Can Computers Learn?

Tom M. Mitchell, "Does Machine Learning Really Work? AI Magazine, Fall 1997 v18 n3 p11(10)
 
 
 
Fredkin Professor of AI and Learning 

Director, Center for Automated Learning and Discovery 

School of Computer Science 

Carnegie Mellon University

  

"Although we do not yet know how to make computers learn nearly as well as people learn, in recent years many succesful machine-learning applications have been developed"

THE NICHE FOR MACHINE LEARNING

Data Mining

Difficult-to-Program Applications

Customized Software Applications
 
 

Data Mining

Use historical databases to improve subsequent decision making



Examples:

Banks: which future loan applicants are likely to be credit worthy?

Hosptials: Which new patients are likely to respond best to new treatments?

Credit Card Companies: Which transactions are likely to be fraudlent?

Businesses: Which new customers are likely to buy new products?

Colleges: Which new students are likely to graduate?

For more on data mining, see http://www.kdnuggets.com/

Example: Predicting Medical Outcomes from Historical Data

http://www.schenley.com

Data: records on 9714 pregnant women

215 attributes over time: health history, measurements, type of delivery, final health of mother and baby.

General Problem: Given time-series data, we want to learn to predict features that occur late in the time series based on features known earlier.

Which future patients are at exceptionally high risk of requiring an emergency caesarean section?
 
Patient 103 time =1

Age: 23

FirstPregnancy: no

Anemia: No

Diabetes: No

PreviousPrematureBirth?

No

Ultrasound: ?

Elective C-Section: ?

Emergency 

C-Section: ?

 Patient 103 time =2

Age: 23

FirstPregnancy: no

Anemia: No

Diabetes: YES

PreviousPrematureBirth?

No

Ultrasound: abnormal

Elective C-Section: no

Emergency 

C-Section: ?

 Patient 103 time =n

Age: 23

FirstPregnancy: no

Anemia: No

Diabetes: No

PreviousPrematureBirth?

No

Ultrasound: ?

Elective C-Section: no

Emergency 

C-Section: Yes

Learned Rule:

If         No previous vaginal delivery and

            Abnormal 2nd Trimester Ultrasound and

            Malpresentation at admissions,

Then

Probability of Emergency C-Section is 0.6

Training set accuracy: 26/41 = .63

Test set accuracy: 12/20 = .60

THE NICHE FOR MACHINE LEARNING

Data Mining

Difficult-to-Program Applications

Customized Software Applications
 
 

Some applications have proved too difficult for traditional manual programming:

Face Recognition

Speech Understanding

http://www.cs.cmu.edu/~tom/faces.html
 


 
 

THE NICHE FOR MACHINE LEARNING

Data Mining

Difficult-to-Program Applications

Customized Software Applications
 
 

Example: News Weeder: Learns the reading interests of its users.

News Weeder develops a general model of a reader's interest from a collection of specific text documents the user has rated for level of interest.

It then automatically examines new articles to produce a "top 20" list for the user.

K. Lang. News Weeder: "Learning to Filter Netnews". In Proc. of the 12th International Conference on Machine Learning ICML95, 1995
 
True

Rating

     Predicted Rating        
  1 2 3 4 5 skip Total
1 0 1 0 0 0 1 2
2 1 15 6 4 0 15 41
3 0 6 31 20 0 15 72
4 0 6 8 42 0 20 76
5 0 0 0 4 0 1 5
skip 0 8 4 5 1 141 159

 

Part 3