Method and apparatus for neural learning of natural multi-spike trains in spiking neural networks

• US 9,111,224 B2
• Filed: 10/19/2011
• Issued: 08/18/2015
• Est. Priority Date: 10/19/2011
• Status: Active Grant

• Current Assignee: Qualcomm, Inc.
• Original Assignee: Qualcomm, Inc.
• Inventors: Hunzinger, Jason Frank
• Primary Examiner(s): Gaffin, Jeffrey A
• Assistant Examiner(s): Misir, David
• Application Number:
  US13/276,812
• Publication Number:
  US 20130103626A1
• Time in Patent Office: 1,399 Days
• US Class Current: 1/1
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning

Abstract

Certain aspects of the present disclosure support a technique for neural learning of natural multi-spike trains in spiking neural networks. A synaptic weight can be adapted depending on a resource associated with the synapse, which can be depleted by weight change and can recover over time. In one aspect of the present disclosure, the weight adaptation may depend on a time since the last significant weight change.

Claims

1. A method of neural learning, comprising:

modulating an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron; and

adapting the weight of the synapse based on a resource associated with the synapse, wherein a value of the resource is modified as a function of time relative to adapting the weight of the synapse, such that the value of the resource is recovered over time towards a previous value before the modification.

2. The method of claim 1, further comprising:

modifying the resource in relation to adapting the weight of synapse; and

recovering the resource over time towards a value of the resource before the modification.

3. The method of claim 2, wherein modifying the resource comprises decreasing the resource.

4. The method of claim 3, wherein the resource is decreased by an amount proportional to an absolute value of a change in the weight of synapse during the adaptation.

5. The method of claim 3, wherein the resource is decreased to zero, if an absolute value of a change in the weight of synapse during the adaptation is above a threshold.

6. The method of claim 2, wherein recovering the resource over time comprises:

recovering, after a time delay, the resource to the value before the modification.

7. The method of claim 2, wherein modifying the resource occurs when committing to adapting the weight.

8. The method of claim 2, wherein:

modifying the resource occurs when determining an aspect of adapting the weight, and

adapting the weight occurs at a later time.

9. The method of claim 8, wherein the aspect of adapting the weight comprises at least one of:

an amount of a change in the weight, or a direction of a change in the weight.

10. The method of claim 1, wherein a change in the weight of synapse during the adaptation is not applied, if the resource is below a threshold.

11. The method of claim 1, wherein a change in the weight of synapse during the adaptation is not applied, if the time since the last change is below a threshold.

12. The method of claim 1, wherein adapting the weight of synapse further comprises:

updating the weight of synapse effective at a time of spiking of the post-synaptic neuron or the pre-synaptic neuron;

storing, at a time of the weight update, a time of updating the weight as a last time the resource was used;

computing the resource at another time of updating the weight as a function of the other time and the last time the resource was used; and

obtaining a change in the weight of synapse as a function of the computed resource.

13. The method of claim 12, wherein a value of the last time the resource was used is not overwritten, if a magnitude of the change in the weight of synapse is below a threshold.

14. The method of claim 12, wherein the change in the weight is zero, if a value of the other time minus the last time the resource was used is smaller than a threshold.

15. An apparatus for neural learning, comprising:

a first circuit configured to modulate an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron; and

a second circuit configured to adapt the weight of the synapse based on a resource associated with the synapse, wherein a value of the resource is modified as a function of time relative to adapting the weight of the synapse, such that the value of the resource is recovered over time towards a previous value before the modification.

16. The apparatus of claim 15, further comprising:

a third circuit configured to modify the resource in relation to adapting the weight of synapse; and

a fourth circuit configured to recover the resource over time towards a value of the resource before the modification.

17. The apparatus of claim 16, wherein modifying the resource comprises decreasing the resource.

18. The apparatus of claim 17, wherein the resource is decreased by an amount proportional to an absolute value of a change in the weight of synapse during the adaptation.

19. The apparatus of claim 17, wherein the resource is decreased to zero, if an absolute value of a change in the weight of synapse during the adaptation is above a threshold.

20. The apparatus of claim 16, wherein the fourth circuit is also configured to:

recover, after a time delay, the resource to the value before the modification.

21. The apparatus of claim 16, wherein modifying the resource occurs when committing to adapting the weight.

22. The apparatus of claim 16, wherein:

modifying the resource occurs when determining an aspect of adapting the weight, and

adapting the weight occurs at a later time.

23. The apparatus of claim 22, wherein the aspect of adapting the weight comprises at least one of:

an amount of a change in the weight, or a direction of a change in the weight.

24. The apparatus of claim 15, wherein a change in the weight of synapse during the adaptation is not applied, if the resource is below a threshold.

25. The apparatus of claim 15, wherein a change in the weight of synapse during the adaptation is not applied, if the time since the last change is below a threshold.

26. The apparatus of claim 15, wherein the second circuit is also configured to:

update the weight of synapse effective at a time of spiking of the post-synaptic neuron or the pre-synaptic neuron;

store, at a time of the weight update, a time of updating the weight as a last time the resource was used;

compute the resource at another time of updating the weight as a function of the other time and the last time the resource was used; and

obtain a change in the weight of synapse as a function of the computed resource.

27. The apparatus of claim 26, wherein a value of the last time the resource was used is not overwritten, if a magnitude of the change in the weight of synapse is below a threshold.

28. The apparatus of claim 26, wherein the change in the weight is zero, if a value of the other time minus the last time the resource was used is smaller than a threshold.

29. An apparatus for neural learning, comprising:

means for modulating an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron; and

means for adapting the weight of the synapse based on a resource associated with the synapse, wherein a value of the resource is modified as a function of time relative to adapting the weight of the synapse, such that the value of the resource is recovered over time towards a previous value before the modification.

30. The apparatus of claim 29, further comprising:

means for modifying the resource in relation to adapting the weight of synapse; and

means for recovering the resource over time towards a value of the resource before the modification.

31. The apparatus of claim 30, wherein the means for modifying the resource comprises means for decreasing the resource.

32. The apparatus of claim 31, wherein the resource is decreased by an amount proportional to an absolute value of a change in the weight of synapse during the adaptation.

33. The apparatus of claim 31, wherein the resource is decreased to zero, if an absolute value of a change in the weight of synapse during the adaptation is above a threshold.

34. The apparatus of claim 30, wherein the means for recovering the resource over time comprises:

means for recovering, after a time delay, the resource to the value before the modification.

35. The apparatus of claim 30, wherein modifying the resource occurs when committing to adapting the weight.

36. The apparatus of claim 30, wherein:

modifying the resource occurs when determining an aspect of adapting the weight, and

adapting the weight occurs at a later time.

37. The apparatus of claim 36, wherein the aspect of adapting the weight comprises at least one of:

an amount of a change in the weight, or a direction of a change in the weight.

38. The apparatus of claim 29, wherein a change in the weight of synapse during the adaptation is not applied, if the resource is below a threshold.

39. The apparatus of claim 29, wherein a change in the weight of synapse during the adaptation is not applied, if the time since the last change is below a threshold.

40. The apparatus of claim 29, wherein the means for adapting the weight of synapse comprises:

means for updating the weight of synapse effective at a time of spiking of the post-synaptic neuron or the pre-synaptic neuron;

means for storing, at a time of the weight update, a time of updating the weight as a last time the resource was used;

means for computing the resource at another time of updating the weight as a function of the other time and the last time the resource was used; and

means for obtaining a change in the weight of synapse as a function of the computed resource.

41. The apparatus of claim 40, wherein a value of the last time the resource was used is not overwritten, if a magnitude of the change in the weight of synapse is below a threshold.

42. The apparatus of claim 40, wherein the change in the weight is zero, if a value of the other time minus the last time the resource was used is smaller than a threshold.

43. A computer program product for neural learning, comprising a non-transitory computer-readable medium comprising code for:

modulating an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron; and

adapting the weight of the synapse based on a resource associated with the synapse, wherein a value of the resource is modified as a function of time relative to adapting the weight of the synapse, such that the value of the resource is recovered over time towards a previous value before the modification.

44. The computer program product of claim 43, wherein the non-transitory computer-readable medium further comprises code for:

modifying the resource in relation to adapting the weight of synapse; and

recovering the resource over time towards a value of the resource before the modification.

45. The computer program product of claim 44, wherein modifying the resource comprises decreasing the resource.

46. The computer program product of claim 45, wherein the resource is decreased by an amount proportional to an absolute value of a change in the weight of synapse during the adaptation.

47. The computer program product of claim 45, wherein the resource is decreased to zero, if an absolute value of a change in the weight of synapse during the adaptation is above a threshold.

48. The computer program product of claim 44, wherein the non-transitory computer-readable medium further comprises code for:

recovering, after a time delay, the resource to the value before the modification.

49. The computer program product of claim 44, wherein modifying the resource occurs when committing to adapting the weight.

50. The computer program product of claim 44, wherein:

modifying the resource occurs when determining an aspect of adapting the weight, and

adapting the weight occurs at a later time.

51. The computer program product of claim 50, wherein the aspect of adapting the weight comprises at least one of:

an amount of a change in the weight, or a direction of a change in the weight.

52. The computer program product of claim 43, wherein a change in the weight of synapse during the adaptation is not applied, if the resource is below a threshold.

53. The computer program product of claim 43, wherein a change in the weight of synapse during the adaptation is not applied, if the time since the last change is below a threshold.

54. The computer program product of claim 43, wherein the non-transitory computer-readable medium further comprises code for:

updating the weight of synapse effective at a time of spiking of the post-synaptic neuron or the pre-synaptic neuron;

storing, at a time of the weight update, a time of updating the weight as a last time the resource was used;

computing the resource at another time of updating the weight as a function of the other time and the last time the resource was used; and

obtaining a change in the weight of synapse as a function of the computed resource.

55. The computer program product of claim 54, wherein a value of the last time the resource was used is not overwritten, if a magnitude of the change in the weight of synapse is below a threshold.

56. The computer program product of claim 54, wherein the change in the weight is zero, if a value of the other time minus the last time the resource was used is smaller than a threshold.

57. A method of neural learning, comprising:

modulating an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron;

adapting the weight of synapse based on a resource consideration associated with the synapse and spike timing of the pre-synaptic neuron and the post-synaptic neuron; and

determining the resource consideration based on one or more timing relations between prior spikes of the pre-synaptic and post-synaptic neurons.

58. The method of claim 57, wherein the one or more timing relations comprise information about a time difference between a pair of the prior spikes of the pre-synaptic and post-synaptic neurons.

59. The method of claim 57, wherein the one or more timing relations comprise information about a time difference between a pair of the prior spikes and another pair of the prior spikes.

60. The method of claim 57, wherein the one or more timing relations comprise information about a time since a pair of the prior spikes occurred.

61. An apparatus for neural learning, comprising:

a first circuit configured to modulate an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron;

a second circuit configured to adapt the weight of synapse based on a resource consideration associated with the synapse and spike timing of the pre-synaptic neuron and the post-synaptic neuron; and

a third circuit configured to determine the resource consideration based on one or more timing relations between prior spikes of the pre-synaptic and post-synaptic neurons.

62. The apparatus of claim 61, wherein the one or more timing relations comprise information about a time difference between a pair of the prior spikes of the pre-synaptic and post-synaptic neurons.

63. The apparatus of claim 61, wherein the one or more timing relations comprise information about a time difference between a pair of the prior spikes and another pair of the prior spikes.

64. The apparatus of claim 61, wherein the one or more timing relations comprise information about a time since a pair of the prior spikes occurred.

65. An apparatus for neural learning, comprising:

means for modulating an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron;

means for adapting the weight of synapse based on a resource consideration associated with the synapse and spike timing of the pre-synaptic neuron and the post-synaptic neuron; and

means for determining the resource consideration based on one or more timing relations between prior spikes of the pre-synaptic and post-synaptic neurons.

66. The apparatus of claim 65, wherein the one or more timing relations comprise information about a time difference between a pair of the prior spikes of the pre-synaptic and post-synaptic neurons.

67. The apparatus of claim 65, wherein the one or more timing relations comprise information about a time difference between a pair of the prior spikes and another pair of the prior spikes.

68. The apparatus of claim 65, wherein the one or more timing relations comprise information about a time since a pair of the prior spikes occurred.

69. A computer program product for neural learning, comprising a non-transitory computer-readable medium comprising code for:

modulating an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron;

adapting the weight of synapse based on a resource consideration associated with the synapse and spike timing of the pre-synaptic neuron and the post-synaptic neuron; and

determining the resource consideration based on one or more timing relations between prior spikes of the pre-synaptic and post-synaptic neurons.

70. The computer program product of claim 69, wherein the one or more timing relations comprise information about a time difference between a pair of the prior spikes of the pre-synaptic and post-synaptic neurons.

71. The computer program product of claim 69, wherein the one or more timing relations comprise information about a time difference between a pair of the prior spikes and another pair of the prior spikes.

72. The computer program product of claim 69, wherein the one or more timing relations comprise information about a time since a pair of the prior spikes occurred.

73. A method of neural learning, comprising:

modulating an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron;

adapting the weight of synapse based on a resource consideration associated with the synapse and spike timing of the pre-synaptic neuron and the post-synaptic neuron; and

determining the resource consideration based on an amount of time since a prior commitment to a change of the weight.

74. The method of claim 73, wherein the change of weight occurs a time after committing to the change of weight.

75. An apparatus for neural learning, comprising:

a first circuit configured to modulate an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron;

a second circuit configured to adapt the weight of synapse based on a resource consideration associated with the synapse and spike timing of the pre-synaptic neuron and the post-synaptic neuron; and

a third circuit configured to determine the resource consideration based on an amount of time since a prior commitment to a change of the weight.

76. The apparatus of claim 75, wherein the change of weight occurs a time after committing to the change of weight.

77. An apparatus for neural learning, comprising:

means for modulating an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron;

means for adapting the weight of synapse based on a resource consideration associated with the synapse and spike timing of the pre-synaptic neuron and the post-synaptic neuron; and

means for determining the resource consideration based on an amount of time since a prior commitment to a change of the weight.

78. The apparatus of claim 77, wherein the change of weight occurs a time after committing to the change of weight.

79. A computer program product for neural learning, comprising a non-transitory computer-readable medium comprising code for:

modulating an input from a pre-synaptic neuron to a post-synaptic neuron by a weight of a synapse connecting the pre-synaptic neuron and the post-synaptic neuron;

adapting the weight of synapse based on a resource consideration associated with the synapse and spike timing of the pre-synaptic neuron and the post-synaptic neuron; and

determining the resource consideration based on an amount of time since a prior commitment to a change of the weight.

80. The computer program product of claim 79, wherein the change of weight occurs a time after committing to the change of weight.