Method, apparatus, and program for evolving neural network architectures to detect content in media information

• US 20040010480A1
• Filed: 07/09/2002
• Published: 01/15/2004
• Est. Priority Date: 07/09/2002
• Status: Abandoned Application

• Current Assignee: Koninklijke Philips Electronics N.V. (Koninklijke Philips N.V.)
• Original Assignee: Koninklijke Philips Electronics N.V. (Koninklijke Philips N.V.)
• Inventors: McGee, Thomas, Dimitrova, Nevenka, Agnihotri, Lalitha, Schaffer, James David, Jeannin, Sylvie
• Application Number:
  US10/190,810
• Publication Number:
  US 20040010480A1
• Time in Patent Office: Days
• US Class Current: 706/20
• CPC Class Codes: G06N 3/086 using evolutionary algorithms, e.g. genetic algorithms or genetic programming

Abstract

A method for operating a neural network, and a program and apparatus that operate in accordance with the method. The method comprises the steps of applying data indicative of predetermined content, derived from an electronic signal including a representation of the predetermined content, to an input of at least one neural network, to cause the at least one network to generate at least one output indicative of either a detection or a non-detection of the predetermined content. Each neural network has an architecture specified by at least one corresponding parameter. The method also comprises a step of evolving the at least one parameter to modify the architecture of the at least one neural network, based on the at least one output, to increase an accuracy at which that at least one neural network detects the predetermined content indicated by the data.

Claims

1. A method for operating at least one neural network, comprising the steps of:

applying data including an indication of predetermined content to an input of the at least one neural network, to cause the at least one network to generate at least one output indicative of either a detection or a non-detection of the predetermined content, wherein each neural network has an architecture specified by at least one corresponding parameter; and

evolving the at least one parameter to modify the architecture of the at least one neural network, based on the at least one output, to increase an accuracy at which that at least one neural network detects the predetermined content indicated by the data.

2. A method as set forth in claim 1, wherein the data represents at least one of an audio signal and a video signal, and the predetermined content includes desired or undesired content.

3. A method as set forth in claim 1, further comprising a step of deriving the data based on a provided electronic signal, prior to the step of applying.

4. A method as set forth in claim 3, wherein the data includes at least one of low-level data, mid-level data, and hi-level data derived from the electronic signal.

5. A method as set forth in claim 1, wherein the step of evolving includes performing a genetic algorithm to evolve the at least one parameter.

6. A method as set forth in claim 1, wherein the at least one parameter specifies at least one of a number of node layers included in the at least one neural network and a number of nodes included in each node layer.

7. A method as set forth in claim 6, wherein the at least one parameter also specifies at least one of a learning rate for use with the at least one neural network and at least one connection weight value for the at least one neural network.

8. A method as set forth in claim 1, further comprising, prior to the step of applying:

training the at least one neural network to detect the predetermined content by;

applying further data including an indiction of the predetermined content, to the input of the at least one neural network to cause the at least one neural network to generate at least one further output, and

varying at least one connection weight value of the at least one neural network based on the at least one further output.

9. A method as set forth in claim 1, wherein the step of evolving comprises the steps of:

determining the accuracy at which the at least one neural network detects the predetermined content in the applying step, for each architecture of the at least one neural network;

selecting at least one parameter, each specifying a corresponding one of the architectures, based on a result of the step of determining the accuracy; and

producing at least one offspring parameter based on the at least one parameter selected in the step of selecting.

10. A method as set forth in claim 9, wherein the step of evolving further comprises the steps of:

applying further data including an indication of the predetermined content, to the input of the at least one neural network while the at least one neural network has an architecture specified by the at least one offspring parameter produced in the producing step, to cause the at least one network to generate at least one further output indicative of either a detection or a non-detection of the predetermined content;

determining the accuracy at which the at least one neural network detects the predetermined content in that step of applying; and

further selecting one or more of at least one parameter selected in the selecting step and at least one offspring parameter produced in the producing step, based on a result of that step of determining.

11. A method as set forth in claim 10, wherein the step of evolving further comprises the steps of:

determining if there is a convergence of parameters remaining after the further selecting step; and

if there is a convergence, storing a record of at least one of the parameters selected in the further selecting step.

12. A method as set forth in claim 9, wherein the step of producing comprises:

pairing randomly-selected ones of parameters selected in the selecting step;

determining if the parameters paired in the pairing step are incestuous; and

for each of the paired parameters determined to be non-incestuous, swapping one or more values of the parameters with one another.

13. A method as set forth in claim 10, wherein the step of evolving further comprises the steps of:

determining if there is a convergence of all parameters remaining after the further selecting step; and

if there is no convergence, mutating at least one value of a most optimum one of all parameters remaining after the further selecting step, to produce plural mutated versions of the most optimum parameter; and

applying even further data including an indication of the predetermined content, to the input of the at least one neural network while the at least one neural network has an architecture specified by the most optimum parameter or one of the mutated versions of the most optimum parameter.

14. A method as set forth in claim 1, further comprising the step of forwarding at least one of an evolved parameter resulting from the step of evolving and information representing the architecture as modified in the step of evolving, towards a predetermined destination.

15. A method for detecting predetermined content represented in a provided electronic signal representing at least one of video and audio information, the method comprising the steps of:

applying data derived from the signal to inputs of respective ones of separate neural networks to cause the neural networks to generate corresponding outputs indicative of either a detection or a non-detection of the predetermined content, wherein each neural network has a corresponding architecture;

determining an accuracy at which individual ones of the neural networks detect the predetermined content, based on the outputs generated in the step of applying; and

based on the step of determining, modifying the architecture of at least one of the neural networks, to substantially maximize the accuracy at which that least one neural network detects the predetermined content represented in the signal from which the data is derived.

16. A method as set forth in claim 15, further comprising the step of encoding the electronic signal to derive the data applied in the step of applying.

17. An apparatus for detecting predetermined content represented in a provided electronic signal, the apparatus comprising:

a feature deriver, arranged for deriving predetermined feature data indicative of the predetermined content from the provided electronic signal; and

a controller, operating under the control of a stored program, for (a) applying the predetermined feature data derived by the feature deriver to at least one input of at least one neural network to cause the at least one neural network to generate at least one output indicative of either a detection or a non-detection of the predetermined content, wherein each neural network has an architecture specified by at least one corresponding parameter, and (b) evolving the at least one parameter to modify the architecture of the at least one neural network, based on the at least one output, to increase an accuracy at which that at least one neural network detects the predetermined content indicated by the predetermined feature data.

18. An apparatus as set forth in claim 17, wherein the predetermined content includes desired or undesired content.

19. An apparatus as set forth in claim 17, wherein the controller operates under the control of the program to evolve the at least one parameter by performing a genetic algorithm.

20. An apparatus as set forth in claim 17, wherein the controller also operates under the control of the program to, prior to the applying, train the at least one neural network to detect the predetermined content by (i) applying at least some of the predetermined feature data derived by the feature deriver to the at least one input of the at least one neural network to cause the at least one neural network to generate at least one further output, and (ii) varying at least one connection weight value of the at least one neural network based on the at least one further output.

21. An apparatus as set forth in claim 17, wherein the controller operates under the control of the program to perform the evolving by:

determining the accuracy at which the at least one neural network detects the predetermined content indicated by the predetermined feature data, for each architecture of the at least one neural network,

selecting one or more of the at least one parameter, each specifying a corresponding one of the architectures, based on a result of the determining, and

producing at least one offspring parameter based on the one or more at least one parameter selected in the selecting.

22. An apparatus as set forth in claim 21, wherein the feature deriver also is arranged to derive further predetermined feature data indicative of the predetermined content, from a further electronic signal, and the controller also operates under the control of the program to perform the evolving by:

applying the further predetermined feature data to the at least one input of the at least one neural network while the at least one neural network has an architecture specified by the at least one offspring parameter produced in the producing, to cause the at least one neural network to generate at least one further output indicative of either a detection or a non-detection of the predetermined content,

determining the accuracy at which the at least one neural network detects the predetermined content indicated by the further predetermined feature data, and

further selecting at least one of the one or more at least one parameter selected in the selecting and the at least one offspring parameter produced in the producing, based on a result of that step of determining.

23. An apparatus as set forth in claim 22, wherein the controller also operates under the control of the program to further perform the evolving by:

determining if there is a convergence of all parameters remaining after the further selecting, and

if there is a convergence, storing a record of at least parameter selected in the further selecting.

24. An apparatus as set forth in claim 17, and further comprising an interface coupled to the controller, wherein the controller is operable for forwarding at least one of an evolved parameter resulting from the evolving and information representing the architecture as modified in the evolving, through the interface towards at least one external destination coupled to that interface.

25. An apparatus as set forth in claim 24, wherein the apparatus includes a server, and the external destination includes an information appliance.

26. An apparatus for detecting predetermined content indicated in data representative of a provided electronic signal, the apparatus comprising:

neural network means, responsive to the data being applied to an input thereof, for generating at least one output indicative of either a detection or a non-detection of the predetermined content, wherein an architecture of the neural network means is specified by at least one corresponding parameter; and

means for evolving the at least one parameter to modify the architecture of the neural network means, based on the at least one output, to increase an accuracy at which the neural network means detects the predetermined content indicated by the data.

27. An apparatus for detecting predetermined content indicated in predetermined feature data derived from a provided electronic signal, the apparatus comprising:

plural neural networks, each being responsive to the predetermined feature data being applied to an input thereof, for generating a corresponding output indicative of either a detection or a non-detection of the predetermined content, wherein each neural network has a corresponding architecture;

means for determining an accuracy at which each neural network detects the predetermined content, based on the output generated by that neural network; and

means for modifying the architecture of at least one of the neural networks, to substantially maximize the accuracy at which that least one neural network detects the predetermined content indicated by the predetermined feature data, based on a determination made by said means for determining.

28. A program product comprising computer readable-code which, when executed, performs a method for operating at least one neural network, the method comprising the steps of:

applying data including a representation of predetermined content, to an input of at least one neural network, to cause the at least one neural network to generate at least one output indicative of either a detection or a non-detection of the predetermined content, wherein each neural network has an architecture specified by at least one corresponding parameter; and

evolving the at least one parameter to modify the architecture of the at least one neural network, based on the at least one output, to increase an accuracy at which that at least one neural network detects the predetermined content.

29. A storage medium storing a program having computer readable-code which, when executed, performs a method for operating at least one neural network, the method comprising the steps of:

applying data including a representation of predetermined content, to an input of at least one neural network, to cause the at least one neural network to generate at least one output indicative of either a detection or a non-detection of the predetermined content, wherein each neural network has an architecture specified by at least one corresponding parameter; and

evolving the at least one parameter to modify the architecture of the at least one neural network, based on the at least one output, to increase an accuracy at which that at least one neural network detects the predetermined content.

30. A system for exchanging information, comprising:

at least one first information apparatus; and

at least one second information apparatus, comprising;

an interface, coupled to said first information apparatus through an external communication interface,

a feature deriver, arranged for deriving predetermined feature data indicative of predetermined content, based on a provided electronic signal that includes a representation of the predetermined content, and

a controller, operating under the control of a stored program, for (a) applying the predetermined feature data derived by the feature deriver to an input of at least one neural network to cause the at least one neural network to generate at least one output indicative of either a detection or a non-detection of the predetermined content, wherein each neural network has an architecture specified by at least one corresponding parameter, (b) evolving the at least one parameter to modify the architecture of the at least one neural network, based on the at least one output, to increase an accuracy at which that at least one neural network detects the predetermined content indicated by the predetermined feature data, and (c) forwarding information representing at least one of an evolved parameter resulting from the evolving and the architecture as modified by the evolving, to the at least one first information apparatus through the interface and the external communication interface.

31. A system as set forth in claim 30, wherein the first information apparatus is an information appliance, and the second information apparatus is a server.

32. A system as set forth in claim 30, wherein the first information apparatus is a server, and the second information apparatus is an information appliance.

33. A system as set forth in claim 32, wherein the first information apparatus comprises:

a further interface, coupled to said interface of said second information apparatus through the external communication interface; and

a further controller coupled to said further interface, said further controller being responsive to said further interface receiving the information from said second information apparatus for at least one of storing the information in an associated memory and applying further data indicative of the predetermined content, derived from a further signal including a representation of the predetermined content, to a further neural network having an architecture specified in accordance with the information, to cause the further neural network to detect the predetermined content.

34. A system as set forth in claim 33, wherein the first information apparatus is an information appliance, and the second information apparatus is a server.

35. A system as set forth in claim 30, wherein the first information apparatus is operable for providing the electronic signal to said at least one second information apparatus through said interface and said external communication interface, for application of the electronic signal to the feature deriver, and wherein said feature deriver derives the predetermined feature data from the electronic signal in response to that application.