MULTI-LAYER ADAPTIVE NETWORKS.
PHILCO-FORD CORP NEWPORT BEACH CALIF AERONUTRONIC DIV
Pagination or Media Count:
The report is concerned with logic networks whose logical properties are adaptable in response to error signals relating to the networks performance. That is, a device is envisioned which receives inputs, possesses certain variable internal parameters, and produces outputs which depend on both the inputs and the internal parameters. If these outputs are not those desired, error indications to the device should cause such a change in the devicess internal parameters as will properly correct these outputs. If such a device is composed of a network of the variable-weight threshold elements described below, their weights being the required internal parameters, then the device will be called a learning network. Learning networks are of considerable importance because of the possibilities they offer for producing networks whose logical properties as given by their weight values are too complex to have been devised by hand. That is, it may transcend the capabilities of a human logical designer to arrive at the logic required to sort out, say, the various letters in hand-written cursive writing where this constitutes the input to a network. Learning networks could also have importance in real time situations where some form of rapid and complex adaptation is required which is beyond the capabilities of the more orthodox methods. Because of the great difficulties involved, no attempt is made in this report at an analysis of these networks and their properties. It is solely an empirical investigation of them based on computer-simulated experiments. Author
- Computer Hardware
- Computer Systems