Outline Wh y Mac hine Learning? What is a w ell-dened - PDF document

Outline � Wh y Mac hine Learning? � What is a w ell-de�ned learning problem? � An example: learning to pla y c hec k ers � What questions should w e ask ab out Mac hine Learning? 1 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

Wh y Mac hine Learning � Recen t progress in algorithms and theory � Gro wing �o o d of online data � Computational p o 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 o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

T ypical Datamining T ask Data: Giv en: Patient103 Patient103 Patient103 ... time=n time=1 time=2 � 9714 patien t records, eac h describing a Age: 23 Age: 23 Age: 23 pregnancy and birth FirstPregnancy: no FirstPregnancy: no FirstPregnancy: no Anemia: no Anemia: no Anemia: no Diabetes: no Diabetes: no Diabetes: YES � Eac h patien t record con tains 215 features PreviousPrematureBirth: no PreviousPrematureBirth: no PreviousPrematureBirth: no Ultrasound: ? Ultrasound: abnormal Ultrasound: ? Learn to predict: Elective C−Section: ? Elective C−Section: no Elective C−Section: no Emergency C−Section: ? Emergency C−Section: ? Emergency C−Section: Yes ... ... ... � Classes of future patien ts at high risk for Emergency Cesarean Section 3 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

Datamining Result Data: One of 18 learned rules: If No previous vaginal delivery, and Patient103 Patient103 Patient103 ... time=n time=1 time=2 Abnormal 2nd Trimester Ultrasound, and Age: 23 Age: 23 Age: 23 FirstPregnancy: no FirstPregnancy: no FirstPregnancy: no Malpresentation at admission Anemia: no Anemia: no Anemia: no Then Probability of Emergency C-Section Diabetes: no is 0.6 Diabetes: no Diabetes: YES PreviousPrematureBirth: no PreviousPrematureBirth: no PreviousPrematureBirth: no Ultrasound: ? Ultrasound: abnormal Ultrasound: ? Elective C−Section: ? Elective C−Section: no Elective C−Section: no Over training data: 26/41 = .63, Emergency C−Section: ? Emergency C−Section: ? Emergency C−Section: Yes ... ... ... Over test data: 12/20 = .60 4 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

Credit Risk Analysis Data: Rules learned from syn thesized data: If Other-Delinquent-A ccoun ... ts > 2, and Customer103: Customer103: Customer103: (time=t0) (time=t1) (time=tn) Years of credit: 9 Years of credit: 9 Years of credit: 9 Number-Delinquent- Billi ng-Cy cles > 1 Loan balance: $2,400 Loan balance: $3,250 Loan balance: $4,500 Then Profitable-Custome r? = No Income: $52k Income: ? Income: ? Own House: Yes Own House: Yes Own House: Yes [Deny Credit Card application] Other delinquent accts: 2 Other delinquent accts: 2 Other delinquent accts: 3 Max billing cycles late: 3 Max billing cycles late: 4 Max billing cycles late: 6 Profitable customer?: No Profitable customer?: ? Profitable customer?: ? ... 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 o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

Other Prediction Problems Customer purc hase b eha vior: Customer reten tion: ... Customer103: Customer103: Customer103: (time=t0) (time=t1) (time=tn) Sex: M Sex: M Sex: M Age: 53 Age: 53 Age: 53 Income: $50k Income: $50k Income: $50k Own House: Yes Own House: Yes Own House: Yes MS Products: Word MS Products: Word MS Products: Word Computer: 386 PC Computer: Pentium Computer: Pentium Pro cess optimization: Purchase Excel?: Yes Purchase Excel?: ? Purchase Excel?: ? ... ... ... ... Customer103: Customer103: Customer103: (time=t0) (time=t1) (time=tn) Sex: M Sex: M Sex: M Age: 53 Age: 53 Age: 53 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?: No Current−customer?: yes ... Current−customer?: yes ... 6 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997 ... Product72: Product72: Product72: (time=t0) (time=t1) (time=tn) Stage: mix Stage: cook Stage: cool Mixing−speed: 60rpm Temperature: 325 Fan−speed: medium Viscosity: 1.3 Viscosity: 3.2 Viscosity: 1.3 Fat content: 15% Fat content: 12% Fat content: 12% Density: 2.8 Density: 1.1 Density: 1.2 Spectral peak: 2800 Spectral peak: 3200 Spectral peak: 3100 Product underweight?: Product underweight?: ?? Product underweight?: ?? Yes ... ... ...

Problems T o o Di�cult to Program b y Hand AL VINN [P omerleau] driv es 70 mph on high w a ys Sharp Straight Sharp Left Ahead Right 30 Output Units 4 Hidden Units 7 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997 30x32 Sensor Input Retina

Soft w are that Customizes to User h ttp://www.wisewi re.com 8 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

Where Is this Headed? T o da y: tip of the iceb erg � First-generation algorithms: neural nets, decision trees, regression ... � Applied to w ell-formated database � Budding industry Opp ortunit 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 ong learning � Programm ing languages with learning em b edded? 9 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

Relev an t Discipli nes � Arti�cial 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 o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

What is the Learning Problem? Learning = Impro ving with exp erience at some task � Impro v e o v er task T , � with resp ect to p erformance measure P , � based on exp erience E . E.g., Learn to pla y c hec k ers � T : Pla y c hec k ers � P : % of games w on in w orld tournamen t � E : opp ortunit y to pla y against self 11 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

Learning to Pla y Chec k ers � T : Pla y c hec k ers � P : P ercen t of games w on in w orld tournamen t � What exp erience? � What exactly should b e learned? � Ho w shall it b e represen ted? � What sp eci�c algorithm to learn it? 12 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

T yp e of T raining Exp erience � Direct or indirect? � T eac her or not? A problem: is training exp erience represen tativ e of p erformance goal? 13 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

Cho ose the T arget F unction � C hooseM ov e : B oar d ! M ov e ?? � V : B oar d ! < ?? � ... 14 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

P ossible De�nition for T arget F unc- tion V � if b is a �nal b oard state that is w on, then V ( b ) = 100 � if b is a �nal b oard state that is lost, then V ( b ) = � 100 � if b is a �nal b oard state that is dra wn, then V ( b ) = 0 � if b is a not a �nal state in the game, then 0 0 V ( b ) = V ( b ), where b is the b est �nal b oard state that can b e ac hiev ed starting from b and pla ying optimally un til the end of the game. This giv es correct v alues, but is not op erational 15 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997

Cho ose Represen tation for T arget F unction � collecti on of rules? � neural net w ork ? � p olynomial function of b oard features? � ... 16 lecture slides for textb o ok Machine L e arning , T. Mitc hell, McGra w Hill, 1997