Oct 31, 2003
-------------

- Recap neural networks and decision trees
        - DTs are greedy, may not lead to
          "optimal tree"
        - NNs are iterative improvement, also
          can get stuck in local minima

- A more formal way to think of learning
        - version spaces
        - piston analogy

- Version spaces
        - maintain most specific boundary and most
          general boundary, and move them toward
          each other

- If data is consistent (given the bias)
        - pistons will move toward each other to
          enclose a set of consistent hypotheses

- If data is inconsistent (for the particular bias)
        - pistons will cross each other

- Can arrange hypotheses as a lattice
        - example bias: boolean conjunctions of at most
          three literals (positive or negative)

- Worked out examples with version spaces

- Next topic: Knowledge in Learning
        - learning when you already know something
	- combine deduction and induction

- Induction
	- proceeding from specific to general
	- e.g., decision trees, neural nets

- Deduction
	- proceeding from general to specific
	- e.g., proving theorems, resolution

- How to combine these two?
        - many flavors here!

- Example 1:
	- Type A
	- Explanation Based Learning
	- "I was wondering if %s'

- Example 2:
	- Type B
	- Relevance Based Learning
	- "In India, elevators are called lifts."

- Example 3:
	- Type C
	- Knowledge-Based Inductive Learning
	- "Naren is a good teacher."