Accession Number : ADA289885

Title :   Multilayer Piecewise Linear Networks. Revision 1.

Descriptive Note : Final rept. 1993-1994,

Corporate Author : NAVAL AIR WARFARE CENTER WEAPONS DIV CHINA LAKE CA

Personal Author(s) : Alltop, W. O.

PDF Url : ADA289885

Report Date : NOV 1994

Pagination or Media Count : 57

Abstract : Feed-forward layered networks are the simplest neural computers. These networks provide fast compact devices for computing discriminant functions and other mappings applicable to pattern recognition problems. This report treats the problem of weight initialization/assignment for feed-forward networks with piecewise linear neuron transfer function. Specific transformations are shown to be impossible with a single hidden layer. Techniques are presented for constructing such transformations with two hidden layers.

Descriptors :   *NEURAL NETS, *LINEARITY, *GEOMETRY, *COMPUTER NETWORKS, IMAGE PROCESSING, TRANSFER FUNCTIONS, COMPUTERS, PROTOTYPES, WEIGHT, COMPUTER APPLICATIONS, PATTERN RECOGNITION, DISCRIMINATE ANALYSIS, TRANSFORMATIONS, CONVEX SETS, LINEAR ALGEBRA, ALGORITMS.

Subject Categories : Computer Systems
      Numerical Mathematics

Distribution Statement : APPROVED FOR PUBLIC RELEASE