System and method for viewing, modifying, storing, and running artificial neural network components

• US 8,903,746 B2
• Filed: 03/22/2012
• Issued: 12/02/2014
• Est. Priority Date: 03/22/2012
• Status: Expired due to Fees

• Current Assignee: Audrey Kudritskiy
• Original Assignee: Audrey Kudritskiy
• Inventors: Kudritskiy, Andrey
• Primary Examiner(s): CHEN, ALAN S
• Application Number:
  US13/427,375
• Publication Number:
  US 20130254138A1
• Time in Patent Office: 985 Days
• Field of Search: 706/15
• US Class Current: 706/11
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A system and method for artificial neural network processing includes, for example, modifying, by a computer processor, a value of a charge of a node of an artificial neural depending on a number of elapsed steps since a prior predefined significant event. A system and method includes, for example, providing by a processor a real-time representation of an artificial neural network and of graphical effects of a running of the neural network. A system and method includes, for example, automatically modifying the behavior of network nodes based on simultaneous occurrences of events.

Claims

1. A computer-implemented artificial neural network processing method, comprising:

modifying, by a computer processor, a value of a charge of a first node of an artificial neural network based on an input to the node, the node being configured to be activated to produce an output responsive to the value of the charge reaching a first threshold value;

determining, by the processor, that a number of steps, that have elapsed since one of (a) a last activation of the node and (b) a last modification of the value of the charge based on input to the node, is greater than a second threshold value; and

responsive to the determination, modifying, by the processor, the charge of the node.

2. The method of claim 1, wherein the charge of the node is reset to 0 in the modifying that is responsive to the determination.

3. The method of claim 1, wherein the charge of the node is reset to a predetermined value in the modifying that is responsive to the determination.

4. The method of claim 1, wherein the charge of the node is decreased by a predetermined amount in the modifying that is responsive to the determination.

5. The method of claim 1, further comprising:

for each occurrence of any of at least one qualifying event, incrementing a value of a first variable; and

one of (a) responsive to each modification of the value of the charge of node based on input to the node and (b) responsive to each activation of the node, updating a value of a second variable to be consistent with the value of the first variable;

wherein the determining includes comparing the values of the first and second variables.

6. The method of claim 1, further comprising:

storing, by the processor, a plurality of node records, each including a respective identifier of a respective one of the nodes, the respective charge value of the respective node, and a respective step identifier identifying a step at which the respective node one of had been activated and had its charge value modified in response to an input by an activated connection to which the respective node connects;

storing, by the processor, a plurality of connection records, each including a respective identifier of a respective connection between a respective pair of nodes corresponding to a pair of the stored plurality of node records, a respective identifier of a first one of the respective pair of nodes from which the respective connection originates, a respective identifier of a second one of the respective pair of nodes at which the respective connection terminates, and a respective weight; and

producing, by the processor, an output based on values identified in the node records and the connection records.

7. The method of claim 6, further comprising:

for each of the plurality of node records;

modifying its respective charge value according to the weight of one of the connection records that specifies the respective identifier of the respective node of the respective node record as the second of the respective pair of nodes of the respective connection record upon activation of the respective connection;

comparing the modified charge value to an activation threshold; and

responsive to determining in the comparing step that the modified charge value satisfied the activation threshold, updating a charge value of another node corresponding to another of the stored plurality of node records, the another node being identified, by another of the plurality of connection records that identifies the respective node as the node from which the respective connection of the another of the plurality of connection records originates, as the node at which the respective connection of the another of the plurality of connection records terminates.

8. The method of claim 7, wherein different activation thresholds are used in the comparing step for different ones of the nodes to which the plurality of node records correspond.

9. The method of claim 6, wherein the modification of the node charge responsive to the determination includes modifying the respective charge value of the node record of the respective node, and the determination of the number of elapsed steps is based on a number of elapsed steps since the step identified by its respective step identifier of the respective record.

10. The method of claim 6, wherein:

the plurality of node records are stored in a first table that includes a first column in which the node identifiers of the node records are stored, a second column in which the step identifiers of the node records are stored, and a third column in which the charge values of the node records are stored; and

the plurality of connection records are stored in a second table that includes a first column in which the connection identifiers of the connection records are stored, a second column in which the identifiers of the originating nodes of the connection records are stored, a third column in which the identifiers of the terminating nodes of the connection records are stored, and a fourth column in which the weights of the connection records are stored.

11. The method of claim 10, wherein each row of the first table corresponds to a respective one of the node records, and each row of the second table corresponds to a respective one of the connection records.

12. The method of claim 6, wherein for each of the connection records, the node identifiers of the respective connection record correspond to the node identifiers of the node records corresponding to the pair of nodes connected by the respective connection of the respective connection record.

13. The method of claim 6, wherein each of the node records further includes an identification of the respective activation threshold used for the node to which the respective node record corresponds.

14. The method of claim 1, further comprising:

displaying, by the processor and in a first portion of a graphical user interface, a real-time graphical representation of the neural network; and

displaying, by the processor and in a second portion of the graphical user interface, real-time graphical effects of a running of the neural network.

15. The method of claim 14, wherein the node is one of a plurality of nodes of the neural network, the neural network further includes connections between pairs of the plurality of nodes, and the method further comprises:

responsive to user-selection from the first portion of the graphical user interface of one of a graphical representation of one of the connections and a graphical representation of one of the nodes, displaying attributes of the one of the connection and the node of the selected graphical representation.

16. The method of claim 14, wherein the node is one of a plurality of nodes of the neural network, the neural network further includes connections between pairs of the plurality of nodes, and the method further comprises:

graphically indicating in the first portion a real-time activation of at least one of a node and a connection of the neural network that is graphically represented in the first portion.

17. The method of claim 14, further comprising:

responsive to user interaction with the graphical representation in the first portion, modifying attributes of one or more components of the neural network, wherein the modification affects in real-time the graphical effects displayed in the second portion of the graphical user interface.

18. The method of claim 17, wherein the user interaction includes editing attributes associated with one of a selected connection and a selected node.

19. The method of claim 17, wherein the user interaction includes manipulating the graphical representation of at least one connection in the first portion to change at least one node connected to the at least one connection.

20. The method of claim 17, further comprising:

storing in a database all modifications made to the neural network via the graphical user interface.

21. The method of claim 1, wherein the node is one of a plurality of nodes of the neural network, the method further comprising:

responsive to simultaneous activation of at least a first input node and a second input node connected to a third node of the neural network, modifying, by the processor, a weight of a connection between the first input node and the third node.

22. The method of claim 21, wherein modifying the weight includes increasing the weight if a connection weight between the second input node and the third node is positive.

23. The method of claim 21, wherein modifying the weight includes decreasing the weight if a connection weight between the second input node and the third node is negative.

24. The method of claim 21, wherein the artificial neural network is integrated into a robotics system, and the weight modification changes the behavior of the robotics system.

25. The method of claim 1, wherein the node is one of a plurality of nodes of the neural network, the neural network further includes connections between respective combinations of the nodes, and the method further comprises:

modifying, by a computer processor, at least one of a) respective weight(s) of at least a connection of the neural network, and b) respective threshold(s) of at least a node of the neural network;

responsive to a plurality of inputs to the neural network, determining whether at least one output of the neural network matches at least one desired output;

upon determining that at least one output of the neural network matches at least one desired output, saving a configuration of the neural network and outputs of the neural network; and

repeating the modifying, the determining, and the saving steps until the at least one output of the neural network matches all desired outputs.

26. The method of claim 1, wherein each modification of a charge value of a second node without any modification of the charge value of the first node is counted by the processor as a discrete one of the steps.

27. The method of claim 1, wherein the charge of the node is increased in response to the input to the node, and the charge of the node is decreased in the modifying that is responsive to the determination.

28. A computer system for artificial neural network processing, the system comprising:

a computer processor configured to;

modify a value of a charge of a node of an artificial neural network based on an input to the node, the node being configured to be activated to produce an output responsive to the value of the charge reaching a first threshold value;

determine that a number of steps, that have elapsed since one of (a) a last activation of the node and (b) a last modification of the value of the charge based on input to the node, is greater than a second threshold value; and

responsive to the determination, modify the charge of the node.

29. A non-transitory computer-readable medium on which are stored instructions that are executable by a processor, the instructions which, when executed by the processor, cause the processor to perform a method for artificial neural network processing, the method comprising:

modifying a value of a charge of a node of an artificial neural network based on an input to the node, the node being configured to be activated to produce an output responsive to the value of the charge reaching a first threshold value;

determining that a number of steps, that have elapsed since one of (a) a last activation of the node and (b) a last modification of the value of the charge based on input to the node, is greater than a second threshold value; and

responsive to the determination, modifying the charge of the node.