Zone-based federated learning

1. A method for training models performed by a user equipment (UE), comprising:

receiving, at the UE associated with a group of UEs of a plurality of groups of UEs, a global model from a zone server associated with a first zone model of a plurality of zone models, the global model associated with a global server, the group of UEs associated with the first zone model, and each group of UEs associated with a respective zone model of the plurality of zone models;

identifying, at the UE, a device for training the global model based on one or both of a current connectivity state of the UE or a current resource use of the UE, the device being the UE based on the current connectivity state of the UE satisfying a connectivity condition and the current resource use satisfying a resource condition, the device being a proxy device based on one or both of the current connectivity state of the UE failing to satisfy the connectivity condition or the current resource use failing to satisfy the resource condition;

transmitting, to the zone server, model updates based on identifying the device; and

receiving, from the zone server, the first zone model based on the transmitted model updates.

2. The method of claim 1 further comprising:

training the model at the UE based on local training data collected at the UE based on identifying the UE; and

generating the model weight updates based on the training.

3. The method of claim 2 , further comprising maintaining a local training state of the model across training iterations of the global model.

4. The method of claim 2 , further comprising:

receiving a set of models at the UE, the global model being one model of the set of models; and

training the global model from of the set of models based on a training priority of the global model being higher than a training priority associated with each other model of the set of models.

5. The method of claim 1 , further comprising:

transmitting the global model and training data collected at the UE to the proxy device for training the global model at the proxy device using the training data collected at the first UE; and

receiving the model updates from the proxy device based on the training performed at the proxy device.

6. The method of claim 5 , in which the proxy device and the UE are intra-network devices or inter-network devices.

7. The method of claim 1 , in which the global model is an untrained model.

8. The method of claim 1 , in which the global model comprises weight updates from one or more zone models of the plurality of zone models.

9. The method of claim 1 , in which:

the global model is a second zone model generated at a second zone server associated with the second zone model;

the global server is the second zone server; and

one or more attributes of the first zone model and the second zone model satisfy a similarity condition.

10. The method of claim 1 , in which the first UE and each UE in the first group of UEs share one or more common attributes.

11. The method of claim 10 , in which the one or more common attributes comprise one or more of a geographic location, a default language, or a user interface theme.

12. A system for training models, the system comprising a user equipment (UE) and a zone server:

the UE comprising:

a processor;

a memory coupled with the processor; and

instructions stored in the memory and operable, when executed by the processor, to cause the UE:

to receive, at the UE associated with a group of UEs of a plurality of groups of UEs, a global model from the zone server associated with a first zone model of a plurality of zone models, the global model associated with a global server, the group of UEs associated with the first zone model, and each group of UEs associated with a respective zone model of the plurality of zone models;

to identify, at the UE, a device for training the global model based on one or both of a current connectivity state of the UE or a current resource use of the UE, the device being the UE based on the current connectivity state of the UE satisfying a connectivity condition and the current resource use satisfying a resource condition, the device being a proxy device based on one or both of the current connectivity state of the UE failing to satisfy the connectivity condition or the current resource use failing to satisfy the resource condition;

to transmit, to the zone server, model updates based on identifying the device; and

to receive, from the zone server, the first zone model based on the transmitted model updates.

13. The system of claim 12 , in which execution of the instructions further cause the UE:

to train the global model at the UE based on local training data collected at the UE based on identifying the UE; and

to generate the model updates based on the training.

14. The system of claim 13 , in which execution of the instructions further cause the UE:

to receive a set of models at the first UE, the global model being one model of the set of models; and

to train the global model from of the set of models based on a training priority of the global model being higher than a training priority associated with each other model of the set of models.

15. The system of claim 12 , in which execution of the instructions further cause the UE:

to transmit the global model and training data collected at the UE to the proxy device for training the global model at the proxy device using the training data collected at the first UE; and

to receive the model updates from the proxy device based on the training performed at the proxy device.