[PDF] - Outline Wh y Mac hine Learning? What is a w ell-dened PDF Document

SLIDE 1 Outline

Wh

y Mac hine Learning?

What

is a w ell-dened learning problem?

An

example: learning to pla y c hec k ers

What

questions should w e ask ab

ut

Mac hine Learning? 1 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 2 Wh y Mac hine Learning

Recen

t progress in algorithms and theory

Gro

wing

d
f
nline

data

Computational

p

w

er is a v ailable

Budding

industry Three nic hes for mac hine learning:

Data

mining : using historical data to impro v e decisions { medical records ! medical kno wledge

Soft

w are applications w e can't program b y hand { autonomous driving { sp eec h recognition

Self

customizing programs { Newsreader that learns user in terests 2 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 3 T ypical Datamining T ask Data:

Patient103 Patient103 Patient103 ...

time=1 time=2 time=n Age: 23 FirstPregnancy: no Anemia: no Diabetes: no PreviousPrematureBirth: no ... Elective C−Section: ? Emergency C−Section: ? Age: 23 FirstPregnancy: no Anemia: no PreviousPrematureBirth: no Diabetes: YES ... Emergency C−Section: ? Ultrasound: abnormal Elective C−Section: no Age: 23 FirstPregnancy: no Anemia: no PreviousPrematureBirth: no ... Elective C−Section: no Ultrasound: ? Diabetes: no

Emergency C−Section: Yes

Ultrasound: ?

Giv en:

9714

patien t records, eac h describing a pregnancy and birth

Eac

h patien t record con tains 215 features Learn to predict:

Classes
f

future patien ts at high risk for Emergency Cesarean Section 3 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 4 Datamining Result Data:

Patient103 Patient103 Patient103 ...

time=1 time=2 time=n Age: 23 FirstPregnancy: no Anemia: no Diabetes: no PreviousPrematureBirth: no ... Elective C−Section: ? Emergency C−Section: ? Age: 23 FirstPregnancy: no Anemia: no PreviousPrematureBirth: no Diabetes: YES ... Emergency C−Section: ? Ultrasound: abnormal Elective C−Section: no Age: 23 FirstPregnancy: no Anemia: no PreviousPrematureBirth: no ... Elective C−Section: no Ultrasound: ? Diabetes: no

Emergency C−Section: Yes

Ultrasound: ?

One

f

18 learned rules: If No previous vaginal delivery, and Abnormal 2nd Trimester Ultrasound, and Malpresentation at admission Then Probability

f

Emergency C-Section is 0.6 Over training data: 26/41 = .63, Over test data: 12/20 = .60 4 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 5 Credit Risk Analysis Data:

Customer103: Customer103: Customer103:

(time=t0) (time=t1) (time=tn)

...

... Own House: Yes Other delinquent accts: 2 Loan balance: $2,400 Income: $52k Max billing cycles late: 3 Years of credit: 9 Profitable customer?: ? ... Own House: Yes Years of credit: 9 Profitable customer?: ? ... Own House: Yes Years of credit: 9 Loan balance: $3,250 Income: ? Other delinquent accts: 2 Max billing cycles late: 4 Loan balance: $4,500 Income: ? Other delinquent accts: 3 Max billing cycles late: 6

Profitable customer?: No

Rules learned from syn thesized data: If Other-Delinquent-A ccoun ts > 2, and Number-Delinquent- Billi ng-Cy cles > 1 Then Profitable-Custome r? = No [Deny Credit Card application] If Other-Delinquent-A ccoun ts = 0, and (Income > $30k) OR (Years-of-Credit > 3) Then Profitable-Custome r? = Yes [Accept Credit Card application] 5 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 6 Other Prediction Problems Customer purc hase b eha vior:

Customer103: Customer103: Customer103:

(time=t0) (time=t1) (time=tn)

...

... Sex: M Age: 53 Income: $50k Own House: Yes MS Products: Word Computer: 386 PC Purchase Excel?: ? ... Sex: M Age: 53 Income: $50k Own House: Yes MS Products: Word ... Sex: M Age: 53 Income: $50k Own House: Yes Purchase Excel?: ? MS Products: Word Computer: Pentium Computer: Pentium

Purchase Excel?: Yes

Customer reten tion:

Customer103: Customer103:

Age: 53 Age: 53 Age: 53 Sex: M Sex: M Sex: M

Customer103:

(time=t0) (time=t1) (time=tn)

...

Income: $50k Income: $50k Income: $50k Own House: Yes Own House: Yes Own House: Yes Checking: $5k Checking: $20k Checking: $0 Savings: $15k Savings: $0 Savings: $0 ... ... Current−customer?: yes

Current−customer?: No

Current−customer?: yes

Pro cess

ptimization:

(time=t0) (time=t1) (time=tn)

... Product72: Product72: Product72:

... Viscosity: 1.3 ... ... Viscosity: 1.3 Product underweight?: ??

Product underweight?:

Viscosity: 3.2

Yes

Fat content: 15% Stage: mix Mixing−speed: 60rpm Density: 1.1 Stage: cook Temperature: 325 Fat content: 12% Density: 1.2 Stage: cool Fan−speed: medium Fat content: 12% Spectral peak: 3200 Density: 2.8 Spectral peak: 2800 Spectral peak: 3100 Product underweight?: ??

6 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 7 Problems T

Dicult

to Program b y Hand AL VINN [P

merleau]

driv es 70 mph

n

high w a ys

Sharp Left Sharp Right

4 Hidden Units 30 Output Units 30x32 Sensor Input Retina

Straight Ahead

7 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 8 Soft w are that Customizes to User h ttp://www.wisewi re.com 8 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 9 Where Is this Headed? T

da

y: tip

f

the iceb erg

First-generation

algorithms: neural nets, decision trees, regression ...

Applied

to w ell-formated database

Budding

industry Opp

rtunit

y for tomorro w: enormous impact

Learn

across full mixed-media data

Learn

across m ultiple in ternal databases, plus the w eb and newsfeeds

Learn

b y activ e exp erimen tation

Learn

decisions rather than predictions

Cum

ulativ e, lifel

ng

learning

Programm

ing languages with learning em b edded? 9 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 10 Relev an t Discipli nes

Articial

in telli gence

Ba

y esian metho ds

Computational

complexit y theory

Con

trol theory

Information

theory

Philosoph

y

Psyc

hology and neurobiology

Statistics
:

: : 10 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 11 What is the Learning Problem? Learning = Impro ving with exp erience at some task

Impro

v e

v

er task T ,

with

resp ect to p erformance measure P ,

based
n

exp erience E . E.g., Learn to pla y c hec k ers

T

: Pla y c hec k ers

P

: %

f

games w

n

in w

rld

tournamen t

E

:

pp
rtunit

y to pla y against self 11 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 12 Learning to Pla y Chec k ers

T

: Pla y c hec k ers

P

: P ercen t

f

games w

n

in w

rld

tournamen t

What

exp erience?

What

exactly should b e learned?

Ho

w shall it b e represen ted?

What

sp ecic algorithm to learn it? 12 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 13 T yp e

f

T raining Exp erience

Direct
r

indirect?

T

eac her

r

not? A problem: is training exp erience represen tativ e

f

p erformance goal? 13 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 14 Cho

se

the T arget F unction

C

hooseM

v

e : B

ar

d ! M

v

e ??

V

: B

ar

d ! < ??

...

14 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 15 P

ssible

Denition for T arget F unction V

if

b is a nal b

ard

state that is w

n,

then V (b) = 100

if

b is a nal b

ard

state that is lost, then V (b) = 100

if

b is a nal b

ard

state that is dra wn, then V (b) =

if

b is a not a nal state in the game, then V (b) = V (b ), where b is the b est nal b

ard

state that can b e ac hiev ed starting from b and pla ying

ptimally

un til the end

f

the game. This giv es correct v alues, but is not

p

erational 15 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 16 Cho

se

Represen tation for T arget F unction

collecti
n
f

rules?

neural

net w

rk

?

p
lynomial

function

f

b

ard

features?

...

16 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 17 A Represen tation for Learned F unction w +w 1 bp(b)+w 2 r p(b)+w 3 bk (b)+w 4 r k (b)+w 5 bt(b)+w 6 r t(b)

bp(b):

n um b er

f

blac k pieces

n

b

ard

b

r

p(b): n um b er

f

red pieces

n

b

bk

(b): n um b er

f

blac k kings

n

b

r

k (b): n um b er

f

red kings

n

b

bt(b):

n um b er

f

red pieces threatened b y blac k (i.e., whic h can b e tak en

n

blac k's next turn)

r

t(b): n um b er

f

blac k pieces threatened b y red 17 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 18 Obtaining T raining Examples

V

(b): the true target function

^

V (b) : the learned function

V

tr ain (b): the training v alue One rule for estimating training v alues:

V

tr ain (b) ^ V (S uccessor (b)) 18 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 19 Cho

se

W eigh t T uning Rule LMS W eigh t up date rule: Do rep eatedly:

Select

a training example b at random 1. Compute er r

r

(b): er r

r

(b) = V tr ain (b)

^

V (b) 2. F

r

eac h b

ard

feature f i , up date w eigh t w i : w i w i + c

f

i

er

r

r

(b) c is some small constan t, sa y 0.1, to mo derate the rate

f

learning 19 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 20 Design Choices

Determine Target Function Determine Representation

f Learned Function

Determine Type

f Training Experience

Determine Learning Algorithm Games against self Games against experts Table of correct moves Linear function

f six features

Artificial neural network Polynomial Gradient descent Board ➝ value Board ➝ move Completed Design

... ...

Linear programming

... ...

20 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997

SLIDE 21 Some Issues in Mac hine Learning

What

algorithms can appro ximate functions w ell (and when)?

Ho

w do es n um b er

f

training examples inuence accuracy?

Ho

w do es complexit y

f

h yp

thesis

represen tation impact it?

Ho

w do es noisy data inuence accuracy?

What

are the theoretical limits

f

learnabilit y?

Ho

w can prior kno wledge

f

learner help?

What

clues can w e get from biological learning systems?

Ho

w can systems alter their

wn

represen tations? 21 lecture slides for textb

k

Machine L e arning, T. Mitc hell, McGra w Hill, 1997