NEURAL COMPUTATION

CMPS 290D, Winter 2001

Lecture times:
TTh 4:00-5:45am, Social Sciences 2, 363

Instructor:
Prof. Manfred Warmuth (manfred@cse.ucsc.edu)
Phone: (831) 459-4950
Office: 331 Applied Sciences
Office Hours: Mo 10-11, Th 12-1

Lecture 1:
General discussion of brains versus artificial neural nets
Examples of how to communicate with brain incuding Armed Rats
Linearly least squares, Widrow Hoff algorithm, gradient descent, perceptron algorithm, feed forward neural nets

Lecture 2:
On-line versus batch, single neuron with arbitrary increasing transfer function,
dot products, margins, Perceptron Convergence Theorem, rotation invariance

Lecture 3:
Logistic regression, convexity, motivation of relative entropy from coding theory,
logistic regression with square loss leads to local minima, entropic loss

Lecture 4:
Matching Loss functions for single neurons, exponenentially many minima if wrong loss is used, motivation of the hinge loss for learning with threshold functions, loss functions for the multiclass case

Homework 1
Data is in /projects/learning/data/290D01

Homework 2
Voted Perceptron Paper

Homework 3
Paper in which updates are derived

Homework 4

Back to the CE / CS Class Home Pages.
Back to the CE / CS Home Page.