Efficient cognitive signal denoising with sparse output layers

• US 10,380,062 B1
• Filed: 08/21/2018
• Issued: 08/13/2019
• Est. Priority Date: 03/19/2015
• Status: Active Grant

• Current Assignee: HRL Laboratories LLC (The Boeing Co.)
• Original Assignee: HRL Laboratories LLC (The Boeing Co.)
• Inventors: Rao, Shankar R., Petre, Peter
• Primary Examiner(s): Fan, Chieh M
• Assistant Examiner(s): Hailegiorgis, Fitwi Y
• Application Number:
  US16/108,041
• Time in Patent Office: 357 Days
• Field of Search: 375130
• CPC Class Codes: G06F 15/7817 Specially adapted for signal processing, e.g. Harvard architectures
• CPC Class Codes: G06F 18/2134 based on separation criteria, e.g. independent component analysis
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06F 18/24143 Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
• CPC Class Codes: G06F 2218/04 Denoising
• CPC Class Codes: G06F 2218/22 Source localisation; Inverse modelling
• CPC Class Codes: G06F 9/3004 to perform operations on memory
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

Described is a system for signal denoising. The system linearly maps a noisy input signal into a high-dimensional reservoir, where the noisy input signal is a time-series of data points from a mixture of waveforms. A high-dimensional state-space representation of the mixture of waveforms is created by combining the noisy input signal with reservoir states. A delay embedded state signal is generated from the reservoir states, and a denoised spectrogram of the noisy input signal is generated.

Claims

1. A system for signal denoising, the system comprising:

one or more processors and a non-transitory computer-readable medium having executable instructions encoded thereon such that when executed, the one or more processors perform operations of;

linearly mapping a noisy input signal into a high-dimensional reservoir, wherein the noisy input signal is a time-series of data points from a mixture of waveforms;

creating a high-dimensional state-space representation of the mixture of waveforms by combining the noisy input signal with reservoir states,

wherein the reservoir is a recurrent neural network having a plurality of nodes, and

wherein a connectivity matrix of the reservoir comprises a block diagonal form optimized such that computation of the reservoir states scales linearly with the number of nodes;

generating a delay embedded state signal from the reservoir states; and

generating a denoised spectrogram of the noisy input signal.

2. The system as set forth in claim 1, wherein the one or more processors further perform an operation of generating a prediction of the noisy input signal at a time-step into the future by adapting the delay embedded state signal.

3. The system as set forth in claim 1, wherein an output of the reservoir is adapted via gradient descent, producing a prediction of the input signal at a time-step in the future.

4. The system as set forth in claim 1, wherein the one or more processors further performs an operation of iteratively tuning output weights of the reservoir based on a difference between the prediction of the input signal and the noisy input signal.

5. The system as set forth in claim 1, wherein the reservoir has a predefined number of outputs, and wherein the predefined number of outputs is generated by continually mapping reservoirs states through a set of distinct linear functions, with one linear function defined for each output.

6. A computer implemented method for signal denoising, the method comprising an act of:

causing one or more processers to execute instructions encoded on a non-transitory computer-readable medium, such that upon execution, the one or more processors perform operations of;

linearly mapping a noisy input signal into a high-dimensional reservoir, wherein the noisy input signal is a time-series of data points from a mixture of waveforms;

creating a high-dimensional state-space representation of the mixture of waveforms by combining the noisy input signal with reservoir states,

wherein the reservoir is a recurrent neural network having a plurality of nodes, and

wherein a connectivity matrix of the reservoir comprises a block diagonal form optimized such that computation of the reservoir states scales linearly with the number of nodes;

generating a delay embedded state signal from the reservoir states; and

generating a denoised spectrogram of the noisy input signal.

7. The method as set forth in claim 6, wherein the one or more processors further perform an operation of generating a prediction of the noisy input signal at a time-step into the future by adapting the delay embedded state signal.

8. The method as set forth in claim 6, wherein an output of the reservoir is adapted via gradient descent, producing a prediction of the input signal at a time-step in the future.

9. The method as set forth in claim 6, wherein the one or more processors further performs an operation of iteratively tuning output weights of the reservoir based on a difference between the prediction of the input signal and the noisy input signal.

10. The method as set forth in claim 6, wherein the reservoir has a predefined number of outputs, and wherein the predefined number of outputs is generated by continually mapping reservoirs states through a set of distinct linear functions, with one linear function defined for each output.

11. A computer program product for signal denoising, the computer program product comprising:

computer-readable instructions stored on a non-transitory computer-readable medium that are executable by a computer having one or more processors for causing the processor to perform operations of;

linearly mapping a noisy input signal into a high-dimensional reservoir, wherein the noisy input signal is a time-series of data points from a mixture of waveforms;

creating a high-dimensional state-space representation of the mixture of waveforms by combining the noisy input signal with reservoir states,

wherein the reservoir is a recurrent neural network having a plurality of nodes, and

wherein a connectivity matrix of the reservoir comprises a block diagonal form optimized such that computation of the reservoir states scales linearly with the number of nodes;

generating a delay embedded state signal from the reservoir states; and

generating a denoised spectrogram of the noisy input signal.

12. The computer program product as set forth in claim 11, wherein the one or more processors further perform an operation of generating a prediction of the noisy input signal at a time-step into the future by adapting the delay embedded state signal.

13. The computer program product as set forth in claim 11, wherein an output of the reservoir is adapted via gradient descent, producing a prediction of the input signal at a time-step in the future.

14. The computer program product as set forth in claim 11, further comprising instructions for causing the one or more processors further performs an operation of iteratively tuning output weights of the reservoir based on a difference between the prediction of the input signal and the noisy input signal.

15. The computer program product as set forth in claim 11, wherein the reservoir has a predefined number of outputs, and wherein the predefined number of outputs is generated by continually mapping reservoirs states through a set of distinct linear functions, with one linear function defined for each output.