Asynchronous parameter aggregation for machine learning

Patent metadata

Field	Value
Publication number	US-11373115-B2
Application number	US-201815948493-A
Country	US
Kind code	B2
Filing date	Apr 9, 2018
Priority date	Apr 9, 2018
Publication date	Jun 28, 2022
Grant date	Jun 28, 2022

Abstract

Systems and methods are provided for training a machine learned model on a large number of devices, each device acquiring a local set of training data without sharing data sets across devices. The devices train the model on the respective device's set of training data. The devices communicate a parameter vector from the trained model asynchronously with a parameter server. The parameter server updates a master parameter vector and transmits the master parameter vector to the respective device.

First claim

The invention claimed is:

1. A navigation device for training a machine learned model, the navigation device comprising
at least one sensor configured to generate training data;
a communication interface configured to communicate with a parameter server; and
a device processor configured to train the machine learned model using the training data; the device processor further configured to transmit a parameter vector of the trained machine learned model to the parameter server and receive in response, an updated central parameter vector from the parameter server; the device processor further configured to retrain the machine learned model using the updated central parameter vector;
wherein the training data is different than training data that is generated and used by other navigation devices that are training the machine learned model, wherein the training data is not available centrally for use in training the machine learned model;
wherein at least one transmission between the navigation device and the parameter server occur asynchronously with respect to the other navigation devices that are training the machine learned model.

2. The navigation device of claim 1 , wherein the training data is image data, and the machine learned model is trained to identify a position of the navigation device.

3. The navigation device of claim 1 , wherein the training data is image data, and the machine learned model is trained to identify an object in the image data.

4. The navigation device of claim 1 , wherein training the machine learned model includes a gradient descent-based process.

5. The navigation device of claim 1 , wherein the at least one sensor is embedded in a vehicle.

6. The navigation device of claim 1 , wherein the machine learned model comprises a generative adversarial network, wherein the device processor is configured to train the machine learned model using an adversarial training process.

7. The navigation device of claim 1 , wherein the navigation device generates different amounts of training data from the other navigation devices.

8. The navigation device of claim 1 , wherein the updated central parameter vector is updated by the parameter server as a function of a weighting value.

9. The navigation device of claim 1 , wherein the parameter vector comprises a randomly selected subset of parameters of the trained machine learned model.

10. The navigation device of claim 1 , wherein the training data is labeled, and the machine learned model is trained using a supervised training process.

11. A method for training a machine learned model using a plurality of distributed worker devices, the method comprising:
training, by a worker device, a machine learned model using locally generated training data and a set of first parameters, wherein the locally generated training data is generated by and remains on the worker device and is not available centrally for training the machine learned model;
transmitting, by the worker device, a set of second parameters of the trained machine learned model to a parameter server;
receiving, by the worker device, a set of third parameters from the parameter server, wherein the set of third parameters is calculated at least partially as a function of the set of second parameters; and
training, by the worker device the machine learned model using the locally generated training data and the set of third parameters.

12. The method of claim 11 , further comprising:
transmitting, by the work device, a set of fourth parameters of the trained machine learned model to the parameter server.

13. The method of claim 11 , wherein the set of third parameters are received asynchronously from the parameter server.

14. The method of claim 11 , wherein the set of first parameters comprises a randomly assigned set of values.

15. The method of claim 11 , wherein the local generated training data is image data and the machine learned model is an image recognition machine learned model.

16. The method of claim 15 , further comprising:
generating, by the worker device, new image data; and
identifying, by the worker device, an object in the new image data using the trained machine learned model.

17. The method of claim 11 , further comprising:
determining a randomly chosen subset of parameters of the set of second parameters; wherein the randomly chosen subset of parameters is transmitted to the parameter server in place of the set of second parameters.

18. A system for training a machine learned model, the system comprising:
a parameter server configured to communicate with a plurality of worker devices; the parameter server configured to receive locally generated sets of parameters of the trained machine learned model from the plurality of worker devices; the parameter server configured to calculate and transmit, in response to a communication from a worker device of the plurality of worker devices, a set of central parameters to a respective worker device from which the communication originated;
wherein the plurality of worker devices are configured to train the machine learned model using the set of central parameters and respective sets of locally generated training data, wherein the respective sets of locally generated training data remain on respective worker devices of the plurality of worker devices and is not shared centrally for use in training the machine learned model.

19. The system of claim 18 , further comprising
a master parameter server configured to communicate with a plurality of parameter servers; the master parameter server configured to receive central parameters from the plurality of parameter servers; the master parameter server configured to calculate and transmit, in response to a communication from the parameter servers of the plurality of parameter servers, a set of global central parameters to a respective parameter server from which the communication originated.

20. The system of claim 19 , wherein the master parameter server is configured to communicate with both the plurality of parameter servers and the plurality of worker devices.

21. The system of claim 18 , wherein the parameter server is configured to communicate asynchronously with the plurality of worker devices.

Assignees

• Here Global Bv

Inventors

• Kopp Michael
• Neun Moritz
• Sprague Michael
• Jalalirad Amir
• Scavuzzo Marco
• Capota Catalin

Classifications

• G01C21/3602 Primary

  Input other than that of destination using image analysis, e.g. detection of road signs, lanes, buildings, real preceding vehicles using a camera · CPC title

• G06V30/19167

  Active pattern learning · CPC title

• G06V30/19147

  Obtaining sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06N20/00 Primary

  Machine learning · CPC title

• G06V20/56

  exterior to a vehicle by using sensors mounted on the vehicle · CPC title