I Why Machine Learning? - Computational Logic · 1 Machine Learning – Introduction IWhy Machine...

1 Machine Learning – Introduction

I Why Machine Learning?

I What is a well-defined learning problem?

I An example: learning to play checkers

I What questions should we ask about Machine Learning?

1 Machine Learning – Introduction

I Why Machine Learning?

I What is a well-defined learning problem?

I An example: learning to play checkers

I What questions should we ask about Machine Learning?

I Tom M. Mitchell: Machine Learning. McGraw-Hill: 1997.

1 Machine Learning – Introduction (4th April 2006) 1

Why Machine Learning?

I Recent progress in algorithms and theory

I Growing flood of online data

I Computational power is available

I Budding industry

Why Machine Learning?

I Recent progress in algorithms and theory

I Growing flood of online data

I Computational power is available

I Budding industry

I Niches for machine learning:

. Data mining: using historical data to improve decisions

. medical records medical knowledge

. learning to program: we can’t program everything by hand

. autonomous driving

. speech recognition

. Self customizing programs

. Newsreader that learns user interests


Typical Datamining Task

I Data:Patient103 Patient103Patient103 ...time=1 time=2 time=n

Age: 23FirstPregnancy: noAnemia: noDiabetes: noPreviousPrematureBirth: no

...

Elective C−Section: ?Emergency C−Section: ?

Age: 23FirstPregnancy: noAnemia: no

PreviousPrematureBirth: noDiabetes: YES

...Emergency C−Section: ?

Ultrasound: abnormal

Elective C−Section: no


PreviousPrematureBirth: no

...


Ultrasound: ?

Diabetes: no

Emergency C−Section: Yes

Ultrasound: ?

I Given:

. 9714 patient records, each describing a pregnancy and birth

. Each patient record contains 215 features

I Learn to predict:

. Classes of future patients at high risk for Emergency Cesarean Section


Datamining ResultI Data:

Patient103 Patient103Patient103 ...time=1 time=2 time=n

Age: 23FirstPregnancy: noAnemia: noDiabetes: noPreviousPrematureBirth: no

...

Elective C−Section: ?Emergency C−Section: ?


PreviousPrematureBirth: noDiabetes: YES

...Emergency C−Section: ?

Ultrasound: abnormal



PreviousPrematureBirth: no

...


Ultrasound: ?

Diabetes: no

Emergency C−Section: Yes

Ultrasound: ?

I One of 18 learned rules:

If No previous vaginal delivery, andAbnormal 2nd Trimester Ultrasound, andMalpresentation at admission

Then Probability of Emergency C-Section is 0.6

Over training data: 26/41 = .63,Over test data: 12/20 = .60


Credit Risk AnalysisI Data:

Customer103: Customer103: Customer103:(time=t0) (time=t1) (time=tn)...

...

Own House: Yes

Other delinquent accts: 2

Loan balance: $2,400Income: $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: 9Loan balance: $3,250Income: ?

Other delinquent accts: 2Max billing cycles late: 4

Loan balance: $4,500Income: ?

Other delinquent accts: 3Max billing cycles late: 6Profitable customer?: No

I Rules learned from synthesized data:

If Other-Delinquent-Accounts > 2, andNumber-Delinquent-Billing-Cycles > 1

Then Profitable-Customer\TR{?} = No[Deny Credit Card application]

If Other-Delinquent-Accounts = 0, and(Income > $30k) OR (Years-of-Credit > 3)

Then Profitable-Customer\TR{?} = Yes[Accept Credit Card application]


Other Prediction Problems

I Customer purchase behavior:


...

Sex: MAge: 53Income: $50kOwn House: Yes

MS Products: WordComputer: 386 PC

Purchase Excel?: ?

...


MS Products: Word

...


Purchase Excel?: ?

MS Products: WordComputer: Pentium Computer: Pentium

Purchase Excel?: Yes

Other Prediction Problems

I Customer purchase behavior:


...


MS Products: WordComputer: 386 PC

Purchase Excel?: ?

...


MS Products: Word

...


Purchase Excel?: ?

MS Products: WordComputer: Pentium Computer: Pentium

Purchase Excel?: Yes

I Process optimization:

(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: mixMixing−speed: 60rpm

Density: 1.1

Stage: cookTemperature: 325

Fat content: 12%

Density: 1.2

Stage: coolFan−speed: medium

Fat content: 12%

Spectral peak: 3200Density: 2.8Spectral peak: 2800 Spectral peak: 3100

Product underweight?: ??


Problems Too Difficult to Program by Hand

I ALVINN [Pomerleau] drives 70 mph on highways

Sharp Left

SharpRight

4 Hidden Units

30 Output Units

30x32 Sensor Input Retina

Straight Ahead


Software that Customizes to User


Where Is this Headed?

I Today: tip of the iceberg

. First-generation algorithms: neural nets, decision trees, regression, etc.

. Applied to well-formated database

. Budding industry

I Opportunity for tomorrow: enormous impact

. Learn across full mixed-media data

. Learn across multiple internal databases, plus the web and newsfeeds

. Learn by active experimentation

. Learn decisions rather than predictions

. Cumulative, lifelong learning

. Programming languages with embedded learning


Relevant Disciplines

I Artificial intelligence

I Bayesian methods

I Computational complexity theory

I Control theory

I Information theory

I Philosophy

I Psychology and neurobiology

I Statistics

I . . .


What is the Learning Problem?

I Learning = Improving with experience at some task.

. Improve over task T ,

. with respect to performance measure P ,

. based on experience E.

I E.g., Learn to play checkers:

. T : play checkers,

. P : percentage of games won in world tournament,

. E: opportunity to play against self.

I E.g., Learning to drive:

. T : driving on public four-lane highway using vision sensors,

. P : average distance traveled before an error,

. E: a sequence of images and steering commandsrecorded while observing a human driver.


Learning to Play Checkers

I T : play checkers

I P : percentage of games won in world tournament

Learning to Play Checkers

I T : play checkers

I P : percentage of games won in world tournament

I What experience?

I What exactly should be learned?

I How shall it be represented?

I What specific algorithm to learn it?


Type of Training Experience

I Direct or indirect?

I The problem of credit assignment.

I Teacher or not?

Type of Training Experience

I Direct or indirect?

I The problem of credit assignment.

I Teacher or not?

I Problem is training experience representative of performance goal?


Choose the Target Function

I ChooseMove : Board→Move ?

I V : Board→ R ?

I ...


Possible Definition for Target Function V

I if b is a final board state that is won, then V (b) = 100

I if b is a final board state that is lost, then V (b) = −100

I if b is a final board state that is drawn, then V (b) = 0

I if b is a not a final state in the game, then V (b) = V (b′), where b′ is the bestfinal board state that can be achieved starting from b and playing optimally untilthe end of the game.

. This gives correct values, but is not operational.

. Ultimate goal Find an operational description of the ideal target function V .

. But we can often only acquire some approximation V̂ .


Choose Representation for Target Function

I Collection of rules?

I Neural networks ?

I Polynomial function of board features?

I etc.


A Representation for a Learned Function

w0 + w1 · bp(b) + w2 · rp(b) + w3 · bk(b) + w4 · rk(b) + w5 · bt(b) + w6 · rt(b)

I bp(b): number of black pieces on board b

I rp(b): number of red pieces on b

I bk(b): number of black kings on b

I rk(b): number of red kings on b

I bt(b): number of red pieces threatened by black(i.e., which can be taken on black’s next turn)

I rt(b): number of black pieces threatened by red


Obtaining Training Examples

I V (b): the true target function

I V̂ (b): the learned function

I Vtrain(b): the training value

Obtaining Training Examples

I V (b): the true target function

I V̂ (b): the learned function

I Vtrain(b): the training value

I One rule for estimating training values:

. Vtrain(b)← V̂ (Successor(b))


Choose Weight Tuning Rule

I Least mean square weight update rule:

I Do repeatedly:

. Select a training example b at random

. Compute error(b):

error(b) = Vtrain(b)− V̂ (b)

. For each board feature fi, update weight wi:

wi ← wi + c · fi · error(b)

I c is some small constant, say 0.1, to moderate the rate of learning


Design Choices

DetermineTarget Function

Determine Representationof Learned Function

Determine Type of Training Experience

DetermineLearning Algorithm

Games against self

Games against experts Table of correct

moves

Linear functionof six features

Artificial neural network

Polynomial

Gradient descent

Board Ý value

BoardÝ move

Completed Design

...

...

Linear programming

...

...


Some Issues in Machine Learning

I What algorithms can approximate functions well (and when)?

I How does the number and distribution of the training examplesinfluence the accuracy?

I How does the complexity of the hypothesis representation impact it?

I How does noisy data influence accuracy?

I What are the theoretical limits of learnability?

I How can prior knowledge help the learner?

I What clues can we get from biological learning systems?

I How can systems alter their own representations?


I Why Machine Learning? - Computational Logic · 1 Machine Learning – Introduction IWhy Machine...

Documents

Transcript of I Why Machine Learning? - Computational Logic · 1 Machine Learning – Introduction IWhy Machine...