WEIGHTED AVERAGE FEDERATED LEARNING BASED ON NEURAL NETWORK TRAINING LOSS

Abstract

A method of wireless communication by a user equipment (UE) includes computing updates to an artificial neural network as part of an epoch of a federated learning process. The updates include gradients or updated model parameters. The method also includes recording a training loss observed while training the artificial neural network at the epoch of the federated learning process. The method further includes transmitting the updates to a federated learning server that is configured to aggregate the gradients based on the training loss.

Description

Claims

1. A method of wireless communication by a user equipment (UE), comprising: computing updates to an artificial neural network as part of an epoch of a federated learning process, the updates comprising gradients or updated model parameters;recording a training loss observed while training the artificial neural network at the epoch of the federated learning process; andtransmitting the updates to a federated learning server that is configured to aggregate the gradients based on the training loss.

2. The method of claim 1, in which the updates are scaled for aggregation based on a previous value of the training loss.

3. The method of claim 1, in which the updates are scaled for aggregation based on a function of the training loss.

4. The method of claim 1, further comprising: receiving, from the federated learning server, a configuration for scaling the updates; andscaling the updates based on the training loss, prior to transmitting the updates to the federated learning server.

5. The method of claim 1, further comprising transmitting the training loss to the federated learning server during each round of the federated learning process.

6. The method of claim 1, further comprising transmitting the training loss to the federated learning server every N rounds of the federated learning process, where N is greater than or equal to two.

7. The method of claim 1, in which transmitting the updates comprises transmitting the updates on a shared uplink resource via analog communication, the method further comprising transmitting the training loss with digital transmission to the federated learning server during each round of the federated learning process in resources orthogonal to the shared uplink resource.

8. The method of claim 7, further comprising receiving a quantity of training samples for computing the updates, the quantity based on the training loss.

9. An apparatus for wireless communication by a user equipment (UE), comprising: a memory; andat least one processor coupled to the memory, the at least one processor configured: to compute updates to an artificial neural network as part of an epoch of a federated learning process, the updates comprising gradients or updated model parameters;to record a training loss observed while training the artificial neural network at the epoch of the federated learning process; andto transmit the updates to a federated learning server that is configured to aggregate the gradients based on the training loss.

10. The apparatus of claim 9, in which the updates are scaled for aggregation based on a previous value of the training loss.

11. The apparatus of claim 9, in which the updates are scaled for aggregation based on a function of the training loss.

12. The apparatus of claim 9, in which the at least one processor is further configured: to receive, from the federated learning server, a configuration for scaling the updates; andto scale the updates based on the training loss, prior to transmitting the updates to the federated learning server.

13. The apparatus of claim 9, in which the at least one processor is further configured to transmit the training loss to the federated learning server during each round of the federated learning process.

14. The apparatus of claim 9, in which the at least one processor is further configured to transmit the training loss to the federated learning server every N rounds of the federated learning process, where N is greater than or equal to two.

15. The apparatus of claim 9, in which the at least one processor is further configured: to transmit the updates on a shared uplink resource via analog communication; andto transmit the training loss with digital transmission to the federated learning server during each round of the federated learning process in resources orthogonal to the shared uplink resource.

16. The apparatus of claim 15, in which the at least one processor is further configured to receive a quantity of training samples for computing the updates, the quantity based on the training loss.

17. A non-transitory computer-readable medium having program code recorded thereon, the program code comprising: program code to compute updates to an artificial neural network as part of an epoch of a federated learning process, the updates comprising gradients or updated model parameters;program code to record a training loss observed while training the artificial neural network at the epoch of the federated learning process; andprogram code to transmit the updates to a federated learning server that is configured to aggregate the gradients based on the training loss.

18. The non-transitory computer-readable medium of claim 17, in which the updates are scaled for aggregation based on a previous value of the training loss.

19. The non-transitory computer-readable medium of claim 17, in which the updates are scaled for aggregation based on a function of the training loss.

20. The non-transitory computer-readable medium of claim 17, in which the program code further comprises: program code to receive, from the federated learning server, a configuration for scaling the updates; andprogram code to scale the updates based on the training loss, prior to transmitting the updates to the federated learning server.

21. The non-transitory computer-readable medium of claim 17, in which the program code further comprises program code to transmit the training loss to the federated learning server during each round of the federated learning process.

22. The non-transitory computer-readable medium of claim 17, in which the program code further comprises program code to transmit the training loss to the federated learning server every N rounds of the federated learning process, where N is greater than or equal to two.

23. An apparatus for wireless communication by a user equipment (UE), comprising: means for computing updates to an artificial neural network as part of an epoch of a federated learning process, the updates comprising gradients or updated model parameters;means for recording a training loss observed while training the artificial neural network at the epoch of the federated learning process; andmeans for transmitting the updates to a federated learning server that is configured to aggregate the gradients based on the training loss.

24. The apparatus of claim 23, in which the updates are scaled for aggregation based on a previous value of the training loss.

25. The apparatus of claim 23, in which the updates are scaled for aggregation based on a function of the training loss.

26. The apparatus of claim 23, further comprising: means for receiving, from the federated learning server, a configuration for scaling the updates; andmeans for scaling the updates based on the training loss, prior to transmitting the updates to the federated learning server.

27. The apparatus of claim 23, further comprising means for transmitting the training loss to the federated learning server during each round of the federated learning process.

28. The apparatus of claim 23, further comprising means for transmitting the training loss to the federated learning server every N rounds of the federated learning process, where N is greater than or equal to two.

29. The apparatus of claim 23, in which the means for transmitting the updates comprises means for transmitting the updates on a shared uplink resource via analog communication, and means for transmitting the training loss with digital transmission to the federated learning server during each round of the federated learning process in resources orthogonal to the shared uplink resource.

30. The apparatus of claim 29, further comprising means for receiving a quantity of training samples for computing the updates, the quantity based on the training loss.