LEARNING
FROM
DATA
The book website AMLbook. com
contains supporting material
for
instructors
and readers.
LEARNING FROM
DATA
A SHORT COURSE
Yaser S . Abu-Mostafa
Institute
California
of Technology
Malik Magdon-Ismail
Polytechnic
Rensselaer
Institute
Hsuan-Tien Lin
National
Taiwan University
AMLbook.com
of Technology
Polytechnic
Institute
Rensselaer
Troy, NY 12180, USA
magdon@cs.rpi.edu
of Electrical
Engineering
Malik Magdon Ismail
Department
of Computer
Science
Yaser S. Abu 1/fostafa
Departments
and Computer
Science
California
Institute
CA 9 1125, USA
Pasadena,
yaser©caltech.edu
HsuanTien Lin
Department
and Information
National
Taipei,
htlin©csie.ntu.edu.tw
of Computer
106, Taiwan
Science
Engineering
Taiwan University
ISBN 1 0: 1 60049 006 9
ISBN 13:9781 60049 006 4
@2012 Yaser S. Abu Mostafa,
Malik Magdon Ismail, Hsuan Tien Lin.
1.10
reserved.
All rights
without
the written
be reproduced,
means-electronic,
written
permission
the 1976 United States
mechanical,
of the authors,
Copyright Act.
This work may not be translated
permission
of the authors.
or copied in whole or in part
stored in a retrieval
system,
or transmitted
No part of this publication
may
in any form or by any
photocopying, scanning,
or otherwise-without
under Section
107 or 108 of
prior
except as permitted
any implied
Limit of Liability
efforts in preparing
spect to the accuracy
disclaim
No warranty
materials.
situation.
shall not be liable
but not limited
/Disclaimer
of Warranty:
While the authors
have used their
best
this book, they make no representation
or warranties
with re
or completeness
of the contents
of this book and specifically
warranties
of merchantability
or fitness
for a particular purpose.
may be created
or extended
by sales representatives
or written
The advice and strategies
You should consult
herein
with a professional
contained
may not be suitable for
where appropriate.
The authors
sales
your
for any loss of profit or any other commercial
damages, including
to special,
incidental
, consequential,
or other damages.
The use in this publication
terms, even if they are not identified
opinion
as to whether
or not they are subject
of tradenames,
trademarks,
marks, and similar
as such, is not to be taken as an expression
service
of
to proprietary
rights.
This book was typeset
States
of America.
by the authors
and was
printed
and bound in the United
To our teachers)
and to our students
Preface
learning.
not a hurried course. From over
for a short course
of teaching
on machine
a decade
It is a short
to be the core topics
know. We chose the title
'learning
this material,
we
that every student
of the
from data' that faithfully
is about, and made it a point to cover the
topics
in
fashion. Our hope is that the reader
can learn all the fundamentals
should
what we believe
This book is designed
course,
have distilled
subject
describes what the
a story-like
of the subject
Learning
by reading
subject
the book cover to cover.
from data has distinct
theoretical
and practical
tracks.
If you
feel that you
subjects
altogether
about two different
and the practical,
read two books that focus on one track or the other, you may
are reading
the theoretical
criterion
framework for
formance
parts are spelled out. Our philosophy
what we don't know, and what we partially know.
the mathematical
and the heurist
e. Theory that establishes
and so are heuristics
Strengths
for inclusion
learning
is relevanc
is included,
is to say it like it is: what we know,
of real learning
that impact the per
systems.
ic. Our
. In this book, we balance
and weaknesses of the different
the conceptual
The book can be taught
in exactly
the order
it is presented. The notable
may be Chapter
2, which is the
of generalization
exception
The theory
from data, and we made an effort to make it accessible
However, instructors
in the practical
over it, or delay it until after the practical methods
chapter
most theoretical
is central
to a wide readership.
side may skim
3 are taught.
of the book.
to learning
who are more interested
that this chapter
of Chapter
covers
exercises (in gray boxes) throughout
the
these exercises
and enhance
of a particular topic being covered.
for separating
Nevertheless,
is to engage the reader
to the logical flow.
are not crucial
Our reason
that we included
You will notice
text. The main purpose of
understanding
the exercises
they contain
even if you don't do them to completion.
exercises
ditional
each chapter.
appropriate
problems of varying
out is that they
useful
information,
homework
difficulty
as 'easy'
in the
and we strongly encourage
Instructors
problems,
Problems section
at the end of
you to read them,
may find some of the
and we also provide
ad
To help instructors
material
provide
supporting
also a forum that covers
their lectures
with preparing
on the book's
additional
topics
based on the book, we
(AMLbook.
corn). There is
website
We will
from data.
in learning
vii