Intelligent control with hierarchical stacked neural networks

• US 7,152,051 B1
• Filed: 09/30/2002
• Issued: 12/19/2006
• Est. Priority Date: 09/30/2002
• Status: Active Grant

• Current Assignee: Michael Lamport Commons, Mitzi Sturgeon White
• Original Assignee: Michael Lamport Commons, Mitzi Sturgeon White
• Inventors: Commons, Michael Lamport, White, Mitzi Sturgeon
• Primary Examiner(s): Patel, Ramesh
• Assistant Examiner(s): Datskovskiy, Sergey
• Application Number:
  US10/261,560
• Time in Patent Office: 1,541 Days
• Field of Search: 706/16, 706/20
• US Class Current: 706/16
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

An intelligent control system based on an explicit model of cognitive development (Table 1) performs high-level functions. It comprises up to O hierarchically stacked neural networks, N_m, . . . , N_m+(O−1), where m denotes the stage/order tasks performed in the first neural network, N_m, and O denotes the highest stage/order tasks performed in the highest-level neural network. The type of processing actions performed in a network, N_m, corresponds to the complexity for stage/order m. Thus N₁ performs tasks at the level corresponding to stage/order 1. N₅ processes information at the level corresponding to stage/order 5. Stacked neural networks begin and end at any stage/order, but information must be processed by each stage in ascending order sequence. Stages/orders cannot be skipped. Each neural network in a stack may use different architectures, interconnections, algorithms, and training methods, depending on the stage/order of the neural network and the type of intelligent control system implemented.

Claims

1. A stacked neural network, consisting essentially of:

a plurality of architecturally distinct, separately trainable, ordered neural networks of varying complexity;

said plurality being organized in a linear hierarchy of complexity from lower to higher order/stages in a model of cognitive development, wherein output from each member at a respectively lower order/stage provides the input for the next respectively higher order/stage member, substantially without skipping order/stages;

each member of said plurality feeding signals forward and back to other members of said plurality;

said signals being defined in terms of actions available to said each member, said each member transforming actions from at least one member at a lower order/stage, to produce nonarbitrary organizations of said actions from said at least one member at a lower order/stage effective for completing new tasks of increased complexity;

said nonarbitrary organizations being fed to at least one member at a higher order/stage; and

said nonarbitrary organizations being modifiable by feedback signals from members at said higher order/stages.

2. The stacked neural network of claim 1 adapted to form an intelligent control system to detect shoplifting.

3. The stacked neural network of claim 2, further comprising five architecturally distinct neural networks.

4. The stacked neural network of claim 3, wherein each of said five architecturally distinct neural networks corresponds respectively to circular sensory-motor, sensory-motor, nominal, sentential, and pre-operational stages of said model of cognitive development.

5. The stacked neural network of claim 4, wherein said model of cognitive development has 14 stages.

6. The stacked neural network of claim 1 adapted to form an intelligent control system to direct customer calls to a correct department in a large organization.

7. The stacked neural network of claim 6, further comprising four architecturally distinct neural networks.

8. The stacked neural network of claim 7, wherein each of said four architecturally distinct neural networks corresponds respectively to circular sensory-motor, sensory-motor, nominal, and sentential stages of said model of cognitive development.

9. The stacked neural network of claim 8 wherein said model of cognitive development has 14 stages.

10. A method for intelligent control of a system, consisting essentially of the steps of:

forming a plurality of architecturally distinct, separately trainable, ordered neural networks of varying complexity;

organizing said plurality in a linear hierarchy of complexity from lower to higher order/stages in a model of cognitive development, wherein output from each member at a respectively lower order/stage provides the input for the next respectively higher order/stage member, substantially without skipping order/stages;

defining signals from each member of said plurality in terms of actions available to said each member;

outputting from at least one of said each member at least one of said actions, said actions being effective to automatically exert a control influence on said system;

feeding said signals forward and backward to other members of said plurality, whereby nonarbitrary organizations of said actions from at least one member at a lower order/stage are formed that carry out new tasks of increased complexity; and

having the capability of modifying said nonarbitrary organizations by feedback signals from at least one member at a higher order/stage.

11. The method of claim 10 adapted to control a system that detects shoplifting.

12. The method of claim 11, comprising the further steps of identifying patterns of change in light intensity that indicate movement of a new physical object into an area, tracking the movement of said object through said area, labeling said object as “

person”

or “

object of a type for sale”

, associating movement of “

object of a type for sale”

with a person, determining whether said person with an “

object of a type for sale”

passes through a checkout counter, and, if said person does not pass through a checkout counter, issuing a security alert.

13. The method of claim 12, wherein said further steps of identifying, tracking, labeling, associating, determining, and issuing are performed with the help of five architecturally distinct neural networks, each corresponding respectively to circular sensory-motor, sensory-motor, nominal, sentential, and pre-operational stages of said model of cognitive development.

14. The method of claim 10 adapted to control a system that directs customer calls to a correct department in a large organization.

15. The method of claim 14, comprising the further steps of separating components of a customer call into words or intervals between words, grouping contiguous words into speech segments, mapping said contiguous words to concept domains, identifying relationships between pairs of concept domains, thereby creating joint concept domains, correlating each of said joint concept domains with a single department in said organization, querying a customer to determine if said single department is correct, and routing said customer call to said single department if correct or determining a new concept domain if incorrect.

16. The method of claim 15, wherein said further steps of separating, grouping, mapping, identifying, correlating, querying, and routing are performed with the help of four architecturally distinct neural networks, each corresponding respectively to circular sensory-motor, sensory-motor, nominal, and sentential stages of said model of cognitive development.

20. The method according to claim 10, wherein successively higher members within the linear hierarchy process information at increasing levels of abstraction.

17. A computer-readable storage medium embodying program instructions for a method for intelligent control, said method consisting essentially of the steps of:

forming a plurality of architecturally distinct, separately trainable, ordered neural networks of varying complexity;

organizing said plurality in a linear hierarchy of complexity from lower to higher order/stages in a model of cognitive development, wherein output from each member at a respectively lower order/stage provides the input for the next respectively higher order/stage member, substantially without skipping order/stages;

defining signals from each member of said plurality in terms of actions available to said each member;

outputting from at least one of said each member at least one of said actions, said actions being effective to automatically exert a control influence on said system;

feeding said signals forward and backward to other members of said plurality, whereby nonarbitrary organizations of said actions from at least one member at a lower order/stage are formed that carry out new tasks of increased complexity; and

having the capability of modifying said nonarbitrary organizations by feedback signals from at least one member at a higher order/stage.

18. The computer-readable storage medium of claim 17, wherein said method for intelligent control directs a system that detects shoplifting.

19. The computer-readable storage medium of claim 17, wherein said method for intelligent control directs a system that directs customer calls to the correct department in a large organization.