Metrics for specifying and/or testing neural networks

• US 5,550,951 A
• Filed: 03/18/1993
• Issued: 08/27/1996
• Est. Priority Date: 03/18/1993
• Status: Expired due to Fees

• Current Assignee: the united states of america as represented by the secretary of the navy
• Original Assignee: the united states of america as represented by the secretary of the navy
• Inventors: Woodall, Roger L.
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Dorvil, Richemond
• Application Number:
  US08/033,225
• Time in Patent Office: 1,258 Days
• Field of Search: 395/21, 395/22, 395/23, 395/24, 364/551.01, 364/578, 364/579
• US Class Current: 706/15
• CPC Class Codes: G01R 31/318505 Test of Modular systems, e.g. Wafers, MCM's
• CPC Class Codes: G06F 11/2273 Test methods
• CPC Class Codes: G06N 3/0675 using electro-optical, acousto-optical or opto-electronic means

Abstract

A method for testing a neural network, and apparatus for carrying out the method. A first embodiment of the method includes the steps of (a) providing a neural network having a set of connection values, (b) stimulating at least one input of the neural network with an input vector to obtain output signals at an output of the neural network, (c) obtaining a plurality of samples of the output signals, wherein at least one of the plurality of samples is delayed in time from another one of the samples, and wherein at least one of the plurality of samples may represent a difference between two samples, (d) generating an image from the plurality of samples, and (e) comparing the image to a reference image to determine an operational characteristic of the neural network. The step of generating generates an image that is comprised of a plurality of points, wherein each of the points is referenced to an x-y coordinate system, wherein a distance along the x-axis is a function of a value of the output of the neural network for a given input vector applied to a first input of the neural network, and wherein a distance along the y-axis is function of a value of the output of the neural network for the given input vector applied a second input of the neural network.

Claims

1. A computer-implemented method of testing a neural network, comprising a set of processing nodes interconnected by a plurality of interconnecting links in a sequence of layers from an input layer to an output layer, each processing node receiving an input signal over each interconnecting link connected thereto and generating a nodal output signal in response thereto and further in response to a respective connection weight value, nodal output signals from processing nodes in the output layer comprising a neural network output, in relation to a reference represented by a reference image, the method comprising the steps of:

stimulating at least one processing node of the input layer of the neural network with an input vector comprising at least one input signal to obtain output signals from processing nodes at the output layer of the neural network;

obtaining a plurality of samples of the output signals, wherein at least one of the plurality of samples is delayed in time from another one of the samples;

generating an image from the plurality of samples; and

comparing the image to the reference image to determine an operational characteristic of the neural network.

2. A method as set forth in claim 1 wherein the at least one of the plurality of samples represents a difference between two samples.

3. A method as set forth in claim 1 wherein the step of generating generates an image comprising a plurality of points, and wherein each of the points represents a value of the output signal resulting from the application of the input vector to one of the inputs of the neural network.

4. A method as set forth in claim 1 wherein the step of generating generates an image comprising a plurality of points, wherein each of the points is referenced to an x-y coordinate system, wherein a distance along the x-axis is a function of a value of the output of the neural network for a given input vector applied to a first input of the neural network, and wherein a distance along the y-axis is function of a value of the output of the neural network for the given input vector applied to a second input of the neural network.

5. A method as set forth in claim 1 wherein the step of stimulating includes the steps of:

generating an input signal having a voltage potential whose value varies with time; and

successively applying a given value of the time varying voltage potential to each of the inputs of the neural network.

6. A computer-implemented method of testing a neural network, comprising a set of processing nodes interconnected by plurality of interconnecting links in a sequence of layers from an input layer to an output layer, each processing node receiving an input signal over each interconnecting link connected thereto and generating a nodal output signal in response thereto and further in response to a respective connection weight value, nodal output signals from processing nodes in the output layer comprising a neural network output, in relation to a reference represented by a reference image, the method comprising the steps of:

stimulating at least one input of the neural network with an input vector to obtain an output signal at an output of the neural network;

obtaining at least one sample of the output signal;

generating an image from the at least one sample of the output signal and from the input vector; and

comparing the image to a reference image to determine an operational characteristic of the neural network.

7. A method as set forth in claim 6 wherein the at least one sample represents a difference between two samples.

8. A method as set forth in claim 6 wherein the step of generating generates an image comprising a plurality of points, and wherein each of the points represents a value of an output resulting from the input vector to one of the inputs of the neural network.

9. A method as set forth in claim 6 wherein the step of generating generates an image comprising a plurality of points, wherein each of the points is referenced to an x-y coordinate system, wherein a distance along the x-axis is a function of a value of the input vector, and wherein a distance along the y-axis is function of a value of the output of the neural network for the input vector.

10. A method as set forth in claim 6 wherein the step of stimulating includes the steps of:

generating an input signal having a voltage potential whose value varies with time; and

successively applying a given value of the time varying voltage potential to each of the inputs of the neural network.

11. Apparatus for testing a neural network, comprising a set of processing nodes interconnected by a plurality of interconnecting links in a sequence of layers from an input layer to an output layer, each processing node receiving an input signal over each interconnecting link connected thereto and generating a nodal output signal in response thereto and further in response to a respective connection weight value, nodal output signals from processing nodes in the output layer comprising a neural network output, in relation to a reference represented by a reference image, comprising:

means for stimulating at least one processing node of the input layer of the neural network with an input vector to obtain output signals at one of said processing nodes of the output layer of the neural network;

means for obtaining a plurality of samples of the output signals, wherein at least one of the plurality of samples is delayed in time from another one of the samples;

means for generating an image from the plurality of samples; and

means for comparing the image to the reference image to determine an operational characteristic of the neural network.

12. Apparatus as set forth in claim 11 and further including means for generating a difference between two samples.

13. Apparatus as set forth in claim 11 wherein said obtaining means comprises analog voltage storage means for storing a plurality of samples of the output of the neural network.

14. Apparatus as set forth in claim 11 wherein said generating means generates an image comprising a plurality of points, wherein each of the points is referenced to an x-y coordinate system, wherein a distance along the x-axis is a function of a value of the output of the neural network for a given input vector applied to a first input of the neural network, and wherein a distance along the y-axis is function of a value of the output of the neural network for the given input vector applied a second input of the neural network.

15. Apparatus as set forth in claim 11 wherein said stimulating means comprises:

means for generating a time varying voltage potential; and

means for successively applying a given value of the time varying voltage potential to each of the inputs of the neural network.

16. Apparatus for testing a neural network, comprising a set of processing nodes interconnected by a plurality of interconnecting links in a sequence of layers from an input layer to an output layer, each processing node receiving an input signal over each interconnecting link connected thereto and generating a nodal output signal in response thereto and further in response to a respective connection weight value, nodal output signals from processing nodes in the output layer comprising a neural network output, in relation to a reference represented by a reference image, comprising:

means for stimulating at least one processing node of the input layer of the neural network with an input vector to obtain an output signal at one of the processing nodes of the output layer of the neural network;

means for obtaining at least one sample of the output signal; and

means for generating an image from the at least one sample of the output signal and from the input vector; and

means for comparing the image to the reference image to determine an operational characteristic of the neural network.

17. Apparatus as set forth in claim 16 and further including means for generating a difference between two samples.

18. Apparatus as set forth in claim 16 wherein said obtaining means comprises analog voltage storage means for storing a plurality of samples of the output of the neural network.

19. Apparatus as set forth in claim 16 wherein said generating means generates an image comprising a plurality of points, wherein each of the points is referenced to an x-y coordinate system, wherein a distance along the x-axis is a function of a value of the input vector, and wherein a distance along the y-axis is function of a value of the output of the neural network.

20. Apparatus as set forth in claim 16 wherein said stimulating means comprises:

means for generating a time varying voltage potential; and

means for successively applying a given value of the time varying voltage potential to each of the inputs of the neural network.