SGN-2506: Introduction to Pattern Recognition

Jussi Tohka
Tampere University of Technology
Institute of Signal Processing
2006

September 1, 2006
ii
Preface

This is an English translation of the lecture notes written originally in Finnish for
the course SGN-2500 Johdatus hahmontunnistukseen. The basis for the original
lecture notes was the course Introduction to Pattern Recognition that I lectured at
the Tampere University of Technology during 2003 and 2004. Especially, the course
in the fall semester of 2003 was based on the book Pattern Classification, 2nd Edition
by Richard Duda, Peter Hart and David Stork. The course has thereafter diverted
from the book, but still the order of topics during the course is much the same as
in the book by Duda et al.
These lecture notes correspond to a four credit point course, which has typically
been lectured during 24 lecture-hours. The driving idea behind these lecture notes
is that the basic material is presented thoroughly. Some additonal information
are then presented in a more relaxed manner. This, for example, if the formal
treatment would require too much mathematical background. The aim is to provide
the student the basic understanding of the foundations of the statistical pattern
recognition and a basis for advanced pattern recognition courses. This necessarily
means that the weight is on the probabilistic foundations of pattern recognition and
specific pattern recognition techniques and applications get less attention. This may
naturally bother some engineering students but I think the choice was a necessary
one.

Tampere summer 2006

Jussi Tohka
iv
Contents

1 Introduction 1

2 Pattern Recognition Systems 3

2.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.1 Optical Character Recognition (OCR) . . . . . . . . . . . . . 3
2.1.2 Irises of Fisher/Anderson . . . . . . . . . . . . . . . . . . . . . 3
2.2 Basic Structure of Pattern Recognition Systems . . . . . . . . . . . . 4
2.3 Design of Pattern Recognition Systems . . . . . . . . . . . . . . . . . 6
2.4 Supervised and Unsupervised Learning and Classification . . . . . . . 7

3 Background on Probability and Statistics 9

3.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Basic Definitions and Terminology . . . . . . . . . . . . . . . . . . . . 9
3.3 Properties of Probability Spaces . . . . . . . . . . . . . . . . . . . . . 10
3.4 Random Variables and Random Vectors . . . . . . . . . . . . . . . . 11
3.5 Probability Densities and Cumulative Distributions . . . . . . . . . . 12
3.5.1 Cumulative Distribution Function (cdf) . . . . . . . . . . . . . 12
3.5.2 Probability Densities: Discrete Case . . . . . . . . . . . . . . . 13
3.5.3 Probability Densities: Continuous Case . . . . . . . . . . . . . 13
3.6 Conditional Distributions and Independence . . . . . . . . . . . . . . 14
3.6.1 Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.6.2 Random Variables . . . . . . . . . . . . . . . . . . . . . . . . 15
3.7 Bayes Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.8 Multivariate Normal Distribution . . . . . . . . . . . . . . . . . . . . 17

4 Bayesian Decision Theory and Optimal Classifiers 19

4.1 Classification Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 Classification Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3 Bayes Minimum Error Classifier . . . . . . . . . . . . . . . . . . . . 22
4.4 Bayes Minimum Risk Classifier . . . . . . . . . . . . . . . . . . . . . 23
4.5 Discriminant Functions and Decision Surfaces . . . . . . . . . . . . . 24
4.6 Discriminant Functions for Normally Distributed Classes . . . . . . . 26
4.7 Independent Binary Features . . . . . . . . . . . . . . . . . . . . . . . 31
vi CONTENTS

5 Supervised Learning of the Bayes Classifier 35

5.1 Supervised Learning and the Bayes Classifier . . . . . . . . . . . . . . 35
5.2 Parametric Estimation of Pdfs . . . . . . . . . . . . . . . . . . . . . . 36
5.2.1 Idea of Parameter Estimation . . . . . . . . . . . . . . . . . . 36
5.2.2 Maximum Likelihood Estimation . . . . . . . . . . . . . . . . 36
5.2.3 Finding ML-estimate . . . . . . . . . . . . . . . . . . . . . . . 37
5.2.4 ML-estimates for the Normal Density . . . . . . . . . . . . . . 37
5.2.5 Properties of ML-Estimate . . . . . . . . . . . . . . . . . . . . 38
5.2.6 Classification Example . . . . . . . . . . . . . . . . . . . . . . 39
5.3 Non-Parametric Estimation of Density Functions . . . . . . . . . . . 40
5.3.1 Histogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3.2 General Formulation of Density Estimation . . . . . . . . . . . 41
5.4 Parzen Windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.4.1 From Histograms to Parzen Windows . . . . . . . . . . . . . . 43
5.4.2 Parzen Estimates . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.4.3 Properties of Parzen Estimates . . . . . . . . . . . . . . . . . 44
5.4.4 Window Width . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.4.5 Parzen Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.4.6 Probabilistic Neural Networks . . . . . . . . . . . . . . . . . . 46
5.5 k-Nearest Neighbors Classifier . . . . . . . . . . . . . . . . . . . . . . 47
5.5.1 kn Nearest Neighbors Density Estimation . . . . . . . . . . . . 47
5.5.2 Nearest Neighbor Rule . . . . . . . . . . . . . . . . . . . . . . 48
5.5.3 k Nearest Neighbor Classification Rule . . . . . . . . . . . . . 48
5.5.4 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.6 On Sources of Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

6 Linear Discriminant Functions and Classifiers 53

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2 Properties of Linear Classifiers . . . . . . . . . . . . . . . . . . . . . . 54
6.2.1 Linear Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.2.2 The Two Class Case . . . . . . . . . . . . . . . . . . . . . . . 54
6.2.3 The c Class Case . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.3 Linearly Separable Training Samples . . . . . . . . . . . . . . . . . . 56
6.4 Perceptron Criterion and Algorithm in 2-Class Case . . . . . . . . . . 58
6.4.1 Perceptron Criterion . . . . . . . . . . . . . . . . . . . . . . . 58
6.4.2 Perceptron Algorithm . . . . . . . . . . . . . . . . . . . . . . 58
6.5 Perceptron for Multi-Class Case . . . . . . . . . . . . . . . . . . . . . 59
6.6 Minimum Squared Error Criterion . . . . . . . . . . . . . . . . . . . . 60

7 Classifier Evaluation 63
7.1 Estimation of the Probability of the Classification Error . . . . . . . . 63
7.2 Confusion Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.3 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
CONTENTS vii

8 Unsupervised Learning and Clustering 67

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
8.2 The Clustering Problem . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.3 K-means Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.3.1 K-means Criterion . . . . . . . . . . . . . . . . . . . . . . . . 68
8.3.2 K-means Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 69
8.3.3 Properties of the K-means Clustering . . . . . . . . . . . . . . 70
8.4 Finite Mixture Models . . . . . . . . . . . . . . . . . . . . . . . . . . 70
8.5 EM Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
viii CONTENTS
Chapter 1

Introduction

The term pattern recognition refers to the task of placing some object to a cor-
rect class based on the measurements about the object. Usually this task is to
be performed automatically with the help of computer. Objects to be recognized,
measurements about the objects, and possible classes can be almost anything in the
world. For this reason, there are very different pattern recognition tasks. A system
that makes measurements about certain objects and thereafter classifies these ob-
jects is called a pattern recognition system. For example, a bottle recycling machine
is a pattern recognition system. The customer inputs his/her bottles (and cans) into
the machine, the machine recognizes the bottles, delivers them in proper containers,
computes the amount of compensation for the customer and prints a receipt for
the customer. A spam (junk-mail) filter is another example of pattern recognition
systems. A spam filter recognizes automatically junk e-mails and places them in a
different folder (e.g. /dev/null) than the user’s inbox. The list of pattern recognition
systems is almost endless. Pattern recognition has a number of applications ranging
from medicine to speech recognition.
Some pattern recognition tasks are everyday tasks (e.g. speech recognition) and
some pattern recognition tasks are not-so-everyday tasks. However, although some
of these task seem trivial for humans, it does not necessarily imply that the re-
lated pattern recognition problems would be easy. For example, it is very difficult
to ’teach’ a computer to read hand-written text. A part of the challenge follow
because a letter ’A’ written by a person B can look highly different than a letter
’A’ written by another person. For this reason, it is worthwhile to model the vari-
ation within a class of objects (e.g. hand-written ’A’s). For the modeling of the
variation during this course, we shall concentrate on statistical pattern recognition,
in which the classes and objects within the classes are modeled statistically. For
the purposes of this course, we can further divide statistical pattern recognition
into two subclasses. Roughly speaking, in one we model the variation within object
classes (generative modeling) and in the other we model the variation between the
object classes (discriminative modeling). If understood broadly, statistical pattern
recognition covers a major part of all pattern recognition applications and systems.
Syntactic pattern recognition forms another class of pattern recognition meth-
ods. During this course, we do not cover syntactic pattern recognition. The basic
2 Introduction to Pattern Recognition

idea of syntactic pattern recognition is that the patterns (observations about the
objects to be classified) can always be represented with the help of simpler and
simpler subpatterns leading eventually to atomic patterns which cannot anymore
be decomposed into subpatterns. Pattern recognition is then the study of atomic
patterns and the language between relations of these atomic patterns. The theory
of formal languages forms the basis of syntactic pattern recognition.
Some scholars distinguish yet another type of pattern recognition: The neural
pattern recognition, which utilizes artificial neural networks to solve pattern recog-
nition problems. However, artificial neural networks can well be included within the
framework of statistical pattern recognition. Artificial neural networks are covered
during the courses ’Pattern recognition’ and ’Neural computation’.
Chapter 2

Pattern Recognition Systems

2.1 Examples

2.1.1 Optical Character Recognition (OCR)

Optical character recognition (OCR) means recognition of alpha-numeric letters

based on image-input. The difficulty of the OCR problems depends on whether
characters to be recognized are hand-written or written out by a machine (printer).
The difficulty level of an OCR problem is additionally influenced by the quality of
image input, if it can be assumed that the characters are within the boxes reserved
for them as in machine readable forms, and how many different characters the system
needs to recognize.
During the course, we will briefly consider the recognition of hand-written nu-
merals. The data is freely available from the machine learning database of the
University of California, Irvine. The data has been provided by E. Alpaydin and C.
Kaynak from the University of Istanbul.

2.1.2 Irises of Fisher/Anderson

Perhaps the best known material used in the comparisons of pattern classifiers is
the so-called Iris dataset. This dataset was collected by Edgar Anderson in 1935. It
contains sepal width, sepal length, petal width and petal length measurements from
150 irises belonging to three different sub-species. R.A. Fisher made the material
famous by using it as an example in the 1936 paper ’The use of multiple measure-
ments in taxonomic problems’, which can be considered as the first article about the
pattern recognition. This is why the material is usually referred as the Fisher’s Iris
Data instead of the Anderson’s Iris Data. The petal widths and lengths are shown
in Figure 2.1.
4 Introduction to Pattern Recognition

Figure 2.1: Petal widths (x-axis) and lengths (y-axis) of Fisher’s irises.

2.2 Basic Structure of Pattern Recognition Sys-

tems
The task of the pattern recognition system is to classify an object into a correct
class based on the measurements about the object. Note that possible classes are
usually well-defined already before the design of the pattern recognition system.
Many pattern recognition systems can be thought to consist of five stages:

1. Sensing (measurement);
2. Pre-processing and segmentation;
3. Feature extraction;
4. Classification;
5. Post-processing;

A majority of these stages are very application specific. Take the sensing stage
as an example: the measurements required by speech recognition and finger print
identification are rather different as are those required by OCR and Anderson’s Iris
measurements.
For the classification stage, the situation is different. There exist a rather general
theory about the classification and an array of general algorithms can be formulated
for the classification stage. For this reason, the main emphasis during this course
is on the classifier design. However, the final aim is to design a pattern recognition
system. Therefore it is important to gain understanding about every stage of the
pattern recognition system although we can give few application non-specific tools
for the design of these stages in a pattern recognition system.
2.2 Basic Structure of Pattern Recognition Systems 5

Sensing refers to some measurement or observation about the object to be clas-

sified. For example, the data can consist of sounds or images and sensing equipment
can be a microphone array or a camera. Often one measurement (e.g. image) in-
cludes information about several objects to be classified. For instance, assume that
we want to recognize the address written on the envelope. We must then classify
several characters to recognize the whole address. The data here is probably an
image of the envelope including the sub-images of all the characters to be classified
and some background that has nothing to do with the pattern recognition task.
Pre-processing refers to filtering the raw data for noise suppression and other
operations performed on the raw data to improve its quality. In segmentation,
the measurement data is partitioned so that each part represents exactly one object
to be classified. For example in address recognition, an image of the whole address
needs to be divided to images representing just one character. The result of the
segmentation can be represented as a vector that is called a pattern vector.
Feature extraction. Especially when dealing with pictorial information the
amount of data per one object can be huge. A high resolution facial photograph (for
face recognition) can contain 1024*1024 pixels. The pattern vectors have then over a
million components. The most part of this data is useless for classification. In feature
extraction, we are searching for the features that best characterize the data for
classification The result of the feature extraction stage is called a feature vector. The
space of all possible feature vectors is called the feature space. In face recognition, a
widely used technique to reduce the number features is principal component analysis
(PCA). (This yields so called eigenfaces.) PCA is a statistical technique to reduce
the dimensionality of a data vector while retaining most of the information that
the data vector contains. In addition to mathematical/computational/statistical
techniques, feature extraction can be performed heuristically by picking such features
from a pattern vector that could be assumed to be useful in classification. For face
recognition, such a feature can be the distance between the two eyes of a person.
Sometimes, dimensionality of the input data is very limited and the pattern vector
can be chosen as the feature vector. For example this is the case with the image
segmentation methods that are based on the pattern recognition principles. Feature
extraction is highly application specific although some general techniques for feature
extraction and selection have been developed. You will hear more about these during
the course ’Pattern Recognition’.
In general, the line between feature extraction and classification is fuzzy. The
task of the feature extractor is to produce a representation about the data that
enables an easy classification. On the other hand, the task of the classifier is to
produce the best classification accuracy given the extracted features. Clearly, these
two stages are interdependent from the application point of view. Also, and perhaps
more importantly, what is the best representation for the data depends on the
classifier applied. There is no such thing as the universally optimal features.
The classifier takes as an input the feature vector extracted from the object
to be classified. It places then the feature vector (i.e. the object) to class that is
the most appropriate one. In address recognition, the classifier receives the features
extracted from the sub-image containing just one character and places it to one of
6 Introduction to Pattern Recognition

the following classes: ’A’,’B’,’C’..., ’0’,’1’,...,’9’. The classifier can be thought as a

mapping from the feature space to the set of possible classes. Note that the classifier
cannot distinguish between two objects with the same feature vector.
The main content of this course is within the classifier design. There is a sim-
ple reason: The abstraction offered by the concept of the feature vector makes it
possible to develop a general, application-independent theory for the design of clas-
sifiers. Therefore, it is possible to use the same underlying principles when designing
classifiers for address recognition systems and bottle recycling machines.
Post-processing. A pattern recognition system rarely exists in a vacuum. The
final task of the pattern recognition system is to decide upon an action based on the
classification result(s). A simple example is a bottle recycling machine, which places
bottles and cans to correct boxes for further processing. Different actions can have
also different costs associated with them. This information can be included to the
classifier design as we will see later on. Also, we can have several interdependent
classification results in our hands. For example in an address recognition task, we
have the classification results for multiple characters and the task is to decide upon
the address that these characters form. Therefore, it is possible to use the context: In
this case, the information about the other classification results to correct a possible
misclassification. If the result of the classifier is ’Hollywoud Boulevard’ then the
address is probably ’Hollywood Boulevard’.

2.3 Design of Pattern Recognition Systems

The design of a pattern recognition system is an iterative process. If the system
is not good enough, we go back to the drawing table and try to improve some of
the five stages of the system (see the previous section 2.2). After that the system is
tested again and it is improved if it still does not meet the requirements placed for it.
The process is repeated until the system meets the requirements. The design process
is founded on the knowledge about the particular pattern recognition application,
the theory of pattern recognition and simply trial and error. Clearly, the smaller
the importance of ’trial and error’ the faster is the design cycle which makes it
important to minimize the amount of trial and error in the design of the pattern
recognition system. The importance of ’trial and error’ can be minimized by having
good knowledge about the theory of pattern recognition and integrating that with
knowledge about the application area. When is a pattern recognition system good
enough? This obviously depends on the requirements of the application, but there
are some general guidelines about how to evaluate the system. We will return to
these later on during this course. How can a system be improved? For example, it
can be possible to add a new feature to feature vectors to improve the classification
accuracy. However, this is not always possible. For example with the Fisher’s Iris
dataset, we are restricted to the four existing features because we cannot influence
the original measurements. Another possibility to improve a pattern recognition
system is to improve the classifier. For this too there exists a variety of different
means and it might not be a simple task to pick the best one out of these. We will
2.4 Supervised and Unsupervised Learning and Classification 7

return to this dilemma once we have the required background.

When designing a pattern recognition system, there are other factors that need
to be considered besides the classification accuracy. For example, the time spent for
classifying an object might be of crucial importance sometimes. Another important
factor is the cost of the system.

2.4 Supervised and Unsupervised Learning and

Classification
The classifier component of a pattern recognition system has to be taught to identify
certain feature vectors to belong to a certain class. The points here are that 1) it
is impossible to define the correct classes for all the possible feature vectors1 and 2)
the purpose of the system is to assign an object which it has not seen previously to
a correct class.
It is important to distinguish between two types of machine learning when con-
sidering the pattern recognition systems. The (main) learning types are supervised
learning and unsupervised learning. We also use terms supervised classification and
unsupervised classification
In supervised classification, we present examples of the correct classification (a
feature vector along with its correct class) to teach a classifier. Based on these
examples, that are sometimes termed prototypes or training samples, the classifier
then learns how to assign an unseen feature vector to a correct class. The generation
of the prototypes (i.e. the classification of feature vectors/objects they represent)
has to be done manually in most cases. This can mean lot of work: After all, it was
because we wanted to avoid the hand-labeling of objects that we decided to design a
pattern recognition system in the first place. That is why the number of prototypes
is usually very small compared to the number of possible inputs received by the
pattern recognition system. Based on these examples we would have to deduct the
class of a never seen object. Therefore, the classifier design must be based on the
assumptions made about the classification problem in addition to prototypes used to
teach the classifier. These assumptions can often be described best in the language
of probability theory.
In unsupervised classification or clustering, there is no explicit teacher nor train-
ing samples. The classification of the feature vectors must be based on similarity
between them based on which they are divided into natural groupings. Whether any
two feature vectors are similar depends on the application. Obviously, unsupervised
classification is a more difficult problem than supervised classification and super-
vised classification is the preferable option if it is possible. In some cases, however,
it is necessary to resort to unsupervised learning. For example, this is the case if
the feature vector describing an object can be expected to change with time.
1
Think about 16 times 16 binary images where each pixel is a feature, which has either value
0 or 1. In this case, we have 2256 ≈ 1077 possible feature vectors. To put it more plainly, we have
more feature vectors in the feature space than there are atoms in this galaxy.
8 Introduction to Pattern Recognition

There exists a third type of learning: reinforcement learning. In this learning

type, the teacher does not provide the correct classes for feature vectors but the
teacher provides feedback whether the classification was correct or not. In OCR,
assume that the correct class for a feature vector is ’R’. The classifier places it into the
class ’B’. In reinforcement learning, the feedback would be that ’the classification is
incorrect’. However the feedback does not include any information about the correct
class.
Chapter 3

Background on Probability and

Statistics

Understanding the basis of the probability theory is necessary for the understand-
ing of the statistical pattern recognition. After all we aim to derive a classifier that
makes as few errors as possible. The concept of the error is defined statistically,
because there is variation among the feature vectors representing a certain class.
Very far reaching results of the probability theory are not needed in an introductory
pattern recognition course. However, the very basic concepts, results and assump-
tions of the probability theory are immensely useful. The goal of this chapter is to
review the basic concepts of the probability theory.

3.1 Examples
The probability theory was born due to the needs of gambling. Therefore, it is
pertinent to start with elementary examples related to gambling.
What is the probability that a card drawn from a well-shuffled deck of cards is
an ace?
In general, the probability of an event is the number of the favorable outcomes
divided by the total number of all outcomes. In this case, there are 4 favorable
outcomes (aces). The total number of outcomes is the number of cards in the deck,
that is 52. The probability of the event is therefore 4/52 = 1/13.
Consider next rolling two dice at the same time. The total number of outcomes
is 36. What is then the probability that the sum of the two dice is equal to 5? The
number of the favorable outcomes is 4 (what are these?). And the probability is
1/9.

3.2 Basic Definitions and Terminology

The set S is a sample space for an experiment (for us a measurement), if every
physical outcome of the experiment refers to a unique element of S. These elements
of S are termed elementary events.
10 Introduction to Pattern Recognition

A subset of S is called an event. An event is a set of possible outcomes of an

experiment. A single performance of an experiment is a trial. At each trial we
observe a single elementary event. We say that an event A occurs in the trial if
the outcome of the trial belongs to A. If two events do not intersect, i.e. their
intersection is empty, they are called mutually exclusive. More generally, the events
E1 , E2 , . . . , En are mutually exclusive if Ei ∩ Ej = ∅ for any i 6= j. Given two
mutually exclusive events, A and B, the occurrence of A in a trial prevents B from
occurring in the same trial and vice versa.
A probability measure assigns a probability to each event of the sample space. It
expresses how frequently an outcome (an elementary event) belonging to an event
occurs within the trials of an experiment. More precisely a probability measure P
on S is a real-valued function defined on the events of S such that:
1. For any event E ⊆ S, 0 ≤ P (E) ≤ 1.

2. P (S) = 1.

3. If events E1 , E2 , . . . , Ei , . . . are mutually exclusive

∞
[
P( Ei ) = P (E1 ) + P (E2 ) + · · ·
i=1

These s.c. Kolmogorov axioms have natural interpretations (frequency interpre-

tations). The axiom 1 gives the bounds for the frequency of occurrence as a real
number. For example, a negative frequency of occurrence would not be reasonable.
The axiom 2 states that in every trial something occurs. The axiom 3 states that if
the events are mutually exclusive, the probability that one of the events occurs in a
trial is the sum of the probabilities of events. Note that in order to this make any
sense it is necessary that the events are mutually exclusive. 1
Example In the example involving a deck of cards, a natural sample space is
the set of all cards. A card is a result of a trial. The event of our interest was the
set of aces. The probability of each elementary event is 1/52 and the probabilities
of other events can be computed based on this and the Kolmogorov axioms.

3.3 Properties of Probability Spaces

Probability spaces and measures have some useful properties, which can be proved
from the three axioms only. To present (and prove) these properties one must
rely mathematical abstraction of the probability spaces offered by their definition.
However, as we are more interested in their interpretation, a little dictionary is
needed: In the following E and F are events and E c denotes the complement of E.
1
More precisely the probability space is defined as a three-tuple (S, F, P ), where F a set of
subsets of S satisfying certain assumptions (s.c. Borel-field) where the mapping P is defined.
The definitions as precise as this are not needed in this course. The axiomatic definition of the
probability was conceived because there are certain problems with the classical definition utilized
in the examples of the previous section. (e.g. Bertnand’s paradox).
3.4 Random Variables and Random Vectors 11

set theory probability

Ec E does not happen
E∪F E or F happens
E∩F E and F both happen

Note that often P (E ∩ F ) is abbreviated as P (E, F ).

And some properties:

Theorem 1 Let E and F be events of the sample space S. The probability measure
P of S has the following properties:

1. P (E c ) = 1 − P (E)

2. P (∅) = 0

3. If E is a sub-event of F :n (E ⊆ F ), then P (F − E) = P (F ) − P (E)

4. P (E ∪ F ) = P (E) + P (F ) − P (E ∩ F )

If the events F1 , . . . , Fn are mutually exclusive and S = F1 ∪ · · · ∪ Fn , then

5. P (E) = ni=1 P (E ∩ Fi ).
P

A collection of events as in the point 5 is called a partition of S. Note that an

event F and its complement F c always form a partition. The proofs of the above
theorem are simple and they are skipped in these notes. The Venn diagrams can be
useful if the logic behind the theorem seems fuzzy.

3.4 Random Variables and Random Vectors

Random variables characterize the concept of random experiment in probability
theory. In pattern recognition , the aim is to assign an object to a correct class
based on the observations about the object. The observations about an object from
a certain class can be thought as a trial.
Formally, a random variable (RV) is defined as a real-valued function X defined
on a sample space S, i.e. X : S → R. The set of values {X(x) : x ∈ S} taken on by
X is called the co-domain or the range of X.
Example Recall our first example, where we wanted to compute the probability
of drawing an ace from a deck of cards. The sample space consisted of the labels
of the cards. A relevant random variable in this case is a map X : labels of cards
→ {0, 1}, where X(ace) = 1 and X(other cards) = 0. When rolling two dice, a
suitable RV is e.g. a map from the sample space to the sum of two dice.
These are only examples of relevant RVs. There is no unique RV that can be
related to a certain experiment. Instead there exist several possibilities, some of
which are better than others.
12 Introduction to Pattern Recognition

Random vectors are generalizations of random variables to multivariate situa-

tions. Formally, these are defined as maps X : S → Rd , where d is an integer.
For random vectors, the definition of co-domain is a natural extension of the defi-
nition of co-domain for random variables. In pattern recognition, the classification
seldom succeeds if we have just one observation/measurement/feature in our dis-
posal. Therefore, we are mainly concerned with random vectors. From now on,
we do not distinguish between a single variate random variable and a random vec-
tor. Both are called random variables unless there is a special need to emphasize
multidimensionality.
We must still connect the concept of the random variable to the concept of
the probability space. The task of random variables is to transfer the probabilities
defined for the subsets of the sample space to the subsets of the Euclidean space:
The probability of a subset A of the co-domain of X is equal to the probability of
inverse image of A in the sample space. This makes it possible to model probabilities
with the probability distributions.
Often, we can just ’forget’ about the sample space and model the probabilities
via random variables. The argument of the random variable is often dropped from
the notation, i.e. P (X ∈ B) is equivalent to P ({s : X(s) ∈ B, s ∈ S, B ⊆ Rd })2 .
Note that a value obtained by a random variable reflects a result of a single
trial. This is different than the random variable itself. These two should under no
circumstances be confused although the notation may give possibility to do that.

3.5 Probability Densities and Cumulative Distri-

butions
3.5.1 Cumulative Distribution Function (cdf )
First we need some additional notation: For vectors x = [x1 , . . . , xd ]T and y =
[y1 , . . . , yd ]T notation x < y is equivalent to x1 < y1 and x2 < y2 and · · · and
xd < yd . That is, if the components of the vectors are compared pairwise then the
components of x are all smaller than components of y.
The cumulative distribution function FX (cdf) of an RV X is defined for all x as

FX (x) = P (X ≤ x) = P ({s : X(s) ≤ x}). (3.1)

The cdf measures the probability mass of all y that are smaller than x.
Lets once again consider the deck of cards and the probability of drawing an ace.
The RV in this case was the map from the labels of the cards to the set {0, 1} with
the value of the RV equal to 1 in the case of a ace and 0 otherwise. Then FX (x) = 0,
when x < 0, FX (x) = 48/52, when 0 ≤ x < 1, and FX (x) = 1 otherwise.

Theorem 2 The cdf FX has the following properties:

2
For a good and more formal treatment of the issue, see E.R. Dougherty: Probability and
Statistics for the Engineering, Computing and Physical Sciences.
3.5 Probability Densities and Cumulative Distributions 13

1. FX is increasing, i.e. if x ≤ y, then FX (x) ≤ FX (y).

2. limx→−∞ FX (x) = 0, limx→∞ FX (x) = 1.

3.5.2 Probability Densities: Discrete Case

An RV X is discrete if its co-domain is denumerable (The elements denumerable set
can be written as a list x1 , x2 , x3 , . . .. Then list may be finite or infinite).
The probability density function (pdf) pX of a discrete random variable X is
defined as

P (X = x) if x belongs to codomain ofX
pX (x) = . (3.2)
0 otherwise

In other words, a value of the pdf pX (x) is the probability of the event X = x. If
an RV X has a pdf pX , we say that X is distributed according to pX and we denote
X ∼ pX .

3.5.3 Probability Densities: Continuous Case

An RV X is continuous if there exists such function pX that
Z x
FX (x) = pX (y)dy. (3.3)
−∞

the function pX is then the probability density function of X. In Figure 3.1, the pdf
and the cdf of the standard normal distribution are depicted.

Figure 3.1: The pdf and the cdf of the standard normal distribution.

The probability that the occurrence x of an RV X belongs to a subset B of Rd

is Z
P (X ∈ B) = pX (y)dy.
B

In the continuous case, the value of pdf pX (x) is not a probability: The value of any
integral over a single point is zero and therefore it is not meaningful to study the
14 Introduction to Pattern Recognition

value of the probability of occurrence of a single value of an RV in the continuous

case.
On integrals.

• During this course, it is assumed that the integrals are multivariate Riemann
integrals.

• The integral Z x
pX (y)dy
−∞

is an abbreviation for
Z x1 Z xd
··· pX (y1 , y2 , . . . , yd )dy1 · · · dyd ,
−∞ −∞

where x = [x1 , . . . , xd ]T and y = [y1 , . . . , yd ]T

• And most importantly: you do not have to evaluate any integrals in practical
pattern recognition. However, the basic properties of integrals are useful to
know.

The following theorem characterizes the continuous pdfs:

Theorem 3 A function p : Rd → R is the probability density for a continuous RV

if and only if

1. p(x) ≥ 0 for all x.

R∞
2. −∞ p(x)dx = 1.

The above theorem is a direct consequence of the Kolmogorov axioms. A similar

theorem is valid for the discrete case. Also, the pdf (if it exists) defines uniquely the
random variable. This means that in most cases all we need is the knowledge about
the pdf of an RV.
Finally, the probability density function is the derivative of the cumulative dis-
tribution function if the cdf is differentiable. In the multivariate case, the pdf is
obtained based on the cdf by differentiating the cdf with respect to all the compo-
nents of the vector valued variate.

3.6 Conditional Distributions and Independence

3.6.1 Events
Assume that the events E and F relate to the same experiment. The events E ⊂ S
and F ⊂ S are (statistically) independent if

P (E ∩ F ) = P (E)P (F ). (3.4)
3.6 Conditional Distributions and Independence 15

If E and F are not independent, they are dependent. The independence of events E
and F means that an occurrence of E does not affect the likelihood of an occurrence
of F in the same experiment. Assume that P (F ) 6= 0. The conditional probability
of E relative to F ⊆ S is denoted by P (E|F ) and defined as

P (E ∩ F )
P (E|F ) = . (3.5)
P (F )

This refers to the probability of the event E provided that the event F has already
occurred. In other words, we know that the result of a trial was included in the
event F .
If E and F are independent, then

P (E ∩ F ) P (E)P (F )
P (E|F ) = = = P (E).
P (F ) P (F )

Accordingly, if P (E|F ) = P (E), then

P (E ∩ F ) = P (E)P (F ), (3.6)

and events E and F are independent. The point is that an alternative definition of
independence can be given via conditional probabilities. Note also that the concept
of independence is different from the concept of mutual exclusivity. Indeed, if the
events E and F are mutually exclusive, then P (E|F ) = P (F |E) = 0, and E and F
are dependent.
Definition (3.4) generalizes in straight-forward manner: Events E1 , . . . , En are
independent if
P (E1 ∩ · · · ∩ En ) = P (E1 ) · · · P (En ). (3.7)
Assume that we have more than two events, say E, F and G. They are pairwise
independent if E and F are independent and F and G are independent and E and G
are independent. However, it does not follow from the pairwise independence that
E, F and G would be independent.

3.6.2 Random Variables

Let us now study the random variables X1 , X2 , . . . , Xn related to the same experi-
ment. We define a random variable

X = [X1T , X2T , . . . , XnT ]T .

(X is a vector of vectors). The pdf

pX (x) = p(X1 ,...,Xn ) (x1 , x2 , . . . , xn )

is the joint probability density function of the RVs X1 , X2 , . . . , Xn . The value

p(X1 ,...,Xn ) (x1 , x2 , . . . , xn )

16 Introduction to Pattern Recognition

refers to the probability density of X1 receiving the value x1 , X2 receiving the value
x2 and so forth in a trial. The pdfs pXi (xi ) are called the marginal densities of X.
These are obtained from the joint density by integration. The joint and marginal
distributions have all the properties of pdfs.
The random variables X1 , X2 , . . . , Xn are independent if and only if

pX (x) = p(X1 ,...,Xn ) (x1 , x2 , . . . , xn ) = pX1 (x1 ) · · · pXn (xn ). (3.8)

Let Ei i = 1, . . . , n and Fj , j = 1, . . . , m be events related to X1 and X2 , respec-

tively. Assume that X1 and X2 are independent, Then also all the events related to
the RVs X1 and X2 are independent, i.e.

P (Ei ∩ Fj ) = P (Ei )P (Fj ),

for all i and j. That is to say that the results of the (sub)experiments modeled by
independent random variables are not dependent on each other. The independence
of the two (or more) RVs can be defined with the help of the events related to the
RVs. This definition is equivalent to the definition by the pdfs.
It is natural to attach the concept of conditional probability to random variables.
We have RVs X and Y . A new RV modeling the probability of X assuming that
we already know that Y = y is denoted by X|y. It is called the conditional random
variable of X given y. The RV X|y has a density function defined by

pX,Y (x, y)
pX|Y (x|y) = . (3.9)
pY (y)

RVs X and Y are independent if and only if

pX|Y (x, y) = pX (x)

for all x, y.

3.7 Bayes Rule

This section is devoted to a very important statistical result to pattern recognition.
This Bayes Rule was named after Thomas Bayes (1702 - 1761). This rule has versions
concerning events and RVs, which both are represented. The derivations here are
carried out only for the case concerning events.
We shall now derive the Bayes rule. Multiplying by P (F ) the both sides of the
definition of the conditional probability for events E and F yields

P (E ∩ F ) = P (F )P (E|F ).

For this we have assumed that P (F ) > 0. If also P (E) > 0, then

P (E ∩ F ) = P (E)P (F |E).
3.8 Multivariate Normal Distribution 17

Combining the two equations yields

P (F )P (E|F ) = P (E ∩ F ) = P (E)P (F |E).

And furthermore
P (F )P (E|F )
P (F |E) = .
P (E)
This is the Bayes rule. A different formulation of the rule is obtained via the point
5 in the Theorem 1. If the events F1 , . . . , Fn are mutually exclusive and the sample
space S = F1 ∪ · · · ∪ Fn , then for all k it holds that
P (Fk )P (E|Fk ) P (Fk )P (E|Fk )
P (Fk |E) = = Pn .
P (E) i=1 P (Fi )P (E|Fi )

We will state these results and the corresponding ones for the RVs in a theorem:

Theorem 4 Assume that E, F, F1 , . . . , Fn are events of S. Moreover F1 , . . . , Fn

form a partition of S, and P (E) > 0, P (F ) > 0. Then
P (F )P (E|F )
1. P (F |E) = P (E)
;

2. P (Fk |E) = PnP (Fk )P (E|Fk ) .

i=1 P (Fi )P (E|Fi )

Assume that X and Y are RVs related to the same experiment. Then
pX (x)pY |X (y|x)
3. pX|Y (x|y) = pY (y)
;
pX (x)pY |X (y|x)
4. pX|Y (x|y) = R
pX (x)pY |X (y|x)dx
.

Note that the points 3 and 4 hold even if one of the RVs was discrete and the
other was continuous. Obviously, it can be necessary to replace the integral by a
sum.

3.8 Multivariate Normal Distribution

The multivariate normal distribution is that typical also in the pattern recognition
that it warrants a more detailed study. First we need the definition for positive
definite matrices: Symmetric d × d matrix A is positive-definite, if for all non-zero
x ∈ Rd it holds that
xT Ax > 0.
A d-component RV X ∼ N (µ, Σ) distributed according the normal distribution has
a pdf
1 1
pX (x) = p exp[− (x − µ)T Σ−1 (x − µ)], (3.10)
(2π)d/2 det(Σ) 2
where the parameter µ is a d component vector (the mean of X) and Σ is a d × d
positive definite matrix (the covariance of X).
18 Introduction to Pattern Recognition

Figure 3.2: The pdf of a 2-dimensional normal distribution

In the univariate case, the pdf is

1 1 x−µ 2
pX (x) = √ exp[− ( ) ].
2πσ 2 σ

The normal distribution has some interesting properties that make it special
among the distributions of continuous random variables. Some of them are worth
of knowing for this course. In what follows, we denote the components of µ by µi
and the components of Σ by σij . If X ∼ N (µ, Σ) then

• sub-variables X1 , . . . , Xd of X are normally distributed: Xi ∼ N (µi , σii );

• sub-variables Xi and Xj of X are independent if and only if σij = 0.

• Let A and B be non-singular d × d matrices. Then AX ∼ N (Aµ, AΣAT ) and

RVs AX and BX are independent if and only if AΣB T is the zero-matrix.

• The sum of two normally distributed RVs is normally distributed.

Chapter 4

Bayesian Decision Theory and

Optimal Classifiers

The Bayesian decision theory offers a solid foundation for classifier design. It tells
us how to design the optimal classifier when the statistical properties of the classi-
fication problem are known. The theory is a formalization of some common-sense
principles, but it offers a good basis for classifier design.
About notation. In what follows we are a bit more sloppy with the notation
than previously. For example, pdfs are not anymore indexed with the symbol of the
random variable they correspond to.

4.1 Classification Problem

Feature vectors x that we aim to classify belong to the feature space. The feature
space is usually d-dimensional Euclidean space Rd . However, the feature space can
be for instance {0, 1}d , the space formed by the d-bit binary numbers. We denote
the feature space with the symbol F.
The task is to assign an arbitrary feature vector x ∈ F to one of the c classes1
The c classes are denoted by ω1 , . . . , ωc and these are assumed to be known. Note
that we are studying a pair of random variables (X, ω), where X models the value
of the feature vector and ω models its class.
We know the

1. prior probabilities P (ω1 ), . . . , P (ωc ) of the classes and

2. the class conditional probability density functions p(x|ω1 ), . . . , p(x|ωc ).

1
To be precise, the task is to assign the object that is represented by the feature vector to
a class. Hence, what we have just said is not formally true. However, it is impossible for the
classifier to distinguish between two objects with the same feature vector and it is reasonable to
speak about the feature vectors instead of the object they represent. This is also typical in the
literature. In some cases, this simplification of the terminology might lead to misconceptions and
in these situations we will use more formal terminology.
20 Introduction to Pattern Recognition

The prior probability P (ωi ) defines what percentage of all feature vectors belong
to the class ωi . The class conditional pdf p(x|ωi ), as it is clear from the notation,
defines the pdf of the feature vectors belonging to ωi . This is same as the density
of X given that the class is ωi . To be precise: The marginal distribution of ω is
known and the distributions of X given ωi are known for all i. The general laws of
probability obviously hold. For example,
c
X
P (ωi ) = 1.
i=1

We denote a classifier by the symbol α. Because the classification problem is

to assign an arbitrary feature vector x ∈ F to one of c classes, the classifier is a
function from the feature space onto the set of classes, i.e. α : F → {ω1 , . . . , ωc }.
The classification result for the feature vector x is α(x). We often call a function α
a decision rule.

Figure 4.1: Decision regions

The consideration of the partition of the feature space generated by a classifier

leads to another interpretation for the classifier. Every feature vector is assigned
to a unique class by a classifier and therefore the classifier defines those feature
vectors that belong to a certain class, see Figure 4.1. The parts of the feature space
corresponding to classes ω1 , . . . , ωc are denoted by R1 , . . . , Rc , respectively. The sets
Ri are called decision regions. Clearly,
Ri = {x : α(x) = ωi }.
The decision regions form a partition of the feature space (i.e. Ri ∩ Rj = ∅ when
i 6= j and R1 ∪ · · · ∪ Rc = F) As the decision rule defines the decision regions,
4.2 Classification Error 21

the decision regions define the decision rule: The classifier can be represented by
its decision regions. However, especially for large d, the representation by decision
regions is not convenient. However, the representation by decision regions often
offers additional intuition on how the classifier works.

4.2 Classification Error

The classification problem is by nature statistical. Objects having same feature
vectors can belong to different classes. Therefore, we cannot assume that it would
be possible to derive a classifier that works perfectly. (The perfect classifier would
assign every object to the correct class). This is why we must study the classification
error: For a specified classification problem our task to derive classifier that makes
as few errors (misclassifications) as possible. We denote the classification error by
the decision rule α by E(α). Because the classifier α assigns an object to a class
only based on the feature vector of that object, we study the probability E(α(x)|x)
that objects with the feature vector x are misclassified.
Let us take a closer look into the classification error and show that it can be
computed if the preliminary assumptions of the previous section hold. From item 1
in Theorem 1, it follows that
E(α(x)|x) = 1 − P (α(x)|x), (4.1)
where P (α(x)|x) is the probability that the class α(x) is correct for the object. The
classification error for the classifier α can be written as
Z Z
E(α) = E(α(x)|x)p(x)dx = [1 − P (α(x)|x)]p(x)dx. (4.2)
F F

We need to still know P (α(x)|x). From the Bayes rule we get

p(x|α(x))P (α(x))
P (α(x)|x) = . (4.3)
p(x)
And hence the classification error is
Z
E(α) = 1 − p(x|α(x))P (α(x))dx, (4.4)
F

where p(x|α(x)), P (α(x)) are known for every x as it was defined in the previous
section. Note that
p(x|α(x))P (α(x)) = p(x, α(x)).
That is, the classification error of α is equal to the probability of the complement
of the events {(x, α(x))}.
With the knowledge of decision regions, we can rewrite the classification error as
Xc Z c Z
X
E(α) = [1 − p(x|ωi )P (ωi )]dx = 1 − p(x|ωi )P (ωi )dx, (4.5)
i=1 Ri i=1 Ri

where Ri , i = 1, . . . , c are decision regions for the decision rule α.

22 Introduction to Pattern Recognition

4.3 Bayes Minimum Error Classifier

In this section, we study Bayes minimum error classifier which, provided that the
assumptions of section 4.1 hold, minimizes the classification error. (The assumptions
were that class conditional pdfs and prior probabilities are known.) This means that
Bayes minimum error classifier, later on just Bayes classifier, is optimal for a given
classification problem.
The Bayes classifier is defined as

αBayes (x) = arg max P (ωi |x). (4.6)

ωi ,i=1,...,c

Notation arg maxx f (x) means the value of the argument x that yields the maximum
value for the function f . For example, if f (x) = −(x − 1)2 , then arg maxx f (x) = 1
(and max f (x) = 0). In other words,

the Bayes classifier selects the most probable class when the observed
feature vector is x.

The posterior probability P (ωi |x) is evaluated based on the Bayes rule, i.e. P (ωi |x) =
p(x|ωi )P (ωi )
p(x)
. Note additionally that p(x) is equal for all classes and p(x) ≥ 0 for all
x and we can multiply the right hand side of (4.6) without changing the classifier.
The Bayes classifier can be now rewritten as

αBayes (x) = arg max p(x|ωi )P (ωi ). (4.7)

ωi ,i=1,...,c

If two or more classes have the same posterior probability given x, we can freely
choose between them. In practice, the Bayes classification of x is performed by
computing p(x|ωi )P (ωi ) for each class ωi , i = 1, . . . , c and assigning x to the class
with the maximum p(x|ωi )P (ωi ).
By its definition the Bayes classifier minimizes the conditional error E(α(x)|x) =
1 − P (α(x)|x) for all x. Because of this and basic properties of integrals, the Bayes
classifier minimizes also the classification error:

Theorem 5 Let α : F → {ω1 , . . . , ωc } be an arbitrary decision rule and αBayes :

F → {ω1 , . . . , ωc } the Bayes classifier. Then

E(αBayes ) ≤ E(α).

The classification error E(αBayes ) of the Bayes classifier is called the Bayes error.
It is the smallest possible classification error for a fixed classification problem. The
Bayes error is Z
E(αBayes ) = 1 − max[p(x|ωi )P (ωi )]dx (4.8)
F i

Finally, note that the definition of the Bayes classifier does not require the as-
sumption that the class conditional pdfs are Gaussian distributed. The class condi-
tional pdfs can be any proper pdfs.
4.4 Bayes Minimum Risk Classifier 23

4.4 Bayes Minimum Risk Classifier

The Bayes minimum error classifier is a special case of the more general Bayes
minimum risk classifier. In addition to the assumptions and terminology introduced
previously, we are given a actions α1 , . . . , αa . These correspond to the actions that
follow the classification. For example, a bottle recycling machine, after classifying
the bottles and cans, places them in the correct containers and returns a receipt to
the customer. The actions are tied to the classes via a loss function λ. The value
λ(αi |ωj ) quantifies the loss incurred by taking an action αi when the true class is ωj .
The decision rules are now functions from the feature space onto the set of actions.
We need still one more definition. The conditional risk of taking the action αi when
the observed feature vector is x is defined as
c
X
R(αi |x) = λ(αi |ωj )P (ωj |x). (4.9)
j=1

The definition of the Bayes minimum risk classifier is then:

The Bayes minimum risk classifier chooses the action with the minimum
conditional risk.

The Bayes minimum risk classifier is the optimal in this more general setting: it
is the decision rule that minimizes the total risk given by
Z
Rtotal (α) = R(α(x)|x)p(x)dx. (4.10)

The Bayes classifier of the previous section is obtained when the action αi is the
classification to the class ωi and the loss function is zero-one loss:

0 if i = j
λ(αi |ωj ) = .
1 if i 6= j

The number of actions a can be different from the number of classes c. This
is useful e.g. when it is preferable that the pattern recognition system is able to
separate those feature vectors that cannot be reliably classified.
An example about spam filtering illustrates the differences between the minimum
error and the minimum risk classification. An incoming e-mail is either a normal
(potentially important) e-mail (ω1 ) or a junk mail (ω2 ). We have two actions: α1
(keep the mail) and α2 (put the mail to /dev/null). Because losing a normal e-mail
is on average three times more painful than getting a junk mail into the inbox, we
select a loss function
λ(α1 |ω1 ) = 0 λ(α1 |ω2 ) = 1
.
λ(α1 |ω2 ) = 3 λ(α2 |ω2 ) = 0
(You may disagree about the loss function.) In addition, we know that P (ω1 ) =
0.4, P (ω2 ) = 0.6. Now, an e-mail has been received and its feature vector is x. Based
on the feature vector, we have computed p(x|ω1 ) = 0.35, p(x|ω2 ) = 0.65.
24 Introduction to Pattern Recognition

The posterior probabilities are:

0.35 · 0.4
P (ω1 |x) = = 0.264;
0.35 · 0.4 + 0.65 · 0.6
P (ω2 |x) = 0.736.
For the minimum risk classifier, we must still compute the conditional losses:
R(α1 |x) = 0 · 0.264 + 0.736 = 0.736,
R(α2 |x) = 0 · 0.736 + 3 · 0.264 = 0.792.
The Bayes (minimum error) classifier classifies the e-mail as spam but the min-
imum risk classifier still keeps it in the inbox because of the smaller loss of that
action.
After this example, we will concentrate only on the minimum error classification.
However, many of the practical classifiers that will be introduced generalize easily
to the minimum risk case.

4.5 Discriminant Functions and Decision Surfaces

We have already noticed that the same classifier can be represented in several dif-
ferent ways, e.g. compare (4.6) and (4.7) which define the same Bayes classifier.
As always, we would like this representation be as simple as possible. In general, a
classifier can be defined with the help of a set of discriminant functions. Each class
ωi has its own discriminant function gi that takes a feature vector x as an input.
The classifier assigns the feature vector x to the class with the highest gi (x). In
other words, the class is ωi if
gi (x) > gj (x)
for all i 6= j.
For the Bayes classifier, we can select discriminant functions as:

gi (x) = P (ωi |x), i = 1, . . . , c

or as
gi (x) = p(x|ωi )P (ωi ), i = 1, . . . , c.
The word ’select’ appears above for a reason: Discriminant functions for a certain
classifier are not uniquely defined. Many different sets of discriminant functions may
define exactly the same classifier. However, a set of discriminant functions define a
unique classifier (but not uniquely). The next result is useful:
Theorem 6 Let f : R → R increasing, i.e. f (x) < f (y) always when x < y. Then
the following sets of discriminant functions
gi (x), i = 1, . . . , c
and
f (gi (x)), i = 1, . . . , c
define the same classifier.
4.5 Discriminant Functions and Decision Surfaces 25

This means that discriminant functions can be multiplied by positive constants,

constants can be added to them or they can be replaced by their logarithms without
altering the underlying classifier.
Later on, we will study discriminant functions of the type
gi (x) = wiT x + wi0 .
These are called linear discriminant functions. Above wi is a vector of the equal
dimensionality with feature vector and wi0 is a scalar. A linear classifier is a clas-
sifier that can be represented by using only linear discriminant functions. Linear
discriminant functions can be conceived as the most simple type of discriminant
functions. This explains the interest in linear classifiers.
When the classification problem has just two classes, we can represent the clas-
sifier using a single discriminant function
g(x) = g1 (x) − g2 (x). (4.11)
If g(x) > 0, then x is assigned to ω1 and otherwise x is assigned to ω2 .
The classifier divides the feature space into parts corresponding classes. The
decision regions can be defined with the help of discriminant functions as follows:
Ri = {x| gi (x) > gj (x), i 6= j}.
The boundaries between decision regions
Rij = {x : gi (x) = gj (x), gk (x) < gi (x), k 6= i, j}
are called decision surfaces. It is easier to grasp the geometry of the classifier (i.e.
the partition of the feature space by the classifier) based on decision surfaces than
based on decision regions directly. For example, it easy to see that the decision
surfaces of linear classifiers are segments of lines between the decision regions.

Example. Consider a two-class classification problem with P (ω1 ) = 0.6, P (ω2 ) =

0.4, p(x|ω1 ) = √12π exp[−0.5x2 ] and p(x|ω2 ) = √12π exp[−0.5(x − 1)2 ]. Find the deci-
sion regions and boundaries for the minimum error rate Bayes classifier.
The decision region R1 is the set of points x for which P (ω1 |x) > P (ω2 |x). The
decision region R2 is the set of points x for which P (ω2 |x) > P (ω1 |x). The decision
boundary is the set of points x for which P (ω2 |x) = P (ω1 |x).
Begin with the decision boundary:
P (ω1 |x) = P (ω2 |x) ⇒ p(x|ω1 )P (ω1 ) = p(x|ω2 )P (ω2 ),
where we used the Bayes formula and multiplied with p(x). Continuing
p(x|ω1 )P (ω1 ) = p(x|ω2 )P (ω2 ) ⇒ ln[p(x|ω1 )P (ω1 )] = ln[p(x|ω2 )P (ω2 )]
⇒ −x2 /2 + ln 0.6 = −(x − 1)2 /2 + ln 0.4 ⇒ x2 − 2 ln 0.6 = x2 − 2x + 1 − 2 ln 0.4.
The decision boundary is x∗ = 0.5 + ln 0.6 − ln 0.4 ≈ 0.91. The decision regions are
(see Figure 4.2)
R1 = {x : x < x∗ }, R2 = {x : x > x∗ }.
26 Introduction to Pattern Recognition

Figure 4.2: Class conditional densities (top) and P (ω1 |x) and P (ω2 |x) and decision
regions (bottom).

4.6 Discriminant Functions for Normally Distributed

Classes
The multivariate normal distribution is the most popular model for class conditional
pdfs. There are two principal reasons for this. First is due to its analytical tractabil-
ity. Secondly, it offers a good model for the case where the feature vectors from a
certain class can be thought as noisy versions of some average vector typical to the
class. In this section, we study the discriminant functions, decision regions and de-
cision surfaces of the Bayes classifier when the class conditional pdfs are modeled by
normal densities. The purpose of this section is to illustrate the concepts introduced
and hence special attention should be given to the derivations of the results.
The pdf of the multivariate normal distribution is
1 1
pnormal (x|µ, Σ) = p exp[− (x − µ)T Σ−1 (x − µ)], (4.12)
(2π)d/2 det(Σ) 2

where µ ∈ Rd and positive definite d × d matrix Σ are the parameters of the distri-
bution.
We assume that the class conditional pdfs are normally distributed and prior
probabilities can be arbitrary. We denote the parameters for the class conditional
pdf for ωi by µi , Σi i.e. p(x|ωi ) = pnormal (x|µi , Σi ).
We begin with the Bayes classifier defined in Eq. (4.7). This gives the discrimi-
4.6 Discriminant Functions for Normally Distributed Classes 27

nant functions
gi (x) = pnormal (x|µi , Σi )P (ωi ). (4.13)
By replacing these with their logarithms (c.f. Theorem 6) and substituting the
normal densities, we obtain2
1 d 1
gi (x) = − (x − µi )T Σ−1
i (x − µi ) − ln 2π − ln det(Σi ) + ln P (ωi ). (4.14)
2 2 2
We distinguish three distinct cases:
1. Σi = σ 2 I, where σ 2 is a scalar and I is the identity matrix;
2. Σi = Σ, i.e. all classes have equal covariance matrices;
3. Σi is arbitrary.

Case 1
Dropping the constants from the right hand side of Eq. (4.14) yields
1 ||x − µi ||2
gi (x) = − (x − µi )T (σ 2 I)−1 (x − µi ) + ln P (ωi ) = − + ln P (ωi ), (4.15)
2 2σ 2
The symbol || · || denotes the Euclidean norm, i.e.
||x − µi ||2 = (x − µi )T (x − µi ).
Expanding the norm yields
1
gi (x) = − (xT x − 2µTi x + µTi µi ) + ln P (ωi ). (4.16)
2σ 2
The term xT x is equal for all classes so it can be dropped. This gives
1 T 1
gi (x) = 2
(µi x − µTi µi ) + ln P (ωi ), (4.17)
σ 2
which is a linear discriminant function with
µi
wi = 2 ,
σ
and
−µTi µ
wi0 = + P (ωi ).
2σ 2
Decision regions in the case 1 are illustrated in Figure 4.3 which is figure 2.10
from Duda, Hart and Stork.
An important special case of the discriminant functions (4.15) is obtained when
P (ωi ) = 1c for all i. Then x is assigned to the class with the mean vector closest to
x:. This is called minimum distance classifier. It is a linear classifier as a special
case of a linear classifier.
2
Note that the discriminant functions gi (x) in (4.13) and (4.14) are not same functions. gi (x)
is just a general symbol for the discriminant functions of the same classifier. However, a more
prudent notation would be a lot messier.
28 Introduction to Pattern Recognition

Figure 4.3: Case 1

Case 2
We obtain a linear classifier even if the features are dependent, but the covariance
matrices for all classes are equal (i.e. Σi = Σ). It can be represented by the
discriminant functions
gi (x) = wiT x + wi0 , (4.18)
where
wi = Σ−1 µi
and
1
wi0 = − µTi Σ−1 µi + ln P (ωi ).
2
Decision regions in the case 2 are illustrated in Figure 4.4 which is figure 2.11
from Duda, Hart and Stork.

Case 3
Now no additional assumptions about the class conditional pdfs are made. In this
case, the discriminant functions

1 d 1
gi (x) = − (x − µi )T Σ−1
i (x − µi ) − ln 2π − ln det(Σi ) + ln P (ωi ).
2 2 2

cannot be simplified much further. Only the constant term d2 ln 2π can be dropped.
Discriminant functions are not anymore linear but quadratic. They have much more
complicated decision regions than the linear classifiers of the two previous cases.
Now, decision surfaces are also quadratic and the decision regions do not have to be
even connected sets.
Decision regions in the case 3 are illustrated in Figure 4.5 which is Figure 2.14
from Duda, Hart and Stork.
4.6 Discriminant Functions for Normally Distributed Classes 29

Figure 4.4: Case 2

30 Introduction to Pattern Recognition

Figure 4.5: Case 3

4.7 Independent Binary Features 31

4.7 Independent Binary Features

We consider still the case where the features are binary-valued. That is, their
value is either zero (or ’no’) or one (or ’yes’). The random variable representing
feature vectors is now discrete and so are the class conditional pdfs. The theory of
classification is not affected by this. Only, integrals need to be replaced by sums.
All the tricks that we have learned this far are applicable.
We make a further assumption that the features are independent and the class
conditional marginal density for each feature is the density of the Bernoulli distri-
bution. This yields
d
x
Y
P (x|ωi ) = qijj (1 − qij )(1−xj ) , (4.19)
j=1

where qij , j = 1, . . . , d are parameters for the class conditional density of the class
ωi . The distribution can be interpreted in the way that jth feature answers ’yes’
to the question asked with the probability of qij . (Compare to the coin tossing
experiment.) It is still worth emphasizing that the features are independent from
each other, i.e. the answer to the jth question does not depend on the answers to
the other questions. For a 2-class pattern recognition problem, if q1j > q2j the value
of the object’s jth feature is 1 more often if the object belongs to the class ω1 .
We now show that the Bayes classifier is linear for this problem. For the Bayes
classifier
αBayes (x) = arg max P (ωi |x),
ωi ,i=1,...,c

discriminant functions are given as

gi (x) = P (x|ωi )P (ωi ).

Substituting Eq. (4.19) to the above equation gives

d
x
Y
gi (x) = qijj (1 − qij )(1−xj ) P (ωi ). (4.20)
j=1

Replacing the right hand side by its logarithm and simplifying give
d
X
gi (x) = [xj ln(1 − qij ) + ln qij − xj ln(1 − qij )] + ln P (ωi ). (4.21)
j=1

These are linear discriminant functions

d
X
gi (x) = wiT x + wi0 = wij xj + wi0 ,
j=1

where
qij
wij = ln ,
1 − qij
32 Introduction to Pattern Recognition

and
d
X
wi0 = ln(1 − qij ) + ln P (ωj ).
j=1

In the case of just two classes, the discriminant function is

d d
X q1j (1 − q2j ) X (1 − q1j ) P (ω1 )
g(x) = g1 (x) − g2 (x) = ln xj + ln + ln .
j=1
(1 − q 1j )q 2j j=1
(1 − q 2j ) P (ω2 )

We will end this section (and chapter) by studying how the value of the clas-
sification error behaves when we add more and more independent features. This
is somewhat more pleasant in the discrete case when we do not need to evaluate
integrals numerically. However, in order to the computations become possible for
large d, we must make some additional simplifying assumptions. Here the results
are more important than their derivation and we will keep derivations as brief as
possible.
The classification error is
X
E(αBayes ) = 1 − max P (x|ωi )P (ωi ), (4.22)
x

where x is summed over all d-bit binary vectors. Again, the greater the probability
of the most probable class the smaller the classification error. Consider the 2-class
case. We model the classes using Binomial densities, that is qij = qi for all features
j and
d!
P (x|ωi ) = P (k|qi ) = q k (1 − qi )d−k , i = 1, 2,
k! + (d − k)! i
where k is the number of 1s in the feature vector x. Let us still assume that the
prior probabilities are equal to 0.5 and q1 > q2 . The class of the feature vector is
decided by how many 1s it contains. The Bayes classifier is linear and it classifies x
to ω1 if the number of 1s in x exceeds a certain threshold t. The threshold t can be
computed from (4.21). The classification error is
t d
X d! k d−k
X d!
E(αBayes ) = 1− 0.5 q2 (1−q2 ) + 0.5 q1k (1−q1 )d−k .
k=0
k! + (d − k)! k=t+1
k! + (d − k)!

Note that we sum over all k and not over all x. (There are many fewer k than there
are x.)
The classification error with respect to d when q1 = 0.3, q2 = 0.2 is plotted
in Figure 4.6. Note that the classification error diminishes when the number of
features d grows. This is generally true for the Bayes classifiers. However, this does
not mean that other, non-optimal classifiers share this property. For example, if we
would have (for some reason) selected t = bd/2c (that does not lead to the Bayes
classifier), then we would obtain a classification error that would grow towards 0.5
with d. (bd/2c refers to the greatest integer that is lesser than equal to d/2.)
4.7 Independent Binary Features 33

Figure 4.6: Classification error when class conditional pdfs are Binomial.
34 Introduction to Pattern Recognition
Chapter 5

Supervised Learning of the Bayes

Classifier

5.1 Supervised Learning and the Bayes Classifier

In the previous chapter, we assumed that the class conditional pdfs p(x|ωi ) and the
prior probabilities P (ωi ) were known. In practice, this is never the case. Designing a
classifier based on the Bayes decision theory implies learning of the class conditional
pdfs and priors. In this chapter, we study supervised learning of class conditional
pdfs.
For supervised learning we need training samples. As we have defined in Chapter
2, training samples are feature vectors for which the correct classes are known. We
refer to the set of training samples as training data or as training set. In the training
set there are feature vectors from each class ω1 , . . . , ωc . We can re-arrange training
samples based on their classes: Training samples from the class ωi are denoted by
Di = {xi1 , . . . , xini },
where ni is the number of training samples from the class ωi . We assume that
the training samples in the sets Di are occurrences of the independent random
variables. (That is: they were measured from different objects.) Further, these
random variables are assumed to be distributed according to p(x|ωi ). In short, we
say that Di is independent and identically distributed (i.i.d.).
The training data may be collected in two distinct ways. These are meaningful
when we need to learn the prior probabilities. In mixture sampling, a set of objects
are randomly selected, their feature vectors are computed and then the objects are
hand-classified to the most appropriate classes. (This results pairs consisting of a
feature vector and the class of the feature vector, but these pairs can be re-arranged
to obtain representation as above for the training data.) In separate sampling, the
training data for each class is collected separately. For the classifier training, the
difference of these two sampling techniques is that based on the mixed sampling we
can deduce the priors P (ω1 ), . . . , P (ωc ) as
ni
P (ωi ) = Pc . (5.1)
j=1 nj
36 Introduction to Pattern Recognition

On the other hand, the prior probabilities cannot be deduced based on the separate
sampling, and in this case, it is most reasonable to assume that they are known.
The estimation of class conditional pdfs is more difficult. We need now to esti-
mate a function p(x|ωi ) based on a finite number of training samples. Next sections
are dedicated to different approaches to estimate the class conditional pdfs.

5.2 Parametric Estimation of Pdfs

5.2.1 Idea of Parameter Estimation
We study first an approach where the estimation of the class conditional pdfs p(x|ωi )
is reduced to the estimation of the parameter vector θi . To do this we must assume
that p(x|ωi ) belongs to some family of parametric distributions. For example, we
can assume that p(x|ωi ) is a normal density with unknown parameters θi = (µi , Σi ).
Of course, then it is necessary to justify the choice of the parametric family. Good
justifications are grounded e.g. on physical properties of the sensing equipment
and/or on properties of the feature extractor.
We denote
p(x|ωi ) = p(x|ωi , θi ),
where the class ωi defines the parametric family, and the parameter vector θi defines
the member of that parametric family. We emphasize that the parametric families
do not need to be same for all classes. For example, the class ω1 can be normally
distributed whereas ω2 is distributed according to the Cauchy distribution. Due
to independence of the training samples, the parameter vectors can be estimated
separately for each class. The aim is now to estimate the value of the parameter
vector θi based on the training samples

Di = {xi1 , . . . , xini }.

5.2.2 Maximum Likelihood Estimation

We momentarily drop the indices describing the classes to simplify the notation.
The aim is to estimate the value of the parameter vector θ based on the training
samples D = {x1 , . . . , xn }. We assume that the training samples are occurrences of
i.i.d. random variables distributed according to the density p(x|θ).
The maximum likelihood estimate or the ML estimate θ̂ maximizes the probability
of the training data with respect to θ. Due to the i.i.d. assumption, the probability
of D is
Yn
p(D|θ) = p(xi |θ).
i=1

Hence
n
Y
θ̂ = arg max p(xi |θ). (5.2)
i=1
5.2 Parametric Estimation of Pdfs 37

The function p(D|θ) - when thought as a function of θ - is termed the likelihood

function.
Often it is more convenient to work with the logarithm of the likelihood function
than with the likelihood function itself. The logarithm is an increasing function and
therefore we can safely replace the likelihood function with its logarithm without
affecting the maximization result. Hence,
n
X
θ̂ = arg max ln p(xi |θ). (5.3)
i=1
Pn
The function ln p(D|θ) = i=1 ln p(xi |θ) is called the log-likelihood function.

5.2.3 Finding ML-estimate

A pleasant feature of the ML-estimate is that we can find it, or at least search
for it, numerically. In some important cases, the ML-estimate can be determined
analytically. This can be done with a following procedure:

1. Derive the gradient of the likelihood function or the log-likelihood function.

2. Solve the zeros of the gradient and search also all the other critical points of
the (log)-likelihood function. (e.g. the points where at least one of the partial
derivatives of the function are not defined are critical in addition to the points
where gradient is zero.)

3. Evaluate the (log)-likelihood function at the critical points and select the crit-
ical point with the highest likelihood value as the estimate.

A word of warning: the ML-estimate does not necessarily exist for all distribu-
tions, for example the likelihood function could grow without limit. However, the
ML-estimates exist for simple parametric families.

5.2.4 ML-estimates for the Normal Density

We derive the ML-estimates for the normal density

1 1
p(x|θ) = p exp[− (x − µ)T Σ−1 (x − µ)].
(2π)d/2 det(Σ) 2

Consider first the mean vector, i.e. θ = µ. The covariance matrix Σ is assumed
to be known and fixed. (We will see that fixing the covariance will not affect the
estimate for mean.) The log-likelihood is
n
X n
X
l(θ) = ln p(xi |µ) = − 0.5(ln[(2π)d/2 det(Σ)] + (xi − µ)T Σ−1 (xi − µ)).
i=1 i=1
38 Introduction to Pattern Recognition

Its gradient is
n
X
∇l(µ) = − Σ−1 (xi − µ).
i=1

Set ∇l(µ) = 0. The matrix Σ is non-singular, and therefore Σ−1 is non-singular.

This implies that ∇l(µ) = 0 has a unique solution. The solution is
n
1X
µ̂ = xi . (5.4)
n i=1

It is easy to see that this is (the only) local maximum of the log-likelihood function,
and therefore it is the ML-estimate. Note that in this case, the ML-estimate is the
sample mean of D. As we already hinted, the exact knowledge of the covariance was
not necessary for obtaining the ML-estimate for the mean. Hence, the result holds
even if both µ and Σ were unknown.
The derivation of the ML-estimate for the covariance matrix is more complicated.
Hence, we will skip it and just present the result of the derivation. The ML -estimate
is n
1X
Σ̂ = (xi − µ̂)(xi − µ̂)T , (5.5)
n i=1
where µ̂ is the ML estimate for µ defined in Eq. (5.4).
A numeric example:

x_1 x_2 x_3 x_4 x_5

13 -115 119 33 -19
29 119 -4 17 73

mu = 6.2000
46.8000

sigma = 5762.6 -3212.2

-3212.2 1937.0

5.2.5 Properties of ML-Estimate

There exist variety of methods for the parameter estimation. Therefore, it is impor-
tant to know some basic properties of the ML-estimates. It is a favourable property
that the ML-estimate can be sought for numerically. In addition, it has some pleas-
ing theoretical properties, which we will briefly describe next.
The ML-estimates are asymptotically unbiased. Less formally, this means that
they give the correct result on average provided that there are infinitely many train-
ing samples available. To be more precise, the expected value of the ML-estimate
(over all training sets of size n) is equal to the correct parameter value when n
approaches infinity. If n is finite, the above result does not hold in general. For
example, the ML-estimate for the covariance matrix of the normal density is biased.
5.2 Parametric Estimation of Pdfs 39

When n approaches infinity, the ML-estimate is the best unbiased estimate in the
sense that the variance of the ML-estimate is the smallest possible. In addition,
the ML-estimate is asymptotically consistent. This means that the ML-estimate is
arbitrarily close to the true parameter value with the probability of 1 when n tends
to infinity.
Almost all optimality properties of the ML-estimate are so called large sample
properties. If n is small, little can be said about the ML-estimates in general.

5.2.6 Classification Example

We will illustrate the classifier training based on ML-estimation with Fisher’s Iris
data. We take the 10 first measurements from each species as training samples. We
model the class conditional pdfs by the normal density, and we assume that the
priors P (ωi ) = 13 for all three classes. We represent the training sets for all classes
in a matrix form: Each row corresponds to one training sample. For the class ω1
(iris setosa), the 10 training samples are:

5.1000 3.5000 1.4000 0.2000

4.9000 3.0000 1.4000 0.2000
4.7000 3.2000 1.3000 0.2000
4.6000 3.1000 1.5000 0.2000
5.0000 3.6000 1.4000 0.2000
5.4000 3.9000 1.7000 0.4000
4.6000 3.4000 1.4000 0.3000
5.0000 3.4000 1.5000 0.2000
4.4000 2.9000 1.4000 0.2000
4.9000 3.1000 1.5000 0.1000

We obtain based on these:

mu1 = 4.8600 3.3100 1.4500 0.2200

Sigma1 =
0.0764 0.0634 0.0170 0.0078
0.0634 0.0849 0.0155 0.0148
0.0170 0.0155 0.0105 0.0040
0.0078 0.0148 0.0040 0.0056

The same for the other classes:

mu2 = 6.1000 2.8700 4.3700 1.3800
Sigma2 =
0.4760 0.1740 0.2910 0.0720
0.1740 0.1041 0.1191 0.0384
0.2910 0.1191 0.2141 0.0584
0.0720 0.0384 0.0584 0.0256
40 Introduction to Pattern Recognition

mu3 = 6.5700 2.9400 5.7700 2.0400

Sigma3 =
0.5821 0.1252 0.4141 0.0782
0.1252 0.1024 0.1052 0.0794
0.4141 0.1052 0.3241 0.0752
0.0782 0.0794 0.0752 0.0764

And that is it! We can now classify e.g. the feature vector x = [ 5.9 4.2 3.0 1.5 ]T
by computing the values of the discriminant functions (4.13) for all three classes.
The classifier assigns x to ω2 , or in other words, the flower in the question is iris
versicolor.

5.3 Non-Parametric Estimation of Density Func-

tions
It is often the case in pattern recognition that assuming class conditional pdfs to be
members of a certain parametric family is not reasonable. The parameter estimation
approach of the previous section is then not possible. Instead, we must estimate
the class conditional pdfs non-parametrically. In non-parametric estimation (or
density estimation), we try to estimate p(x|ωi ) in each point x whereas in parametric
estimation we tried to estimate some unknown parameter vector. It is also possible
to estimate directly the posterior probabilities P (ωi |x) assuming that the training
data was collected using mixed sampling.
We formulate the basic problem of density estimation in the following way: We
are given the training data D = {x1 , . . . , xn } ,where the samples are i.i.d. , and are
all drawn from the unknown density p(x). The aim is to find an estimate p̂(x) for
p(x) in every point x.
We will consider two different methods for non-parametric estimation, Parzen
windows and k-nearest neighbors method. But before that, we introduce histograms
and consider density estimation in a general level.

5.3.1 Histogram
Histograms are the simplest approach to density estimation. The feature space is
divided into m equal sized cells or bins Bi . (See Figure 5.1.) Then, the number of
the training samples ni , i = 1, . . . n falling into each cell is computed. The density
estimate is
ni
p̂(x) = , when x ∈ Bi , (5.6)
Vn
where V is the volume of the cell. (All the cells have equal volume and an index is
not needed.)
The histogram estimate - with a definition as narrow as this one - is not a
good way to estimate densities, especially so when there are many (continuous-
5.3 Non-Parametric Estimation of Density Functions 41

Figure 5.1: Histogram. Aterisks (*) denote the training samples.

type) features. Firstly, the density estimates are discontinuous. Secondly, a good
size for the cells is hard to select.

5.3.2 General Formulation of Density Estimation

Let us generalize the concept of the histogram by allowing cells of varying shape
and size. Particularly, at the proximity of x, we are interested in how to select a
suitable cell size/shape producing an accurate p̂(x). Consider the estimate p̂(x) at
the point x and the set (neighborhood) B surrounding x. The probability that a
certain training sample xj is in B is
Z
P = p(x)dx. (5.7)
B

We still need the probability that k from n training samples fall into the set B.
The training samples are independent, and each of them is in the set B with the
probability P . Hence, we can apply the binomial density to the situation. That is,
there are exactly k samples in the set B with the probability
 
n
Pk = P k (1 − P )n−k , (5.8)
k
where  
n n!
=
k k!(n − k)!
is the number of different subsets of D of the size k.
Based on the Binomial theorem, we obtain the expected value of k. It is
n n  
X X n−1
E[k] = kPk = P k−1 (1 − P )n−k nP = nP. (5.9)
k−1
k=0 k=1
42 Introduction to Pattern Recognition

(We have skipped some algebra.) Now, we substitute E[k] by the observed number
k̂ of training samples falling into B. We obtain an estimate for P :

P̂ = k̂/n. (5.10)

If p(x) is continuous and B is small enough,

Z
p(x)dx ' p(x)V,
B

where V is the volume of B. This gives the estimate (or an interpretation for the
estimate)
R
k̂ p(x)dx
p̂(x) = ' BR . (5.11)
nV B
dx
We can draw the following conclusions:

• The obtained density estimate is a space averaged version of the true density.
The smaller the volume V the more accurate the estimate is. However, if n is
fixed, diminishing V will lead sooner or later to B which does not contain any
training samples and the density estimate will become useless.

• In more general level two fact need to be accepted. 1) The density estimate is
always a space averaged version of the true density. 2) The estimate k̂ is only
an estimate of nP when n is finite.

The principal question is how to select B and V . Let us assume that we have
unlimited number of training data. We can then approach this problem in the
following manner: To estimate p(x), we form a series of cells B1 , B2 , . . . containing
x. The cell B1 is used when we have just one training sample, the cell B1 is used
when we have two training samples and so-forth. The volume of Bn is denoted by
Vn and the number of training samples falling into Bn is denoted by kn . The density
estimate based on these is
kn
pn (x) = .
nVn
If we want that
lim pn (x) = p(x)
n→∞

then three conditions must be fulfilled

1. limn→∞ Vn = 0

2. limn→∞ kn = ∞

3. limn→∞ kn /n = 0.
5.4 Parzen Windows 43

The conditions may seem rather abstract but they are in fact natural. The first
condition assures us that the space averaged pn (x) will converge to p(x) if there
are enough training samples. The second one assures us that the frequency ratio
kn /n will be a good enough estimate for P . The third one states that the number
of samples falling to a region Bn is always a negligibly small portion of the total
number of samples. This is required if pn (x) is to converge at all.
How to guarantee the conditions in the limit of unlimited number of training
data? There are two alternatives. We may fix Vn and estimate kn in the best way
possible. Or we can fix kn and estimate Vn . The precise definition of Bn is not
necessary in either alternatives (though it may be implied.)
The optimality results about the density estimators are necessarily of asymptotic
nature. Rosenblatt showed already in 1956 that in the continuous case density
estimates based on finite n are necessarily biased.

5.4 Parzen Windows

5.4.1 From Histograms to Parzen Windows
As an introduction to Parzen windows, we consider a certain generalization of the
histogram. Assume that the region Bn is a d-dimensional hypercube. If hn is the
length of the side of the hypercube, its volume is given by Vn = hdn . We define still a
function that receives value 1 inside the hypercube centered at the origin, and value
0 outside the hypercube:

1 if |uj | ≤ 1/2
ϕ(u) = (5.12)
0 otherwise
It follows that ϕ((x − xi )/hn ) = 1 if xi falls in the hypercube of volume Vn centered
at x. And ϕ((x − xi )/hn ) = 0 otherwise. The number of training samples in this
hypercube is therefore given by
n
X x − xi
kn = ϕ( ).
i=1
hn

This leads to the density estimate

n
kn 1 X 1 x − xi
pn (x) = = ϕ( ) (5.13)
nVn n i=1 Vn hn

Above estimate differs from the histogram defined in section 5.3.1 in that we do not
define the cells a-priori. Rather the cells are defined only after observing x and they
can also intersect each other for different x.

5.4.2 Parzen Estimates

If we consider also other ϕ functions than above, we obtain a more general approach
to density estimation. The Parzen-window density estimate at x using n training
44 Introduction to Pattern Recognition

samples and the window function ϕ is defined by

n
1 X 1 x − xi
pn (x) = ϕ( ). (5.14)
n i=1 Vn hn

The estimate pn (x) is an average of values of the window function at different points.
(See figure 5.2.) Typically the window function has its maximum at the origin and
its values become smaller when we move further away from the origin. Then each
training sample is contributing to the estimate in accordance with its distance from
x. The normal density ϕ(u) = (2π)1d/2 exp[−0.5uT u] is the most widely applied
window function. Usually Vn = hdn , but the geometric interpretation of Vn as the
volume of a hypercube is not completely valid anymore. We will discuss the selection
of window length hn in section 5.4.4.

Figure 5.2: Parzen estimate

Sometimes Parzen estimates are called kernel estimates. The Parzen estimates
were introduced by Emanuel Parzen in 1962. However, Rosenblatt studied similar
estimates already in 1956.

5.4.3 Properties of Parzen Estimates

Parzen estimates have several interesting properties, some of which are studied now.
ForR pn (x) to be a legitimate density, it is required that 1) pn (x) ≥ 0 for all x and
2) pn (x)dx = 1. If we maintain the relation hdn = Vn , this is guaranteed if the
window function is a legitimate density:

1. ϕ(x) ≥ 0 for all x;

R
2. ϕ(x)dx = 1.
5.4 Parzen Windows 45

The proof is an exercise.

With an unlimited number of training samples, it is possible to let Vn approach
zero, and have pn (x) to converge to p(x) (for example in the point-wise mean-square
sense). For this, it is required that:

1. The density function p must be continuous.

2. The window function must be bounded and a legitimate density.

3. The values of the window function must be negligible at infinity.

4. Vn → 0 when n → ∞.

5. nVn → ∞ when n → ∞.

In addition to the assumptions about the window function and the window length,
we needed to make assumptions about the true density. The requirements concerning
the window function are natural and e.g. the normal density is in accordance with
them.
For pattern recognition, these optimality/convergence properties are important,
because with these properties we are able to show that the classification error of
a classifier based on Parzen windows tends to the Bayes error when n approaches
infinity. This holds also more generally: When the density estimates converge (in
a certain sense) to the true class conditional pdfs, and the prior probabilities are
properly selected, then the error of the resulting classifier tends to the Bayes error.
We do not go into detail with the conditions for the convergence, since their proper
treatment would require too much effort.

5.4.4 Window Width

The window width satisfying the requirements in the previous paragraph can be
chosen as √
hn = h1 / n.
In practice, we need to select h1 . This is not easy. The choice often is application
specific and it is rooted in the properties of the density estimate required by the
application. If h1 (hn ) is too large, then the density estimate pn (x) will be very
smooth and ‘out-of-focus’. If h1 is too small, then the estimate pn (x) will be just
superposition of n sharp pulses centered at training samples, i.e. an erratic and
noisy estimate of the true density.

5.4.5 Parzen Classifiers

It is straight-forward to apply Parzen estimates to the classification problems. For
each class we have the training samples D1 , . . . , Dn . Then, if possible, we first
estimate the priors P (ω1 ), . . . , P (ωc ). Thereafter, we estimate the class conditional
pdfs p(·|ωi ) based on Parzen windows, and we construct the Bayes classifier based
46 Introduction to Pattern Recognition

on the estimates. The hard part is that it is not possible to estimate p(·|ωi ) for
all possible feature vectors x. For each test point x to be classified, we must first
compute the Parzen estimate of the each class conditional density, and only after
that we can classify the test point. The Parzen-classifier can be written simply as
ni
1 X 1 x − xij
αP arzen (x) = arg max P̂ (ωi ) ϕ( ), (5.15)
ωi ,i=1,...,c ni j=1 hdi hi

where P̂ (ωi ) are the estimates for prior probabilities, xi1 , . . . , xini are training sam-
ples from ωi , ϕ is the window function (usually the same window function is used
for all classes), and hi is the window width for the class ωi .
As can be seen based on Eq. (5.15), Parzen classifier demand much more com-
putation than the classifiers based on the ML-estimation. Every classification with
Parzen classifiers requires n evaluations of a pdf, where n is the total number of
training samples.
Note that by substituting P̂ (ωi ) = ni /n in Eq. (5.15), the Parzen classifier can
be rewritten as
ni
1X 1 x − xij
αP arzen (x) = arg max d
ϕ( ).
ωi ,i=1,...,c n h h i
j=1 i

5.4.6 Probabilistic Neural Networks

Parzen classification can be also implemented in an neural network configuration.
Such implementation is termed Probabilistic neural network (PNN). Then it is as-
sumed that the training samples have been normalized to the unit length, i.e.
||xij || = 1.
Also the test point has to be normalized similarly. This is not reasonable in all
applications. We denote the normalized training samples by yi . We further assume
that the window function is the normal density.
A PNN consists of d input units comprising the input layer, n pattern units
comprising of the pattern layer, and c category units. Each input unit is connected
to each pattern unit. There is a pattern unit for each training sample. Each pattern
unit is connected to one and only one category unit corresponding the category of
the training sample. For the normalized test point x, each pattern unit computes
the inner product
neti = yiT x
and feeds the value exp[(neti − 1)/h2n ] to the category unit to which it is connected
to. Each output unit sums up the contributions of the pattern units connected to
it and assigns x to the class with the greatest sum. This indeed implements the
Parzen classifier because
x − yi
ϕ( ) ∝ exp[−(x − yi )T (x − yi )/(2h2n )] (5.16)
hn
= exp[−(xT x + yiT yi − 2xT yi )/(2h2n )] (5.17)
2
= exp[(neti − 1)/hn ], (5.18)
5.5 k-Nearest Neighbors Classifier 47

where the last equality is due to our normalization assumptions. Notation a(x) ∝
b(x) means that a(x) = C · b(x) for some constant C and for all x.

5.5 k-Nearest Neighbors Classifier

5.5.1 kn Nearest Neighbors Density Estimation
With the Parzen classification, the design of the classifier involved selecting window
functions and suitable window widths. This is not an easy task. One possibility is
let them depend on the training data. With respect to the general formulation of the
density estimation, this means fixing kn and computing of suitable (small enough)
Vn based on the selected kn . To be more precise, we place the center of the cell Bn
into the test point x and let the cell grow until it encircles kn training samples to
estimate p(x). These kn training samples are kn nearest neighbors of x. Here, kn is
a given parameter. The kn nearest neighbor (KNN) density estimate is given by
kn
pn (x) = , (5.19)
nVn
where Vn is the volume of the smallest possible x centered ball that contains kn
training samples, and n is the total number of training samples. If the true density
is high in the proximity of x, the cell Bn is small and the density estimate is accurate.
If the true density is low in the proximity of x, the cell Bn is large and the density
estimate is not that accurate.
Let p(x) be continuous at x. The KNN density estimate converges if
1. kn → ∞ when n → ∞;
2. kn /n → 0 when n → ∞.
A surprising property of the KNN density estimate is that although it is continuous,
its partial derivatives are necessarily not continuous. Typically, KNN density esti-
mates are quite ’spiky’ and the density estimates by Parzen windows more smooth.
It is also possible to estimate posterior probabilities P (ωi |x) directly with the
KNN method. This leads to more interesting result than in the case of Parzen
windows. We assume that the training samples D1 , . . . , Dc have been collected by
the mixture sampling. Also, we drop the indices describing the number of training
samples from all estimates and parameters. This is to avoid confusion in notation.
Now, center a cell encircling k training samples at the test point x. The k training
samples can be from any class. ki of these belongs to the class ωi . Then, the estimate
for the joint probability of x and ωi is
ki
p(x, ωi ) = .
nV
Using the definition of the conditional densities, we obtain an estimate for the pos-
terior probability
p(x, ωi ) p(x, ωi ) ki ki
P (ωi |x) = = Pc = Pc = . (5.20)
p(x) j=1 p(x, ωj ) j=1 kj k
48 Introduction to Pattern Recognition

This posterior probability estimate is maximal for the class with the highest ki and
x is assigned to that class. This is the k nearest neighbor (KNN) classification rule.

5.5.2 Nearest Neighbor Rule

We consider first the nearest neighbor (NN) classification, where x is assigned to
the class of its nearest neighbor. That is, k = 1. We can write this rule also as a
traditional function:

αnn (x) = arg min min ||x − xij ||. (5.21)

ωi ,i=1,...,c j

It was required that kn → ∞ when n → ∞ in order to KNN density estimates to

converge to the true density function. With the NN rule, kn = 1 for all n. Therefore,
the NN classification rule is not optimal. However, we can show that the error of the
NN classifier is at most twice the Bayes error when the number of training samples
tends to infinity. More precisely,
c
E(αbayes ) ≤ E(αnn ) ≤ E(αbayes )(2 − E(αbayes )).
c−1

The upper bound is reached when the true class conditional pdfs are identical for
all classes.
The rather positive error bound above is a theoretical one. In real life, we
do not have infinite number of training samples and there are no corresponding
results for the case when we have a finite number of training samples. However, the
NN classifier is very simple, and in practice it often works well. We explain this
heuristically. Denote the nearest neighbor of the test point x by x∗ . Let the class of
x∗ be ω ∗ . Because we assume that the training samples have been correctly classified,
ω ∗ maximizes the posterior P (ωi |x∗ ) with respect to classes ωi , i = 1, . . . , c with high
probability. Then, if x is close to x∗ , ω ∗ maximizes also the posterior P (ωi |x) with
respect to classes ωi , i = 1, . . . , c with high probability if the posteriors are smooth
enough.
We finally consider the geometry of the decision regions of the N N rule. The
decision region for class ωi is formed by the cells whose points are closer to some
training sample in Di than any other training sample. This kind of partition of the
feature space is termed Voronoi diagram. See Figure 5.3 due to Duda, Hart and
Stork.

5.5.3 k Nearest Neighbor Classification Rule

The KNN classification rule consists of finding k-nearest training samples to the
test point x and classifying x to the class which is most frequent in those k nearest
neighbors of x. Note that the KNN rule is - as all decision rules are - a function
from the feature space to the set of classes, although it is not easy to write the KNN
rule as a ’traditional’ function.
5.5 k-Nearest Neighbors Classifier 49

Figure 5.3: Decision regions for the nearest neighbor rule

In the two class case, we following error estimate holds in the case of unlimited
number of training samples:
r
2E(αnn )
E(αbayes ) ≤ E(αknn ) ≤ E(αbayes ) + .
k
Also, from this error estimate it follows that the KNN rule is optimal when k → ∞.
In real life, the situation is once again different: Especially when the feature space is
high-dimensional, the required number of training samples grows very rapidly with
k.
The KNN classifier and the NN classifier suffer from the same practical problem
as the Parzen classifier. The computational complexity of classification grows lin-
early with the number of training samples, because the distance from each training
sample to the test point must be computed before the classification is possible. This
may be a problem if the classifier needs to be very fast. Let us note, however, that
there exist several techniques to speed up KNN classifiers.

5.5.4 Metrics
We have assumed until now that the distance between the points a and b in the
feature space (feature vectors) is measured with the Euclidean metric:
v
u d
uX
L(a, b) = ||a − b|| = t (ai − bi )2 .
i=1

However, many other distance measures or metrics can be applied. The selection
of a metric obviously affects the KNN classification results. Especially, the scaling
50 Introduction to Pattern Recognition

of the features has a fundamental effect to the classification results although this
transformation merely accounts to a different choice of units for features.
Formally, a metric is a function L(·, ·) from Rd × Rd to R. For all vectors a, b, c,
a metric L must satisfy

1. L(a, b) ≥ 0

2. L(a, b) = 0 if and only if a = b.

3. L(a, b) = L(b, a)

4. L(a, b) + L(b, c) ≥ L(a, c).

The only one of these properties which we occasionally want to sacrifice is the
property 2. Other properties are important for pattern recognition applications.
Some examples of useful metrics:

• Minkowski metrics:
d
X
Lm (a, b) = ( (|ai − bi |m )1/m .
i=1

• L-infinity metric
L∞ (a, b) = max |ai − bi |.
i

• Mahalanobis-distance
p
LM ahalanobis,C (a, b) = (a − b)T C −1 (a − b),

where C is a given positive definite matrix.

5.6 On Sources of Error

We have mentioned repeatedly in this chapter that the optimality properties of the
classifiers hold only if we have an infinite number of training samples. However, we
always have only a finite number of them. In this section, we discuss the sources of
errors when the number of training samples is finite.
Before discussing error sources, it is important to note that in pattern recognition
two error types can be distinguished: the training error and the test error. The
training error refers to the classification errors for the training samples, i.e. how
large portion of the training samples is misclassified by the designed classifier. Here,
we are not concerned with this error type. The important error type is the the test
error, which described how large portion of all possible objects are misclassified.
Note that the Bayes classifier minimizes the test error when the assumptions made
about class conditional pdfs and priors hold. The estimation of the test error is a
challenging task, and we will return to the issue in Chapter 7.
5.6 On Sources of Error 51

Before studying the sources of the test error, we recall what inputs did different
classifiers require for their design. First, all supervised classifiers need training
samples. Second, the classifiers based on the ML-estimation need knowledge about
the parametric families of the class conditional pdfs. Parzen classifiers need to be
given the window functions and window lengths. For KNN classifiers, we need to
select the parameter k.
The Bayes error is a necessary error source of every pattern classification prob-
lem. This error can never be reduced without changing the classification problem.
If we take into account the whole pattern recognition system, we can change the
classification problem e.g. by adding an extra feature. This can possibly reduce the
Bayes error.
The second source of classification error with practical classifiers is that the
probabilistic model for the problem is incorrect. This is termed model error. In
ML-based classification, these are due to having incorrect (parametric) model for
class conditional densities, i.e. the data from the class ωi is not distributed according
to any of the densities p(·|ωi , θi ). In the non-parametric case, we interpret a bad
selection of window functions, window lengths, or the parameter k as model errors.
The model error is more important with the parametric ML-based classifiers than
with the non-parametric classifiers. This type of error can be reduced by selection
a better probabilistic model for the problem. Again, this is rarely straight-forward.
The third and final error source is the estimation error. This error is due having
only a finite number of training samples. This error is more pronounced in the
non-parametric case where more training samples are required than with ML-based
classifiers. In the ML-estimation, the larger the parametric family the lesser the
model error but the greater the estimation error, i.e. there is a trade-off between
the model and estimation errors. The estimation error can be reduced by adding
more training samples.
In general, if we add features to the classifier, we also increase the training
error. Further, the required number of training samples grows exponentially with
the number of features in the non-parametric case. Hence, adding one more feature
to the pattern recognition system might not always be a good idea. The phenomenon
is often termed the curse of dimensionality.
52 Introduction to Pattern Recognition
Chapter 6

Linear Discriminant Functions and

Classifiers

6.1 Introduction
To this point, we have designed classifiers by estimating the class conditional pdfs
and prior probabilities based on training data. Based on the estimated class condi-
tional pdfs and prior probabilities, we have then derived the Bayes classifier. This
way the classifier minimizes the classification error when the estimated pdfs and
priors are close to the true ones. The problem was that the estimation problems
are challenging, and with the finite number of training samples, it is impossible to
guarantee the quality of estimates.
In this Chapter, we consider another way to derive the classifier. We assume that
we know the parametric forms of the discriminant functions in a similar way as with
the ML estimation. The estimation will be formulated as a problem of minimizing a
criterion function - again much in a same way than with the ML technique. However,
now we are aiming directly to find the discriminant functions without first estimating
the class conditional pdfs. An example illustrates positive features of this type of
classifier design: Consider a two-category classification problem, where the class
conditional pdfs are normal densities with equal covariances. Then, as stated in
Section 4.6, the Bayes classifier is linear, and it can be represented with the following
discriminant function:

g(x) = g1 (x) − g2 (x) = wT x + w0 ,

where
w = Σ−1 (µ1 − µ2 )
and
1
w0 = − (µ1 − µ2 )T Σ−1 (µ1 − µ2 ) + ln P (ω1 ) − ln P (ω2 ).
2
For deriving this discriminant function, it was necessary to estimate 2d+(d2 +d)/2+1
parameter values. The first term is due to the two mean vectors, the second is
due to the single (symmetric) covariance matrix, and the third one is due to the
54 Introduction to Pattern Recognition

priors which sum to one. Hence, it might be reasonable to estimate directly the
discriminant function, which has only d + 1 parameter values: d of them from
the weight vector w, and one of them from the threshold weight w0 . In addition,
the parameterization of the discriminant functions may sometimes be more natural
than the parameterization of the class conditional pdfs.
The estimation of w and w0 will be performed by minimizing a criterion function
as it was already mentioned. The criterion can be e.g. the number of misclassified
training samples but more often it is some related function. Note, however, that
in this case we will lose the direct contact with the test-error which is much more
important quantity than the training error.

6.2 Properties of Linear Classifiers

6.2.1 Linear Classifiers
Recall first how the classifier was defined with the help of discriminant functions.
For each class, we had a discriminant function gi , i = 1, . . . , c whose input was a
feature vector x. The feature vector x was assigned to ωi if

gi (x) > gj (x)

for all i 6= j. (If gi (x) = gj (x) we can use some arbitrary rule such as classifying
x to ωi if i < j. As with the Bayes classifier, the selection of such a rule bears no
importance.)
A discriminant function is said to be linear if it can written as
d
X
gi (x) = wiT x + wi0 = wij xj + wi0 ,
j=1

where wi = [wi1 , . . . , wid ]T is the weight vector and the scalar wi0 is threshold
weight. The classifier relying only on the linear discriminant functions is called
linear classifier or linear machine.

6.2.2 The Two Class Case

Consider a 2-class classification problem and a linear classifier. The points on the
decision surface satisfy

g1 (x) = w1T x + w10 = w2T x + w20 = g2 (x).

From this, we get

(w1 − w2 )T x + w10 − w20 = 0.
Denote w = w1 − w2 and w0 = w10 − w20 , and represent the classifier with a single
discriminant function. We obtain the decision surface

R12 = {x|wT x + w0 = 0}. (6.1)

6.2 Properties of Linear Classifiers 55

This is a hyper-plane. A hyper-plane is a point when d = 1, a line when d = 2, and

so-forth.
We choose two arbitrary points laying on the decision surface, say x1 and x2 .
Now
wT x1 − wT x2 − w0 + w0 = wT (x1 − x2 ) = 0.
Hence, the weight vector w is normal to the decision surface. We can write
w
x = xp + r ,
||w||
where xp is the normal projection of the feature vector x to the decision surface
R12 . Then, because g(xp ) = 0, we get
w
g(x) = wT x + w0 = wT (xp + r ) + w0 = r||w||.
||w||
The scalar r is greater than zero if x belongs to ω1 and less than zero if x belongs to
ω2 . Also, a high absolute value of g(x) means that x lies far away from the decision
surface. Figure 6.1 clarifies. It is Fig. 5.2 from Duda, Hart and Stork.

Figure 6.1: Geometry of linear 2-class classifier

6.2.3 The c Class Case

The results of the previous paragraph obviously hold if we consider just 2 of the c
classes. When we take all classes into account, we note that the boundary between
Ri and Rj is a portion of the hyper-plane

Hij = {x|gi (x) = gj (x)}.

Moreover, we can characterize also decision regions. Namely, these are convex.
56 Introduction to Pattern Recognition

A set R ⊂ Rd is convex (Figure 6.2), if for all r1 , r2 ∈ R, also (1 − λ)r1 + λr2 ∈

R,when λ ∈ [0, 1]. In other words, if R is convex each line segment between its
points belongs in whole to R. For example, a ball is convex set in the 3-dimensional
Euclidean space. Convex sets are also connected, and therefore a decision region of
a linear classifier is connected.

Figure 6.2: Only the left-most of above sets is convex and the others are not convex.

To prove the convexity of the decision regions of linear classifiers, consider two
points r1 , r2 belonging to the decision region Ri . Because these points belong to Ri ,
it holds that
gi (r1 ) > gj (r1 ), gi (r2 ) > gj (r2 ) (6.2)
for all j 6= i. Now, consider the point (1 − λ)r1 + λr2 , where λ ∈ [0, 1]. Due to
linearity of the discriminant functions

gi ((1 − λ)r1 + λr2 ) = wiT ((1 − λ)r1 + λr2 ) + wi0 > gj ((1 − λ)r1 + λr2 ). (6.3)

This means that the decision regions are convex. It is an exercise to prove that (6.3)
follows from (6.2).
The convexity of the decision regions naturally limits the number of classification
tasks which can effectively be solved by linear machines.

6.3 Linearly Separable Training Samples

Consider first the 2-class case. We assume that the classification task is such that
the classification error is very small for some linear classifier. We have training
samples from the classes ω1 and ω2 . We denote these as before

Di = {xi1 , . . . , xini }, i = 1, 2,

where ni is the number of samples belonging to ωi . Now, if a such linear classifier

exists that it classifies all the training samples correctly, i.e.

g(x1j ) > 0,

for all j = 1, . . . , n1 and

g(x2j ) < 0
6.3 Linearly Separable Training Samples 57

for all j = 1, . . . , n2 , we say that the training sets/samples D1 and D2 are linearly
separable.
Two-class linear classifiers and the above condition can be presented in a more
compact form if we make two modifications to the notation:
1) We form the augmented feature vector y based on x:
y = [1, x1 , . . . , xd ]T .
Similarly, the augmented weight vector is obtained from w and w0 :
a = [w0 , w1 , . . . , wd ]T = [w0 , wT ]T .
Linear discriminant functions can now be written as
g(x) = wT x + w0 = aT y = g(y).
2) We denote by yij the augmented feature vector generated from the training
sample xij . In the 2-class case, we reduce the two training sets D1 , D2 to a single
training set by simply replacing the training samples from ω2 by their negatives.
This because
aT y2j < 0 ⇔ aT (−y2j ) > 0.
Hence, we can forget about ω2 , replace its training samples by their negatives, and
attach them to the training set for ω1 . We denote the resulting training set by
D = {y1 , . . . , yn1 +n2 }. Note that ’replacement by negatives’ must be performed
expressly for augmented feature vectors. (Why?).
As an example, consider the training samples
D1 = {[1, 1]T , [2, 2]T , [2, 1]T }
and
D2 = {[1, −1]T , [1, −2]T , [2, −2]T ]
The modified training set is then
D = {[1, 1, 1]T , [1, 2, 2]T , [1, 2, 1]T , [−1, −1, 1]T , [−1, −1, 2]T , [−1, −2, 2]T }.
Given that these two modifications have been performed, we can re-define linear
separability: The training set (samples) D on linearly separable if there exists such
augmented weight vector a that
aT yj > 0 (6.4)
for all y1 , . . . , yn1 +n2 .
If we assume that the training samples are linearly separable, then we can try
to solve the system of inequalities (6.4) to design the linear classifier. One possible
solution method is introduced in the next section. However, a solution - if exists -
is not unique. 1) If we have a solution vector a which is a solution to the system
(6.4), then we can always multiply this vector a positive constant to obtain another
solution. (Remember that the weight vectors were normals to the decision surfaces).
2) It is possible to have multiple decision surfaces that separate the training sets.
Then, we have many (genuinely different) augmented weight vectors that solve the
system (6.4).
58 Introduction to Pattern Recognition

6.4 Perceptron Criterion and Algorithm in 2-Class

Case
6.4.1 Perceptron Criterion
To find a weight vector a that satisfies

aT yj > 0

for all j, we formulate an optimization problem. We search the minimum argument

of the function X
Jp (a) = −aT yj . (6.5)
j:aT yj <0

The function Jp is termed Perceptron criterion. In other words, we sum the inner
products of the weight vector a and those training samples that are incorrectly
classifier. The Perceptron criterion is the negative of that sum.
The value Jp (a) is always greater than equal to zero. The Perceptron criterion
obtains its minimum (that is zero) when a solves the system of inequalities (6.4).
(There is also a degenerate case. When a = 0, Jp (a) = 0. However, this obviously is
not the solution we are searching for.) To sum up, we have converted the problem
of solving inequalities
P into an optimization problem.
The criterion j:aT yj <0 −1 would also fulfill the above requirements. However,
this criterion function is not continuous which would lead to numerical difficulties.
Instead, Jp is continuous.

6.4.2 Perceptron Algorithm

Usually, the analytic (i.e. by paper and pen) minimization of the criterion Jp is not
possible. Therefore, we must rely on numerical techniques to solve the minimization
problem. The most simple method for the numerical minimization is the steepest
descent method, which is well applicable here. In the steepest descent method, we
choose an initial value (denoted by a(0)) for a and start updating a little by little to
the direction of the negative gradient of Jp at a(t) at each iteration t = 0, 1, 2, . . ..
The idea behind the steepest descent method is the value of any continuous function
f at x decreases most rapidly to the direction of its negative gradient −∇f (x). The
gradient of Jp at a is X
∇Jp (a) = −yj . (6.6)
j:aT yj <0

This yields the Perceptron algorithm:

Perceptron algorithm
Set t ← 0,
PInitialize a(0), η, .
T
while − j:a(t)T yj <0 a yj >  do
P
a(t + 1) = a(t) + η j:a(t)T yj <0 yj .
6.5 Perceptron for Multi-Class Case 59

Set t ← t + 1
end while
Return a(t).

Two parameters - in addition to the initial value a(0) - need to be set: The
stopping condition  and the step length η. Sometimes the step length is called the
learning rate. If training samples indeed are linearly separable, we can set  to zero.
The reason for this is that the Perceptron algorithm converges to the solution in
a finite number of steps as long as the training samples are linearly separable. If
the training samples are not linearly separable, the Perceptron algorithm does not
converge and it is not a good choice in the linearly non-separable case. The step
length can be set to 1.
We illustrate the Perceptron algorithm by continuing the example of Section 6.3
and designing a linear classifier for those training samples. We select a(0) = [1, 1, 1]T .
In the first iteration, samples y4 , y5 , and y6 are incorrectly classified. Hence,

a(1) = a(0) + y4 + y5 + y6 = [−2, −3, 6]T .

By a(1), the sample number 3 is misclassified. Hence,

a(2) = a(1) + y3 = [−1, −1, 7]T .

This is also the final result since all the training samples are now correctly classified.

6.5 Perceptron for Multi-Class Case

We have concentrated only on the design of 2-class linear classifiers up to this point.
Now, we would like to design a linear classifier for a c-class problem. We have
augmented training samples from c classes:

Di = {yi1 , . . . , yini }, i = 1, . . . , c.

We say that these training sets are linearly separable if there exists a linear clas-
sifier which classifies all the training samples correctly. The task is to find such
(augmented) weight vectors a1 , . . . , ac that always when y ∈ Di , it holds that

aTi y > aTj y (6.7)

for all j 6= i. In other words

aTi yik > aTj yik ,
for all i = 1, . . . , c, j 6= i and k = 1, . . . , ni . These inequalities can be manipulated
to a 2-class and c · d feature problem by the Kesler’s construction. This is a very
useful theoretical tool but its use leads to computationally ineffective algorithms.
The Perceptron algorithm for the c-class case is:
60 Introduction to Pattern Recognition

c-class Perceptron algorithm

Set t ← 0, initialize a1 (0), . . . , ac (0).
while Misclassified training samples exist do
for all misclassified yik do
ai (t + 1) = ai (t) + yik
aj (t + 1) = aj (t) − yik jos ai (t)T yik ≤ aj (t)T yik
al (t + 1) = al (t) jos ai (t)T yik > al (t)T yik
Set t ← t + 1
end for
end while
Return a1 (t), . . . , ac (t).

As in the 2-class case, this algorithm works only if the training samples are
linearly separable.

6.6 Minimum Squared Error Criterion

We still need an algorithm for designing linear classifiers when the training samples
are not linearly separable. We will base the algorithm to the minimum squared error
criterion and we will briefly introduce the criterion in the 2-class case.
We modify the training samples again as in Section 6.3. As a result we have a
single training set of augmented feature vectors. We aim to minimize the criterion:
nX
1 +n2

Js (a) = (aT yi − 1)2 . (6.8)

i=1

The criterion Js can be written in the matrix-form as

Js (a) = ||Y a − 1||2 , (6.9)
where Y is the n1 + n2 × d + 1 matrix consisting of the training samples, and 1 is
n1 + n2 component vector consisting of only ones. Precisely
 
y1
 y2 
Y =
 
.. 
 . 
yn1 +n2
and  
1
 1 
1= .
 
..
 . 
1
This criterion has the unique minimum at
a = (Y T Y )−1 Y T 1. (6.10)
6.6 Minimum Squared Error Criterion 61

Note that we do not need to assume that Y would be non-singular or square. The
selection of the vector 1 is based on the relation with the Bayes classifier. For more
information on this, see section 5.8.3 in Duda, Hart and Stork.
62 Introduction to Pattern Recognition
Chapter 7

Classifier Evaluation

7.1 Estimation of the Probability of the Classifi-

cation Error
It is important to estimate the error probability of the classifier when designing
practical pattern recognition systems. This provides answers to important questions
such as ’is the system good enough or does it need a further improvement?’ and ’is
the classifier A better than the classifier B?’.
Recall once more that we are interested in the probability that a test point which
does not belong to the training set is mis-classified. This was termed test error in
Section 5.5. Next, we introduce three ways to estimate the test error.
The training error is the frequency of error for the training data. The training
error is a poor and overly optimistic estimate to the test error. For example, the
training error of the NN-classifiers is automatically zero. This is obviously not true
for the test error. The estimation of the test error by the training error is termed
as the resubstitution method.
A much better estimate for the test error is obtained by dividing the training
data
D = D 1 ∪ D 2 ∪ · · · ∪ Dc

into two sets Dtraining and Dtest . We use Dtraining for the training of the classifier
and Dtest is solely used to estimate the test error. This method is called the holdout
method. It is a good method when we have large number of training samples, and
hence a modest decrease in the number of training samples does not essentially
decrease the quality of the classifier. Note that the division into sets Dtraining and
Dtest must be randomized. For example, selecting Dtest = D1 is not a good idea.
In the case that we have only a few training samples, it is the best to use some
cross-validation method to estimate the test error. The most simple of these is
the leave-one-out method. In the leave-one-out method, we drop a single training
sample from the training set and design a classifier based on the other training
samples. Then, it is studied whether the dropped sample is classified correctly or
not by the designed classifier. This is repeated for all n training samples. The test
64 Introduction to Pattern Recognition

error estimate for the classifier α is the obvious

v
Ê(α) = ,
n
where v is the number of misclassified training samples, and n = ci=1 ni is the total
P

number of training samples. (The hat-notation refers to an estimate, i.e. Ê(α) is

an estimate of the true classification error E(α).) This method yields an accurate
estimate for the test error. Its main disadvantage is the computational complexity:
to estimate the test error for one classifier, we must design n classifiers.

7.2 Confusion Matrix

It is often worthwhile to study the confusion matrix in addition to the classification
error. The element sij of the confusion matrix
 
s11 s12 · · · s1c
S =  ... .. . 
. .. 

sc1 sc2 · · · scc

is the number of test samples that are assigned to the class ωi but whose correct
class would be ωj . The order of indices i, j can be selected also differently, and it
varies from one source reference to another. Hence, it pays off to be attentive. The
confusion matrix can be estimated by the same methods as the test error.
Based on the confusion matrix, the classification (test) error can be obtained as
Pc Pc Pc
i=1 j=1 sij − i=1 sii
Ê(α) = Pc Pc .
i=1 j=1 sij

Based on the confusion matrix, also other important quantities for classifier design
can be computed.

7.3 An Example
Let us illustrate the leave-one-out-method in the case of Fisher’s Iris dataset. We
study the classification based on ML-estimation, and we assume that the class condi-
tional pdfs are Gaussian. The data was collected by separate sampling and therefore
we assume that the prior probabilities for all classes are equal. This data can be
loaded into Matlab by

>> load fisheriris

command. (At least, if the Statistics Toolbox is available). There should now be
variables ’meas’ and ’species’ in Matlab. The variable ’meas’ contains the feature
vectors from 150 Irises and the variable ’species’ contains the class of each feature
vector. The feature vectors numbered from 1 to 50 belong to the class ω1 (iris
7.3 An Example 65

setosa), the feature vectors numbered from 51 to 100 belong to the class ω2 (iris
versicolor), and the feature vectors numbered from 101 to 150 belong to the class ω3
(iris virginica). This simplifies the implementation of the leave-one-out method in
Matlab. The following piece of Matlab-code estimates the confusion matrix by the
leave-one-out method:

% initialize the confusion matrix

s = zeros(3,3);
% class 1, feature vectors 1 - 50
% class 2, feature vectors 51 - 100
% class 3, feature vectors 101 - 150
for i = 1:150
trueclass = ceil(i/50);
test = meas(i,:); % test point is feature vector number i
% to the class ’trueclass’ training data contains all feature
% vectors from that class except the test point
% for other classes all training data is used.
train1 = [meas(1:(i - 1),:);meas((i + 1):50,:)];
train2 = [meas(51:(i - 1),:);meas((i + 1):100,:)];
train3 = [meas(101:(i - 1),:);meas((i + 1):150,:)];
% Train the classifier, i.e. compute the means vectors and
% covariance matrices for each class
mu1 = mean(train1);
cov1 = cov(train1);
mu2 = mean(train2);
cov2 = cov(train2);
mu3 = mean(train3);
cov3 = cov(train3);
% the classifier has been trained. Now compute the posterior
% probabilities and see to what class the test point is classified
% to
posterior(1) = mvnpdf(test,mu1,cov1)*(1/3);
posterior(2) = mvnpdf(test,mu2,cov2)*(1/3);
posterior(3) = mvnpdf(test,mu3,cov3)*(1/3);
[maxposterior,testclass] = max(posterior);
% and increment the corresponding element of the confusion matrix by
% one.
s(testclass,trueclass) = s(testclass,trueclass) + 1;
end

Above, ’ceil(x)’ returns the least integer that is larger than the real number
x. We have used also some Matlab specific programming tricks: for example
meas(56:50,:) is an empty matrix. About ’mvnpdf’, Matlab’s help states the
following
Y = MVNPDF(X,MU,SIGMA) returns the density of the multivariate normal
66 Introduction to Pattern Recognition

distribution with mean MU and covariance SIGMA, evaluated at each row

of X.

The confusion matrix is

s =

49 2 2
1 41 6
0 7 42

And the estimate for the test error is accordingly:

Ê(α) = 18/150 ≈ 0.12.

Chapter 8

Unsupervised Learning and

Clustering

8.1 Introduction
Until now, we have assumed that we have a set of correctly labeled training sam-
ples for training the classifier. These procedures using training samples are part
of supervised classification (see Section 2.4). However, in some circumstances it is
necessary to resort to unsupervised classification that is also termed as clustering.
These procedures do not use labeled training data. Instead, we have a collection of
unlabeled samples, and we try to classify them based only on the features in data.
In other words: there is no explicit teacher. Clearly, clustering is a more difficult
problem than supervised classification. However, there a certain situations where
clustering is either useful or necessary. These include:
1. The collection and classification of training data can be costly and time con-
suming. Therefore, it can be impossible to collect a training set. Also, when
there are very many training samples, it can be that all of these cannot be
hand-labeled. Then, it is useful to train a supervised classifier with a small
portion of training data and use then clustering procedures to fine tune the
classifier based on the large, unclassified dataset.
2. For data mining, it can be useful to search for natural clusters/groupings
among the data, and then recognize the clusters.
3. The properties of feature vectors can change over time. Then, supervised
classification is not reasonable, because sooner or later the test feature vectors
would have completely different properties than the training data had. For
example, this problem is commonplace in the processing of medical images.
4. The clustering can be useful when searching for good parametric families for
the class conditional densities for the supervised classification.
Above situations only are examples of the situations where clustering is worth-
while. There are many more situations requiring clustering.
68 Introduction to Pattern Recognition

8.2 The Clustering Problem

We define the clustering problem more precisely. We have a set of feature vectors

D = {x1 , . . . , xn }.

The task is to place each of these feature vectors into one of the c classes. In other
words, the task is to find c sets D1 , . . . , Dc so that
c
[
Di = D
i=1

and
Dj ∩ Di = ∅
for all i 6= j.
Naturally, we need some further assumptions for the problem to be sensible. (An
arbitrary division of the feature vector between different classes is not likely to be
useful.) Here, we assume that we can measure the similarity of any two feature
vectors somehow. Then, the task is to maximize the similarity of feature vectors
within a class. For this, we need to be able to define how similar are the feature
vectors belonging to some set Di .
In the following sections, we will introduce two alternatives for defining how
similar a certain set of feature vectors is. We have already seen their counterparts in
supervised classification. The first one, the k-means clustering, reminds minimum
distance classification. The second one, the expectation maximization algorithm, is
rooted on ML-estimates and finite mixture models

8.3 K-means Clustering

8.3.1 K-means Criterion
Recall that the minimum distance classifier assigned a test point to the class with
the nearest mean to the test point. Denote now the number of feature vectors in
the class Di by ni . (We do not know the value of ni yet.).
The problem is that we do not know the mean vectors. Despite of this, define
the similarity of Di as X
s(Di ) = − ||x − µi ||2 ,
x∈Di
1
P
where µi = ni x∈Di x. We solve the unknown mean vectors by maximizing the
sum of similarities of all classes. Equivalently, we can minimize the negative of this
sum: c X
X
J(D1 , . . . , Dc ) = ||x − µi ||2 . (8.1)
i=1 x∈Di

This is the k-means criterion.

8.3 K-means Clustering 69

8.3.2 K-means Algorithm

Note that based on the knowledge of clusters D1 , . . . , Dc , we can compute the mean
vectors. Conversely, based on the knowledge of mean vectors, we can compute the
clusters. This observation is important when minimizing the criterion (8.1). The
criterion J is a function of the partition of the discrete set D, and therefore the
steepest descent method of Chapter 6 cannot be used. The algorithm to minimize
the criterion (8.1) is termed the k-means algorithm. The basic principle is to compute
the mean vectors based on the current clustering, re-classify the feature vectors based
on updated mean vectors, and repeat these two steps until the clustering does not
change anymore.

k-means algorithm
Initialize µ1 (0), . . . , µc (0), set t ← 0
repeat
Classify each x1 , . . . , xn to the class Dj (t) whose mean vector µj (t) is the nearest
to xi .
for k = 1 to c do
update the mean vectors µk (t + 1) = |Dk1(t)| x∈Dk (t) x
P
end for
Set t ← t + 1
until clustering did not change
Return D1 (t − 1), . . . , Dc (t − 1).

Notes: 1) Dj (t − 1) is returned because the iteration counter was incremented

by one just before the termination of the algorithm . 2) The notation |Dk (t)| refers
to the number of elements in the set Dk (t).
Typically, the mean vectors are initialized as randomly selected feature vectors.
For example, we can set µj (0) ← xj . Also, the result of the algorithm depends
on the initialization. This is because the k-means algorithm is a local optimization
algorithm and the minimization problem in this case has several local minima. (See
Figure 8.1). In the case of the k-means criterion, these local minima are not too
numerous.

Figure 8.1: Local vs. global maximum

70 Introduction to Pattern Recognition

8.3.3 Properties of the K-means Clustering

The K-means clustering is ’an application’ of the minimum distance classifier to
clustering. Therefore, the k-means clustering suffers from the same drawbacks as
the minimum distance classification. If the clusters are of different sizes (i.e. they
contain a different number of feature vectors), the k-means clustering is not suc-
cessful. If the clusters have very different covariances or scales of different features
are different, then the clustering by k-means is not appropriate. See figure 8.2.
Note that these problems are due to the criterion function and not the algorithm
minimizing it. Because of these problems, we need a more complicated clustering
criterion/algorithm to solve clustering problems such as those depicted in Figure
8.2. Instead, k-means performs well when clusters form compact clouds that are
well separated from one another.

Figure 8.2: Examples of problematic clustering tasks for k-means. Top row: The
scale of features is different (note the scaling of the axis). Bottom row: There are
many more feature vectors in one class than in the other. ’True classifications’ are
shown on left and k-means results are on right.

8.4 Finite Mixture Models

We consider Gaussian finite mixture models (FMMs)1 for solving clustering problems
that are out of the scope of K-means. The task is now to learn the Bayes classifier
1
Finite mixture models can be derived also for other families of parametric pdfs, but we will
just consider FMMs based on normal densities during this course.
8.4 Finite Mixture Models 71

in the unsupervised manner (without any training samples) when the parametric
families of the class conditional densities are known. We will assume that:
1. The class conditional densities are modeled by normal densities, i.e. p(x|ωj ) =
pnormal (x|µJ , Σj ) for all j = 1, . . . , c.
2. Parameters µ1 , . . . , µc , Σ1 , . . . , Σc are not known.
3. Priors P (ω1 ), . . . , P (ωc ) are not known.
4. There are no training samples: i.e. the classes of feature vectors x1 , . . . , xn are
not known, and the task is to classify the feature vectors.
We assume that feature vectors are occurrences of independent random variables.
Each random variable is distributed according to some class conditional density but
we do not know according to which one. We can, however, write the probability
density that xi is observed and it belongs to some class. Firstly,

p(xi , ωj ) = P (ωj )p(xi |ωj )

for all i and j. Secondly

c
X
p(xi ) = p(xi , ωj ),
j=1

because xi has to belong to exactly one of the classes ω1 , . . . , ωc . (The left hand side
of the above equation is a marginal density.) By combining these, we obtain
c
X
p(xi ) = p(xi |ωj )P (ωj ). (8.2)
j=1

This is a mixture density. The priors P (ωi ) in mixture densities are called mixing
parameters, and the class conditional densities are called component densities.
We assumed that the component densities are normal densities. Hence, we obtain
the parametric mixture density
c
X
p(xi |θ) = pnormal (xi |µj , Σj )P (ωj ), (8.3)
j=1

where the parameter vector

θ = (µ1 , . . . , µc , Σ1 , . . . , Σc , P (ω1 ), . . . , P (ωc ))T .

We can then derive an algorithm to solve the parameter vector θ̂ based on ML

estimation. We will introduce the algorithm in the next section. Right now, we go
a bit further and study how xi can be classified when we have ML estimates for
all the parameters. We will use the Bayes classifier as in the section 5.2, that is
we substitute the unknown parameters by their estimated values. This leads to the
decision rule
αmixture (xi ) = arg max p(xi |µ̂j , Σ̂j )P̂ (ωj ).
ωj ,j=1,...,c
72 Introduction to Pattern Recognition

Figure 8.3: Finite mixture models. On top component densities are shown. On bot-
tom the resulting mixture density when component densities and mixing parameters
are as shown in the top panel.

Figure 8.4: A finite mixture model in the case of two features.

8.5 EM Algorithm 73

A problem with FMMs is that they are not always identifiable. This means that
two different parameter vectors θ1 and θ2 may result in exactly the same mixture
density, i.e.

p(x|θ1 ) = p(x|θ2 )

for all x. However, the classifier derived based on the parameter vector θ1 can still
be different than the classifier that is based on θ2 . And therefore, this is a serious
problem because the parameter vectors cannot be distuinguished from each other
based on ML estimation.. Switching problems are special cases of identifiability
problems. If we switch the classes ωi and ωj , we end up with exactly the same
mixture density as long as component densities for ωi and ωj belong to the same
parametric family. Therefore, the classes must be named (or given interpretation)
after the classification. This is problematic, if we are after a completely unsupervised
classifier.

8.5 EM Algorithm

We return to solving the ML estimation problem. Because x1 , . . . , xn were occur-

rences of independent random variables, we can write the likelihood function as

n
Y n X
Y c
p(xi |θ) = pnormal (xi |µj , Σj )P (ωj ).
i=1 i=1 j=1

Maximizing this with respect to the parameter vector θ results in the ML estimate
θ̂. Once again, it is more convenient to deal with the log-likelihood function, i.e.

n
X Xc
θ̂ = arg max ln[ pnormal (xi |µj , Σj )P (ωj )]. (8.4)
θ
i=1 j=1

This is a challenging maximization task. (Actually, we are interested in the largest

local maximum, because the (log)-likelihood function is not bounded from above,
see e.g. Duda, Hart, Stork, Chapter 10). However, we can derive a expectation-
maximization (EM) algorithm for solving (8.4). We will skip the derivation. The
algorithm can be given as:
74 Introduction to Pattern Recognition

EM algorithm

1. Initialize µ0j , Σ0j , P 0 (ωj ), set t ← 0.

2. (E-step) Compute the posterior probabilities that xi belongs to the class ωj

P t (ωj )pnormal (xi |µtj , Σtj )

pt+1
ij = Pc t t t
.
k=0 P (ωk )pnormal (xi |µk , Σk )

3. (M-step) Update the parameter values

X
P t+1 (ωj ) = (1/n) pt+1
ij (8.5)
i
P t+1
pij xi
µt+1
j = Pi t+1 (8.6)
i pij
P t+1
i pij (xi − µt+1 t+1 T
j )(xi − µj )
Σt+1
j = P t+1 (8.7)
i pij

4. Set t ← t + 1.

5. Stop if some convergence criterion is fulfilled, otherwise return to step 2.

In most cases, The EM algorithm finds a local maximum of the likelihood func-
tion. However, there are usually several local maxima and hence the final result is
strongly dependent on the initialization.