Systems and Methods for Distributed On-Device Learning with Data-Correlated Availability

Patent metadata

Field	Value
Publication number	US-2019171978-A1
Application number	US-201715834001-A
Country	US
Kind code	A1
Filing date	Dec 6, 2017
Priority date	Dec 6, 2017
Publication date	Jun 6, 2019
Grant date	—

Abstract

The present disclosure provides systems and methods for distributed training of machine learning models. In one example, a computer-implemented method is provided for training machine-learned models. The method includes obtaining, by one or more computing devices, a plurality of regions based at least in part on temporal availability of user devices; selecting a plurality of available user devices within a region; and providing a current version of a machine-learned model associated with the region to the plurality of selected user devices within the region. The method includes obtaining, from the plurality of selected user devices, updated machine-learned model data generated by the plurality of selected user devices through training of the current version of the machine-learned model associated with the region using data local to each of the plurality of selected user devices and generating an updated machine-learned model associated with the region based on the updated machine-learned model data.

First claim

What is claimed is:

1 . A computer-implemented method for training machine-learned models, the method comprising:
obtaining, by one or more computing devices, a plurality of regions based at least in part on temporal availability of user devices; selecting, by the one or more computing devices, a plurality of available user devices within a region; providing, by the one or more computing devices, a current version of a machine-learned model associated with the region to the plurality of selected user devices within the region; obtaining, by the one or more computing devices from the plurality of selected user devices, updated machine-learned model data generated by the plurality of selected user devices through training of the current version of the machine-learned model associated with the region using data local to each of the plurality of selected user devices; and generating, by the one or more computing devices, an updated machine-learned model associated with the region based on the updated machine-learned model data.

2 . The computer-implemented method of claim 1 , wherein obtaining, by one or more computing devices, a plurality of regions based at least in part on temporal availability of user devices further comprises generating the plurality of regions based at least in part on one or more of:
time zones, latitude ranges, longitude ranges, semantic boundaries, user population, or diurnal availability patterns.

3 . The computer-implemented method of claim 2 , wherein each region is generated such that each region comprises a user population having a similar diurnal cycle.

4 . The computer-implemented method of claim 1 , further comprising associating, by the one or more computing devices, a copy of a global machine-learned model with each region wherein the machine-learned model associated with the region is trained using federated learning based on users in the region.

5 . The computer-implemented method of claim 1 , further comprising:
providing, by the one or more computing devices, a regularization term to the plurality of selected user devices within the region, wherein the regularization term is added to the loss function for training of the current version of the machine-learned model associated with the region, and wherein the regularization term represents a sum of distances, measured in parameter space, of the model associated with the region from at least one model of at least one other region.

6 . The computer-implemented method of claim 1 , wherein generating, by the one or more computing devices, the updated machine-learned model associated with the region based on the updated machine-learned model data further comprises performing, by the one or more computing devices, multitask learning to bias the machine-learned model associated with the region toward at least one machine-learned model associated with at least one other region.

7 . The computer-implemented method of claim 1 , further comprising:
computing, by the one or more devices, a centroid of at least one model of at least one other region in parameter space; and providing, by the one or more computing devices, the centroid to the plurality of selected user devices within the region, wherein each of the plurality of selected user devices computes a distance, measured in parameter space, of the model associated with the region from the centroid as a regularization term that is added to the loss function for training of the current version of the machine-learned model associated with the region.

8 . A computing device comprising:
one or more processors; and one or more non-transitory computer-readable media that store instructions that, when executed by the one or more processors, cause the computing device to:
generate a plurality of regions based at least in part on temporal availability of user devices;
select a plurality of available user devices within a region;
provide a current version of a machine-learned model associated with the region to the plurality of selected user devices within the region, wherein each of the plurality of selected user devices performs training of the current version of the machine-learned model associated with the region using data local to each of the plurality of selected user devices;
obtain, from the plurality of selected user devices, updated machine-learned model data; and
generate an updated machine-learned model associated with the region based on the updated machine-learned model data.

9 . The computing device of claim 8 , wherein generating a plurality of regions based at least in part on temporal availability of user devices further comprises generating the plurality of regions based at least in part on one or more of:
time zones, latitude ranges, longitude ranges, semantic boundaries, user population, or diurnal availability patterns.

10 . The computing device of claim 9 , wherein each region is generated such that each region comprises a user population having a similar diurnal cycle.

11 . The computing device of claim 8 , further comprising instructions that, when executed by the one or more processors, cause the computing device to associate a copy of a global machine-learned model with each region wherein the machine-learned model associated with the region is trained using federated learning based on users in the region.

12 . The computing device of claim 8 , further comprising instructions that, when executed by the one or more processors, cause the computing device to:
compute a regularization term which represents a sum of distances, measured in parameter space, of the model associated with the region from at least one other model of at least one other region; and provide the regularization term to the plurality of selected user devices within the region, wherein the regularization term is added to the loss function for training of the current version of the machine-learned model associated with the region.

13 . The computing device of claim 8 , wherein generating the updated machine-learned model associated with the region based on the updated machine-learned model data further comprises performing multitask learning to bias the machine-learned model associated with the region toward at least one machine-learned model associated with at least one other region.

14 . The computing device of claim 8 , further comprising instructions that, when executed by the one or more processors, cause the computing device to:
compute a centroid of at least one model of at least one other region in parameter space; and provide the centroid to the plurality of selected user devices within the region, wherein each of the plurality of selected user devices computes a distance, measured in parameter space, of the model associated with the region from the centroid as a regularization term that is added to the loss function for training of the current version of the machine-learned model associated with the region.

15 . A system comprising:
a server; and a plurality of user devices; the server comprising:
one or more processors; and
one or more non-transitory computer-readable media that store instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, the operations comprising:
obtaining a plurality of regions based at least in part on temporal availability of user devices;
selecting, from the plurality of user devices, a plurality of available user devices within a region;
providing a current version of a machine-learned model associated with the region to the plur

Assignees

• Google Llc

Inventors

• Bonawitz Keith

Classifications

• G06N20/00 Primary

  Machine learning · CPC title

• G06N99/005 Primary

  Physics · mapped topic

• G06F16/95 Primary

  Retrieval from the web · CPC title

• G06N20/20

  Ensemble learning · CPC title