CSE255  THURSDAY JANUARY 22, 1998



GARY COTTRELL ON NEURAL NETWORKS (OUTLINE)

The motivation behind neural nets, some intuition about human
computation being a constraint satisfaction process, the Necker cube
example.

How to train a perceptron using an intuitive, non-mathematical rule.
Example of training the perceptron to learn OR.

What a linear discriminant is, how a perceptron is a linear
discriminant.  Motivation for the generalization to the logistic by
showing that the logistic can be interpreted as the posterior
probability of the class given the data.

Multi-layer nets, and explained back prop in two ways, first showing
descent along an error surface (the board is weight space, the error
surface floats in front of the board).

Alternative derivation of backpropagation via calculus from the chain
rule, showing that the resulting rule is nearly equivalent to the
non-mathematical perceptron rule given previously.

The NETTalk network, the NETTalk audio tape.