Self-organizing neural network for classifying pattern signatures with `a posteriori` conditional class probability

• US 5,384,895 A
• Filed: 08/28/1992
• Issued: 01/24/1995
• Est. Priority Date: 08/28/1992
• Status: Expired due to Fees

• Current Assignee: The United States of America As Represented By The Secretary of Agriculture
• Original Assignee: the united states of america as represented by the secretary of the navy
• Inventors: Poston, Wendy L., Rogers, George W., Priebe, Carey E., Solka, Jeffrey L.
• Primary Examiner(s): Downs, Robert W.
• Assistant Examiner(s): DAVIS, GEORGE B
• Application Number:
  US07/936,369
• Time in Patent Office: 879 Days
• Field of Search: 395/22, 395/23, 395/24
• US Class Current: 706/20
• CPC Class Codes: G06F 18/24155 Bayesian classification
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks

Abstract

A self-organizing neural network and method for classifying a pattern signature having N-features is provided. The network provides a posteriori conditional class probability that the pattern signature belongs to a selected class from a plurality of classes with which the neural network was trained. In its training mode, a plurality of training vectors is processed to generate an N-feature, N-dimensional space defined by a set of non-overlapping trained clusters. Each training vector has N-feature coordinates and a class coordinate. Each trained cluster has a center and a radius defined by a vigilance parameter. The center of each trained cluster is a reference vector that represents a recursive mean of the N-feature coordinates from training vectors bounded by a corresponding trained cluster. Each reference vector defines a fractional probability associated with the selected class based upon a ratio of i) a count of training vectors from the selected class that are bounded by the corresponding trained cluster to ii) a total count of training vectors bounded by the corresponding trained cluster. In the exercise mode, an input vector defines the pattern signature to be classified. The input vector has N-feature coordinates associated with an unknown class. One of the reference vectors is selected so as to minimize differences with the N-feature coordinates of the input vector. The fractional probability of the selected one of the reference vectors is the a posteriori conditional class probability that the input vector belongs to the selected class.

Claims

1. A self-organizing neural network for classifying an N-feature pattern signature with a posteriori conditional class probability that the pattern signature belongs to a selecting class from a plurality of classes, comprising:

a first plurality of processing layers for storing each of a first plurality of training vectors as non-overlapping clusters, each of said first plurality of training vectors having N-feature coordinates and a class coordinate, each of said non-overlapping clusters having a center and a radius defined by a vigilance parameter, wherein the center of each of said non-overlapping clusters is a reference vector, wherein the reference vector associated with each of said non-overlapping clusters is successively indexed and is equal to a corresponding one of said first plurality of training vectors,

said first plurality of processing layers updating one cluster from said non-overlapping clusters and storing said first plurality of training vectors when one of a second plurality of training vectors has corresponding N-feature coordinates bounded by the radius of said one cluster, wherein N-feature coordinates of a cluster'"'"'s reference vector are recursively combined with N-feature coordinates of said one of said second plurality of training vectors having corresponding N-feature coordinates bounded by the radius of said one cluster, wherein a plurality of trained clusters are generated using said first and second plurality of training vectors,

said first plurality of processing layers determining i) a count of training vectors associated with the selected class that are bounded within each of said plurality of trained clusters, and ii) a total count of training vectors bounded within each of said plurality of trained clusters, wherein each of said plurality of trained clusters has a reference vector that defines a fractional probability associated with the selected class based upon a ratio of i) said count of training vectors associated with the selected class to ii) said total count of training vectors bounded by a corresponding one of said plurality of trained clusters; and

means for processing an input vector defining the pattern signature to be classified, the input vector having N-feature coordinates associated with an unknown class, said processing means for selecting one of reference vectors associated with said plurality of training vectors that minimizes differences with the N-feature coordinates of the input vector, wherein a fractional probability of the selected one of the reference vectors is the a posteriori conditional class probability that the input vector belongs to the selected class.

2. A neural network as in claim 1 wherein said processing means minimizes differences with the N-feature coordinates of the input vector using a distance metric.

3. A neural network as in claim 2 wherein the distance metric is a Euclidean distance metric.

4. A neural network as in claim 2 wherein the distance metric is a Manhattan distance metric.

5. A neural network as in claim 2 wherein the distance metric is a Mahalanobis distance metric.

6. A neural network as in claim 1 wherein said processing means comprises:

means for generating feature difference measurements for each of the reference vectors associated with said plurality of trained clusters, wherein a feature difference measurement is a distance between the N-feature coordinates of each of the reference vectors associated with said plurality of trained clusters and the NS feature coordinates of the input vector; and

means for selecting a minimum feature distance measurement for the input vector.

7. A neural network as in claim 6, wherein said means for selecting the minimum feature difference measurement is a Maxnet processor.

8. A neural network as in claim 1 wherein said first plurality of processing layers includes a composite processor for determining a Euclidean distance between each of said second plurality of training vectors and each of said non-overlapping clusters.

9. A neural network as in claim 1 further comprising:

an input layer of staging nodes for storing the N-feature coordinates and the class coordinate for each of said first and second plurality of training vectors; and

an output layer of nodes, each output layer node for identifying the selected class and the a posteriori conditional class probability.

10. A neural network as in claim 1 wherein said processing means comprises:

said first plurality of processing layers for determining differences between the N-feature coordinates of the input vector and each of the reference vectors associated with said plurality of trained clusters; and

a second plurality of processing layers for selecting the one reference vector that minimizes the differences.

11. A method, utilizing a self-organizing neural network, for classifying an N-feature pattern signature with a posteriori conditional class probability that the pattern signature belongs to a selecting class from a plurality of classes, comprising the steps of:

storing each of a first plurality of training vectors as non-overlapping clusters in a first plurality of processing layers, each of said first plurality of training vectors having N-feature coordinates and a class coordinate, each of said non-overlapping clusters having a center and a radius defined by a vigilance parameter, wherein the center of each of said non-overlapping clusters is a reference vector, wherein the reference vector associated with each of said non-overlapping clusters is successively indexed and is equal to a corresponding one of said first plurality of training vectors;

updating one cluster from said non-overlapping clusters and storing said first plurality of training vectors when one of a second plurality of training vectors has corresponding N-feature coordinates bounded by the radius of said one cluster, wherein said step of updating is defined as N-feature coordinates of a cluster'"'"'s reference vector being recursively combined with N-feature coordinates of said one of second plurality of training vectors having corresponding N-feature coordinates bounded by the radius of said one clusters, wherein a plurality of trained clusters are generated using said first and second plurality of training vectors;

determining i) a count of training vectors associated with the selected class that are bounded within each of said plurality of trained clusters, and ii) a total count of training vectors bounded within each of said plurality of trained clusters, wherein each of said plurality of trained clusters has a reference vector that defines a fractional probability associated with the selected class based upon a ratio of i) said count of training vectors associated with the selected class to ii) said total count of training vectors bounded by a corresponding one of said plurality of trained clusters;

determining, using the first plurality of processing layers, a difference between N-feature coordinates of an input vector and each of the reference vectors associated with said plurality of trained clusters; and

selecting, using a second plurality of processing layers, one of reference vectors associated with said plurality of training vectors that minimizes said difference, wherein a fractional probability of the selected one of the reference vectors is the a posteriori conditional class probability that the input vector belongs to the selected class.

12. A method according to claim 11 wherein said step of selecting minimizes differences with the N-feature coordinates of the input vector using a distance metric.

13. A method according to claim 12 wherein the distance metric is a Euclidean distance metric.

14. A method according to claim 12 wherein the distance metric is a Manhattan distance metric.

15. A method according to claim 12 wherein the distance metric is a Mahalanobis distance metric.