Self-organizing pattern classification neural network system

• US 5,239,594 A
• Filed: 02/06/1992
• Issued: 08/24/1993
• Est. Priority Date: 02/12/1991
• Status: Expired due to Term

• Current Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Inventors: Yoda, Fumio
• Primary Examiner(s): Mancuso, Joseph
• Assistant Examiner(s): Fox, David
• Application Number:
  US07/832,678
• Time in Patent Office: 565 Days
• Field of Search: 382/14, 382/15, 382/36, 382/37, 382/38, 382/39, 364/274.9, 364/916.2, 395/21, 395/24, 395/27
• US Class Current: 382/158
• CPC Class Codes: G06F 18/24137 Distances to cluster centroïds
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning

Abstract

A self-organizing pattern classification neural network system includes means for receiving incoming pattern of signals that were processed by feature extractors that extract feature vectors from the incoming signal. These feature vectors correspond to information regarding certain features of the incoming signal. The extracted feature vectors then each pass to separate self-organizing neural network classifiers. The classifiers compare the feature vectors to templates corresponding to respective classes and output the results of their comparisons. The output from the classifier for each class enter a discriminator. The discriminator generates a classification response indicating the best class for the input signal. The classification response includes information indicative of whether the classification is possible and also includes the identified best class. Lastly, the system includes a learning trigger for transferring a correct glass signal to the self-organizing classifiers so that they can determine the validity of their classification results.

Claims

1. A self-organizing pattern classification neural network system for classifying incoming pattern signals into their classes, comprising:

(a) feature extractors for extracting different feature vectors from an incoming pattern signal;

(b) self-organizing neural network classifiers, one for each of said features extractors, for receiving the feature vectors and generating response vectors, each response vector comprising a plurality of response scalars corresponding to the respective classes wherein each said self-organizing classifier comprises;

i) input nodes for receiving feature scalars of each of said feature vectors;

ii) a plurality of intermediate nodes for receiving said feature scalars for said input nodes and for generating a plurality of intermediate outputs;

iii) a plurality of output nodes for receiving intermediate outputs of intermediate nodes of a class, for determining a smallest intermediate output, and for transferring, to said discriminator, said smallest intermediate output as a response scalar;

iv) a self-organizing selector for receiving said smallest intermediate output and a node number of said intermediate node which gives said smallest intermediate output and for determining a weights update signal based on node number and said intermediate output from said intermediate node, and a correct class signal from said learning trigger; and

c) a discriminator for receiving said response vectors and generating a classification response, which includes information indicative of whether classification is possible and also includes an identified class; and

d) a learning trigger for transferring a correct class signal to said self-organizing classifiers based on a class of said incoming pattern signal and based on said classification response.

2. The self-organizing pattern classification neural network system of claim 1, wherein each said intermediate node comprises:

element comparators for comparing elements of a feature vector with a template to produce a squared output indicative of a square of a difference between said feature vector and said template;

an adder for adding outputs of said element comparators to produce an adder output; and

a square root computer for calculating the square root of said adder output.

3. The self-organizing pattern classification neural network system of claim 2, wherein each said element comparator squares a difference between an element of the feature vector and a scalar of said template and transfers the results to said adder.

4. The self-organizing pattern classification neural network system of claim 2, wherein said templates of said element signals are updated by an update signal from said self-organizing selector.

5. The self-organizing pattern classification neural network system of claim 1, wherein said output node detects a smallest one among said intermediate outputs from intermediate nodes of a class and transfers a node number of said smallest output to said self-organizing selector and also transfers, said smallest output to said discriminator as a response scalar.

6. The self-organizing pattern classification neural network system of claim 1, wherein said self-organizing selector modifies weighting vectors of intermediate node based on node numbers and outputs of intermediate nodes of respective classes which are transferred from said output nodes and a correct class signal from said learning trigger.