Neural network adaptation to current computational resources

• US 10,140,573 B2
• Filed: 05/02/2014
• Issued: 11/27/2018
• Est. Priority Date: 03/03/2014
• Status: Active Grant

• Current Assignee: Qualcomm, Inc.
• Original Assignee: Qualcomm, Inc.
• Inventors: Julian, David Jonathan, Chang, Ilwoo
• Primary Examiner(s): Vincent, David R
• Application Number:
  US14/268,372
• Publication Number:
  US 20150248609A1
• Time in Patent Office: 1,670 Days
• Field of Search: 706 15, 706 45
• CPC Class Codes: G06N 3/048 Activation functions
• CPC Class Codes: G06N 3/082 modifying the architecture, e.g. adding, deleting or silencing nodes or connections

Abstract

Methods and apparatus are provided for processing in an artificial nervous system. According to certain aspects, resolution of one or more functions performed by processing units of a neuron model may be reduced, based at least in part on availability of computational resources or a power target or budget. The reduction in resolution may be compensated for by adjusting one or more network weights.

Claims

1. A method for processing in a Neural Processing Unit (NPU) cortex of a robotic device having a system of artificial neurons, comprising:

reducing resolution of one or more functions performed by at least one processing unit of the NPU cortex, based at least in part on availability of computational resources, the at least one processing unit including at least one neuromorphic processor;

compensating for the reduction in resolution by adjusting one or more network weights associated with the system; and

operating at least one sensor of the robotic device, using the at least one processing unit, at the reduced resolution and the adjusted one or more network weights for performing the one or more functions.

2. The method of claim 1, wherein:

reducing resolution comprises sub-sampling; and

compensating comprises increasing weights in proportion to the reduction in resolution.

3. The method of claim 2, wherein:

the sub-sampling comprise sampling 1 out of N pixels; and

compensating comprise increasing weights by a factor of N.

4. The method of claim 1, wherein reducing resolution further comprises reducing a number of the at least one processing unit of the neuron model.

5. The method of claim 1, further comprising:

dynamically adjusting the resolution of the one or more functions based at least in part on at least one of a computational load or available computational resources associated with other functions performed by other processing units in the system.

6. The method of claim 1, further comprising:

increasing resolution of a subset of the one or more functions performed by a pertinent subset of the at least one processing unit, if results obtained by the at least one processing unit trigger the system to a particular operational state.

7. The method of claim 1, further comprising:

adjusting resolution of a subset of the one or more functions performed by a pertinent subset of the at least one processing unit to preferred resolution, if results obtained by the at least one processing unit trigger the system to a particular operational state.

8. The method of claim 1, further comprising:

adding, in the system, one or more other functions performed by one or more other processing units, if the one or more functions provide a desired result.

9. The method of claim 1, further comprising:

increasing resolution of a subset of the one or more functions performed by a subset of the at least one processing unit, if the at least one processing unit performing the one or more functions detects a context switch; and

scaling the one or more network weights, if the at least one processing unit performing the one or more functions detects the context switch.

10. The method of claim 1, wherein reducing resolution comprises sub-sampling of auditory neurons in the system.

11. The method of claim 10, wherein the sub-sampling of the auditory neurons comprises adjusting a frequency bandwidth associated with the auditory neurons.

12. The method of claim 10, further comprising:

adjusting a size of Fast Fourier Transform (FFT) based at least in part on availability of the computational resources and operational state of the system, wherein an output of the FFT of audio frames is input into the auditory neurons.

13. The method of claim 12, wherein adjusting the size of FFT comprises adjusting resolution of frequency bin sizes associated with the FFT.

14. The method of claim 12, wherein adjusting the size of FFT comprises adjusting a dynamic range of audio.

15. The method of claim 1, further comprising:

sampling accelerometer or Global Positioning System (GPS) data at an adjustable rate, wherein

reducing resolution comprises varying a number of the at least one processing unit that process the sampled data based at least in part on the adjustable rate.

16. An apparatus for processing in a Neural Processing Unit (NPU) cortex of a robotic device having a system of artificial neurons, comprising:

a processing system configured to;

reduce resolution of one or more functions performed by at least one processing unit of the NPU cortex, based at least in part on availability of computational resources, the at least one processing unit including at least one neuromorphic processor;

compensate for the reduction in resolution by adjusting one or more network weights associated with the system;

operate at least one sensor of the robotic device, using the at least one processing unit, at the reduced resolution and the adjusted one or more network weights for performing the one or more functions; and

a memory coupled to the processing system.

28. The apparatus of claim 16, wherein the processing system is configured to:

reduce resolution comprises sub-sampling; and

compensate comprises increasing weights in proportion to the reduction in resolution.

29. The apparatus of claim 16, wherein the processing system is configured to reduce the resolution by reducing a number of the processing units of the neuron model.

30. The apparatus of claim 16, wherein the processing system is further configured to:

dynamically adjust the resolution of the one or more functions based at least in part on at least one of a computational load or available computational resources associated with other functions performed by other processing units in the system.

31. The apparatus of claim 16, wherein the processing system is further configured to:

increase resolution of a subset of the one or more functions performed by a pertinent subset of the at least one processing unit, if results obtained by the at least one processing unit trigger the system to a particular operational state.

32. The apparatus of claim 16, wherein the processing system is further configured to:

adjust resolution of a subset of the one or more functions performed by a pertinent subset of the at least one processing unit to preferred resolution, if results obtained by the at least one processing unit trigger the system to a particular operational state.

17. An apparatus for processing in a Neural Processing Unit (NPU) cortex of a robotic device having a system of artificial neurons, comprising:

means for reducing resolution of one or more functions performed by at least one processing unit of the NPU cortex, based at least in part on availability of computational resources, the at least one processing unit including at least one neuromorphic processor;

means for compensating for the reduction in resolution by adjusting one or more network weights associated with the system; and

means for operating at least one sensor of the robotic device, using the at least one processing unit, at the reduced resolution and the adjusted one or more network weights for performing the one or more functions.

18. A non-transitory computer-readable medium for processing in a Neural Processing Unit (NPU) cortex of a robotic device having a system of artificial neurons, the non-transitory computer-readable medium having instructions executable to:

reduce resolution of one or more functions performed by at least one processing unit of the NPU cortex, based at least in part on availability of computational resources, the at least one processing unit including at least one neuromorphic processor;

compensate for the reduction in resolution by adjusting one or more network weights associated with the system; and

operate at least one sensor of the robotic device, using the at least one processing unit, at the reduced resolution and the adjusted one or more network weights for performing the one or more functions.

19. A method for processing in a Neural Processing Unit (NPU) cortex of a robotic device having a system of artificial neurons, comprising:

reducing retina resolution of an artificial retina device of the robotic device, based at least in part on availability of computational resources;

compensating for the reduction in retina resolution by adjusting one or more network weights associated with the system; and

operating the artificial retina device at the reduced retina resolution and the adjusted one or more network weights.

20. The method of claim 19, wherein the reducing comprises reducing at least one of:

the number of cones, the number of Retinal Ganglion Cells (RGCs), or subsampling pixels.

21. The method of claim 20, wherein the compensation comprises increasing the weights of at least one of:

remaining cones or remaining RGCs.

22. The method of claim 19, further comprising:

dynamically adjusting the retina resolution based at least in part on at least one of a computational load or available computational resources associated with other sensors operating in the system.

23. The method of claim 19, further comprising:

increasing resolution of a subset of sensors associated with the retina device, if results obtained by the sensors associated with the retina device trigger the system to a particular operational state.

24. The method of claim 19, further comprising:

adding, in the system, one or more other sensors, if sensors associated with the retina device provide a desired result.

25. An apparatus for processing in a Neural Processing Unit (NPU) cortex of a robotic device having a system of artificial neurons, comprising:

a processing system configured to;

reduce retina resolution of an artificial retina device of the robotic device, based at least in part on availability of computational resources;

compensate for the reduction in retina resolution by adjusting one or more network weights associated with the system;

operate the artificial retina device at the reduced retina resolution and the adjusted one or more network weights; and

a memory coupled to the processing system.

33. The apparatus of claim 25, wherein the reducing comprises reducing at least one of:

the number of cones, the number of Retinal Ganglion Cells (RGCs), or subsampling pixels.

34. The apparatus of claim 33, wherein the compensation comprises increasing the weights of at least one of:

remaining cones or remaining RGCs.

35. The apparatus of claim 25, wherein the processing system is further configured to:

increase resolution of a subset of sensors associated with the retina device, if results obtained by the sensors associated with the retina device trigger the system to a particular operational state.

26. An apparatus for processing in a Neural Processing Unit (NPU) cortex of a robotic device having a system of artificial neurons, comprising:

means for reducing retina resolution of an artificial retina device of the robotic device, based at least in part on availability of computational resources;

means for compensating for the reduction in retina resolution by adjusting one or more network weights associated with the system; and

operating the artificial retina device at the reduced retina resolution and the adjusted one or more network weights.

27. A non-transitory computer-readable medium for processing in a Neural Processing Unit (NPU) cortex of a robotic device having a system of artificial neurons, the non-transitory computer-readable medium having instructions executable to:

reduce retina resolution of an artificial retina device of the robotic device, based at least in part on availability of computational resources;

compensate for the reduction in retina resolution by adjusting one or more network weights associated with the system; and

operate the artificial retina device at the reduced retina resolution and the adjusted one or more network weights.