Predictive self-organizing neural network

• US 5,214,715 A
• Filed: 01/31/1991
• Issued: 05/25/1993
• Est. Priority Date: 01/31/1991
• Status: Expired due to Term

• Current Assignee: Trustees of Boston University
• Original Assignee: Trustees of Boston University
• Inventors: Reynolds, John W., Grossberg, Stephen, Carpenter, Gail A.
• Primary Examiner(s): Boudreau, Leo H.
• Assistant Examiner(s): CAMMARATA, MICHAEL
• Application Number:
  US07/648,653
• Time in Patent Office: 845 Days
• Field of Search: 382/14, 382/15, 382/10, 382/21, 382/49, 395/21, 395/23, 395/24, 395/26, 307/201
• US Class Current: 382/157
• CPC Class Codes: G06F 18/23211 with adaptive number of clusters
• CPC Class Codes: G06N 3/0409 Adaptive resonance theory [ART] networks

Abstract

An A pattern recognition subsystem responds to an A feature representation input to select A-category-representation and predict a B-category-representation and its associated B feature representation input. During learning trials, a predicted B-category-representation is compared to that obtained through a B pattern recognition subsystem. With mismatch, a vigilance parameter of the A-pattern-recognition subsystem is increased to cause reset of the first-category-representation selection. Inputs to the pattern recognition subsystems may be preprocessed to complement code the inputs.

Claims

1. A pattern recognition system comprising:

an A pattern recognition subsystem for searching, selecting and learning an A-category-representation in response to an A input pattern;

means for predicting a B-category-representation from a selected A-category-representation;

means for providing a control B-category-representation;

means for detecting a mismatch between a predicted B-category-representation and a control B-category representation; and

means responsive to a mismatch between a predicted B-category-representation and a control B-category-representation to cause selection of a new A-category-representation.

2. A system as claimed in claim 1 wherein the A pattern recognition subsystem comprises:

a feature representation field of nodes for receiving input signals, defining an A input pattern, and template signals;

means for selecting an A-category-representation in an A-category-representation field of nodes based on a pattern from the feature representation field;

means for generating the template signals based on the selected A-category-representation;

means for adapting A-category-representation selection and the template signals to the input signals; and

first reset means for resetting A-category-representation selection with an insufficient match between the input pattern and the template signal.

3. A system as claimed in claim 2 further comprising a B pattern recognition subsystem for providing the control B-category-representation from a B input pattern.

4. A system as claimed in claim 1 wherein the A pattern recognition subsystem resets A-category-representation selection with an insufficient match within the A pattern recognition subsystem between the A-category-representation and the A-input pattern with a first level of matching vigilance, and the means responsive to a mismatch between a predicted B-category-representation and a control B-category representation increases the matching vigilance within the A pattern recognition subsystem to cause reset of the A-category-representation selection, increased vigilance being maintained with subsequent search for and selection of an A-category-representation and matching of a selected A-category-representation to said A input pattern.

5. A system as claimed in claim 4 wherein vigilance is increased to a minimum level required to reset the A-category-representation.

6. A system as claimed in claim 1 wherein the means for predicting comprises a mapping field of nodes having a one-to-one correspondence between nodes of a B-category-representation field and nodes of the mapping field and adaptive mapping from nodes of the A-category-representation field to nodes of the mapping field.

7. A system as claimed in claim 1 wherein the input pattern comprises a vector of feature representations and complements of the feature representations.

8. A system as claimed in claim 1 further comprising:

a B pattern recognition subsystem for selecting the control B-category-representation in response to a B input pattern and for learning the B-category-representation as a response to the B input pattern.

9. A system as claimed in claim 8 wherein the B-category-representation selected by the B pattern recognition subsystem is learned by the means for predicting as a predicted B-category-representation associated with a selected A-category-representation.

10. A system as claimed in claim 8 wherein the predicted B category representation associated with a selected A-category-representation primes the B pattern recognition subsystem to initially select the predicted B-category-representation as the B-category-representation.

11. A system as claimed in claim 8 wherein the a pattern recognition subsystem resets A-category-representation selection with an insufficient match within the A pattern recognition subsystem between the A-category-representation and the A-input pattern with a first level of matching vigilance, and the means responsive to a mismatch between a predicted B-category-representation and a control B-category representation increases the matching vigilance within the A pattern recognition subsystem to cause reset of the A-category-representation selection, increased vigilance being maintained with subsequent search for and selection of an A-category-representation and matching of a selected A-category-representation to said A input pattern.

12. A system as claimed in claim 11 wherein the A pattern recognition subsystem comprises:

a feature representation field of nodes for receiving input signals, defining an A input pattern, and template signals;

means for selecting an A-category-representation in an A-category-representation field of nodes based on a pattern from the feature representation field;

means for generating the template signals based on the selected A-category-representation;

means for adapting A-category-representation selection and the template signals to the input signals; and

first reset means for resetting A-category-representation selection with an insufficient match between the input pattern and the template signals.

13. A pattern recognition system comprising:

A. an A pattern recognition subsystem comprising;

an A feature representation field of nodes for receiving A input signals, defining an A input pattern, and A template signals;

means for selecting an A-category-representation in an A-category-representation field of nodes based on a pattern from the feature representation field;

means for generating the A template signals based on the selected A-category-representation;

means for adapting A-category-representation selection and the A template signals to the A input signals; and

A reset means for resetting A-category-representation selection with an insufficient match between the A input pattern and the A template signals at a first level of matching vigilance;

B. a B pattern recognition subsystem comprising;

a B feature representation field of nodes for receiving B input signals, defining a B input pattern, and B template signals;

means for selecting a B-category-representation in a B-category-representation field of nodes based on a pattern from the B feature representation field;

means for generating the B template signals based on the selected B-category-representation;

means for adapting B-category-representation selection with an insufficient match between the B input pattern and the B template signals;

C. means for associating the selected A-category-representation with a predicted B-category-representation comprising a mapping field of nodes having a one-to-one correspondence with nodes of the B-category-representation field and adaptive mapping from nodes in the A-category-representation field, the predicated B-category-representation associated with a selected first-category-representation being learned as the B-category-representation selected by the B pattern recognition subsystem; and

D. system reset means for resetting an A-category-representation selection in the A pattern recognition subsystem where a predicted B-category-representation associated with the selected A-category-representation fails to match a B-category-representation selected by the B pattern recognition subsystem, the system reset means increasing vigilance of the A reset means to cause reset of the A-category-representation selection, increased vigilance being maintained with subsequent selection of an A-category-representation with said A input signals applied to the A feature representation field.

14. A system as claimed in claim 13 wherein the predicted B-category-representation associated with a selected A-category-representation primes the B pattern recognition subsystem to initially select the predicted B-category-representation.

15. A system as claimed in claim 14 wherein the input signals to each pattern recognition subsystem comprise a vector of feature representations and complements of the feature representations.

16. A pattern recognition method comprising:

searching, selecting and learning an A-category-representation in response to an A input pattern;

predicting a B-category-representation from a selected A-category-representation;

providing a control B-category-representation;

detecting a mismatch between the predicted B-category-representation and the control B-category-representation; and

selecting a new A-category-representation when a mismatch between the predicted B-category-representation and the control B-category-representation is detected.

17. A method as claimed in claim 16 further comprising:

based on an A input pattern, selecting an A-category-representation;

generating A template signals based on the selected A-category-representation;

adapting A-category-representation selection and the A template signals to the A input signals with a sufficient match between the A input pattern and the A template signal; and

resetting A-category-representation selection with an insufficient match between the A input pattern and the A template signal.

18. A method as claimed in claim 17 further comprising providing the control B-category-representation from a B input pattern.

19. A method as claimed in claim 16 wherein the A-category-representation selection is reset with a first level of matching vigilance with an insufficient match between the A input pattern and the A-category-representation and, where a predicted B-category-representation associated with a selected A-category-representation fails to match a control B-category-representation, the matching vigilance is increased to cause reset of the A-category-representation selection, increased vigilance being maintained with subsequent selection of an A-category-representation and matching of a selected A-category-representation to said A input pattern.

20. A method as claimed in claim 19 wherein vigilance is increased to a minimum level required to reset the A-category-representation.

21. A method as claimed in claim 16 further comprising mapping between control B-category-representation to A-category-representations through a one-to-one mapping of control B-category-representations to predicted B-category-representations and adaptive mapping of A-category-representations to predicted B-category-representations.

22. A method as claimed in claim 16 further comprising providing as the input pattern a vector of feature representations and complements of the feature representations.

23. A method as claimed in claim 16 further comprising:

selecting the control B-category-representation in response to a B input pattern and learning the B-category-representation as a response to the second input pattern.

24. A method is claimed in claim 23 further comprising learning the selected B-category-representation as a predicted B-category-representation associated with a selected A-category-representation.

25. A method is claimed in claim 23 further comprising initially selecting the predicted B-category-representation as the B-category-representation.

26. The method as claimed in claim 23 wherein the A-category-representation selection is reset with a first level of matching vigilance with an insufficient match between the A input pattern and the A-category-representation and, where a predicted B-category-representation associated with a selected A-category-representation fails to match a control B-category-representation, the matching vigilance is increased to cause reset of the A-category-representation selection, increased vigilance being maintained with subsequent selection of an A-category-representation and matching of a selected A-category-representation to said A input patterns.