Prof. K. Rajan

8/2/2019 Prof. K. Rajan

http://slidepdf.com/reader/full/prof-k-rajan 1/65

1/29/2012 1



K.RajanDepartment

of

Electrical

&

Electronics

Engineering

u a o y ec n c o ege

Annamalainagar

õí è¢ñ¢

Department of Electrical & Electronics Engineering

Muthiah Polytechnic College, Annamalainagar1/29/2012 2



Learning

Learning is a general term that denotes the way in which peopleand machine enrich their knowledge and improve their skills.

Learning is one of the component of Intelligence.





Learning Algorithm

Learning algorithm is a procedure to learn systematically.

Al orithm s ecifies how to learn or ac uire the knowled e.





Artificial Intelligence (AI)

AI is a branch of Computer Science that deals with thestudy and creation of computer systems that exhibit

some form of intelligence.

,

and use knowledge in a meaningful way.

The term Artificial Intelligence was coined by John

McCarthy, in 1956 at MIT





Components of Intelligence

Learning – It is the process of acquiring knowledge, skills, experience by study

and training.

Reasoning –It refers to the ability of drawing conclusions that are appropriate

to the

situation

in

hand.

Understanding – It refers to the identification of the significance,

interpretation, or explanation for certain data or information. It is the ability to

.

Creativity‐ It is the ability to generate new ideas or to conceive new perspectives

on existing data.

Department of Electrical & Electronics EngineeringMuthiah Polytechnic College, Annamalainagar

1/29/2012 6



Domain areas of AI

ame p ay ng earc

Speech recognition

Computer Vision

Expert systems

Heuristic classification

Robotics


1/29/2012 7



Knowledge based systems

now e ge s e ne as e remem er ng o prev ous y earne ma er a . s

the result

of

learning

and

reasoning.

,

cells) in the brain, which contains approximately 10 12 neurons.

Knowled e la s an im ortant role in buildin intelli ent knowled e based

systems.

Knowledge needs to be acquired from different sources such as procedures,

rules and

facts.

Knowledge is to be stored/organised in such a way that retrieval of information


1/29/2012 8

can be easy.



Knowledge based systems

In knowledge based systems, decisions and actions are based on the

manipulation of knowledge. (facts are compared and altered in some

manner). Manipulation is computational equivalent of reasoning.

Knowledge can be represented using various schemes. Eg. FOPL (First Order

Predicate Logic), and associative network.


1/29/2012 9



The knowledge base contains predefined concepts, domain constraints,

.

Knowledge representation approaches that can be used in learning systems are

•production rules

•Frames

•semantic networks

•predicate calculus

•vectors and matrices

•Graphs

•formal grammar

and

•procedural encoding


1/29/2012 10



FOPL‐First order predicate logic

‐is a representation scheme for reasoning.

‐it comprises of symbols to represent statements.

Predicate – This s mbol returns a value of true or false. Ca ital

letters and words are used to represent predicates.

(eg.

A,

B,

NOT,

EQUAL) –

(eg. Has‐a , part‐of ,Father‐of)

Variables

ConstantsLogical quantifiers (Existential‐ there exist an x, Universal‐ for all x)

Connectors v‐dis unction ~ ‐ ne ation ^ ‐ con unction


1/29/2012 11



Structure of machine learning system

The

machine

learning

system•the environment

•t e earning e ement

•the knowledge base

earn ng emen now e ge ase er ormance

Element

nv ronmen


1/29/2012 12



Structure of machine learning system

Learning

Element Knowledge

Base Performance

ElementEnvironment

The environment supplies some piece of information to the learning element

explicit knowledge base

the erformance element uses the knowled e base to erform its task.

Finally, the information gained during performing the task can serve as a

feedback to the learning element.


1/29/2012 13



earn ng sys em may acqu re ru es o e av our, escr p ons o

physical objects, problem‐solving heuristics, classification

taxonomies over a sample space, and many other types of

knowledge useful in the performance of a wide variety of tasks.

Complex tasks require more knowledge than simple ones.

The knowledge base grows more in size, the problems of

complicated.


1/29/2012 14



Definition for Learning

A computer program is said to learn from experienceE

withrespect to some class of tasks T and performance measure P, if

, ,

experience E.


1/29/2012 15



A handwriting recognition learning problem

Task T: recognizing and classifying handwritten words within images

Performance measure P: Percent of words correctly classified

ra n ng exper ence : ata ase o an wr tten wor s w t g ven

classifications


1/29/2012 16



A robot driving learning problem

Task T: driving on public four-lane highways using vision sensors.

Performance measure P: average distance travelled before an error

(as judged by human overseer)

Training experience E: a sequence of images and steering

commands recorded while observing a human driver.


1/29/2012 17



Some major issues

ome ma or ssues w c concern e process o earn ng now e ge n

general are:

data do not contain enough information a system cannot "learn" much

from it.

What should a system learn? In order to solve a particular problem, a

system has to learn specific features, dependencies relevant to the

solution, but not learn everything(Some irrelevant features only)


1/29/2012 18



Some major issues

How to test how well the system has learned appropriate knowledge?

Testing the learning process is usually done through measuring thelearning error. The main approaches are:

Partitioning of data. A part of the data, say 70%, is used for training and

the other part for testing.

The leaving‐one‐out method means that we train n times the system

with (n ‐ 1) examples and check the system's reaction to the left‐out

exam le. After doin this n times we can calculate the correct answer of

the system as the ratio between the number of correctly processedexamples and the number n of all the examples


1/29/2012 19



Machine learning

application areas of artificial intelligence (AI).

Machine learning is the study of making machines acquire new

knowledge, and recognize existing knowledge.

Machine learning is the capability of a computer to learn from

examples.


1/29/2012 20



Machine learning

performance criterion using example data or past experience.

With a model using some parameters, learning is the execution

of a computer program to optimize the parameters of the model.

future, or descriptive to gain knowledge from data, or both.


1/29/2012 21



Machine learning ‐ Goal

The goal of machine learning is to design programs that learnand/or discover, i.e. automatically improve their performance on

Successful learner

.

Makes general conclusions about the data it is trained on.

Act appropriately in new situations.


1/29/2012 22



Machine learning

There are certain patterns in the data.

,

construct a good and useful approximation.

at approx mat on may not exp a n everyt ng, ut may st

be able to account for some part of the data.

We believe that though identifying the complete process maynot be possible, we can still detect certain patterns or


1/29/2012 23

.



Machine learning draws on concepts and results from

Statistics

Artificial intelli ence

Information theory

BiologyCognitive science

Computational complexity and

on ro

eory


1/29/2012 24



Major types of Machine Learning

•Supervised learning

• nsuperv se earn ng

•Reinforcement learning


1/29/2012 25



In supervised learning, there is a “teacher” that provides

the learner with a set of input‐output pairs.


1/29/2012 26



Process of supervised learning


1/29/2012 27



Process of supervised learning

Training data


1/29/2012 28



Decision Tress Classifier

Training data


1/29/2012 29



In unsupervised learning, there is no teacher, providingdesired answers, but since the data are not entirely random,

can be applied in new cases.

Eg. Clustering

Algorithms


1/29/2012 30



Reinforcement learning

Reinforcement learning corresponds to something between supervised andunsupervised approaches.

It differs from supervised learning in the sense that explicit input‐output pairs

are not available.

An agent explores environment and is able to take actions. Depending on the

outcome of the series of actions taken, the agent is rewarded or penalized.

Reinforcement learning is called “learning with critic”, as opposed to learningwith a teacher which is the supervised learning.


1/29/2012 31



Reinforcement learning

Reinforcement learning is called “learning with critic”, as opposed to learningwith a teacher which is the supervised learning.

Eg. Game playing is an important research area in both artificial intelligence

and machine

learning.


1/29/2012 32



Learning approaches are also classified as

Statistical (probabilistic or stochastic) methods

Connectionist methods/neural networks

Symbolic machine learning algorithms

ene c me o s an

Other hybrid approaches


1/29/2012 33



Statistical methods

.

Perform some analysis which uses primarily the text characteristics without

.

Statistical techniques are

N‐

ram techni ues

Unsupervised clustering and

Hidden Markov model

that have been used for corpus‐based language analysis, probabilistic

grammar learning and lexicon‐building.


1/29/2012 34



Statistical methods

Naïve Bayes classifier

K-Nearest neighbour

Hidden Markov Model

xpec a on ax m sa on a gor m orwar - ac war

Maximum Entropy models


1/29/2012 35



A neural network

A neural network is based on concepts of how the human brain is organised

and how it learns.

The nodes correspond to the neurons in the brain, and the links correspond

to the connections between neurons.

n or er o ma e pre c on, e neura ne wor accep s e va ues or e

predictors on the input nodes. These values are then multiplied by values

that are stored in the links, called weights.

These values are then added together at the output node, and a specialthreshold function is applied to get the prediction.


1/29/2012 36



Symbolic learning methods

Do not use probabilities explicitly

ec s on trees

Transformation based learning (TBL)

Inductive logic programming (ILP) and

.


1/29/2012 37



Minimum description length

MDL approaches aim to minimize the description of the set of words in the

input corpus.

,

defined as reducing the total length of a set of data.

Introducin a theor which can enerate certain data and thus serves as anabbreviation of the data set.

The implementation uses a learning mechanism which decreases the total

description length in each step.


1/29/2012 38



Genetic Algorithms (GAs)

Randomized search and optimization techniques

Guided by the principles of evolution and natural genetics

Efficient, adaptive and robust search processes


1/29/2012 39



Hybrid Methods ‐ Ensemble Learning

examples.

,

the so called base classifier, by changing the training set or the input features

or the parameters of the classifier.

The predictions of all base classifiers are combined into a single final

prediction.

The idea builds on the assumption that combining the output of multipleexperts is better than the output of any single expert


1/29/2012 40



Hybrid Methods ‐ Ensemble Learning


1/29/2012 41



Selection of features

ac ne earn ng requ res se ec on o

Samples and features

Choice of algorithm.

Features are usuall re- rocessed.Dimensional reduction will be used to identify a subset of features, or

mathematical combinations of features that greatly reduces the size of the

machine learning problem.

A distance metric represents how far samples are separated from one another

in `feature space'. Proper representation schemes should be used for better


1/29/2012 42

learning.



About Features

The input vector is called by a variety of names, some of these are input

vector, pattern vector, feature vector.

e componen s o e npu vec or are var ous y ca e ea ures,

attributes, input variables, and components.

The values of the components can be of three main types: real valued

numbers, discrete valued numbers, or categorical values .


1/29/2012 43



About Features

The number of features in the instances determine the search space that needs

to be explored by the learning algorithm for a given classification task.

e presence o a arge num er o rre evan ea ures unnecessar y ncreases e

size of the search space, thus increasing the time needed for classification.

, ,

it difficult to extract knowledge such as classification rules in a way that is

comprehensible to humans. Conversely, the rules based on a small number of

relevant features are often concise easier to understand and use.

The most important reason behind feature selection is that it can eliminate the

effects of the curse of dimensionality.


1/29/2012 44



Applications of

machine

learning

Problems that can be solved by Machine learning are classified as

Association rule learnin

Classification

Prediction

Regression

Pattern recognition


1/29/2012 45



Applications of

machine

learning

Association rule learning

One application of machine learning is basket analysis , which isfinding associations between products bought by customers:

If people who buy X typically also buy Y, and if there is a customer who

buys X and does not buy Y, he or she is a potential Y customer.

Department of Electrical & Electronics EngineeringMuthiah Polytechnic College, Annamalainagar1/29/2012 46



Applications of

machine

learning


Finding an association rule‐

is learning a conditional probability of theform P(YIX)

where Y is the product we would like to condition on X, which is the

product or the set of products which we know that the customer has




Applications of

machine

learning


We may want to make a distinction among customers and toward this,estimate P(YIX, D)

where D is the set of customer attributes, for example, gender, age, marital

status, and so on.

If this is a bookseller instead of a supermarket, products can be books or

authors.

In the case of a Web portal, items correspond to links to Web pages, and wecan estimate the links a user is likely to click and use this information to

download such pages in advance for faster access.




Applications of

machine

learning

Classification

In

credit scoring,

the

bank

calculates

the

risk

given

the

amount

of

credit

and

the information about the customer.

This is an example of a classification problem where there are two classes:

low‐risk and high‐risk customers.

The information about a customer makes up the input to the classifier whose task is to assign the input to one of the two classes.




Applications of

machine

learning

Classification & Prediction

After training with the past data, a classification rule learned may be of the form

IF income>

p1

AND

savings>

p2

THEN

low

‐risk

ELSE

high

‐risk

It is a function that separates the examples of different classes.

predictions for novel instances , , if the future is similar to the past.




Applications of

machine

learning

Regression

In a system that can predict the price of a used car.

‐ , , , ,

information‐that we believe affect a car's worth.

.

Such problems where the output is a number are REGRESSION problems.




Applications of

machine

learning

Clustering

Clustering is grouping of input.

with a data of customers, a clustering model allocates customers similar in

their attributes to the same group.

Input : demographic information as well as the past transactions with the

.

the company may decide strategies, services and products, specific to

.

Such a grouping also allows identifying those who are outliers, namely, those

who are different from other customers.




Applications of

machine

learning

in Pattern Recognition

2

1




Applications of

machine

learning

in Pattern Recognition

is recognizing characters from their images

Character recognition

• Printed character recognition (OCR) – Collection of dots

• Handwritten character recognition – Collection of dots and strokes

There are multiple classes - as many as the number of characters

A character image is not just a collection of random dots; it is a collection of strokes and has a regularity that we can capture by a learning program.





Applications of

machine

learning

Sequence learning

A word is a sequence of characters and successive characters are notindependent but are constrained by the words of the language.

This has the advantage that even if we cannot recognize a character, we can

still read t?e word.

Such contextual dependencies may also occur in higher levels, between

words and sentences, through the syntax and semantics of the language.

There are machine learning algorithms to learn sequences and model suchdependencies



equence earn ng me o s are a so use n o n orma cs



Applications of

machine

learning

Face Recognition

In Face recognition

•the classes are people to be recognized

•the learning program should learn to associate the face images to

identities.





Applications of

machine

learning

The input is acoustic

The classes are words that can be uttered

This time the association to be learned is from an acoustic signal to a word

.





Applications of

machine

learning

Medical diagnosis The inputs are the relevant information about the patient

The inputs contain the patient's age, gender, past medical history, and

.





Applications of

machine

learning

Outlier detection

Another use of machine learning is outlier detection, .

After learning the rule, we are not interested in the rule but the exceptions

,

.





Applications of

machine

learning

Outlier detection

Learning a rule from data allows knowledge extraction. ,

we have an explanation about the process underlying the data.

‐ ‐,

risk customers (Credit scoring) we have the knowledge of the properties of

low‐risk customers.

We

can

then

use

this

information

to

target

potential

low‐

risk

customers

more efficiently.





Applications of

machine

learning

Compression

Learning also performs compression by fitting a rule to the data.

,

to store and less computation to process.

.

,

sum of every possible pair of numbers.





Applications of

machine

learning

Networking & Communications

In networking and telecommunications, call patterns and traffic data are

analyzed for network optimization and maximizing the quality of service.





Applications of machine learning in

Natural Lan ua e Processin

Machine learning also helps us find solutions for many natural language

processing tasks.

Text categorisation

Text summarisation

Word segmentation

tagg ng

Parsing

Word sense disambiguation

Unknown word recognitionSpeech recognition and





Applications of machine learning in

Natural Lan ua e Processin

Machine learning also helps us find solutions for many natural language

processing tasks.

anguage en ca on

From text

From Speech

Speaker Diarization

To identify the speaker changes and the speaker clusters, and to estimate the

number of speakers involved in the document

To be

able

to

process

speech

documents

as

well

as

documents

containing

music, silence, and other sounds.





Speaker Diarization

•To identify the speaker turns and the speaker clusters, and to estimate the

number of

speakers

involved

in

the

document,

without

any

priori

information.

• To e a e to process speec ocuments as we as ocuments containing music,

silence, and other sounds.



Prof. K. Rajan

Documents