Machine Learning Presentation Outline

A. A subfield of AI concerned with how programs can learn from experience

B. Deals with using percepts to improve an agent’s ability to act in the future

C. Involves additions to an agent’s knowledge base

II. Types of Learning

A. From Examples

1. Good for concept-learning

2. Uses induction

3. Examples and near-misses are both important

B. By Being Told

1. Rote learning

2. Inference learning

C. From Experimentation

1. Improves with practice

2. Learns by discovery

III. The Importance of Learning

A. Helps us to understand and improve how humans learn

B. Can discover new information or make predictions unable to be obtained by humans

C. Can fill in incomplete (“brittle”) domains

IV. Past Successes

A. Oil Industry

Who: British Petroleum

Problem: Separating crude oil and natural gas

ML used to: Determine control parameters for separation process

Result: Time savings - more than a day  ten minutes

B. Chemical Process Control

Who: Westinghouse

Problem: Manufacturing nuclear fuel pellets

ML used to: Determine control parameters

Result: Saved $10,000,000 per year

C. Loan Application Screening

Who: American Express (UK)

Problem: Humans had only a 50% success rate in predicting whether borderline loan applicants would default

ML used to: Predict whether borderline loan applicants would default

Result: Success rate went up to 70%

D. Cataloging Faint Objects

Who: second Palomar Observatory Sky Survey

Problem: Objects in 3000 images needed to be identified and classified. Previously, this had been done by hand, but that was impossible here because there was too much data and some objects were too faint.

ML used to: Develop automatic cataloging system

Result: Automatic system was more than 92% accurate, which was good enough

V. Current Developments

A. Data mining - systematically gathering and analyzing very large amounts of information

B. Medical - diagnostic systems

C. Financial - targeted marketing

D. Government - illegal activities, security threats

VI. Current Problems

A. Finding good methods of representing generalizations

B. Making better use of expectations based on prior knowledge

C. Dealing with inconsistencies

D. Dealing with partially-learned generalizations

E. Integrating logic with probability

VII. The Future of Learning

A. Increasing generalization of learning algorithms

B. More accurate predictive abilities