IP Library › Granted Patent US 12,271,806
Granted Patent B2
US 12,271,806 · App. 16/678,474 · Granted Apr 8, 2025

Artificial neural network training

Inventor: John E. Mixter (Tucson, AZ)
Assignee: Raytheon Company
G06N3/047G06F17/18G06N3/084
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,271,806
App. No.
16/678,474
Granted
Apr 8, 2025
Kind
B2
Abstract

An artificial neural network receives data for the inputs of a perceptron in the artificial neural network. The network determines an average of the data for each of the inputs of the perceptron, determines a standard deviation of the average for each of the inputs of the perceptron, and determines an average of the standard deviations for the perceptron. The network then sets a learning rate for the perceptron equal to the average of the standard deviations, and trains the artificial neural network using the learning rate for the perceptron.

Claims (42)

1. A process for training an artificial neural network comprising:

(a) receiving input data for inputs of a perceptron in the artificial neural network;

(b) determining an average of the input data for each of the inputs of the perceptron;

(c) determining a standard deviation of the average for each of the inputs of the perceptron;

(d) determining an average of the standard deviations for the perceptron;

(e) setting a learning rate for the perceptron equal to the average of the standard deviations; and

(f) training the artificial neural network using the learning rate for the perceptron.

2. The process of claim 1 , comprising setting an initial weight for perceptrons in the artificial neural network before executing operations (a)-(e).

3. The process of claim 2 , wherein the initial weight is set to the same value for all the perceptrons in the artificial neural network.

4. The process of claim 3 , wherein the same value is equal to a maximum value that does not cause any perceptron in the network to saturate its activation function.

5. The process of claim 1 , comprising randomizing a plurality of weights prior to training the artificial neural network after the perceptron learning rates have been calculated.

6. The process of claim 1 , comprising executing operations (a)-(e) for each perceptron in each layer of the artificial neural network, thereby generating an individualized learning rate for each perceptron in the artificial neural network; and

training the artificial neural network using the individualized learning rate determined for each perceptron.

7. The process of claim 6 , wherein the same input data are used for each layer of the artificial neural network.

8. A non-transitory computer readable medium comprising instructions that when executed by a processor executes a process comprising:

(a) receiving input data for inputs of a perceptron in an artificial neural network;

(b) determining an average of the input data for each of the inputs of the perceptron;

(c) determining a standard deviation of the average for each of the inputs of the perceptron;

(d) determining an average of the standard deviations for the perceptron;

(e) setting a learning rate for the perceptron equal to the average of the standard deviations; and

(f) training the artificial neural network using the learning rate for the perceptron.

9. The non-transitory computer readable medium of claim 8 , comprising instructions for setting an initial weight for perceptrons in the artificial neural network before executing operations (a)-(e).

10. The non-transitory computer readable medium of claim 9 , wherein the initial weight is set to the same value for all the perceptrons in the artificial neural network.

11. The non-transitory computer readable medium of claim 10 , wherein the same value is equal to a maximum value that does not cause any perceptron in the network to saturate its activation function.

12. The non-transitory computer readable medium of claim 8 , comprising instructions for randomizing a plurality of weights prior to training the artificial neural network after the perceptron learning rates have been calculated.

13. The non-transitory computer readable medium of claim 8 , comprising instructions for executing operations (a)-(e) for each perceptron in each layer of the artificial neural network, thereby generating an individualized learning rate for each perceptron in the artificial neural network; and training the artificial neural network using the individualized learning rate determined for each perceptron.

14. The non-transitory computer readable medium of claim 13 , wherein the same input data are used for each layer of the artificial neural network.

15. A system comprising:

a computer processor; and

a computer memory coupled to the computer processor;

wherein the computer processor is operable for:

(a) receiving input data for inputs of a perceptron in an artificial neural network;

(b) determining an average of the input data for each of the inputs of the perceptron;

(c) determining a standard deviation of the average for each of the inputs of the perceptron;

(d) determining an average of the standard deviations for the perceptron;

(e) setting a learning rate for the perceptron equal to the average of the standard deviations; and

(f) training the artificial neural network using the learning rate for the perceptron.

16. The system of claim 15 , wherein the computer processor is operable for setting an initial weight for perceptrons in the artificial neural network before executing operations (a)-(e).

17. The system of claim 16 , wherein the initial weight is set to the same value for all the perceptrons in the artificial neural network; and wherein the same value is equal to a maximum value that does not cause any perceptron in the network to saturate its activation function.

18. The system of claim 15 , wherein the computer processor is operable for randomizing a plurality of weights prior to training the artificial neural network after the perceptron learning rates have been calculated.

19. The system of claim 15 , wherein the computer processor is operable for executing operations (a)-(e) for each perceptron in each layer of the artificial neural network, thereby generating an individualized learning rate for each perceptron in the artificial neural network; and training the artificial neural network using the individualized learning rate determined for each perceptron.

20. The system of claim 19 , wherein the same input data is used for each layer of the artificial neural network.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Dec 3, 2019
From: MIXTER, JOHN E.
To: RAYTHEON COMPANY
Reel/Frame 051158/0878 →
Continuity (1)
Related Publication 20210142151A1 · May 13, 2021
References Cited (13)
US 5161204A · Hutcheson · 1992 [cited by examiner]
US 5479576A · Watanabe · 1995 [cited by examiner]
US 20120303562A1 · Paguio · 2012 [cited by examiner]
US 20190159735A1 · Rundo · 2019 [cited by examiner]
WO WO2021091614A1 · 2021 [cited by applicant]
“International Application Serial No. PCT US2020 049674, International Search Report mailed Dec. 11, 2020”, 3 pgs. [cited by applicant]
“International Application Serial No. PCT US2020 049674, Written Opinion mailed Dec. 11, 2020”, 8 pgs. [cited by applicant]
Dhaval, Dholakia, “Activation Functions”, (Apr. 27, 2018), 1-9. [cited by applicant]
Nicol, Schraudolpk N, “Tempering Backpropagation Networks: Not All Weights are Created Equal”, Advances In Neural Information Processing vol. 8, (Nov. 27, 1995), 563-569. [cited by applicant]
“International Application Serial No. PCT/US2020/049674, International Preliminary Report on Patentability mailed May 19, 2022”, 10 pgs. [cited by applicant]
“European Application Serial No. 20775778.2, Summons to Attend Oral Proceedings mailed May 16, 2024”, 9 pgs. [cited by applicant]
“European Application Serial No. 20775778.2, Communication Pursuant to Article 94(3) EPC mailed Aug. 21, 2023”, 9 pgs. [cited by applicant]
“European Application Serial No. 20775778.2, Response filed Oct. 7, 2024 to Summons to Attend Oral Proceedings mailed May 16, 2024”, 4 pgs. [cited by applicant]