Neural network and system

• US 4,954,963 A
• Filed: 03/02/1989
• Issued: 09/04/1990
• Est. Priority Date: 03/02/1989
• Status: Expired due to Term

• Current Assignee: Texas Instruments, Inc.
• Original Assignee: Texas Instruments, Inc.
• Inventors: Gately, Michael T., Penz, Perry A.
• Primary Examiner(s): MACDONALD, ALLEN R
• Application Number:
  US07/318,038
• Time in Patent Office: 551 Days
• Field of Search: 364/513, 364/900, 364/200, 364/274.9, 364/276.6, 364/917.95, 382/14, 382/15
• US Class Current: 706/25
• CPC Class Codes: G01S 7/292 Extracting wanted echo-signals
• CPC Class Codes: G01S 7/417 involving the use of neural networks
• CPC Class Codes: G06F 18/2135 based on approximation criteria, e.g. principal component analysis
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks

Abstract

Neural network systems (100) with learning and recall are applied to clustered multiple-featured data (122,124,126) and analog data.

Claims

1. An information processor, comprising:

(a) an encoder, said encoder encoding input information into a closeness code format and providing an output in response thereto; and

(b) a clustering neural network providing an output therefrom, said clustering network having an input receiving said output of said encoder and said output of said clustering network, said clustering neural network clustering said input thereto with thresholded analog neurons in conjunction with a learning rule and a recall rule to preserve analog aspects.

2. The processor of claim 1, further comprising:

(a) a classifying neural network having stored information in a format of vectors recalled from said clustering network using said recall rule, said classifying network comparing said stored information recalled from said clustering network with other information stored therein with a second recall rule.

3. The processor of claim 1, wherein:

(a) said learning rule is a Widrow-Hoff learning; and

(b) said recall rule is a BSB-type recall rule.

4. A method of information processing, comprising the steps of:

(a) encoding said information into closeness code format;

(b) clustering said encoded information with a thresholded analog neural network;

(c) comparing recalled versions of said encoded information with stored information.

5. The method of claim 4, wherein;

(a) said stored information is stored in a second thresholded analog neural network; and

(b) said comparing is by inputting said recalled versions of said encoded information into said second thresholded network and recalling said stored information.

6. The method of claim 4, comprising the further steps of:

(a) deleting a portion of each of said recalled versions; and

(b) adding further information to each of said recalled versions

7. A method of processing multiple-feature information, comprising the steps of:

(a) encoding information by concatenated closeness codes of features of said information;

(b) learning without saturation said encoded information in a thresholded analog neural network; and

(c) inputting said encoded information into said network and recalling processed versions of said encoded information.