DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
This action is responsive to the Application filed on November 2, 2023. Claims 1-20 are pending in the case. Claims 1, 17, and 19 are the independent claims.
This action is non-final.
Claim Rejections – 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries set forth in Graham v. John Deere Co., 383 U.S. 1, 148 USPQ 459 (1966), that are applied for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims under pre-AIA 35 U.S.C. 103(a), the examiner presumes that the subject matter of the various claims was commonly owned at the time any inventions covered therein were made absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and invention dates of each claim that was not commonly owned at the time a later invention was made in order for the examiner to consider the applicability of pre-AIA 35 U.S.C. 103(c) and potential pre-AIA 35 U.S.C. 102€, (f) or (g) prior art under pre-AIA 35 U.S.C. 103(a).
Claims 1, 4-8, 11, 12, 15, and 17 are rejected under 35 U.S.C. 103 as being unpatentable over K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman, V. Ivanov, C. Kiddon, J. Konecny, S. Mazzocchi, H. McMahan, T. Van Overveldt, D. Perou, D. Ramage, J. Roselander. Towards Federated Learning at Scale: System Design. Proc. 2nd SysML Conference. 2019. Retrieved from the Internet: https://proceedings.mlsys.org/paper_files/paper/2019/file/7b770da633baf74895be22a8807f1a8f-Paper.pdf. (hereinafter Bonawitz). in view of Rehman, Y.A.U., Gao, Y., Shen, J., de Gusmão, P.P.B., Lane, N. (2022). Federated Self-supervised Learning for Video Understanding. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds) Computer Vision – ECCV 2022. ECCV 2022. Lecture Notes in Computer Science, vol 13691. Springer, Cham. https://doi.org/10.1007/978-3-031-19821-2_29. Retrieved from the Internet: https://link.springer.com/chapter/10.1007/978-3-031-19821-2_29. (hereinafter Rehman).
With respect to claim 1, Bonawitz teaches a method implemented by one or more remote processors of a remote system, the method comprising:
initiating a given round of decentralized learning of a global machine learning (ML) model (e.g. second page, Fig. 1 and section 2.1, first paragraph, indicating that a given round of federated learning begins/is initiated with a selection step in which devices check in with the FL server, announcing that they are ready to run an FL task);
during the given round of decentralized learning of the global ML model:
receiving, from a plurality of corresponding client devices, a plurality of client updates for the global ML model, wherein each of the plurality of client updates is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding client data using a corresponding on-device ML model that are corresponding on-device counterparts of the global ML model (e.g. second page, Fig. 1, step 4, performing on-device training, model update is reported back; second page, section 2.1, third paragraph, each participant performs local computation based on the global state (i.e. FL checkpoint for that round) and its local dataset, and sends an update in the form of an FL checkpoint back to the server);
identifying a checkpoint version of the global ML model that is stored remotely at the remote system (e.g. second page, Fig. 1, step 2, server reads model checkpoint from persistent storage; i.e. for a given round of federated learning, a current checkpoint version of the global model is read from the server for distribution to client devices and for subsequent updating);
generating, based on the checkpoint version of the global ML model and based on the plurality of client updates received from the plurality of corresponding client devices, a decentralized version of the global ML model (e.g. second page, Fig. 1, step 5, server aggregates updates into the global model as they arrive; second page, section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, as updates are received, the server aggregates them using Federated Averaging; i.e. as each update arrives from a client device, the server integrates/aggregates the received update into the current global model checkpoint, resulting in an updated/decentralized version of the global ML model which is based on the checkpoint version (i.e. the global model checkpoint utilized for the given round) and based on the client updates which have been received up to that point; this appears to be an intermediate version of the global model which exists as the updates are incorporated and prior to the completion of the round of federated learning); and
generating, based on the checkpoint version of the global ML model and based on the decentralized version of the global ML model, a new version of the global ML model (e.g. second page, Fig. 1, step 6, server writes global model checkpoint into persistent storage; section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, if enough devices report in time, the round will be successfully completed and the server will update its global model; i.e. once the round of federated learning is completed, the final version of the updated global model, which is based on both the original global model checkpoint for the round and the latest version incorporating/aggregating received updates during the round, is generated and stored as the new/updated global model checkpoint to be utilized in subsequent rounds of federated learning); and
in response to determining one or more conditions are satisfied: causing the new version of the global ML model or an additional averaged version of the global ML model to be deployed as the global ML model (e.g. second page, section 2.1, third paragraph, once a round is established, the server next sends to each participant the current global model parameters and any other necessary state as an FL checkpoint; process repeats; i.e. in response to determining that a next/subsequent round has been established, the new version of the global model checkpoint resulting from the previous round of federated learning is deployed as the global model for the next/subsequent round of federated learning).
Bonawitz does not explicitly disclose that the new version of the global ML model is an averaged version of the global ML model. However, Rehman teaches generating, based on the checkpoint version of the ML model and based on an updated version of the ML model, an averaged version of the ML model (e.g. page 512, first through fourth paragraphs describing FedVSSL Aggregation, indicating that, in a federated learning process, FL aggregation strategies are integrated based on weighted averaging under a common framework and then knowledge from past global models is induced while performing model parameter aggregation; performing averaging over global models, where Equation 2 represents stochastic weight averaging of the global models, and the simple averaging of the multiple checkpoints of training models obtains better generalization than conventional training; the process also includes use of different/distinct aggregation strategies for client models; page 513, Algorithm 1 showing the process in context, i.e. where local training occurs (step 4) and corresponding local/client aggregation methods are employed (steps 6-8) resulting in updated global model weights (step 9; analogous to a decentralized version of the global ML model as claimed); subsequent to this aggregation of client/local data to generate a new global model/checkpoint, averaging of the plurality of global models is performed (step 10, showing computation of Equation 2, previously described as stochastic weighted averaging of the global models), where this averaging of the global models (using at least the most recent updated global model and at least one previous global model/checkpoint) is analogous to generating an averaged version of the global model based on the checkpoint version of the global model and the decentralized version of the global model, as claimed).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz and Rehman in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), to incorporate the teachings of Rehman (directed to federated self-supervised learning for video understanding) to include the capability to generate an averaged version of the global model based on both the decentralized version of the global model (i.e. an updated version of the global model including aggregated client/local updates) and the checkpoint version of the global model (i.e. at least one previous checkpoint of the global model, such as the checkpoint utilized as the basis for the federated learning round). One of ordinary skill would have been motivated to perform such a modification in order to induce knowledge from past global models while performing model parameter aggregation, obtaining better generalization than conventional training, as described in Rehman (page 512, first through fourth paragraphs).
With respect to claim 17, Bonawitz teaches a method implemented by one or more remote processors of a remote system, the method comprising:
initiating N rounds of decentralized learning of a global machine learning (ML) model, wherein N is a positive integer greater than one (e.g. second page, Fig. 1, showing that a federated learning session consists of multiple (i.e. more than one) rounds, and is initiated with a selection step in a first/initial round in which devices check in with the FL server);
during a given round of decentralized learning of the global ML model of the N rounds of decentralized learning of the global ML model (e.g. second page, Fig. 1 and section 2.1, first paragraph, indicating various steps occurring within a given round of federated learning using a global model):
receiving a plurality of client updates from a plurality of corresponding client devices, wherein each of the plurality of client updates is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding client data using a corresponding on-device ML model that are corresponding on-device counterparts of the global ML model (e.g. second page, Fig. 1, step 4, performing on-device training, model update is reported back; second page, section 2.1, third paragraph, each participant performs local computation based on the global state (i.e. FL checkpoint for that round) and its local dataset, and sends an update in the form of an FL checkpoint back to the server);
identifying a checkpoint version of the global ML model that is stored remotely at the remote system (e.g. second page, Fig. 1, step 2, server reads model checkpoint from persistent storage; i.e. for a given round of federated learning, a current checkpoint version of the global model is read from the server for distribution to client devices and for subsequent updating); and
updating, based on the checkpoint version of the global ML model and based on the plurality of client updates received from the plurality of corresponding client devices, a decentralized version of the global ML model (e.g. second page, Fig. 1, step 5, server aggregates updates into the global model as they arrive; second page, section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, as updates are received, the server aggregates them using Federated Averaging; i.e. as each update arrives from a client device, the server integrates/aggregates the received update into the current global model checkpoint, resulting in an updated/decentralized version of the global ML model which is based on the checkpoint version (i.e. the global model checkpoint utilized for the given round) and based on the client updates which have been received up to that point; this appears to be an intermediate version of the global model which exists as the updates are incorporated and prior to the completion of the round of federated learning);
subsequent to the N rounds of decentralized learning of the global ML model (i.e. following completion of a final round of the N rounds of federated learning as shown in Fig. 1, such as performing step 6 at the end of the final round):
generating, based on the checkpoint version of the global ML model and based on the decentralized version of the global ML model, a new version of the global ML model (e.g. second page, Fig. 1, step 6, server writes global model checkpoint into persistent storage; section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, if enough devices report in time, the round will be successfully completed and the server will update its global model; i.e. once the round, such as a final round, of federated learning is completed, the final version of the updated global model, which is based on both the original global model checkpoint for the round and the latest version incorporating/aggregating received updates during the round, is generated and stored as the new/updated global model checkpoint to be utilized a later federated learning session, or as the global model for prediction); and
in response to determining the one or more conditions are satisfied: causing the new version of the global ML model or an additional new version of the global ML model to be deployed as the global ML model (e.g. second page, section 2.1, third paragraph, once a round is established, the server next sends to each participant the current global model parameters and any other necessary state as an FL checkpoint; process repeats; i.e. in response to determining that a new round has been established, such as a new round in a new federated learning session, the new version of the global model checkpoint resulting from the previous round of federated learning is deployed as the global model for the new round of federated learning).
Bonawitz does not explicitly disclose that the new version of the global ML model is an averaged version of the global ML model. However, Rehman teaches generating, based on the checkpoint version of the ML model and based on an updated version of the ML model, an averaged version of the ML model (e.g. page 512, first through fourth paragraphs describing FedVSSL Aggregation, indicating that, in a federated learning process, FL aggregation strategies are integrated based on weighted averaging under a common framework and then knowledge from past global models is induced while performing model parameter aggregation; performing averaging over global models, where Equation 2 represents stochastic weight averaging of the global models, and the simple averaging of the multiple checkpoints of training models obtains better generalization than conventional training; the process also includes use of different/distinct aggregation strategies for client models; page 513, Algorithm 1 showing the process in context, i.e. where local training occurs (step 4) and corresponding local/client aggregation methods are employed (steps 6-8) resulting in updated global model weights (step 9; analogous to a decentralized version of the global ML model as claimed); subsequent to this aggregation of client/local data to generate a new global model/checkpoint, averaging of the plurality of global models is performed (step 10, showing computation of Equation 2, previously described as stochastic weighted averaging of the global models), where this averaging of the global models (using at least the most recent updated global model and at least one previous global model/checkpoint) is analogous to generating an averaged version of the global model based on the checkpoint version of the global model and the decentralized version of the global model, as claimed).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz and Rehman in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), to incorporate the teachings of Rehman (directed to federated self-supervised learning for video understanding) to include the capability to generate an averaged version of the global model based on both the decentralized version of the global model (i.e. an updated version of the global model including aggregated client/local updates) and the checkpoint version of the global model (i.e. at least one previous checkpoint of the global model, such as the checkpoint utilized as the basis for the federated learning round). One of ordinary skill would have been motivated to perform such a modification in order to induce knowledge from past global models while performing model parameter aggregation, obtaining better generalization than conventional training, as described in Rehman (page 512, first through fourth paragraphs).
With respect to claim 4, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed, and Bonawitz further teaches wherein generating the decentralized version of the global ML model based on the checkpoint version of the global ML model and based on the plurality of client updates received from the plurality of corresponding client devices comprises: updating, based on the plurality of client updates received from the plurality of corresponding client devices, the checkpoint version of the global ML model to generate the decentralized version of the global ML model for the given round of decentralized learning; and storing, in remote memory of the remote system, the decentralized version of the global ML model (e.g. second page, Fig. 1, step 5, server aggregates updates into the global model as they arrive; second page, section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, as updates are received, the server aggregates them using Federated Averaging; i.e. as each update arrives from a client device, the server integrates/aggregates the received update into the current global model checkpoint, resulting in an updated/decentralized version of the global ML model which is based on the checkpoint version (i.e. the global model checkpoint utilized for the given round) and based on the client updates which have been received up to that point; this appears to be an intermediate version of the global model which exists as the updates are incorporated and prior to the completion of the round of federated learning).
With respect to claim 5, Bonawitz in view of Rehman teaches all of the limitations of claim 4 as previously discussed, and Bonawitz and Rehman further teaches wherein the plurality of client updates correspond to a plurality of client gradients generated locally at the corresponding client devices, and wherein updating the checkpoint version of the global ML model to generate the decentralized version of the global ML model for the given round of decentralized learning comprises: updating, based on the plurality of client gradients, the checkpoint version of the global ML model to generate the decentralized version of the global ML model for the given round of decentralized learning (e.g. Bonawitz second page, Fig. 1, step 5, server aggregates updates into the global model as they arrive; second page, section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, as updates are received, the server aggregates them using Federated Averaging; Rehman page 512, first through fourth paragraphs, discussing collection of participating client’s local gradients, computing overall local gradients from clients, etc. in order to aggregate the gradients with the global model to produce the updated global model for the training round; i.e. both Bonawitz and Rehman teach utilizing a plurality of client updates, generated locally at client devices, to update the checkpoint version of the global ML model for a given learning round, where Rehman specifically teaches that the client updates include local client gradients).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz and Rehman in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), to incorporate the teachings of Rehman (directed to federated self-supervised learning for video understanding) to include the capability to generate an averaged version of the global model based on both the decentralized version of the global model (i.e. an updated version of the global model including aggregated client/local updates, including gradients) and the checkpoint version of the global model (i.e. at least one previous checkpoint of the global model, such as the checkpoint utilized as the basis for the federated learning round). One of ordinary skill would have been motivated to perform such a modification in order to induce knowledge from past global models while performing model parameter aggregation, obtaining better generalization than conventional training, as described in Rehman (page 512, first through fourth paragraphs).
With respect to claim 6, Bonawitz in view of Rehman teaches all of the limitations of claim 4 as previously discussed, and Bonawitz further teaches wherein the plurality of client updates correspond to a plurality of client weights determined locally at the corresponding client devices, and wherein updating the checkpoint version of the global ML model to generate the decentralized version of the global ML model for the given round of decentralized learning comprises: replacing, based on the plurality of client weights, global weights of the checkpoint version of the global ML model to generate the decentralized version of the global ML model for the given round of decentralized learning (e.g. second page, Fig. 1, step 5, server aggregates updates into the global model as they arrive; second page, section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, as updates are received, the server aggregates them using Federated Averaging; i.e. as each update arrives from a client device, the server integrates/aggregates the received update into the current global model checkpoint, resulting in an updated/decentralized version of the global ML model which is based on the checkpoint version (i.e. the global model checkpoint utilized for the given round) and based on the client updates which have been received up to that point; this appears to be an intermediate version of the global model which exists as the updates are incorporated and prior to the completion of the round of federated learning).
With respect to claim 7, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed, and Rehman further teaches wherein generating the averaged version of the global ML model based on the checkpoint version of the global ML model and based on the decentralized version of the global ML model comprises: averaging corresponding weights of the checkpoint version of the global ML model and the decentralized version of the global ML model to generate averaged weights for the averaged version of the global ML model; and storing, in remote memory of the remote system, the averaged version of the global ML model (e.g. page 512, first through fourth paragraphs describing FedVSSL Aggregation, indicating that, in a federated learning process, FL aggregation strategies are integrated based on weighted averaging under a common framework and then knowledge from past global models is induced while performing model parameter aggregation; performing averaging over global models, where Equation 2 represents stochastic weight averaging of the global models, and the simple averaging of the multiple checkpoints of training models obtains better generalization than conventional training; the process also includes use of different/distinct aggregation strategies for client models; page 513, Algorithm 1 showing the process in context, i.e. where local training occurs (step 4) and corresponding local/client aggregation methods are employed (steps 6-8) resulting in updated global model weights (step 9; analogous to a decentralized version of the global ML model as claimed); subsequent to this aggregation of client/local data to generate a new global model/checkpoint, averaging of the plurality of global models is performed (step 10, showing computation of Equation 2, previously described as stochastic weighted averaging of the global models), where this averaging of the global models (using at least the most recent updated global model and at least one previous global model/checkpoint) is analogous to averaging corresponding weights of the checkpoint version of the global ML model (i.e. at least one previous global model checkpoint) and the decentralized version of the global ML model (i.e. the updated global model checkpoint aggregating local updates), as claimed).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz and Rehman in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), to incorporate the teachings of Rehman (directed to federated self-supervised learning for video understanding) to include the capability to generate an averaged version of the global model based on both the decentralized version of the global model (i.e. an updated version of the global model including aggregated client/local updates) and the checkpoint version of the global model (i.e. at least one previous checkpoint of the global model, such as the checkpoint utilized as the basis for the federated learning round). One of ordinary skill would have been motivated to perform such a modification in order to induce knowledge from past global models while performing model parameter aggregation, obtaining better generalization than conventional training, as described in Rehman (page 512, first through fourth paragraphs).
With respect to claim 8, Bonawitz in view of Rehman teaches all of the limitations of claim 7 as previously discussed and Rehman further teaches wherein the averaged weights for the averaged version of the global ML model are weighted averaged weights of the corresponding weights of the checkpoint version of the global ML model and the decentralized version of the global ML model (e.g. page 512, third and fourth paragraphs, performing stochastic weight averaging of the global models; averaging of the multiple checkpoints; local gradients of clients computed from weighted combination of aggregation strategies, using weighted combination of FedAvg and Loss).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz and Rehman in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), to incorporate the teachings of Rehman (directed to federated self-supervised learning for video understanding) to include the capability to generate an averaged version of the global model using weighted averaged weights. One of ordinary skill would have been motivated to perform such a modification in order to induce knowledge from past global models while performing model parameter aggregation, obtaining better generalization than conventional training, as described in Rehman (page 512, first through fourth paragraphs).
With respect to claim 11, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed, and Bonawitz further teaches wherein causing the averaged version of the global ML model or the updated averaged version of the global ML model to be deployed as the global ML model comprises utilizing the averaged version of the global ML model or the updated averaged version of the global ML model at the remote system (e.g. second page, section 2.1, third paragraph, once a round is established, the server next sends to each participant the current global model parameters and any other necessary state as an FL checkpoint; process repeats; i.e. in response to determining that a next/subsequent round has been established, the new version of the global model checkpoint resulting from the previous round of federated learning is read from storage and transmitted to participant devices as the global model for the next/subsequent round of federated learning, where the acts of reading the new version of the global model checkpoint from storage and transmitting it both appear to be ways in which the server/remote system may utilize the new model).
With respect to claim 12, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed, and Bonawitz further teaches wherein causing the averaged version of the global ML model or the updated averaged version of the global ML model to be deployed as the global ML model comprises transmitting the averaged version of the global ML model or the updated averaged version of the global ML model to the plurality of client devices and a plurality of additional client devices (e.g. second page, section 2.1, third paragraph, once a round is established, the server next sends to each participant the current global model parameters and any other necessary state as an FL checkpoint; process repeats; i.e. in response to determining that a next/subsequent round has been established, the new version of the global model checkpoint resulting from the previous round of federated learning is read from storage and transmitted to participant devices as the global model for the next/subsequent round of federated learning, where client device selection occurs in each round of federated learning, such that in the subsequent round, the new version of the global model checkpoint may be sent to both devices which participated in the previous round and additional devices which are selected for the next round).
With respect to claim 15, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed, and Rehman further teaches wherein the global ML model is a vision-based global ML model, and wherein each of the plurality of client updates is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding vision data generated locally at the given one of the plurality of corresponding client devices and using a corresponding on-device vision-based ML model counterpart that is corresponding on-device counterpart of the vision-based global ML model (e.g. page 507, first full paragraph through second paragraph, combining video self supervised learning and federated learning to enable large scale decentralized feature learning from real world data, practically improving performance and enabling vast majority of vision models for video applications; video SSL training in cross-device FL environment; page 509, section 3.1, integrating FL system with video SSL learning; dataset distributed among decentralized clients, each learning intermediate features collaboratively by training the video SSL approach on their respective local data before performing synchronization through the sever, including receiving the client’s model parameters, performing model aggregation, and sending back the global model again to the clients).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz and Rehman in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), to incorporate the teachings of Rehman (directed to federated self-supervised learning for video understanding) to include the capability to implement the global model as a vison based global ML model where the locally generated client updates are based on processing corresponding local vision data using a corresponding local vison model counterpart of the global model. One of ordinary skill would have been motivated to perform such a modification in order to induce knowledge from past global models while performing model parameter aggregation, obtaining better generalization than conventional training, as described in Rehman (page 512, first through fourth paragraphs).
Claims 14 and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Bonawitz in view of Rehman, further in view of Xu et al. (US 20240303504 A1).
With respect to claim 14, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed. Bonawitz and Rehman do not explicitly disclose wherein the global ML model is an audio-based global ML model, and wherein each of the plurality of client updates is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding audio data generated locally at the given one of the plurality of corresponding client devices and using a corresponding on-device audio-based ML model counterpart that is corresponding on-device counterpart of the audio-based global ML model.
However, Xu teaches wherein the global ML model is an audio-based global ML model, and wherein each of the plurality of client updates is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding audio data generated locally at the given one of the plurality of corresponding client devices and using a corresponding on-device audio-based ML model counterpart that is corresponding on-device counterpart of the audio-based global ML model (e.g. paragraphs 0060-0061, models, including both local and global model can be any model/neural network suitable for natural language processing/NLP; neural network to perform NLP operations including speech recognition; paragraph 0063, federated server sending global updates to clients, and receiving local updates from clients and aggregating the local updates to generate global updates; paragraph 0066, each client in federated learning framework stores its own private data, including audio data).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, and Xu in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale) and Rehman (directed to federated self-supervised learning for video understanding), to incorporate the teachings of Xu (directed to federated learning) to include the capability to implement the global model as an audio based global ML model where the locally generated client updates are based on processing corresponding local audio data using a corresponding local audio model counterpart of the global model. One of ordinary skill would have been motivated to perform such a modification in order to provide neural networks that work better for clients participating in federated learning as described in Xu (paragraph 0003).
With respect to claim 16, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed. Bonawitz and Rehman do not explicitly disclose wherein the global ML model is a text-based global ML model, and wherein each of the plurality of client updates is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding textual data generated locally at the given one of the plurality of corresponding client devices and using a corresponding on-device text-based ML model counterpart that is corresponding on-device counterpart of the text-based global ML model.
However, Xu teaches wherein the global ML model is a text-based global ML model, and wherein each of the plurality of client updates is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding textual data generated locally at the given one of the plurality of corresponding client devices and using a corresponding on-device text-based ML model counterpart that is corresponding on-device counterpart of the text-based global ML model (e.g. paragraphs 0060-0061, models, including both local and global model can be any model/neural network suitable for natural language processing/NLP; neural network to perform NLP operations including text summarization, text classification, text generation; paragraph 0063, federated server sending global updates to clients, and receiving local updates from clients and aggregating the local updates to generate global updates; paragraph 0066, each client in federated learning framework stores its own private data, including text data).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, and Xu in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale) and Rehman (directed to federated self-supervised learning for video understanding), to incorporate the teachings of Xu (directed to federated learning) to include the capability to implement the global model as an text based global ML model where the locally generated client updates are based on processing corresponding local text data using a corresponding local text model counterpart of the global model. One of ordinary skill would have been motivated to perform such a modification in order to provide neural networks that work better for clients participating in federated learning as described in Xu (paragraph 0003).
Claims 2, 3, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Bonawitz in view of Rehman, further in view of Ghose et al. (US 20230004872 A1).
With respect to claim 2, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed, and Bonawitz further teaches prior to initiating the given round of decentralized learning of the global ML model: storing, in remote memory of the remote system, the global ML model as the checkpoint version of the global ML model (e.g. second page, as shown in Fig. 2, at the start of training round I, the global model checkpoint has already been stored in the persistent storage, since it is then read from the persistent storage during round i and then sent to participating devices; moreover, as also shown in Fig. 1, the updated global model checkpoint is written to persistent storage at step 6 prior to the initiation of the next round i+1 of federated learning).
Bonawitz and Rehman do not explicitly disclose training, based on server data that is accessible by the remote system, the global ML model. However, Ghose teaches prior to initiating the given round of decentralized learning of the global ML model: training, based on server data that is accessible by the remote system, the global ML model; and storing, in remote memory of the remote system, the global ML model as the checkpoint version of the global ML model (e.g. paragraph 0028, Fig. 1, continuous federated learning scheme including global/central site including a global model 16 trained on a primary dataset/global dataset 17; paragraph 0040, generating a trained global model using an actual global dataset at a global site (central or main site) and then providing the trained global model to multiple local sites or nodes, establishing a continuous federated learning framework).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, and Ghose in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale) and Rehman (directed to federated self-supervised learning for video understanding), to incorporate the teachings of Ghose (directed to continuous federated learning) to include the capability to, prior to initiating a given round of decentralized learning, train the global model using server data, and store it in memory for use as the global model checkpoint in the given round of learning. One of ordinary skill would have been motivated to perform such a modification in order to resolve issues of data sharing privacy while enabling local learning at the site level to account for site-specific preference while maintaining global performance, mitigating the issue of catastrophic forgetting, as described in Ghose (paragraph 0025).
With respect to claim 3, Bonawitz in view of Rehman, further in view of Ghose teaches all of the limitations of claim 2 as previously discussed, and Bonawitz further teaches prior to receiving the plurality of client updates for the global ML model from the plurality of corresponding client devices: transmitting, to each of the plurality of corresponding client devices and over one or more networks, the global ML model or weights of the global ML model, wherein transmitting the global ML model or the weights of the global ML model to each of the plurality of corresponding client devices causes each of the plurality of corresponding client devices to store the global ML model or the weights of the global ML model in corresponding on-device storage as the corresponding on-device ML model (e.g. first page, second column, second full paragraph, indicating that the described system trains the deep neural network/model on a participant device/phone using local data which will never leave the device, and involves combining weights in the cloud (i.e. indicative of weights of both the global model as transmitted to the device as part of the global model parameters, and of updated weights of the local model which are send back to the server as updates); second page, Fig. 1, showing that, for a given round of federated learning, prior to the participant devices reporting back with updates, in step 3, the model and configuration are sent to the selected devices; second page, section 2.1, third paragraph, once a round is established, the server next sends to each participant the current global model parameters and any other necessary state as an FL checkpoint and each participant performs a local computation based on the global state; seventh page, section 7.2, second paragraph, indicating that the FL plan for the participant device includes computations such as loading and saving weights).
With respect to claim 19, Bonawitz teaches a method implemented by one or more remote processors of a remote system, the method comprising:
initiating N rounds of decentralized learning of a global machine learning (ML) model, wherein N is a positive integer greater than one (e.g. second page, Fig. 1, showing that a federated learning session consists of multiple (i.e. more than one) rounds, and is initiated with a selection step in a first/initial round in which devices check in with the FL server);
during a given round of decentralized learning of the global ML model of the N rounds of decentralized learning of the global ML model (e.g. second page, Fig. 1 and section 2.1, first paragraph, indicating various steps occurring within a given round of federated learning using a global model):
receiving a plurality of client updates from a plurality of corresponding client devices, wherein each of the plurality of client updates is generated locally at a given one of the plurality of corresponding client devices based on processing corresponding client data using a corresponding on-device ML model that are corresponding on-device counterparts of the global ML model (e.g. second page, Fig. 1, step 4, performing on-device training, model update is reported back; second page, section 2.1, third paragraph, each participant performs local computation based on the global state (i.e. FL checkpoint for that round) and its local dataset, and sends an update in the form of an FL checkpoint back to the server);
identifying a checkpoint version of the global ML model that is stored remotely at the remote system (e.g. second page, Fig. 1, step 2, server reads model checkpoint from persistent storage; i.e. for a given round of federated learning, a current checkpoint version of the global model is read from the server for distribution to client devices and for subsequent updating); and
updating, based on the checkpoint version of the global ML model and based on the plurality of client updates received from the plurality of corresponding client devices, a decentralized version of the global ML model (e.g. second page, Fig. 1, step 5, server aggregates updates into the global model as they arrive; second page, section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, as updates are received, the server aggregates them using Federated Averaging; i.e. as each update arrives from a client device, the server integrates/aggregates the received update into the current global model checkpoint, resulting in an updated/decentralized version of the global ML model which is based on the checkpoint version (i.e. the global model checkpoint utilized for the given round) and based on the client updates which have been received up to that point; this appears to be an intermediate version of the global model which exists as the updates are incorporated and prior to the completion of the round of federated learning);
subsequent to the N rounds of decentralized learning of the global ML model (i.e. following completion of a final round of the N rounds of federated learning as shown in Fig. 1, such as performing step 6 at the end of the final round):
generating, based on the decentralized version of the global ML model and the server/checkpoint version of the global ML model, a new version of the global ML mode (e.g. second page, Fig. 1, step 6, server writes global model checkpoint into persistent storage; section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, if enough devices report in time, the round will be successfully completed and the server will update its global model; i.e. once the round, such as a final round, of federated learning is completed, the final version of the updated global model, which is based on both the original global model checkpoint for the round and the latest version incorporating/aggregating received updates during the round, is generated and stored as the new/updated global model checkpoint to be utilized a later federated learning session, or as the global model for prediction)l; and
in response to determining the one or more conditions are satisfied: causing the averaged version of the global ML model to be deployed as the global ML model (e.g. second page, section 2.1, third paragraph, once a round is established, the server next sends to each participant the current global model parameters and any other necessary state as an FL checkpoint; process repeats; i.e. in response to determining that a new round has been established, such as a new round in a new federated learning session, the new version of the global model checkpoint resulting from the previous round of federated learning is deployed as the global model for the new round of federated learning).
Bonawitz does not explicitly disclose that the new version of the global ML model is an averaged version of the global ML model. However, Rehman teaches generating, based on the checkpoint version of the ML model and based on an updated version of the ML model, an averaged version of the ML model (e.g. page 512, first through fourth paragraphs describing FedVSSL Aggregation, indicating that, in a federated learning process, FL aggregation strategies are integrated based on weighted averaging under a common framework and then knowledge from past global models is induced while performing model parameter aggregation; performing averaging over global models, where Equation 2 represents stochastic weight averaging of the global models, and the simple averaging of the multiple checkpoints of training models obtains better generalization than conventional training; the process also includes use of different/distinct aggregation strategies for client models; page 513, Algorithm 1 showing the process in context, i.e. where local training occurs (step 4) and corresponding local/client aggregation methods are employed (steps 6-8) resulting in updated global model weights (step 9; analogous to a decentralized version of the global ML model as claimed); subsequent to this aggregation of client/local data to generate a new global model/checkpoint, averaging of the plurality of global models is performed (step 10, showing computation of Equation 2, previously described as stochastic weighted averaging of the global models), where this averaging of the global models (using at least the most recent updated global model and at least one previous global model/checkpoint) is analogous to generating an averaged version of the global model based on the checkpoint version of the global model and the decentralized version of the global model, as claimed).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz and Rehman in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), to incorporate the teachings of Rehman (directed to federated self-supervised learning for video understanding) to include the capability to generate an averaged version of the global model based on both the decentralized version of the global model (i.e. an updated version of the global model including aggregated client/local updates) and the checkpoint version of the global model (i.e. at least one previous checkpoint of the global model, such as the checkpoint utilized as the basis for the federated learning round). One of ordinary skill would have been motivated to perform such a modification in order to induce knowledge from past global models while performing model parameter aggregation, obtaining better generalization than conventional training, as described in Rehman (page 512, first through fourth paragraphs).
Bonawitz and Rehman do not explicitly disclose
initiating M rounds of centralized learning of the global ML model, wherein M is a positive integer greater than one, such that the server/checkpoint version of the global ML model is a centralized version of the global ML model, and the generating of the averaged version of the global ML model and causing the averaged version of the global ML model to be deployed are performed subsequent to the M rounds of centralized training of the global ML model, and:
during a given round of centralized learning of the global ML model of the M rounds of centralized learning of the global ML model:
obtaining corresponding server data that is accessible by the remote system; processing, using a centralized version of the global ML model that initially corresponds to the checkpoint version of the global ML model, the corresponding server data to generate a corresponding remote update for the global ML model; updating, based on the checkpoint version of the global ML model and based on the corresponding remote update generated remotely at the remote system, the centralized version of the global ML model.
However, Ghose teaches
initiating M rounds of centralized learning of the global ML model, wherein M is a positive integer greater than one, such that the server/checkpoint version of the global ML model is a centralized version of the global ML model, and the generating of the averaged version of the global ML model and causing the averaged version of the global ML model to be deployed are performed subsequent to the M rounds of centralized training of the global ML model (e.g. paragraph 0028, describing continuous federated learning scheme 10 of Fig. 1, in which a global site 12 includes as global model 16 which is trained on a primary dataset 17 and which is then used for incremental site-specific tuning to create local models, the local models then being set back for fine tuning of the global model; i.e. the federated learning process is configured as a continuous/repeated process which includes continuous/repeated rounds of federated learning including both centralized training of the global on a primary dataset at a global/central site and local training at local sites and subsequent updates of the global model using local updates and (in a subsequent round of the continuous process) further training of the global model at the global/central site; therefore, M rounds, M being greater than one, of centralized learning of the global ML model may be included within the overall federated learning process, ultimately resulting in a centralized version of the global ML model as the checkpoint and generation and deployment of the final version of the global model occurring following both the rounds of centralized training and the rounds of decentralized/local training), and:
during a given round of centralized learning of the global ML model of the M rounds of centralized learning of the global ML model:
obtaining corresponding server data that is accessible by the remote system (e.g. paragraph 0028, using primary dataset 17);
processing, using a centralized version of the global ML model that initially corresponds to the checkpoint version of the global ML model, the corresponding server data to generate a corresponding remote update for the global ML model; and updating, based on the checkpoint version of the global ML model and based on the corresponding remote update generated remotely at the remote system, the centralized version of the global ML model (e.g. paragraph 0028, global model 16 at global site 12 trained on primary/global dataset 17; i.e. where training of the global model consists of utilizing a version of the global model generated prior to the training (such as an untrained or previously trained global model), analogous to a checkpoint version of the global ML model, and training the global model/checkpoint version using the global training data, resulting in an update of the global model which generates/updates the previous version/checkpoint and results in a centralized version of the global ML model (i.e. the global model which is trained on the global dataset)).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, and Ghose in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale) and Rehman (directed to federated self-supervised learning for video understanding), to incorporate the teachings of Ghose (directed to continuous federated learning) to include the capability to configure the federated learning process as a continuous/repeated process which includes continuous/repeated rounds of federated learning including both centralized training of the global on a primary dataset at a global/central site and local training at local sites and subsequent updates of the global model using local updates and (in a subsequent round of the continuous process) further training of the global model at the global/central site; such that multiple rounds of centralized learning of the global ML model may be included within the overall federated learning process, ultimately resulting in a centralized version of the global ML model as the checkpoint and generation and deployment of the final version of the global model occurring following both the rounds of centralized training and the rounds of decentralized/local training. One of ordinary skill would have been motivated to perform such a modification in order to resolve issues of data sharing privacy while enabling local learning at the site level to account for site-specific preference while maintaining global performance, mitigating the issue of catastrophic forgetting, as described in Ghose (paragraph 0025).
Claims 9 and 13 are rejected under 35 U.S.C. 103 as being unpatentable over Bonawitz in view of Rehman, further in view of Mu et al. (US 20230004872 A1).
With respect to claim 9, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed, and Bonawitz teaches the method further comprising:
initiating a given additional round of decentralized learning of the global ML model (e.g. second page, Fig. 1; i.e. as shown, the federated learning protocol includes multiple rounds of federated learning such that, after a given round of federated learning, an additional round of federated learning may be initiated);
during the given additional round of decentralized learning of the global ML model:
receiving, from a plurality of corresponding additional client devices, a plurality of additional client updates for further updating the global ML model, wherein each of the plurality of additional client updates is generated locally at a given one of the plurality of corresponding additional client devices based on processing corresponding additional client data using a corresponding additional on-device ML model that are corresponding on-device counterparts of the global ML model (e.g. second page, Fig. 1, step 4, performing on-device training, model update is reported back; second page, section 2.1, third paragraph, each participant performs local computation based on the global state (i.e. FL checkpoint for that round) and its local dataset, and sends an update in the form of an FL checkpoint back to the server; i.e. where this step is executed in each round (including an additional round) of federated learning such that the updates received in additional rounds are additional client updates for further updating the global model);
generating, based on the decentralized version of the global ML model and based on the plurality of additional client updates received from the plurality of corresponding additional client devices, an updated decentralized version of the global ML model (e.g. second page, Fig. 1, step 5, server aggregates updates into the global model as they arrive; second page, section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, as updates are received, the server aggregates them using Federated Averaging; i.e. as each update arrives from a client device, the server integrates/aggregates the received update into the current global model checkpoint, resulting in an updated/decentralized version of the global ML model which is based on the checkpoint version (i.e. the global model checkpoint utilized for the given round) and based on the client updates which have been received up to that point; this appears to be an intermediate version of the global model which exists as the updates are incorporated and prior to the completion of the round of federated learning; where this step is executed in each round (including an additional round) of federated learning, such that the decentralized version generated in the additional round would be based on the decentralized version of the global ML model (updated version from the previous round aggregating client updates from that round) and additional client updates);
generating, based on the checkpoint version of the global ML model and based on the updated decentralized version of the global ML model, an additional new version of the global ML model (e.g. second page, Fig. 1, step 6, server writes global model checkpoint into persistent storage; section 2.1, third paragraph, the server incorporates received updates into its global state; third page, first column, second paragraph, if enough devices report in time, the round will be successfully completed and the server will update its global model; i.e. once the round of federated learning is completed, the final version of the updated global model, which is based on both the original global model checkpoint for the round and the latest version incorporating/aggregating received updates during the round, is generated and stored as the new/updated global model checkpoint to be utilized in subsequent rounds of federated learning; where this step is executed in each round (including an additional round) of federated learning, such that the final version in the additional round would be generated based on the checkpoint version of the global ML model (since the given round and additional rounds are ultimately each based on the checkpoint version used as the basis for the original given round) and the updated decentralized version); and
in response to determining the one or more conditions are satisfied: causing the new version of the global ML model or the additional new version of the global ML model to be deployed as the global ML model (e.g. second page, section 2.1, third paragraph, once a round is established, the server next sends to each participant the current global model parameters and any other necessary state as an FL checkpoint; process repeats; i.e. in response to determining that a next/subsequent round has been established, the new version of the global model checkpoint resulting from the previous round of federated learning is deployed as the global model for the next/subsequent round of federated learning; where this step is executed in each round (including an additional round) of federated learning, such that the new/updated final version of the global ML model resulting from the additional round may be deployed as the global ML model to be utilized in subsequent training rounds).
Bonawitz does not explicitly disclose that the additional new version of the global ML model is an averaged version of the global ML model. However, Rehman teaches generating, based on the checkpoint version of the ML model and based on an updated version of the ML model, an averaged version of the ML model (e.g. page 512, first through fourth paragraphs describing FedVSSL Aggregation, indicating that, in a federated learning process, FL aggregation strategies are integrated based on weighted averaging under a common framework and then knowledge from past global models is induced while performing model parameter aggregation; performing averaging over global models, where Equation 2 represents stochastic weight averaging of the global models, and the simple averaging of the multiple checkpoints of training models obtains better generalization than conventional training; the process also includes use of different/distinct aggregation strategies for client models; page 513, Algorithm 1 showing the process in context, i.e. where local training occurs (step 4) and corresponding local/client aggregation methods are employed (steps 6-8) resulting in updated global model weights (step 9; analogous to a decentralized version of the global ML model as claimed); subsequent to this aggregation of client/local data to generate a new global model/checkpoint, averaging of the plurality of global models is performed (step 10, showing computation of Equation 2, previously described as stochastic weighted averaging of the global models), where this averaging of the global models (using at least the most recent updated global model and at least one previous global model/checkpoint) is analogous to generating an averaged version of the global model based on the checkpoint version of the global model and the decentralized version of the global model, as claimed).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz and Rehman in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), to incorporate the teachings of Rehman (directed to federated self-supervised learning for video understanding) to include the capability to generate an averaged version of the global model based on both the decentralized version of the global model (i.e. an updated version of the global model including aggregated client/local updates) and the checkpoint version of the global model (i.e. at least one previous checkpoint of the global model, such as the checkpoint utilized as the basis for the federated learning round). One of ordinary skill would have been motivated to perform such a modification in order to induce knowledge from past global models while performing model parameter aggregation, obtaining better generalization than conventional training, as described in Rehman (page 512, first through fourth paragraphs).
Bonawitz and Rehman do not explicitly disclose:
in response to determining the one or more conditions are not satisfied: initiating a given additional round of decentralized learning of the global ML model.
However, Mu teaches
in response to determining the one or more conditions are not satisfied: initiating a given additional round of decentralized learning of the global ML model (e.g. paragraph 0097-0098, information for stopping model training sent in response to case that model learning result of the global model meets the model training request; determining that model learning result of current global model meets model training request, such as meeting a specific requirement for model accuracy; where the current global model has reached the sufficient accuracy, ending the training task is determined; paragraph 0106, determining that terminal continues participating in training; completing one round of local model training; completing one round of global model learning; signaling whether terminal continues training; i.e. during learning, it is determined whether a stopping condition is met, such as the global model reaching a requested accuracy level; if the stopping condition/predetermined accuracy is not satisfied, the federated learning continues, initiating another round of learning); and
in response to determining the one or more conditions are satisfied: causing the averaged version of the global ML model or the additional averaged version of the global ML model to be deployed as the global ML model (e.g. paragraph 0097-0098, information for stopping model training sent in response to case that model learning result of the global model meets the model training request; determining that model learning result of current global model meets model training request, such as meeting a specific requirement for model accuracy; where the current global model has reached the sufficient accuracy, ending the training task is determined, and the current global model is obtained; i.e. during learning, it is determined whether a stopping condition is met, such as the global model reaching a requested accuracy level; if the stopping condition/predetermined accuracy is satisfied, the federated learning is stopped, and the current version of the global model is obtained as the version to be deployed).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, and Mu in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale) and Rehman (directed to federated self-supervised learning for video understanding), to incorporate the teachings of Mu (directed to model learning) to include the capability to utilize, as the condition for determining whether to stop or continue federated learning in an additional round, a stopping condition such as the global model reaching a particular accuracy, such that when the stopping condition is not met, learning continues and the next round is initiated and when the stopping condition is met, learning stops and the current global model (i.e. most recent version of the global model aggregating local updates from the most recent round of learning as taught by Bonawitz, where Rehman further teaches that this version may be based on averaging of the updated version with previous global model checkpoints) is used/deployed as the global model. One of ordinary skill would have been motivated to perform such a modification in order to effectively ensure information security during big data exchange and protect terminal data and personal data privacy as described in Mu (paragraph 0044).
With respect to claim 13, Bonawitz in view of Rehman teaches all of the limitations of claim 1 as previously discussed. Bonawitz and Rehman do not explicitly disclose wherein the one or more conditions include one or more of: whether performance of the averaged version of the global ML model or the additional averaged version of the global ML model satisfies a performance threshold, whether a threshold quantity of averaged versions of the global ML model have been generated, whether a threshold quantity of rounds of decentralized learning of the global ML model have been performed, a time of day, or a day of week.
However, Mu teaches wherein the one or more conditions include one or more of: whether performance of the averaged version of the global ML model or the additional averaged version of the global ML model satisfies a performance threshold, whether a threshold quantity of averaged versions of the global ML model have been generated, whether a threshold quantity of rounds of decentralized learning of the global ML model have been performed, a time of day, or a day of week (e.g. paragraph 0097-0098, information for stopping model training sent in response to case that model learning result of the global model meets the model training request; determining that model learning result of current global model meets model training request, such as meeting a specific requirement for model accuracy; where the current global model has reached the sufficient accuracy, ending the training task is determined, and the current global model is obtained; i.e. during learning, it is determined whether a stopping condition is met, such as the global model reaching a requested accuracy level, where this is analogous to a condition of whether the performance of the global ML model satisfies a performance threshold; if the stopping condition/predetermined accuracy is satisfied, the federated learning is stopped and the current global model is used/deployed; if the stopping condition/predetermined accuracy is not satisfied, the federated learning continues, including initiating an additional round of learning).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, and Mu in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale) and Rehman (directed to federated self-supervised learning for video understanding), to incorporate the teachings of Mu (directed to model learning) to include the capability to utilize, as the condition for determining whether to stop or continue federated learning in an additional round, a stopping condition such as the global model reaching a particular accuracy, such that when the stopping condition is not met, learning continues and the next round is initiated and when the stopping condition is met, learning stops and the current global model (i.e. most recent version of the global model aggregating local updates from the most recent round of learning as taught by Bonawitz, where Rehman further teaches that this version may be based on averaging of the updated version with previous global model checkpoints and therefore the accuracy may be with respect to the averaged version of the global model) is used/deployed as the global model. One of ordinary skill would have been motivated to perform such a modification in order to effectively ensure information security during big data exchange and protect terminal data and personal data privacy as described in Mu (paragraph 0044)
Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Bonawitz in view of Rehman, further in view of Mu, further in view of Chawla et al. (US 20220343167 A1).
With respect to claim 10, Bonawitz in view of Rehman, further in view of Mu teaches all of the limitations of claim 9 as previously discussed. Bonawitz, Rehman, and Mu do not explicitly disclose evaluating the averaged version of the global ML model and the additional averaged version of the global ML model to determine corresponding performance measures for the averaged version of the global ML model and the additional averaged version of the global ML model; and selecting, based on the corresponding performance measures, the averaged version of the global ML model or the additional averaged version of the global ML model to be deployed as the global ML model.
However, Chawla teaches evaluating the averaged version of the global ML model and the additional averaged version of the global ML model to determine corresponding performance measures for the averaged version of the global ML model and the additional averaged version of the global ML model; and selecting, based on the corresponding performance measures, the averaged version of the global ML model or the additional averaged version of the global ML model to be deployed as the global ML model (e.g. paragraph 0043, evaluating model performance; paragraph 0044-0045, selecting modified final weights/previous set of model weights for use in operating neural network based on model performance; paragraphs 0047-0050, comparing performance of neural network using final model weights, neural network using modified final neural network weights, and intermediate neural network weights, and selecting from among the final weights, modified final weights, and intermediate weights in response to the comparison; method may also including combining final weights with final modified weights and/or intermediate weights to obtain a combined set of neural network weights; combining neural network weights my include generating a weighted average of the neural network weights; i.e. evaluating/comparing a first version/checkpoint of a global model (as defined by a set of final weights) with an additional version/checkpoint of a global model (as defined by a modified set of final weights), determining corresponding performance measures, and selecting either the first version or the additional version to be deployed, where model weights may be averaged weights (as also taught by Rehman, as previously cited)).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, Mu, and Chawla in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), Rehman (directed to federated self-supervised learning for video understanding), and Mu (directed to model learning) to incorporate the teachings of Chawla (directed to machine learning model lifecycles) to include the capability to evaluate the performance of various versions of the global model (as indicated by corresponding weights for those versions/checkpoints, and where those versions may be averaged versions/checkpoints as taught by Rehman), including at least an averaged version of the global ML model (resulting from a given training round) and an additional averaged version of the global ML model (resulting from an additional training round), and selecting one of these for deployment as the global ML model based on their respective performance . One of ordinary skill would have been motivated to perform such a modification in order to provide more accurate and efficient selection of features, tuning, and selection of topologies in federated learning systems as described in Chawla (paragraph 0016)
Claim 18 is rejected under 35 U.S.C. 103 as being unpatentable over Bonawitz in view of Rehman, further in view of Chu et al. (US 20210374617 A1).
With respect to claim 18, Bonawitz in view of Rehman teaches all of the limitations of claim 17 as previously discussed. Bonawitz and Rehman do not explicitly disclose wherein N is a configurable parameter. However, Chu teaches wherein N is a configurable parameter (e.g. paragraph 0045, indicating that an administrator may define a number of training rounds).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, and Chu in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale) and Rehman (directed to federated self-supervised learning for video understanding) to incorporate the teachings of Chu (directed to horizontal federated learning using non-IID data) to include the capability for a user to configure the number of training rounds. One of ordinary skill would have been motivated to perform such a modification in order to enable learning of customized models for clients holding non-IID datasets while also ensuring data privacy as described in Chu (paragraph 0005-0007).
Claim 20 is rejected under 35 U.S.C. 103 as being unpatentable over Bonawitz in view of Rehman, further in view of Ghose, further in view of Chu, further in view of Liu et al. (US 20230083116 A1).
With respect to claim 20, Bonawitz in view of Rehman, further in view of Ghose teaches all of the limitations of claim 19 as previously discussed. Bonawitz, Rehman, and Ghose do not explicitly disclose wherein N is a configurable parameter. However, Chu teaches wherein N is a configurable parameter (e.g. paragraph 0045, indicating that an administrator may define a number of training rounds).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, Ghose, and Chu in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), Ghose (directed to continuous federated learning), and Rehman (directed to federated self-supervised learning for video understanding) to incorporate the teachings of Chu (directed to horizontal federated learning using non-IID data) to include the capability for a user to configure the number of training rounds. One of ordinary skill would have been motivated to perform such a modification in order to enable learning of customized models for clients holding non-IID datasets while also ensuring data privacy as described in Chu (paragraph 0005-0007).
Bonawitz, Rehman, Ghose, and Chu do not explicitly disclose wherein M is a separate configurable parameter, and wherein M is larger than N. However, Liu teaches wherein M is a separate configurable parameter, and wherein M is larger than N (e.g. paragraph 0050, during given round of training of global model, receiving parameters from edge devices, aggregating the received parameters and updating the global model; paragraphs 0051 and 0054, performing training iterations on current aggregation model using globally shared dataset; specified number of iterations/number of server-side iterations; i.e. within a given overall federated training round, analogous to a round of decentralized learning, multiple rounds/iterations of centralized training at the central server may also be performed, analogous to rounds of centralized learning; this number of centralized iterations/rounds may be specified and is therefore configurable, and is distinct from the overall number of federated training rounds; moreover, since multiple iterations of centralized training occur for each given round of federated training, the number M of centralized training iterations/rounds is larger than the number N of federated/decentralized rounds).
Accordingly, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention having the teachings of Bonawitz, Rehman, Ghose, Chu, and Liu in front of him to have modified the teachings of Bonawitz (directed to federated learning at scale), Ghose (directed to continuous federated learning), Chu (directed to horizontal federated learning using non-IID data), and Rehman (directed to federated self-supervised learning for video understanding) to incorporate the teachings of Liu (directed to federated learning) to include the capability for a user to configure the number of training rounds for both an overall federated/decentralized learning process and for a number of centralized training iterations/rounds occurring within each round of the federated/decentralized learning process, such that the total configured number of centralized iterations/rounds is larger than the configured number of federated/decentralized rounds. One of ordinary skill would have been motivated to perform such a modification in order to enable federated learning while complying with provisions of relevant laws and regulations, taking essential confidentiality measures, and not violating public order and good custom as described in Liu (paragraph 0020).
It is noted that any citation to specific pages, columns, lines, or figures in the prior art references and any interpretation of the references should not be considered to be limiting in any way. “The use of patents as references is not limited to what the patentees describe as their own inventions or to the problems with which they are concerned. They are part of the literature of the art, relevant for all they contain,” In re Heck, 699 F.2d 1331, 1332-33, 216 USPQ 1038, 1039 (Fed. Cir. 1983) (quoting in re Lemelson, 397 F.2d 1006, 1009, 158 USPQ 275, 277 (GCPA 1968)). Further, a reference may be relied upon for all that it would have reasonably suggested to one having ordinary skill the art, including nonpreferred embodiments. Merck & Co, v. Biocraft Laboratories, 874 F.2d 804, 10 USPQ2d 1843 (Fed. Cir.), cert, denied, 493 U.S. 975 (1989). See also Upsher-Smith Labs. v. Pamlab, LLC, 412 F,3d 1319, 1323, 75 USPQ2d 1213, 1215 (Fed. Cir, 2005): Celeritas Technologies Ltd. v. Rockwell International Corp., 150 F.3d 1354, 1361, 47 USPQ2d 1516, 1522-23 (Fed. Cir. 1998).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JEREMY L STANLEY whose telephone number is (469)295-9105. The examiner can normally be reached on Monday-Friday from 9:00 AM to 5:00 PM CST.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Abdullah Al Kawsar, can be reached at telephone number (571) 270-3169. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from Patent Center and the Private Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from Patent Center or Private PAIR. Status information for unpublished applications is available through Patent Center and Private PAIR for authorized users only. Should you have questions about access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free).
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) Form at https://www.uspto.gov/patents/uspto-automated- interview-request-air-form.
/JEREMY L STANLEY/
Primary Examiner, Art Unit 2127
Prosecution Insights