Mixed client-server federated learning of machine learning model(s)
Implementations disclosed herein are directed to federated learning of machine learning (“ML”) model(s) based on gradient(s) generated at corresponding client devices and a remote system. Processor(s) of the corresponding client devices can process client data generated locally at the corresponding client devices using corresponding on-device ML model(s) to generate corresponding predicted outputs, generate corresponding client gradients based on the corresponding predicted outputs, and transmit the corresponding client gradients to the remote system. Processor(s) of the remote system can process remote data obtained from remote database(s) using global ML model(s) to generate additional corresponding predicted outputs, generate corresponding remote gradients based on the additional corresponding predicted outputs. Further, the remote system can utilize the corresponding client gradients and the corresponding remote gradients to update the global ML model(s) or weights thereof. The updated global ML model(s) and/or the updated weights thereof can be transmitted back to the corresponding client devices.
1. A method implemented by one or more processors of a remote system, the method comprising:
receiving a plurality of client gradients from a plurality of corresponding client devices, wherein each of the plurality of client gradients is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding audio data that captures at least part of a corresponding spoken utterance of a corresponding user of the given one of the plurality of corresponding client devices;
generating a plurality of remote gradients, wherein generating each of the plurality of remote gradients comprises:
obtaining additional audio data that captures at least part of an additional spoken utterance of an additional user;
processing, using a global machine learning (ML) model stored remotely at the remote system, the additional audio data to generate predicted output; and
generating an additional gradient, for inclusion in the plurality of remote gradients, based on comparing the additional predicted output to ground truth output corresponding to the additional audio data;
selecting a set of client gradients from among the plurality of client gradients;
selecting an additional set of remote gradients from among the plurality of remote gradients; and
utilizing the set of client gradients and the additional set of remote gradients to update weights of the global ML model.
2. The method of claim 1 , wherein utilizing the set of client gradients and the additional set of remote gradients to update the weights of the global ML model comprises:
utilizing the set of client gradients to update the weights of the global ML model; and
subsequent to utilizing the set of client gradients to update the weights of the global ML model:
utilizing the additional set of remote gradients to further update the weights of the global ML model.
3. The method of claim 1 , wherein utilizing the set of client gradients and the additional set of remote gradients to update the weights of the global ML model comprises:
utilizing the additional set of remote gradients to update the weights of the global ML model; and
subsequent to utilizing the additional set of remote gradients to update the weights of the global ML model:
utilizing the set of client gradients to further update the weights of the global ML model.
4. The method of claim 1 , wherein utilizing the set of client gradients and the additional set of remote gradients to update the weights of the global ML model comprises:
utilizing the set of client gradients to update first weights of a first instance the global ML model;
utilizing, in parallel, the additional set of remote gradients to update second weights of a second instance of the global ML model; and
utilizing the updated first weights of the first instance of the global ML model and the updated second weights of the second instance of the global ML model to update the weights of the global ML model.
5. The method of claim 4 , wherein utilizing the updated first weights of the first instance of the global ML model and the updated second weights of the second instance of the global ML model to update the weights of the global ML model comprises:
determining an average of the updated first weights of the first instance of the global ML model and the updated second weights of the second instance of the global ML model; and
utilizing the average of the updated first weights and the updated second weights to update the weights of the global ML model.
6. The method of claim 1 , wherein the additional set of remote gradients includes a threshold quantity of additional gradients, and wherein the threshold quantity of additional gradients is based on a quantity of the plurality of client gradients received from the plurality of corresponding client devices.
7. The method of claim 1 , wherein selecting the set of client gradients comprises:
selecting one or more of the plurality of client gradients for inclusion in the set of client gradients based on the one or more of the plurality of client gradients being a particular type of gradient.
8. The method of claim 7 , wherein selecting the additional set of remote gradients comprises:
selecting one or more of the plurality of remote gradients for inclusion in the additional set of remote gradients based on the one or more of the plurality of remote gradients being the particular type of gradient.
9. The method of claim 1 , wherein selecting the set of client gradients comprises:
selecting one or more of the plurality of client gradients for inclusion in the set of client gradients based on the one or more of the plurality of client gradients being a particular type of gradient.
10. The method of claim 9 , wherein selecting the additional set of remote gradients comprises:
selecting one or more of the plurality of remote gradients for inclusion in the additional set of remote gradients based on the one or more of the plurality of remote gradients being the particular type of gradient.
11. The method of claim 1 , further comprising:
analyzing a distribution of remote gradients stored in one or more databases that are accessible by the remote system to identify a particular type of gradient, and
wherein the plurality of remote gradients that are generated conform to the particular type of gradient.
12. The method of claim 11 , wherein identifying the particular type of gradient comprises:
determining, based on analyzing the distribution of remote gradients, that the distribution of remote gradients does not include a threshold quantity of remote gradients of the particular type of gradient.
13. The method of claim 1 , further comprising:
transmitting, to each of the plurality of corresponding client devices, the updated global ML model or the updated global weights of the global ML model, wherein transmitting the global ML model or the updated global weights is subsequent to the remote system updating the global weights of the global ML model based on the set of client gradients and the additional set of remote gradients; and
wherein transmitting the updated global ML model or the updated global weights to each of the plurality of corresponding client devices causes each of the plurality of corresponding client devices to replace, in corresponding local storage, an on-device ML model with the updated global ML model or local weights of the on-device ML model with the updated global weights of the global ML model.
14. The method of claim 1 , wherein the ground truth output corresponding to the additional audio data is based on the additional spoken utterance captured in the additional audio data.
15. The method of claim 14 , wherein the additional spoken utterance captured in the additional audio data includes a particular word or phrase, and wherein the ground truth output corresponding to the additional audio data includes an indication of whether the additional spoken utterance captures the particular word or phrase.
16. The method of claim 14 , wherein the additional spoken utterance captured in the additional audio data includes one or more terms, and wherein the ground truth output corresponding to the additional audio data includes a transcription of the one or more terms.
17. A method comprising:
by one or more client processors of a client device:
detecting, via one or more microphones of the client device, audio data that captures at least part of a spoken utterance of a user of the client device;
processing, using an on-device machine learning (ML) model stored locally on the client device, the audio data to generate predicted output;
generating a client gradient based on the predicted output; and
transmitting, to a remote system and from the client device, the client gradient;
by one or more remote processors of the remote system:
obtaining additional audio data that captures at least part of an additional spoken utterance of an additional user;
processing, using a global ML model stored remotely at the remote system, the additional audio data to generate additional predicted output;
generating a remote gradient based on the additional predicted output; and
utilizing the client and the remote gradient to update weights of the global ML model.
18. The method of claim 17 , further comprising:
by one or more remote processors of the remote system:
transmitting, to the client device, the updated global ML model or the updated global weights of the global ML model, wherein transmitting the global ML model or the updated global weights is subsequent to the remote system updating the global weights of the global ML model based on the client gradient and the remote gradient.
19. The method of claim 18 , further comprising:
by one or more of the processors of the client device:
receiving, from the remote system, the updated global ML model or the updated global weights of the global ML model; and
causing the client device to replace, in local storage of the client device, the on-device ML model with the updated global ML model or local weights of the on-device ML model with the updated global weights of the global ML model.
20. A remote system comprising:
one or more hardware processors; and
memory storing instructions that, when executed by the one or more hardware processors, cause the one or more hardware processor to:
receive a plurality of client gradients from a plurality of corresponding client devices, wherein each of the plurality of client gradients is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding audio data that captures at least part of a corresponding spoken utterance of a corresponding user of the given one of the plurality of corresponding client devices;
generate a plurality of remote gradients, wherein the instructions to generate each of the plurality of remote gradients comprise instructions to:
obtain additional audio data that captures at least part of an additional spoken utterance of an additional user;
process, using a global machine learning (ML) model stored remotely at the remote system, the additional audio data to generate predicted output; and
generate an additional gradient, for inclusion in the plurality of remote gradients, based on comparing the additional predicted output to ground truth output corresponding to the additional audio data;
select a set of client gradients from among the plurality of client gradients;
select an additional set of remote gradients from among the plurality of remote gradients; and
utilize the set of client gradients and the additional set of remote gradients to update weights of the global ML model.