GRAPHICAL DATA PROCESSING RESEARCH STUDY AND EXPERIMENTAL INVESTIGATION
Abstract
A pattern may be presented for classification either as a code word of ones and zeros or of plus ones and minus ones; weight changing in the learning machine is effected by adding the pattern vector to the weight vector. It has been found, by digital computer simulation, that convergence is secured much more rapidly when the (1, - 1) representation of the input pat terns is used. An intensive study has been made of the variability of the second-harmonic weights using a test procedure closely approximating the manner in which the cores will operate in the learning machine. Satisfactory performance was obtained over a useful range of drive levels. A discussion is presented of the machine logic and timing, which have now been worked out in detail.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 1963
- Accession Number
- AD0404623
Entities
People
- A. E. Brain
- G. E. Forsen
Organizations
- SRI International