Neural network with modification of neuron weights and reaction coefficient

• US 5,222,194 A
• Filed: 02/27/1991
• Issued: 06/22/1993
• Est. Priority Date: 02/27/1990
• Status: Expired due to Term

• Current Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Original Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Inventors: Nishimura, Kazuo
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/661,194
• Time in Patent Office: 846 Days
• Field of Search: 395/23, 395/24, 395/27
• US Class Current: 706/25
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology

Abstract

A neural network computation apparatus having a plurality of layers, each of the plurality of layers has at least an input layer and an output layer, each layer having a plurality of units, a plurality of links, each of the plurality of links connecting units on the plurality of layers, and changing means for changing input and output characteristics of a particular unit of the plurality of units and/or the weight of a particular link of the plurality of links in accordance with an output of the output layer after learning an example and with a particular rule. After the neural network computation apparatus learns an example, the changing means changes input and output characteristics of units and weights of links in accordance with outputs of the output layer and a particular rule. Thus, a mutual operation between a logical knowledge and a pattern recognizing performance can be accomplished and thereby a determination close to that of a specialist can be accomplished. In other words, a proper determination in accordance with an experience can be made so as to deal with unknown patterns with high flexibility.

Claims

1. A trained neural network, comprising:

processing neuron layer means connected between an input neuron layer and an output neuron layer;

connection means between neurons of different layers having predetermined weights;

means for determining if an output from said output neuron layer resulting from an input to said input neuron layer and a processing by said processing neuron layer means matches any of a plurality of possible outputs, and if said output does not match any of a plurality of possible outputs, modifying at least one of said weights of connections between said neurons and a reaction coefficient of at least one neuron such that said neural network reprocesses said input and calculates another output which matches one of said plurality of possible outputs.

2. A train neural network according to claim 1, wherein said trained neural network further comprises

a characteristic layer connected to said input neuron layer;

an associative layer connected to said characteristic layer; and

a logic unit group including said output neuron layer, said logic unit group connected to said associative layer.

3. A method of operating a trained neural network having a plurality of neural layers including a processing layer connected between an input layer and an output layer, comprising the steps:

inputting data to be processed by the trained neural network by the input layer;

processing the inputted data by the processing layer of the trained neural network;

examining a result of said processing step at the output layer and if the result does not match any of a plurality of predetermined possible outputs, using previously determined expert knowledge to modify at least one of weights of connections between neurons and a reaction coefficient of at least one neuron;

reprocessing said input data by said modified neural network and obtaining one of said predetermined possible outputs.

4. A method of operating a trained neutral network having a plurality of neural layers including a processing layer connected between an input layer and an output layer, comprising the steps:

inputting data to be processed by the trained neural network by the input layer;

processing the inputted data by the processing layer of the trained neural network;

examining a result of said processing step at the output layer and if the result requires an interpolation, using previously determined expert knowledge to modify at least one of weights of connections between neurons and a reaction coefficient of at least one neuron;

reprocessing said input data by said modified neural network in obtaining one of said predetermined possible outputs.