Nov 19 - Lesson Plan

1. Introduction to Learning

Exercise: Write down a learning problem.

Whiteboard: Kinds of learning:

What is the task?
- Learning to classify data ==> classification learning
- Learning a probabilistic model of the world ==> Bayesian learning
- Discovering "interesting" properties of data ==> data-mining
- Learning to parse a language ==> grammar learning
- Learning to act in the world ==> reinforcement learning
- Learning to "operationalize" knowledge ==> speed-up learning

Not distinct!  Much overlap

How is the thing being learned represented?

- A set of formulas in first-order logic
- A set of formulas in propositional logic
- A decision tree **
- A policy or universal plan
- A Bayesian network **
- A neural network **

Basic concepts:
- Training data
	- Attributes or Features -- how is data described?
	- Complete or Incomplete -- are all attributes of all examples known?
	- Labeled vs unlabeled   -- are categories given, or discovered?
- Test data
- Generalization
- Bias I: Prior knowledge 
- Bias II: Occum's razor, minimum description length
- Error

Students switch papers.  Discuss some of the learning problems in
terms of these concepts.


2. Learning Bayesian Networks

Simpliest (but important!) case: Parameter Learning with Complete Data

Given:
Bayes net structure
Training set - each item has value for EVERY variable in the network

Maximum likelihood estimate of parameters (values in the CPT's)
== values of CPT's that maximizes likelihood of the observations
== actual statistics of the training data!

3. Assignment: Spam Filtering by Machine Learning

4. Beyond MLE:
(a) Combining data with prior beliefs about the parameters 
(b) Parameter learning where the data is incomplete (does not include
	all variables in every example)