions as the following constrains f (xi ) = α · kS (xi ) + b = +1 f (xi ) = α · kS (xi ) + b = −1 for i ∈ {j: yj = 1. −1 if yi = 2 . The set of support vectors is denoted as S = {xi : i ∈ ISV }. The vector γ [l × 1] and the matrix H [l × l] are constructed as follows γi = 1 if yi = 1 . .3: Datatype used to describe multiclass SVM classiﬁer. .b [1 × 1] Bias b.b [c × 1] Vector of biased b = [b1 . bias b and the set of support vectors S. . . . The weight vectors are compute as αi = βi if yi = 1 . Binary SVM classiﬁer (structure array): . . for i ∈ {j: yj = 2. . The bias b is computed as an average over all the constrains such that 1 γi − α · kS (xi ) .fun = ’svmclass’ Identiﬁes function associated with this data type. −k(xi . . . for i ∈ ISV . . . xj ) if yi = yj .ker [string] Kernel identiﬁer. . The 0 [l × 1] is vector of zeros and the 1 [l × 1] is vector of ones. sd }. . .options.options. . . . bc ]T .sv. must hold for the optimal solution.j = k(xi . sd }. where k: X × X → R is selected kernel function. Hi.X [n × d] Support vectors S = {s1 . αc ]. .sv.arg [1 × p] Kernel argument(s). where the set of indices is ISV = {i: βi > 0}. Table 5.options.2: Datatype used to describe binary SVM classiﬁer. .fun = ’svmclass’ Identiﬁes function associated with this data type. . Multiclass SVM classiﬁer (structure array): .Table 5. .arg [1 × p] Kernel argument(s).Alpha [d × 1] Weight vector α. 0 < βj < C} . .X [n × d] Support vectors S = {s1 . 0 < βj < C} . −βi if yi = 2 . . b= Ib  i∈I b 60 .options. xj ) if yi = yj . .
7).b [g × 1] Vector of biased b = [b1 .options. . . The bias b can be computed from the KKT conditions as the following constrains f (xi ) = α · kS (xi ) + b = +1 − αi 2C αi f (xi ) = α · kS (xi ) + b = −1 + 2C for i ∈ {i: yj = 1.fun = ’mvsvmclass’ Identiﬁes function associated with this data type. . . yg ] and the bin y(2. βj > 0} . for i ∈ {i: yj = 2. y2 . References: The Sequential Minimal Optimizer (SMO) is described in the books [23. 27.8) subject to β·γ = 1. . . . .4: Datatype used to describe the majority voting strategy for multiclass SVM classiﬁer. The training of the nonlinear (kernel) SVM with L2soft margin corresponds to solving the following QP task β ∗ = argmax β · 1 − β 1 1 β · (H + E)β .arg [1 × p] Kernel argument(s). . .X [n × d] Support vectors S = {s1 .:) contains 2 2 2 [y1 . . . 61 . .options. . yg ]. bg ]T . αg ]. . .bin y [2 × g] Targets for the binary rules. 2 2C (5. . The weight vector α is given by (5.Alpha [d × g] Weight vectors A = [α1 . . An interactive demo on the binary SVM is implemented in demo svm. .Table 5.ker [string] Kernel identiﬁer. The bias b is computed as an average over all the constrains thus 1 αi b= ) − α · kS (xi ) . y2 .5. . .:) 1 1 1 contains [y1 . . . Majority voting SVM classiﬁer (structure array): . where Ib = {i: 0 < βi < C} are the indices of the vectors on the boundary. The vector bin y(1. must hold at the optimal solution. The set of support vectors is set to S = {xi : i ∈ ISV } where the set of indices is ISV = {i: βi > 0}.sv. . γi (1 − ISV  i∈I 2C SV The binary SVM solvers implemented in the STPRtool are enlisted in Table 5. sd }. β ≥ 0. . The SV M light is described for instance in the book [27]. Example: Training binary SVM. where E is the identity matrix. . βj > 0} . 30].
options).arg = 1.org/.joachims.ker = ’rbf’. The trained SVM classiﬁer is evaluated on the testing data riply tst. trn = load(’riply_trn’).options). options. svmlight(trn. options.tst. It useful for small problems only. The SV M light is very powerful optimizer which can handle large problems.X.Function smo svmlight svmquadprog Table 5. It can handle moderate size problems.C = 10. tst = load(’riply_tst’). options. Matlab interface to SV M light software (Version: 5. The executable ﬁle svm learn for the Linux OS must be installed. SVM solver based on the QP solver quadprog of the Matlab optimization toolbox.00) which trains the binary SVM with L1soft margin. % visualization figure. psvm(model). description Implementation of the Sequential Minimal Optimizer (SMO) used to train the binary SVM with L1soft margin.1. The SMO solver is used in the example but other solvers (svmlight.options). The decision boundary is visualized in Figure 5.5: Implemented binary SVM solvers. cerror(ypred. ppatterns(trn). ypred = svmclass(tst.model).y) % load testing data % classify data % compute error 62 . It trains both L1soft and L2soft margin binary SVM. The binary SVM is trained on the Riply’s training set riply trn. The function quadprog is a part of the Matlab Optimization toolbox. svmquadprog(trn. svmquadprog) can by used as well. % train model = % model = % model = % % % % load training data use RBF kernel kernel argument regularization constant SVM classifier smo(trn. It can be downloaded from http://svmlight.
2 −1. yl )} contain the modiﬁed hidden states deﬁned as 1 for y = yi . The problem can be solved by the OneAgainstAll (OAA) decomposition.3 OneAgainstAll decomposition for SVM The input is a set TXY = {(x1 . The function allows to specify any binary SVM 63 . .2 0 −0. The discriminant functions fy (x) = αy · kS (x) + by .6 0. . The goal is to train the multiclass rule q: X ⊆ Rn → Y deﬁned by (5.0920 1. (xl . . y1 ).4 0. Let the training set TXY = {(x1 . y1 ).8 0. The OAA decomposition is implemented in the function oaasvm.5 −1 −0. yl )} of training vectors xi ∈ X ⊆ Rn and corresponding hidden states yi ∈ Y = {1. .5 0 0. .3). c}. y∈Y. yi = 2 for y = yi . . . .ans = 0. The OAA decomposition transforms the multiclass problem into a series of c binary subtasks that can be trained by the binary y SVM. .1: Example showing the binary SVM classiﬁer trained on the Riply’s data.2 1 0. . . y ∈ Y. (xl . . 5. y are trained by the binary SVM solver from the set TXY . 2.5 1 Figure 5.
. The goal is to train the multiclass rule q: X ⊆ Rn → Y based on the majority voting strategy deﬁned by (5. The SMO binary solver is used to train the binary SVM subtasks. yl )} of training vectors xi ∈ X ⊆ Rn and corresponding hidden states yi ∈ Y = {1. . c}. . The result is the multiclass SVM classiﬁer (5. ylj )} contain the training vectors xi ∈ I j = {i: yi = y 1 yi = y 2} and the modiﬁed the hidden states deﬁned as yi = 1 1 for yj = yi .arg = 1. pboundary(model). data = load(’pentagon’).2. 2. % % % % % % load data use SMO solver use RBF kernel kernel argument regularization constant training 5. y1). options. . options. The binary SVM rules qj .4 OneAgainstOne decomposition for SVM The input is a set TXY = {(x1 . .ker = ’rbf’. . Example: Multiclass SVM using the OAA decomposition The example shows the training of the multiclass SVM using the OAA decomposition.solver (see Section 5. model = oaasvm(data. . j = 1. . . options. ppatterns(model.2) used to solve the binary problems. 2 2 for yj = yi . References: The OAA decomposition to train the multiclass SVM and comparison to other methods is described for instance in the paper [12]. .sv. . . % visualization figure.’ko’. Let the training set TXY = {(x1 . i ∈ Ij .3.3) represented by the datatype described in Table 5. (xl .C = 10. .4). ppatterns(data). y1 ). The training data and the decision boundary is visualized in Figure 5.12).solver = ’smo’. g are trained on j the data TXY . .X. . options. . . (xlj . The OAO decomposition for multiclass SVM is implemented in function 64 . . g is constructed for all g = c(c − 1)/2 combinations of 1 2 1 classes yj ∈ Y and yj ∈ Y \ {yj }. .options ). j = 1. The OAO decomposition transforms the multiclass problem into a series of g = c(c − 1)/2 j binary subtasks that can be trained by the binary SVM. . j The training set TXY . The oneagainstone (OAO) decomposition strategy can be used train such a classiﬁer.
8 −1 −1 −0.2: Example showing the multiclass SVM classiﬁer build by oneagainstall decomposition.ker = ’rbf’.4 0. ppatterns(model. options.X. Example: Multiclass SVM using the OAO decomposition The example shows the training of the multiclass SVM using the OAO decomposition.’ko’.2 0 0. The training data and the decision boundary is visualized in Figure 5.4 −0.12).6 −0. % % % % % % load data use SMO solver use RBF kernel kernel argument regularization constant training 65 .solver = ’smo’. The function oaosvm allows to specify an arbitrary binary SVM solver to train the subtasks. data = load(’pentagon’). oaosvm.options).4 0. The SMO binary solver is used to train the binary SVM subtasks.1 0. References: The OAO decomposition to train the multiclass SVM and comparison to other methods is described for instance in the paper [12].2 −0. model = oaosvm(data.sv.3. options.arg = 1.2 0.2 0 −0.8 1 Figure 5. pboundary(model).8 −0. % visualization figure. options.8 0. options.6 0.6 0.4 −0.6 −0.C = 10. ppatterns(data).
the optimization problem (5.1 0. ξ∗ ) = argmin ( wy 2 + b2 ) + C (ξi )p . ξ is a vector of slack variables and I = [1. . i ∈ I . (xl . . where W = [w 1 . . . . y1 ). ξiy ≥ 0 . the parameters of the multiclass rule can be found by solving the multiclass BSVM formulation which is deﬁned as 1 (W∗ .8 0. c}. . The letter B in BSVM stands for the bias term added to the objective (5.b i∈I y∈Y\{y} F (W. .4 −0.5 Multiclass BSVM formulation The input is a set TXY = {(x1 . . .4 0.4 0.9) can be 66 . . bc ]T is vector of biases. .2 0 −0. l] is set of indices. .6 −0.2 0. . .6 0.6 0. . 5. 2. The goal is to train the multiclass rule q: X ⊆ Rn → Y deﬁned by (5.9). The L1soft margin is used for p = 1 and L2soft margin for p = 2.3: Example showing the multiclass SVM classiﬁer build by oneagainstone decomposition. . In the linear case.8 −1 −1 −0.4 −0.b. y ∈ Y \ {yi } . b = [b1 .9) y 2 y∈Y W.ξ) subject to w yi · xi + byi − ( w y · xi + by ) ≥ 1 − ξiy .2 −0. In the case of the L2soft margin p = 2. b∗ .3). yl )} of training vectors xi ∈ X ⊆ Rn and corresponding hidden states yi ∈ Y = {1.2 0 0.8 −0. i ∈ I . y ∈ Y \ {yi } . . .6 −0. . (5.8 1 Figure 5. wc ] is a matrix of normal vectors. .
y∈Y. yi) − δ(u. j) . The nonlinear (kernel) classiﬁer is obtained by substituting the selected kernel k: X × X → R for the dot products x·x to (5. u)−δ(yi .expressed in its dual form A∗ = argmax A i∈I y∈Y\{yi } y αi − 1 2 y u αi αj h(y. The upper bound FU B and the lower bound FLB on the optimal value F (W∗. u)−δ(yj . i. FLB + 1 67 .10) subject to y αi ≥ 0 . y ∈ Y \ {yi} . Kozinec’s algorithm (solver = ’kozinec’).12). . i∈I y∈Y\{yi } j∈I u∈Y\{yj } (5. . (5. The bounds are used in the algorithms to deﬁne the following two stopping conditions: Relative tolerance: FU B − FLB ≤ εrel . b∗ . αc ] are optimized weight vectors. y)) + by .9). where A = [α1 . u)δ(i.11) 2C The criterion (5.3) is composed of the set of discriminant functions fy : X → R which are computed as fy (x) = i∈I xi · x u∈Y\{yi } u αi (δ(u. u. The multiclass rule (5. y∈Y. (5. u. .10) can be solved by most the optimizers. The function h: Y×Y×I×I → R is deﬁned as h(y. y))+ δ(y. The optimization task (5. i. j) . ξ ∗ ) of the criterion (5.11) and (5.9) can be computed. The training of the multiclass BSVM classiﬁer with L2soft margin is implemented in the function bsvm2. The optimizers implemented in the function bsvm2 are enlisted bellow: MitchellDemyanovMalozemov (solver = ’mdm’).10) corresponds to the singleclass SVM criterion which is simple to optimize. y ∈ Y is given by by = i∈I y∈Y\{yi } u αi (δ(y. yj )+δ(y. Nearest Point Algorithm (solver = ’npa’). i ∈ I . The above mentioned algorithms are of iterative nature and converge to the optimal solution of the criterion (5. yi) − δ(y.12) where the bias by . u)) . j) = ( xi ·xj +1)(δ(yi . .
References: The multiclass BSVM formulation is described in paper [12]. ppatterns(model. options.13) .4. . The aim is to ﬁnd a direction ψ = l αi φ(xi ) in the feature space i=1 F given by weights α = [α1 . data = load(’pentagon’). α · (N + µE)α α · (N + µE)α 68 (5.6 Kernel Fisher Discriminant The input is a set TXY = {(x1 . pboundary(model).’ko’. The MitchellDemyanovMalozemov and Nearest Point Algorithm are described in paper [14]. x ) = φ(x) · φ(x ) . (xl . options. . ppatterns(data). y1 ).solver = ’npa’. y1). The ordinary FLD is applied on the mapped data training data TΦY = {(φ(x1 ). % visualization figure. The transformation of the multiclass BSVM criterion to the singleclass criterion (5. Example: Multiclass BSVM with L2soft margin. ’mdm’). yl )}. model = bsvm2(data. . The training data and the decision boundary is visualized in Figure 5. options.sv.3).Absolute tolerance: FU B ≤ εabs .arg = 1. % % % % % % load data use NPA solver use RBF kernel kernel argument regularization constant training 5.12). αl ]T such that the criterion F (α) = α·m 2 α · Mα = .10) implemented in the STPRtool is published in paper [9]. options. . The example shows the training of the multiclass BSVM using the NPA solver.options). 2}. The Kernel Fisher Discriminant is the nonlinear extension of the linear FLD (see Section 2. . . yl )} of training vectors xi ∈ X ⊆ Rn and corresponding values of binary state yi ∈ Y = {1.C = 10. The mapping is performed eﬃciently by means of kernel functions k: X × X → R being the dot products of the mapped vectors k(x. . (φ(xl ). .X. The Kozinec’s algorithm based SVM solver is described in paper [10]. . .ker = ’rbf’. The input training vectors are assumed to be mapped into a new feature space F by a mapping function φ: X → F . The other solvers can be used instead by setting the variable options. . .solver (other options are ’kozinec’.
6 0. . The vector m [l × 1]. . Iy my mT . .4 0. .6 −0. (5. . M = (m1 − m2 )(m1 − m2 )T . .8 −0.13) serves as the regularization term. .4: Example showing the multiclass BSVM classiﬁer with L2soft margin. The KFD training algorithm is implemented in the function kfd. The discriminant function (5.j = k(xi .4 −0. Iy  N = KKT − y∈Y y∈Y.13) deﬁned in the input space X .2 0.14) is known up to the bias b which is determined by the linear SVM with L1soft margin is applied on the projected data TZY = {(z1 . matrices N [l × l] M [l × l] are constructed as my = 1 K1y .8 0.6 0. The projections {z1 . 69 . xj ). is maximized.1 0.2 0 −0.2) is given by the found vector α as f (x) = ψ · φ(x) + b = α · k(x) + b . . zl } are computed as zi = α · k(xi ) . The criterion (5. x)]T is vector of evaluations of the kernel. k(xl .8 1 Figure 5.2 −0.4 −0. The kernel matrix K [l × l] contains kernel evaluations Ki. y where the vector 1y has entries i ∈ Iy equal to and remaining entries zeros.8 −1 −1 −0.6 −0.4 0. yl )}. .2 0 0.13) is the feature space counterpart of the criterion (2.14) where the k(x) = [k(x1 . y1). x). The diagonal matrix µE in the denominator of the criterion (5. m = m1 − m2 . . The discriminant function f (x) of the binary classiﬁer (5. . (zl . .
y ) ans = 0.7 Preimage problem Let TX = {x1 . trn = load(’riply_trn’). . The mapping is performed implicitly by using the kernel functions k: X × X → R which are dot products of the input data mapped into the feature space F such that k(xa . options.5. the isolines with f (x) = +1 and f (x) = −1 are visualized and the support vectors (whole training set) are marked as well. The found classiﬁer is evaluated on the testing data riply tst. cerror( ypred. % evaluation on testing data tst = load(’riply_tst’).e. The kernel methods works with the data nonlinearly mapped φ: X → F into a new feature space F .C = 10. .15) . options. model ). ypred = svmclass( tst.001. Let ψ ∈ F be given by the following kernel expansion l ψ= i=1 αi φ(xi ) . The detailed analysis can be found in book [30]. ppatterns(trn).0960 % % % % % % load training data use RBF kernel kernel argument regularization of linear SVM regularization of KFD criterion KFD training 5.X. . The classiﬁer is visualized as the SVM classiﬁer.References: The KFD formulation is described in paper [19]. tst. options. xb ) = φ(xa ) · φ(xb ) . options. model = kfd(trn. psvm(model). xl } be a set of vector from the input space X ⊆ Rn .arg = 1.ker = ’rbf’. 70 (5. i. The training data and the found kernel classiﬁer is visualized in Figure 5.mu = 0.mat.. options) % visualization figure. Example: Kernel Fisher Discriminant The binary classiﬁer is trained by the KFD on the Riply’s training set riply trn.mat. .
6.5 0 0.5 −1 −0.1.2 0 −0. Let the Radial Basis Function (RBF) k(xa . the preimage problem solves mapping from the feature space F back to the input space X . However.2 1 0.6 0.17) which are enlisted in Table 5.5 x − xi 2 /σ 2 ) .5 xa −xb 2 /σ 2 ) be assumed. The preimage problem can be stated as the following optimization task x = argmin φ(x) − ψ x l 2 = argmin φ(x ) − x i=1 αi φ(xi ) l 2 (5.17) The STPRtool provides three optimization algorithms to solve the RBF preimage problem (5. .16) for the particular case of the RBF kernel leads to the following task l x = argmax x i=1 αi exp(−0.17). .2 −1. all the methods are guaranteed to ﬁnd only a local optimum of the task (5.5 1 Figure 5. αl ]T is a weight vector of real coeﬃcients.8 0. (5.15). 71 . its feature space image φ(x) ∈ F would well approximate the expansion point Ψ ∈ F given by (5.4 0. The preimage problem (5. xi ) + i=1 j=1 αi αj k(xi .5: Example shows the binary classiﬁer trained based on the Kernel Fisher Discriminant. xj ) . . .16) l l = argmin k(x . xb ) = exp(−0. The preimage problem aims to ﬁnd an input vector x ∈ X . where α = [α1 . In other words. x ) − 2 x i=1 αi k(x .
% compute preimage problem x = rbfpreimg(expansion). 72 . expansion. expansion.13).options.X. rbfpreimg3 A method exploiting the correspondence between distances in the input space and the feature space proposed in [16]. % load input data data = load(’riply_trn’).6: Methods to solve the RBF preimage problem implemented in the STPRtool.Table 5.Alpha = ones(num_data. 16]. expansion. rbfpreimg2 A standard gradient optimization method using the Matlab optimization toolbox for 1D search along the gradient. ppatterns(x. References: The preimage problem and the implemented algorithms for its solution are described in [28. % visualization figure. % define the mean in the feature space num_data = size(data.1)/num_data. expansion.X).’or’. Example: Preimage problem The example shows how to visualize a preimage of the data mean in the feature induced by the RBF kernel.arg = 1.X = data.ker = ’rbf’.2). rbfpreimg An iterative ﬁxed point algorithm proposed in [28].sv. The result is visualized in Figure 5. ppatterns(data.6.options.
Let φ: X → F be a mapping from the input space X to the feature space F associated with kernel function such that k(xa . The weight vector β = [β1 . Let the function f : X ⊆ Rn → R be given as l f (x) = i=1 αi k(xi .5 1 Figure 5. . . . k: X × X → R is a kernel function. .2 0 −0. xl } are vectors from Rn . and well approximates the original one (5. βm ]T and the vectors TS = {s1 . . . The reduced set method aims to ﬁnd a new function m m ˜ f (x) = i=1 βi k(si . α = [α1 . . si ∈ Rn determine the reduced kernel expansion (5. i. .6: Example shows training data and the preimage of their mean in the feature space induced by the RBF kernel. . .6 0. m < l.18). . . xb ) = φ(xa ) · φ(xb ) .18) where TX = {x1 . The problem of ﬁnding the 73 .19) is shorter. ˜ ψ (5. αl ]T is a weight vector of real coeﬃcients and b ∈ R is a scalar.19) such that the expansion (5.8 Reduced set method The reduced set method is useful to simplify the kernel classiﬁer trained by the SVM or other kernel machines. .. . (5. x) + b = ψ · φ(x) + b . The function f (x) can be expressed in the feature space F as the linear function f (x) = ψ · φ(x) + b.5 −1 −0.2 −1.5 0 0.8 0. . .1. 5.19).4 0.e.2 1 0. sm }. x) + b = φ(x) · i=1 βi φ(si ) + b .
% evaluate SVM classifiers tst = load(’riply_tst’).20) Let the Radial Basis Function (RBF) k(xa . h2 = pboundary(red_model. The reduced set method for the RBF kernel is implemented in function rsrbf. The original and reduced SVM are evaluated on the testing data riply tst.TS i=1 βi φ(si ) − i=1 αi φ(xi ) .’Reduced SVM’). TS ) = argmin ψ − ψ β .’r’)).’C’.struct(’line_style’. red_ypred = svmclass(tst.struct(’nsv’.X.19) can be stated as the optimization task m l 2 ˜ (β. % load training data trn = load(’riply_trn’).tst. The algorithm converts the task (5. 74 . The SVM classiﬁer is trained from the Riply’s training set riply trn/mat. ypred = svmclass(tst. legend([h1(1) h2(1)].X. In this case.5 xa −xb 2 /σ 2 ) be assumed. xb ) = exp(−0. (5. h1 = pboundary(model. Decision boundaries of the original and the reduced SVM are visualized in Figure 5.y).struct(’line_style’. the optimization task (5.10)). err = cerror(ypred.1.20) into a series of m preimage tasks (see Section 5.’rbf’. The trained kernel expansion (discriminant function) is composed of 94 support vectors. Example: Reduced set method for SVM classiﬁer The example shows how to use the reduced set method to simplify the SVM classiﬁer.10)).model). The reduced expansion with 10 support vectors is found.mat. % visualize decision boundaries figure.model). References: The implemented reduced set method is described in [28]. ppatterns(trn).20) can be solved by an iterative greedy algorithm.TS 2 = argmin β .7). % compute the reduced rule with 10 support vectors red_model = rsrbf(model.’b’)).reduced kernel expansion (5.’arg’. % train SVM classifier model = smo(trn.’Original SVM’.struct(’ker’.7.
4f\n’ ..red_err = cerror(red_ypred.nsv.y). err. err = %. The original decision rule involves 94 support vectors while the reduced one only 10 support vectors. fprintf([’Original SVM: nsv = %d. .. red_err).4f\n’].4 0..0960 Reduced SVM: nsv = 10.tst.0960 1.5 0 0. model. err = 0.8 0.. err = 0.2 Original SVM Reduced SVM 1 0. err = %.6 0.2 −1.2 0 −0.5 1 Figure 5. red_model.nsv.7: Example shows the decision boundaries of the SVM classiﬁer trained by SMO and the approximated SVM classiﬁer found by the reduced set method. ’Reduced SVM: nsv = %d. 75 .5 −1 −0. Original SVM: nsv = 94.
mat /gmm samles/*. Linearly separable multiclass data in 2D. Testing part of the Riply’s data set [24].1. Well known Fisher’s data set.mat /ocr data/*.mat /anderson task/*. The datasets stored in the Matlab data ﬁle are enlisted in Table 6.mat /iris.mat /svm samples/*.1. Input data for the demo on the ExpectationMaximization algorithm for the Gaussian mixture model demo emgmm.Chapter 6 Miscellaneous 6. The functions which generate synthetic data and scripts for converting external datasets to the Matlab are given in Table 6. Input data for the demo on training linear classiﬁers from ﬁnite vector sets demo linclass. Input data for the demo on the Generalized Anderson’s task demo anderson. /multi separable/*.1 Data sets The STPRtool provides several datasets and functions to generate synthetic data. Input data for the demo on the Support Vector Machines demo svm.mat /binary separable/*.mat stored in /data/ directory of the STPRtool.mat 76 . Table 6.1: Datasets /riply trn.2. A description is given in Section 7. Examples of handwritten numerals.mat /riply tst. Training part of the Riply’s data set [24].
It works for 1D. pkernelproj Plots isolines of kernel projection. pandr 77 . Generates data from the multivariate Gaussian distribution.S.3: Functions for visualization. plane3 Plots plane in 3D. pboundary Plots decision boundary of given classiﬁer in 2D. pgmm Visualizes the bivariate Gaussian mixture model. It operates in 2dimensional space only. pline Plots line in 2D.3. pgauss Visualizes set of bivariate Gaussians. The list of functions for visualization is given in Table 6. 2D and 3D. 6. Generates data on circle corrupted by the Gaussian noise. Each function contains an example of using. Converts U. GUI which allows to create interactively labeled vectors sets and labeled sets of Gaussian distributions. Visualizes solution of the Generalized Anderson’s task. ppatterns Visualizes patterns as points in the feature space. showim Displays images entered as column vectors.createdata genlsdata gmmsamp gsamp gencircledata usps2mat Table 6.2 Visualization The STPRtool provides tools for visualization of common models and data. Generates linearly separable binary data with prescribed margin. Generates data from the Gaussian mixture model. psvm Plots decision boundary of binary SVM classiﬁer in 2D. Table 6. Postal Service (USPS) database of handwritten numerals to the Matlab ﬁle.2: Data generation.
The x and y are realizations of random variables with joint probability distribution PXY (x.4) Wε (q(x). (6. (6.3) = argmax PXY (xy)PY (y) . The 0/1loss function W0/1 (q(x).e. The Bayesian risk (6. y∈Y Classiﬁcation with rejectoption: The set of decisions D is assumed to be D = Y ∪ {dont know}.1) with the 0/1loss function corresponds to the expectation of misclassiﬁcation.1) is referred to as the Bayesian rule q ∗ (x) = argmin y y∈Y PXY (x. y) = 0 for q(x) = y . The STPRtool implements the Bayesian rule for two particular cases: Minimization of misclassiﬁcation: The set of decisions D coincides to the set of hidden states Y = {1. . y) . .6. y)W (q(x). y∈Y y∈Y (6. . Let X ⊆ Rn and the sets Y and D be ﬁnite. y).. ∀x ∈ X .2) is used.3 Bayesian classiﬁer The object under study is assumed to be described by a vector of observations x ∈ X and a hidden state y ∈ Y. The rule q: X → Y which minimizes the expectation of misclassiﬁcation is deﬁned as q(x) = argmax PY X (yx) . (6. y∈Y (6.4) is deﬁned as ⎧ ⎨ argmax PXY (xy)PY (y) if 1 − max PY X (yx) < ε . y∈Y 78 . 1 for q(x) = y . Let W : D × Y → R be a loss function which penalizes the decision q(x) ∈ D when the true hidden state is y ∈ Y. y) dx . A decision rule q: X → D takes a decision d ∈ D based on the observation x ∈ X . The rule q: X → Y which minimizes the Bayesian risk with the loss function (6. y) = ⎩ ε for q(x) = dont know . The Bayesian risk R(q) is an expectation of the value of the loss function W when the decision rule q is applied. y)W (q(x).5) q(x) = ⎩ dont know if 1 − max PY X (yx) ≥ ε . The loss function is deﬁned as ⎧ ⎨ 0 for q(x) = y . R(q) = X y∈Y PXY (x. c}. 1 for q(x) = y . i. . where ε is penalty for the decision dont know.1) The optimal rule q ∗ which minimizes the Bayesian risk (6.
Pclass{1} = emgmm(trn. bayes_model.2)).y==2).inx2).X(:.4. % load input training data trn = load(’riply_trn’). cerror(ypred. The Bayesian classiﬁers (6. % Estimation of priors n1 = length(inx1). y ∈ Y.tst.mat.fun = ’bayescls’ Identiﬁes function associated with this data type.X.struct(’ncomp’. Bayesian classiﬁer (structure array): . The distribution PXY (xy) is given by Pclassy which uses the datatype described in Table 4.Prior = [n1 n2]/(n1+n2).4: Datatype used to describe the Bayesian classiﬁer.Pclass{2} = emgmm(trn. inx2 = find(trn. The datatype used by the STPRtool is described in Table 6. The a priori probabilities PY (y).2)). The designed Bayesian classiﬁer is evaluated on the testing data riply tst. 2} are estimated by the relative occurrences (it corresponds to the ML estimate). The classconditional distributions PXY are modeled by the Gaussian mixture models (GMM) with two components. . n2 = length(inx2). References: The Bayesian decision making with applications to PR is treated in details in book [26].y) 79 .mat.Prior [1 × c] Discrete distribution PY (y). Table 6. . y ∈ {1. % Evaluation on testing data tst = load(’riply_tst’). % Estimation of classconditional distributions by EM bayes_model.struct(’ncomp’. bayes_model.To apply the optimal classiﬁcation rules one has to know the classconditional distributions PXY and a priory distribution PY (or their estimates).2.X(:.5) are implemented in function bayescls. ypred = bayescls(tst.3) and (6.y==1).inx1).bayes_model). Example: Bayesian classiﬁer The example shows how to design the plugin Bayesian classiﬁer for the Riply’s training set riply trn. The GMM are estimated by the EM algorithm emgmm. inx1 = find(trn.Pclass [cell 1 × c] Class conditional distributions PXY .
The visualization is given in Figure 6. 2. . y ∈Y y ∈Y (6. 6. % Vislualization of rejetoption rule pboundary(reject_model. the second class and the region in between corresponding to the dont know decision. c} is composed of a set of discriminant functions fy (x) = x · Ay x + by · x + cy . which are quadratic with respect to the input vector x ∈ Rn .fun = ’bayescls’. ppatterns(trn). . The Bayesian classiﬁer splits the feature space into two regions corresponding to the ﬁrst class and the second class. % Penalization for don’t know decision reject_model = bayes_model. . The rejectoption rule splits the feature space into three regions corresponding to the ﬁrst class.ans = 0.struct(’line_style’. 2}. The decision boundary of the rejectoptions rule (dashed line) is displayed for comparison. The quadratic discriminant function fy is determined by a matrix Ay [n × n].eps = 0.6) In the particular binary case Y = {1. % Visualization figure. ∀y ∈ Y . . 80 . a vector by [n × 1] and a scalar cy [1 × 1]. bayes_model.4 Quadratic classiﬁer The quadratic classiﬁcation rule q: X ⊆ Rn → Y = {1.1. pboundary(bayes_model). reject_model. the quadratic classiﬁer is represented by a single discriminant function f (x) = x · Ax + b · x + c .1. The input vector x ∈ Rn is assigned to the class y ∈ Y its corresponding discriminant function fy attains maximal value y = argmax fy (x) = argmax ( x · Ay x + by · x + cy ) . hold on.’k’)).0900 The following example shows how to visualize the decision boundary of the found Bayesian classiﬁer minimizing the misclassiﬁcation (solid black line).
8 0. Example: Quadratic classiﬁer The example shows how to train the quadratic classiﬁer using the Fisher Linear Discriminant and the quadratic data mapping.4 0. Table 6. The implementation of the quadratic classiﬁer itself provides function quadclass.5) and the multiclass (see Table 6. 2 if f (x) = x · Ax + b · x + c < 0 .6) quadratic classiﬁer.5: Datatype used to describe binary quadratic classiﬁer.7) The STPRtool uses a speciﬁc structure array to describe the binary (see Table 6.1 (dashed line).5 0 0.1. The classconditional probabilities are modeled by the Gaussian mixture models with two components estimated by the EM algorithm.1: Figure shows the decision boundary of the Bayesian classiﬁer (solid line) and the decision boundary of the rejectoption rule with ε = 0. . The classiﬁer is trained from the Riply’s 81 . . (6. .A [n × n] Parameter matrix A. a vector b [n × 1] and a scalar c [1 × 1].fun = ’quadclass’ Identiﬁes function associated with this data type.2 −1. Binary linear quadratic (structure array): .c [1 × 1] Parameter scalar c. 2} as follows q(x) = 1 if f (x) = x · Ax + b · x + c ≥ 0 .b [n × 1] Parameter vector b.6 0.5 −1 −0. given by a matrix A [n × n]. The input vector x ∈ Rn is assigned to class y ∈ {1.5 1 Figure 6.2 1 0.2 0 −0.
The found quadratic classiﬁer is evaluated on the testing data riply trn.X. . % visualization figure. c}. lin_model = fld(quad_data). . (xl . . . .y) ans = 0.2).6: Datatype used to describe multiclass quadratic classiﬁer. Multiclass quadratic classiﬁer (structure array): . quad_data = qmap(trn). y1 )..quad_model).fun = ’quadclass’ Identiﬁes function associated with this data type. .b [n × c] Parameter vectors by .e. {xi : xi ∈ Rn (x)} = K. l}.c [1 × c] Parameter scalars cy . . The Knearest neighbor (KNN) classiﬁcation rule q: X → Y is deﬁned as q(x) = argmax v(x. y ∈ Y. yl )} be a set of prototype vectors xi ∈ X ⊆ Rn and corresponding hidden states yi ∈ Y = {1.5 Knearest neighbors classiﬁer Let TXY = {(x1 . . y ∈ Y.mat. .Table 6. y) . % evaluation tst = load(’riply_tst’). . i ∈ {1. y ∈ Y. % make quadratic classifier quad_model = lin2quad(lin_model). . training set riply trn.tst. The boundary of the quadratic classiﬁer is visualized (see Figure 6. cerror(ypred. trn = load(’riply_trn’). . . (6. ppatterns(trn).0940 % load input data % quadratic mapping % train FLD 6. i. .8) y∈Y 82 . Let Rn (x) = {x : x − x ≤ r 2 } be a ball centered in the vector x in which lie K prototype vectors xi . pboundary(quad_model). .mat.A [n × n × c] Parameter matrices Ay . ypred = quadclass(tst.
2: Figure shows the decision boundary of the quadratic classiﬁer. . where v(x. Table 6.2 −1. . Example: Knearest neighbor classiﬁer The example shows how to create the (K = 8)NN classiﬁcation rule from the Riply’s training data riply trn.8). . .7.3.5 −1 −0. . KNN classiﬁer (structure array): . y) is number of prototype vectors xi with hidden state yi = y which lie in the ball xi ∈ Rn (x). . The classiﬁer is evaluated on the testing data riply tst.mat. .2 1 0. .y [1 × l] Labels {y1 . .1. for instance.5 0 0. yl }. The decision boundary is visualized in Figure 6.fun = ’knnclass’ Identiﬁes function associated with this data type. References: The Knearest neighbors classiﬁer is described in most books on PR. The classiﬁcation rule (6.6 0.8) is computationally demanding if the set of prototype vectors is large.7: Datatype used to describe the Knearest neighbor classiﬁer. % load training data and setup 8NN rule trn = load(’riply_trn’). 83 .2 0 −0. The datatype used by the STPRtool to represent the Knearest neighbor classiﬁer is described in Table 6.4 0.X [n × l] Prototype vectors {x1 .5 1 Figure 6.8 0. in book [6]. xl }. . model = knnrule(trn.mat.
3: Figure shows the decision boundary of the (K = 8)nearest neighbor classiﬁer. % evaluate classifier tst = load(’riply_tst’).8 0.y) ans = 0.2 1 0. pboundary(model).6 0.X. 84 .tst.5 0 0.5 −1 −0. cerror(ypred.2 −1.% visualize decision boundary and training data figure. ypred = knnclass(tst.5 1 Figure 6.model).1160 1. ppatterns(trn).4 0.2 0 −0.
The GUI for drawing numerals is implemented in a function mpaper. The drawn numerals are normalized to size 16 × 16 pixels (diﬀerent size can be speciﬁed). The training stage of the OCR system involves (i) collection of training examples of numerals and (ii) training the multiclass SVM. The STPRtool contains examples of numerals collected 85 . To see the trained OCR system run the demo script demo ocr.2).1a). 7. The set of training examples has to be collected before the OCR is trained. In particular. The STPRtool provides GUI which allows to use standard computer mouse as an input device to draw images of numerals. The function mpaper is designed to invoke a speciﬁed function whenever the middle mouse button is pressed. The applications selected are the optical character recognition (OCR) system based on the multiclass SVM (Section 7. The function mpaper creates a form composed of 50 cells to which the numerals can be drawn (see Figure 7. A numeral to be classiﬁed is represented as a vector x containing pixel values taken from the numeral image.1 Optical Character Recognition system The use of the STPRtool is demonstrated on a simple Optical Character Recognition (OCR) system for handwritten numerals from 0 to 9.1) and image denoising system using the kernel PCA (Section 7. The multiclass SVM are used as the base classiﬁer of the numerals. the functions save chars and ocr fun are called from the mpaper to process drawn numerals. The GUI is intended just for demonstration purposes and the scanner or other device would be used in practice instead.Chapter 7 Examples of applications We intend to demonstrate how methods implemented in the STPRtool can be used to solve a more complex applications. The function collect chars invokes the form for drawing numerals and allows to save the numerals to a speciﬁed ﬁle. The function save chars saves drawn numerals in the stage of collecting training examples and the function ocr fun is used to recognize the numerals if the trained OCR is applied.
if a speed of classiﬁcation is the major objective the multiclass BSVM formulation is often a better option as it produce less number of support vectors. (xl . The string Directory speciﬁes the directory where the ﬁles name xx. After the parameters (C ∗ . the OneAgainstOne (OAO) decomposition was used as it requires the shortest training time.mat. . σk )}.1b shows numerals after normalization which are saved to the speciﬁed ﬁle honza cech.mat are stored and the string FileName is a ﬁle name to which the resulting training set is stored. The training procedure for OCR classiﬁer is implemented in the function train ocr. The directory /demo/ocr/models/ contains the SVM model trained by OAO decomposition strategy (function oaosvm) and model trained based on the multiclass BSVM formulation (function bsvm2). .from 6 diﬀerent persons. Figure 7.1a shows the form ﬁlled with the examples of the numeral ’1’. The RBF kernel k(xa . The images of normalized numerals are stored as vectors to a matrix X of size [256 × 50]. A common method to tune the free parameters is to use the crossvalidation to select the best parameters from a preselected set Θ = {(C1 . y1 ). The methods to train the multiclass SVM are described in Chapter 5. Each person drew 50 examples of each of numerals from ’0’ to ’9’. The STPRtool contains a function evalsvm which evaluates the set of SVM parameters Θ based on the crossvalidation estimation of the classiﬁcation error. .mat (the character can be omitted from the name) is used. For example.mat is met then the function mergesets(Directory. The parameter tuning is implemented in the script tune ocr. . the free SVM parameters are C and σ.5 xa − xb 2 /σ 2 ) determined by the parameter σ was used in the experiment. If the ﬁle name format name xx. Having the labeled training set TXY = {(x1 . The resulting ﬁle contains a matrix X of all examples and a vector y containing labels are assigned to examples according the number xx. . xb ) = exp(−0. σ ∗ ) are tuned the SVM classiﬁer is trained using all training data available. σ1 ). σ ∗ ) turned out to be have the lowest crossvalidation error over the given set Θ. However. . The labels from y = 1 to y = 9 corresponds to the numerals from ’1’ to ’9’ and the label y = 10 is used for numeral ’0’. The string name is name of the person who drew the particular numeral and xx stands for a label assigned to the numeral. xb ) with its argument. The parameters C = ∞ (data are separable) and σ = 5 were found to be the best. The Figure 7. (Ck . Therefore. The function evalsvm returns the SVM model its parameters (C ∗ . The numerals are stored in the directory /data/ocr numerals/. The vector dimension n = 256 corresponds to 16 × 16 pixels of each image numeral and l = 50 is the maximal number of numerals drawn to a single form. to collect examples of numeral ’1’ from person honza cech collect_chars(’honza_cech1’) was called.DataFile) can be used to merge all the ﬁle to a single one. In particular. The ﬁle name format name xx. The SVM has two free parameters which have to be tuned: the regularization constant C and the kernel function k(xa . yl )} created the multiclass SVM classiﬁer can be trained. . . 86 .
Postal Service (USPS) database of images of handwritten numerals [17] was used as the class of images to model. The aim is to train a model of images of all numerals at ones. xb ) = φ(xa ) · φ(xb ) . To run the resulting OCR type demo ocr or call mpaper(struct(’fun’. Let TΦ = {φ(x1 ). Having the kernel PCA model of the subspace Fimg it can be ˆ used to map the images x ∈ X into the subspace Ximg . The example which produced the ﬁgure is implemented in demo kpcadenois. The STPRtool contains a script which converts the USPS database to the Matlab data ﬁle.2 Image denoising The aim is to train a model of a given class of images in order to restore images corrupted by noise. A link between the input space X and the feature space F is given by a mapping φ: X → F which is implicitly determined by a selected kernel function k: X × X → R. . The subspace Fimg is modeled by the kernel PCA trained from the set TX mapped into the feature space F by a function φ: X → F .1. A noisy image x is assumed to lie outside the subspace Ximg . The kernel PCA was used to model the image class. . The idea of using the kernel PCA for image denoising is demonstrated in Figure 7.S.mat stored in the directory /demo/ocr/. z i ∈ Rm of the mapped images 87 . φ(xl )} denote the feature space representation of the training images TX . . . Thus the unlabeled training data set TX = {x1 . 7. The U. A snapshot of running OCR is shown in Figure 7.. . It is implemented in the function usps2mat. The additive Gaussian noise was assumed as the source of image degradation. i.mat its contains is summarized in Table 7.e. . The kernel function corresponds to the dot product in the feature space. The script make noisy data produces the Matlab ﬁle usps noisy. . The kernel PCA computes the lower dimensional representation TZ = {z 1 . . The corruption of the USPS images by the Gaussian noise is implemented in script make noisy data. z l }.3.’ocr_fun’)) If one intends to modify the classiﬁer or the way how the classiﬁcation results are displayed then modify the function ocr fun. The denoising procedure is based on mapping the noisy ˆ image x into the subspace Ximg . xl } is used. . The model should describe the nonlinear subspace Ximg ⊆ Rn where the images live. . This method was proposed in [20] and further developed in [15]. It is assumed that the nonlinear subspace Ximg forms a linear subspace Fimg of a new feature space F . The kernel PCA can be used to ﬁnd the basis vectors of the subspace Fimg from the input training data TX .The particular model used by OCR is speciﬁed in the function ocr fun which implicitly uses the ﬁle ocrmodel.2. . k(xa . .
. Structure array trn containing the training data: gnd X [256 × 7291] Training images of numerals represented as column vectors.1) to ˜ compute z which determines φ(x) and (ii) solving the preimage problem (7. X [256 × 2007] Training images corrupted by the Gaussian noise with signaltonoise ration SNR = 1 (it can be changed). The input are 7291 grayscale images of size 16 × 16. the kernel PCA allows to model the nonlinearity which is very likely in the case of images. (7.5 xa − xb 2 /σ 2 ) was used here as the preimage problem (7. . The matrix A [l × m] and the vector b [m × 1] are trained by the ˜ kernel PCA algorithm. The whole ˆ procedure of reconstructing the corrupted image x ∈ Rn involves (i) using (7.Table 7. . This procedure is implemented in function kpcarec.2) which yields the resulting x .mat produced by the script make noisy data. xb ) = exp(−0. . X [256 × 7291] Training images corrupted by the Gaussian noise with signaltonoise ration SNR = 1 (it can be changed). In contrast to the linear PCA. However. . . . . i=1 β = A(z − b) . the nonlinearity hidden in 88 . . k(xl . TΦ to minimize the mean square reconstruction error l εKM S = i=1 ˜ φ(xi ) − φ(xi ) 2 . x)]T contains kernel functions evaluated at the training data TX . . 10} of training images. Given a projection φ(x) onto the kernel PCA subspace an n image x ∈ R of the input space can be recovered by solving the preimage problem ˜ x = argmin φ(x) − φ(x) x 2 . Structure array tst containing the testing data: gnd X [256 × 2007] Testing images of numerals represented as column vectors.1) The vector k(x) = [k(x1 . y [1 × 2007] Labels y ∈ Y = {1. . (7. 10} of training images. z = AT k(x) + b . where ˜ φ(x) = l βi φ(x) . y [1 × 7291] Labels y ∈ Y = {1.2) The Radial Basis Function (RBF) kernel k(xa .2) can be well solved for this case. x).1: The content of the ﬁle usps noisy. The input are 2007 grayscale images of size 16 × 16. .
1) and (7. Figure 7. . The training for linear PCA is implemented in the script train lin denois. Figure 7. (σk . images restored by the linear PCA and kernel PCA. The same idea was used to tune the output dimension of the linear PCA. Having the parameters tuned the resulting kernel PCA and linear PCA are trained on the whole training set.4 shows the idea of tuning the parameters of the kernel PCA. The linear PCA was used to solve the same task for comparison. An example of denoising of the testing data of USPS is implemented in script show denoising. .4 is used for training the model. . the function pcarec was used. The ﬁgure was produced by a script show denois tuning. The results of tuning the parameters are displayed in Figure 7. 89 .2). Another parameter is the dimension m of the subspace kernel PCA subspace to be trained. The whole training procedure of the kernel PCA is implemented in the script train kpca denois. It produces sparse model and can deal with large data sets. εM S (σ.6 shows examples of ground truth images. . In the case of the linear PCA. The x stands for the ground truth image and x denotes the image ˆ restored from noisy one x using (7. The best parameters (σ ∗ . The function kpcarec is used to denoise images using the kernel PCA model. mk )} such that the crossvalidation estimate of the mean square error PX (x) x − x 2 dx . m∗ ) was selected out of a prescribed set Θ = {(σ1 .the parameter σ of the used RBF kernel has to be tuned properly to prevent overﬁtting. m) = X was minimal. noisy images.5. The images are assumed to be generated from distribution PX (x). m1 ). The greedy kernel PCA Algorithm 3.
9 1 (a) (b) Figure 7.5 0. 90 .2 0. 0.15 0.7 0.05 0 0 0. right_button − erase.45 0.3 0.8 0.4 0.Control: left_button − draw.1 0.4 0.35 0.25 0.5 0.6 0. middle_button − classify.1: Figure (a) shows handwritten training examples of numeral ’1’ and Figure (b) shows corresponding normalized images of size 16 × 16 pixels.3 0.1 0.2 0.
4 0.4 0.1 0. middle_button − classify.7 0.8 0.2 0.9 1 (a) 0.1 0.4 0.05 0 0 1 2 3 4 5 6 7 8 9 0 4 5 6 0.9 1 (b) Figure 7.1 0. Figure (a) shows handwritten numerals and Figure (b) shows recognized numerals.6 0.3 0.35 0.3 0.6 0.5 0. 0.4 0 3 2 0.5 0.2 0. right_button − erase.05 0 0 0.15 0.3 0.1 0. 91 .2 0.25 0.7 0.15 0.25 0.8 1 9 8 0.2 0.5 0.Control: left_button − draw.2: A snapshot of running OCR system.45 0.5 0.45 0.3 0.35 0.
Distortion model KPCA model θ∈Θ Figure 7. Preimage problem error θ Data generator Error assessment 92 .3: Demonstration of idea of image denoising by kernel PCA.7 6 5 4 3 2 1 0 −1 −2 −3 −4 −4 Ground truth Noisy examples Corrupted Reconstructed −3 −2 −1 0 1 2 3 4 5 6 7 Figure 7.4: Schema of tuning parameters of kernel PCA model used for image restoration.
6: The results of denosing the handwritten images of USPS database.5 7 7. Ground truth Numerals with added Gaussian noise Linear PCA reconstruction Kernel PCA reconstruction Figure 7.5 13 12 11 10 0 10 20 30 40 50 dim 60 70 80 90 100 4 4.18 20 dim = 50 dim = 100 dim = 200 dim = 300 17 18 16 16 15 14 εMS 14 εMS 12 10 8 6 3.5 8 (a) (b) Figure 7.5: Figure (a) shows plot of εM S with respect to output dimension of linear PCA.5 5 5. Figure(b) shows plot of εM S with respect to output dimension and kernel argument of the kernel PCA.5 σ 6 6. 93 .
• Discrete probability models estimated by the EM algorithm. We welcome anyone who likes to contribute to the STPRtool in any way. • Eﬃcient optimizers to for large scale Support Vector Machines learning problems. • Probabilistic Principal Component Analysis (PPCA) and mixture of PPCA estimated by the ExpectationMaximization (EM) algorithm. • Model selection methods for SVM. suggestion or other feedback which would help in the development of the STPRtool. The future planes involve the extensions of the toolbox by the following methods: • Feature selection methods. • AdaBoost and related learning methods. Therefore we are grateful for any comment.Chapter 8 Conclusions and future planes The STPRtool as well as its documentation is being updating. 94 . There are many relevant methods and proposals which have not been implemented yet. Please send us an email and your potential contribution.
G. June 1962. Classiﬁcation into two multivariate normal distributions with diﬀerrentia covariance matrices. Pattern Classiﬁcation. John Wiley & Sons. Rubin.W. Baudat and F. Neural Computation. ShaweTaylor. [5] A. 2000. Oxford. [2] G. Cesk´ vysok´ uˇen´ technick´. editors. Cambridge University Press. An iterative algorithm learning the maximal margin classiﬁer. ac [10] V.. volume 2. [3] C. Prtools: A matlab toolbox for pattern recognition. Laird. Laurendeau.W. 2000. Katedra kybernetiky. [7] R. 2002. P. 2nd. Maximum likelihood from incomplete data via the EM Algorithm.B. 33:420–431. 1997. Hart. Support Vector Machines. Cristianini and J. and D. Dempster. Neural Networks for Pattern Recognition. 1977. [8] V. Generalized discriminant analysis using a kernel approach. 16th International Conference on Pattern Recognition. Franc and V. 36(9):1985–1996. 39:185– 197. Hlav´ˇ. 95 . Fakulta e c ı e elektrotechnick´. Clarendon Press. Multiclass support vector machine. pages 236–239. IEEE Computer Society. and D. Franc and V. Anouar. edition.E. Pattern Recognition. ac D.O. Journal of the Royal Statistical Society. In Czech). [4] N. Master’s thesis. 2001. Franc. Bahadur. 12(10):2385–2404. N. Bishop. Duda. Anderson and R.R. February 2000. Great Britain. Programov´ n´stroje pro rozpozn´v´n´ (Pattern Recognition Programe a a a ı ˇ e ming Tools. Anals of Mathematical Statistics. 3th edition. 2003. Stork. Duin.M. Kasturi.P. Hlav´ˇ. and Suen C. 2000.M.Bibliography [1] T. In R. a [9] V.P. [6] R.
ECCV Workshop. In A. B. Computer Analysis of Images and Patterns. editors. Hlav´ˇ. Germany.T.C.S. J. editors. Shevade. Sch¨lkopf. P.A. M¨ ller. 96 . In. A. Westenberg. May 2004. [16] J. Kernel o u a pca and denoising in feature spaces. and K. and K. S. Backpropagation applied to handwritten zip code recognition. John Wiley & Sons.W. O. pages 41–48. S. [12] C. Probabilities for sv machines. March 2002. pages 536 – 542. R¨tsch. s editors. B. Smola. Scholz. Nabney. In N. Solla. [15] Kim Kwang. and D. Advances in pattern recognition. [18] G. Petkov and M.. Tsang. Murthy.J. Kernel hebbian o algorithm for singlefram superresolution.[11] V. New York. 1999. C. LeCun. IEEE. and L. Sch¨lkpf. Boser. In Y. [22] J. Kearns. USA.J. M¨ ller. Fisher discriminant analysis with kernel.R. Berlin. SpringerVerlag. 1997.S. Advances in Large Margin Classiﬁers (Neural Information Processing Series). IEEE Transactions on Neural Networks. MA. B. Mika. Neural Networks for Signal Processing. Neural Computation. R¨tsch. J. The EM Algorithm and Extensions. Cambridge. Greedy algorithm for a training set reduction in the ac kernel methods. In Proceedings of the Twentieth International Conference on Machine Learning (ICML2003).T. [17] Y. D.K. Springer. 1986. A fast iterative nearest point algorithm for support vector machine classiﬁer design. August 2003. 1989. and Sch¨lkopf Bernhard. Advances in Neural Information Processing Systems 11. pages 408–415. Cohn. Washington. Hu. [13] I. editors. Franz Matthias. MIT Press. M. Bartlett. Statisical Learning in Computer Vision. January 2000. Douglas. [21] I. Lin.S. In M.H. Krishnan. Kwok and I.J Jackel. K. 11(1):124–136.K. New York. Springer. 2000. Springer. B. pages 426–433. The preimage problem in kernel methods. editors. Hsu and C. Hubbard. In Leonardis Aleˇ and Bischof Horst.. 13(2). Larsen. and G. R. 2003. Bhattacharya. MIT Press. Franc and V.A. W. Howard. a o u [19] S. and D.E. A comparison of methods for multiclass support vector machins. 1999.A.T. D. IEEE Transactions on Neural Networks. [20] S. [14] S. Schuurmans. Scholkopf. NETLAB : algorithms for pattern recognition. Weston. 1:541–551. and S. Platt. 2002. E. London. Principal Component Analysis. Keerthi.W. J.J.R. McLachlan and T. Jollife. Wilson. Denker. Smola. Henderson. Mika. G.
Schlesinger. Kibernetika. The MIT Press. P. [28] B. Smola. Knirsch. pages 124–132. Neural networks and related methods for classiﬁcation (with discusion). J. Ahler. MA.[23] J. Advances in Kernel Methods – Support o Vector Learning. A. 1998. A connection between learning and selflearning in the pattern recognition (in Russian). 56:409–456. 1994. Fast approximation of support o vector kernel expansions. SpringerVerlag. editors. and A. and F. Learning with Kernels. and an interpretation of clustering as approximation in feature spaces. Burges. Royal Statistical Soc. Levi. In P. Neural Computation. [30] B. Schanz.J. 1995. Redmond. Germany. [24] B. C.J. 97 . 2002. Muller. Schlesinger and V. Riply.R. [29] B. [25] M. and K. Smola. [27] B. 2:81–88. 1998.I. A. Burges. Ten lectures on statistical and structural pattern ac recognition. [26] M.D. Microsoft Research.. 1998. Cambridge.C. Series B. Vapnik. Platt. May. Kluwer Academic Publishers. Smola. R. Springer Verlag. Sch¨lkopf. o [31] V. Sequential minimal optimizer: A fast algorithm for training support vector machines. Sch¨lkopf.I. 1968. Berlin. Hlav´ˇ. The nature of statistical learning theory. Sch¨lkopf and A. Technical Report MSRTR9814. DAGM. M. Mustererkennung 199820. Smola. 1999. and C. MIT Press. 10:1299–1319. Nonlinear component analysis as a kernel eigenvalue problem. 2002.J. Scholkopf.
This action might not be possible to undo. Are you sure you want to continue?
Use one of your book credits to continue reading from where you left off, or restart the preview.