DETAILED ACTION
This action is in response to the filing on 01/30/2026. Claims 1-20, are pending and have been considered below.
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 .
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
Claims 1, 6-10, 11, and 16-19 are rejected under 35 U.S.C. 103 as being unpatentable over XU et al. (CN 114418085 A), hereinafter Xu, in view of Turow et al. (US 11,348,029 B1), hereinafter Turow, and further in view of Pushkin et al. (US 12,061,956 B1), hereinafter Pushkin.
Regarding claim 1, Xu teaches a method comprising:
receiving pruned candidate models from source nodes in a distributed computing environment, wherein the pruned candidate models are trained in source nodes, and are stored (Xu discloses a distributed learning environment in which edge nodes upload their locally trained and pruned models to the central server [see Xu, n0011], and the central server collects all edge device models to be assessed for parameter aggregation [see Xu, n0089 and Figure 3]);
the pruned candidate models (Xu discloses pruned local models [see Xu, n0011]).
However, Xu fails to teach receiving associated loss values from source nodes in a distributed computing environment, wherein the pruned candidate models are trained in source nodes with respective local datasets stored therein, and stored in an assessment structure; selecting test candidates from the candidate models; testing parallelly each of the test candidates at generalization nodes with respective local datasets stored therein in the distributed computing environment, wherein the respective local datasets stored at the generalization nodes are different from the respective local datasets stored at the source nodes and a respective local dataset stored at a first generalization node is different from a respective local dataset stored at a second generalization node different from the first generalization node; receiving loss values for the test candidates from the generalization nodes; selecting a winning candidate from the test candidates based on aggregated loss values of the test candidates; and deploying the winning candidate to one or more target nodes.
In the same field of endeavor, Turow teaches:
receiving associated loss values from source nodes in a distributed computing environment, wherein the pruned candidate models are stored in an assessment structure (Turow discloses that performance results may include accuracy level of the candidate model among other criteria [see Turow, Col. 8, lines 13-22], and providing performance results for the candidate models from the ML reduction service [see Turow, Col. 15, lines 64-66] implemented by a plurality of computing instances, which can be run on a plurality of devices, and data store(s) including candidate models and model performance results [see Turow, Col. 11, lines 27-37 and FIG. 2]. Thus, the data store would receive the associated accuracy from the computing instances that determine the performance results, the data store(s) would then store the candidate models and associated model performance results. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on);
selecting test candidates from the candidate models (Turow discloses generating a plurality of candidate models and validating the models to evaluate performance of the candidates [see Turow, Col. 15, lines 46-48 and 55-58]. Thus, the test candidates selected is the whole set of candidate models for the current iteration);
testing parallelly each of the test candidates at generalization nodes with respective local datasets stored therein in the distributed computing environment (Turow discloses validating the candidate models to evaluate their performance [see Turow, Col. 15, lines 55-58] with a plurality of computing instances, which can be run on a plurality of devices [see Col. 11, lines 27-37 and FIG. 2]);
receiving loss values for the test candidates from the generalization nodes (Turow discloses that performance results may include accuracy level of the candidate model among other criteria [see Turow, Col. 8, lines 13-22], and providing performance results for the candidate models from the ML reduction service [see Turow, Col. 15, lines 64-66] implemented by a plurality of computing instances, which can be run on a plurality of devices, and data stores including a model performance results [see Turow, Col. 11, lines 27-37 and FIG. 2]. Thus, performance results data store would receive the associated accuracy from the computing instances that determine the performance results. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on);
selecting a winning candidate from the test candidates based on aggregated loss values of the test candidates (Turow discloses selecting a winning model based on the performance criteria of the ML model, and the performance representation of the candidate models, [see Turow, Col. 16, lines 39-44] which includes accuracy level of the candidates [see Turow, Col. 8, lines 26-30; Col. 16, lines 6-14], such that the performance results are aggregated in a performance result data store [see Turow, FIG. 2]. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on);
deploying the winning candidate to one or more target nodes (Turow discloses deploying the selected model to a computing hub in a distributed computing environment [see Turow, Col. 16, lines 15-22]).
It would have been obvious to one of ordinary skill, in the art at the time before the effective filing date of the invention to incorporate receiving associated loss values from source nodes in a distributed computing environment, wherein the pruned candidate models are stored in an assessment structure; selecting test candidates from the candidate models; testing the test candidates at generalization nodes in the distributed computing environment; receiving loss values for the test candidates from the generalization nodes; selecting a winning candidate from the test candidates based on aggregated loss values of the test candidates; and deploying the winning candidate to one or more target nodes as suggested in Turow into Xu because both methods perform machine learning (see Xu, Abstract; see Turow, Abstract). Incorporating the teaching of Turow into Xu would provide efficient candidate ML models (in terms of the performance criteria) that satisfy the relative threshold, while inefficient candidate ML models that do not satisfy the relative threshold may be removed (see Turow, Col. 3, lines 55-58).
However, the combination of Xu and Turow fails to teach wherein the pruned candidate models are trained in source nodes with respective local datasets stored therein; and wherein the respective local datasets stored at the generalization nodes are different from the respective local datasets stored at the source nodes and a respective local dataset stored at a first generalization node is different from a respective local dataset stored at a second generalization node different from the first generalization node.
In the same field of endeavor, Pushkin teaches:
wherein the pruned candidate models are trained in source nodes with respective local datasets stored therein (Pushkin discloses training a ML model by a plurality of clients where each client trains the model using local data [see Pushkin, Col. 9, ln 8-15]);
wherein the respective local datasets stored at the generalization nodes are different from the respective local datasets stored at the source nodes and a respective local dataset stored at a first generalization node is different from a respective local dataset stored at a second generalization node different from the first generalization node (Pushkin discloses testing the model at edge devices with local testing datasets such that the different devices have their own datasets local to themselves [see Pushkin, Col. 9, ln 19-24]).
It would have been obvious to one of ordinary skill, in the art at the time before the effective filing date of the invention to incorporate wherein the pruned candidate models are trained in source nodes with respective local datasets stored therein; and wherein the respective local datasets stored at the generalization nodes are different from the respective local datasets stored at the source nodes and a respective local dataset stored at a first generalization node is different from a respective local dataset stored at a second generalization node different from the first generalization node as suggested in Pushkin into the combination of Xu and Turow because both methods are directed to distributed learning (see Xu, Abstract; see Pushkin, Abstract). Incorporating the teaching of Pushkin into the combination of Xu and Turow would improve data privacy by not sharing the local training data with the central server (see Pushkin, Col. 1, ln 58-61 and Col. 6, ln 11-26).
Regarding claim 6, the combination of Xu, Turow, and Pushkin as applied in claim 1 above teaches all the limitations of claim 1 and further teaches:
storing the pruned candidate models and their loss values in the assessment structure; and storing loss values determined by the generalization nodes with the loss values in the assessment structure (Turow discloses that performance results may include accuracy level of the candidate model among other criteria [see Turow, Col. 8, lines 13-22], and providing performance results for the candidate models from the ML reduction service [see Turow, Col. 15, lines 64-66] implemented by a plurality of computing instances, which can be run on a plurality of devices, and data store(s) including candidate models and model performance results [see Turow, Col. 11, lines 27-37 and FIG. 2]. Thus, the data store would receive the associated accuracy from the computing instances that determine the performance results, the data store(s) would then store the candidate models and associated model performance results, adding to the models and performance results already in the data store for any additional candidates after the first candidate. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on).
Regarding claim 7, the combination of Xu, Turow, and Pushkin as applied in claim 1 above teaches all the limitations of claim 1 and further teaches:
determining an aggregated loss for each of the test candidates from the pruned candidate models stored in the assessment structure (Turow discloses that performance results may include accuracy level of the candidate model among other criteria [see Turow, Col. 8, lines 13-22], and providing performance results for the candidate models from the ML reduction service [see Turow, Col. 15, lines 64-66] implemented by a plurality of computing instances, which can be run on a plurality of devices, and data store(s) including candidate models and model performance results [see Turow, Col. 11, lines 27-37 and FIG. 2]. Thus, the data store would receive the associated accuracy from the computing instances that determine the performance results, the data store(s) would then store the candidate models and associated model performance results. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on).
Regarding claim 8, the combination of Xu, Turow, and Pushkin as applied in claim 1 above teaches all the limitations of claim 1 and further teaches:
eliminating test candidates whose aggregated loss is greater than a threshold loss (Turow discloses that candidates are validated to obtain performance results and evaluate performance criteria for the models, such as accuracy, wherein the models that do not satisfy the threshold criteria for accuracy are removed [see Turow, Col. 16, lines 31-44]. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on. Thus, for a threshold criteria where accuracy must be above a certain threshold, when the loss of a candidate model is above a threshold loss that puts the accuracy below the accuracy threshold, the candidate would be removed).
Regarding claim 9, the combination of Xu, Turow, and Pushkin as applied in claim 1 above teaches all the limitations of claim 1 and further teaches:
determining the winning candidate as the test candidate with a lowest aggregated loss (Turow discloses selecting a winning model based on the performance criteria of the ML model, and the performance representation of the candidate models, [see Turow, Col. 16, lines 39-44] which includes accuracy level of the candidates [see Turow, Col. 8, lines 26-30; Col. 16, lines 6-14]. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on. Thus, the candidate model would be selected based on the accuracy performance result, such that the model with the highest accuracy is selected, which is the model with the lowest loss).
Regarding claim 10, the combination of Xu, Turow, and Pushkin as applied in claim 1 above teaches all the limitations of claim 1 and further teaches:
wherein the pruned candidate models are generated in a parallel manner at multiple source nodes (Xu discloses that each edge device prunes their local model [see Xu, n0010], displays that the edge devices all receive the global model and upload their pruned models together [see Xu, Figure 2], and that the central server waits for all edge device models to be collected, or a predetermined time [see Xu, n0089]. Thus, the edge devices prune their local models in parallel so that the central server can collect all the pruned local models around the same time, or collect the models that finish within the given time) and wherein the test candidates are tested in a parallel manner at multiple generalization nodes (Turow discloses validating the candidate models to evaluate their performance [see Turow, Col. 15, lines 55-58] with a plurality of computing instances, which can be run on a plurality of devices [see Col. 11, lines 27-37 and FIG. 2]. Thus, it would have been obvious to use parallelization, similar to pruning the local models in Xu [see Xu, n0010, n0089, and Figure 2], for testing the pruned models as disclosed by Turow).
Regarding claim 11, claim 11 contains substantially similar limitations to those found in claim 1. Therefore it is rejected for the same reason as claim 1 above. Additionally, the combination of Xu, Turow, and Pushkin further teaches:
A non-transitory storage medium having stored therein instructions that are executable by one or more hardware processors to perform operations comprising (Turow discloses a computer readable storage medium storing computer readable instructions and that the disclosed modules may be implemented in software for execution by processors [see Turow, Col. 20, lines 11-12 and 32-44]).
Regarding claim 16, claim 16 contains substantially similar limitations to those found in claim 6 above. Consequently, claim 16 is rejected for the same reasons.
Regarding claim 17, claim 17 contains substantially similar limitations to those found in claim 7 above. Consequently, claim 17 is rejected for the same reasons.
Regarding claim 18, claim 18 contains substantially similar limitations to those found in claim 8 above. Consequently, claim 18 is rejected for the same reasons.
Regarding claim 19, the combination of Xu, Turow, and Pushkin as applied in claim 11 above teaches all the limitations of claim 11 and further teaches:
determining the winning candidate as the test candidate with a lowest aggregated loss (Turow discloses selecting a winning model based on the performance criteria of the ML model, and the performance representation of the candidate models, [see Turow, Col. 16, lines 39-44] which includes accuracy level of the candidates [see Turow, Col. 8, lines 26-30; Col. 16, lines 6-14]. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on. Thus, the candidate model would be selected based on the accuracy performance result, such that the model with the highest accuracy is selected, which is the model with the lowest loss), wherein the pruned candidate models are generated in a parallel manner at multiple source nodes (Xu discloses that each edge device prunes their local model [see Xu, n0010], displays that the edge devices all receive the global model and upload their pruned models together [see Xu, Figure 2], and that the central server waits for all edge device models to be collected, or a predetermined time [see Xu, n0089]. Thus, the edge devices prune their local models in parallel so that the central server can collect all the pruned local models around the same time, or collect the models that finish within the given time) and wherein the test candidates are tested in a parallel manner at multiple generalization nodes (Turow discloses validating the candidate models to evaluate their performance [see Turow, Col. 15, lines 55-58] with a plurality of computing instances, which can be run on a plurality of devices [see Col. 11, lines 27-37 and FIG. 2]. Thus, it would have been obvious to use parallelization, similar to pruning the local models in Xu [see Xu, n0010, n0089, and Figure 2], for testing the pruned models as disclosed by Turow).
Claims 2 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over XU et al. (CN 114418085 A), hereinafter Xu, in view of Turow et al. (US 11,348,029 B1), hereinafter Turow, and further in view of Pushkin et al. (US 12,061,956 B1), hereinafter Pushkin, as applied in claim 1 above, and further in view of RODRIGUEZ MULET et al. (US 2024/0289635 A1), hereinafter Rodriguez.
Regarding claim 2, the combination of Xu, Turow, and Pushkin as applied in claim 1 above teaches all the limitations of claim 1.
However, the combination of Xu and Turow fails to teach initializing the source nodes with parameters of initial candidate models, a number of epochs of training, and a learning rate.
In the same field of endeavor, Rodriguez teaches:
initializing the source nodes with parameters of initial candidate models, a number of epochs of training, and a learning rate (Rodriguez discloses that each local device includes local training information including a preliminary local training condition and preliminary local training result [see Rodriguez, para. 12-13], wherein the local training condition includes a learning rate, initialization method or model weight, [see Rodriguez, para. 14], and number of epochs of training [see Rodriguez, para. 60]).
It would have been obvious to one of ordinary skill, in the art at the time before the effective filing date of the invention to incorporate initializing the source nodes with parameters of initial candidate models, a number of epochs of training, and a learning rate as suggested in Rodriguez into the combination of Xu and Turow because both methods perform distributed learning (see Xu, Abstract; see Rodriguez, Abstract). By incorporating the techniques of Rodriguez into the combination of Xu, Turow, and Pushkin, an enormous number of combinations of balancing parameters can be efficiently selected (see Rodriguez, para. 42).
Regarding claim 12, claim 12 contains substantially similar limitations to those found in claim 2 above. Consequently, claim 12 is rejected for the same reasons.
Claims 3-5 and 13-15 are rejected under 35 U.S.C. 103 as being unpatentable over XU et al. (CN 114418085 A), hereinafter Xu, in view of Turow et al. (US 11,348,029 B1), hereinafter Turow, and further in view of Pushkin et al. (US 12,061,956 B1), hereinafter Pushkin, and further in view of RODRIGUEZ MULET et al. (US 2024/0289635 A1), hereinafter Rodriguez, as applied in claim 2 above, and further in view of T. Wang et al., (DATASET DISTILLATION), hereinafter Wang.
Regarding claim 3, the combination of Xu, Turow, Pushkin, and Rodriguez as applied in claim 2 above teaches all the limitations of claim 2 and further teaches:
at each source node, generating an initial model and training the initial model with a dataset to generate a candidate model and pruning the candidate model to generate a pruned candidate model (Xu discloses generating an initial global model and distributing it to each edge device such that the devices have their own local model [see Xu, n0008], then training the local model using local data [see Xu, n0009 and n0014], and finally pruning the local model [see Xu, n0010]).
However, the combination of Xu, Turow, Pushkin, and Rodriguez fails to teach training the model with a distilled dataset.
In the same field of endeavor, Wang teaches:
training the model with a distilled dataset (Wang discloses distilling the training dataset and then training a model with a distilled dataset [see Wang, Sect. 1, para. 1]).
It would have been obvious to one of ordinary skill, in the art at the time before the effective filing date of the invention to incorporate Wang discloses distilling the training dataset and then training a model with a distilled dataset [see Wang, Sect. 1, para. 1] as suggested in Wang into the combination of Xu, Turow, and Rodriguez because both methods perform machine learning (see Xu, Abstract; see Wang, Abstract). Incorporating the teaching of Wang into the combination of Xu, Turow, Pushkin, and Rodriguez would provide fast model training in only a few gradient steps (see Wang, Sect. 1, para. 3).
Regarding claim 4, the combination of Xu, Turow, Pushkin, Rodriguez, and Wang as applied in claim 3 above teaches all the limitations of claim 3 and further teaches:
retraining and repruning the pruned candidate model one or more times (Xu discloses edge devices receiving the global model from the central server, training and pruning a local model, sending the pruned local model to the central server, and the central server aggregating all the pruned parameters [see Xu, n0008-n0010]. Rodriguez discloses multiple rounds of training with communication between the local devices and central server [see Rodriguez, para. 59 and Steps SA4-SA14 of FIG. 2]. Thus, the combination of the methods of Xu and Rodriguez would result in multiple rounds of receiving the global model, training and pruning a local model, sending the pruned local model to the central server, the central server aggregating the pruned parameters, and then repeating with sending out the updated global model).
Regarding claim 5, the combination of Xu, Turow, Pushkin, Rodriguez, and Wang as applied in claim 3 above teaches all the limitations of claim 3 and further teaches:
communicating the pruned candidate model to the central node along with a loss value based on a local dataset of the source node (Xu discloses edge nodes upload their locally pruned models to the central server [see Xu, n0010]. Turow discloses that performance results may include accuracy level of the candidate model among other criteria [see Turow, Col. 8, lines 13-22], and providing performance results for the candidate models from the ML reduction service [see Turow, Col. 15, lines 64-66] implemented by a plurality of computing instances, which can be run on a plurality of devices, and data stores including a model performance results [see Turow, Col. 11, lines 27-37 and FIG. 2]. Thus, performance results data store would receive the associated accuracy from the computing instances that determine the performance results. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on. Thus, the combination of Xu and Turow would upload the pruned model and the associated loss value based on the local data to the central server).
Regarding claim 13, claim 13 contains substantially similar limitations to those found in claim 3 above. Consequently, claim 13 is rejected for the same reasons.
Regarding claim 14, claim 14 contains substantially similar limitations to those found in claim 4 above. Consequently, claim 14 is rejected for the same reasons.
Regarding claim 15, claim 15 contains substantially similar limitations to those found in claim 5 above. Consequently, claim 15 is rejected for the same reasons.
Claim 20 is rejected under 35 U.S.C. 103 as being unpatentable over XU et al. (CN 114418085 A), hereinafter Xu, in view of BALEVI et al. (US 2023/0180152 A1), hereinafter Balevi, and further in view of Turow et al. (US 11,348,029 B1), hereinafter Turow, and further in view of Pushkin et al. (US 12,061,956 B1), hereinafter Pushkin and further in view of T. Wang et al., (DATASET DISTILLATION), hereinafter Wang,
Regarding claim 20, Xu teaches a method comprising:
parallelly training, by each source node, the respective initial model with a respective dataset to generate a respective candidate model (Xu discloses generating an initial global model and distributing it to each edge device such that the devices have their own initial local model [see Xu, n0008], then training the local model using local data [see Xu, n0009 and n0014]);
pruning, by each source node, the respective candidate model to generate a respective pruned candidate model (Xu discloses each edge device pruning their local model [see Xu, n0010]);
the respective pruned candidate models (Xu discloses pruned local models [see Xu, n0011]);
transmitting, by the source nodes, at least one of the pruned candidate models to the central node (Xu discloses a distributed learning environment in which edge nodes upload their locally pruned models to the central server [see Xu, n0011], and the central server collects all edge device models to be assessed for parameter aggregation [see Xu, n0089 and Figure 3]).
However, Xu fails to teach
receiving, by source nodes, model parameters and a learning rate from a central node;
sampling, by source nodes, the model parameters to obtain a respective initial model for each source node;
training the respective initial model with a respective distilled dataset, wherein the respective distilled dataset of a first source node is different from a respective distilled dataset of a second source node different from the first source node;;
evaluating, by each source node, a loss of the respective pruned candidate model;
discarding, by each source node, the respective pruned candidate models whose loss is greater than a threshold; and
transmitting, at least one of the candidate models whose loss is less than or equal to the threshold.
In the same field of endeavor, Balevi teaches:
receiving, by source nodes, model parameters and a learning rate from a central node (Balevi discloses a plurality of user equipment (UE) devices that perform distributed learning, comprising a federated learning component, [see Balevi, para. 35 and FIG. 1], and that the federated learning component comprises a receiver component that receives communications from a base station, including the global model and learning rate [see Balevi, para. 94-95 and FIG. 8]. Thus, the UE device can receive the model parameters because it receives the global model including its parameters, and receives the learning rate as part of configuration information);
It would have been obvious to one of ordinary skill, in the art at the time before the effective filing date of the invention to incorporate receiving model parameters and a learning rate at a source node from a central node and other pruned candidate models generated at the source node both methods perform distributed learning (see Xu, Abstract; see Balevi, Abstract). By incorporating the techniques of Balevi into Xu the users are able to keep the local dataset private and reduce communication costs because the local dataset is not transmitted (see Balevi, para. 24).
However, the combination of Xu and Balevi fails to teach
sampling, by source nodes, the model parameters to obtain a respective initial model for each source node;
training the respective initial model with a respective distilled dataset, wherein the respective distilled dataset of a first source node is different from a respective distilled dataset of a second source node different from the first source node;
evaluating, by each source node, a loss of the respective pruned candidate model;
discarding, by each source node, the respective pruned candidate models whose loss is greater than a threshold; and
transmitting, at least one of the candidate models whose loss is less than or equal to the threshold.
In the same field of endeavor, Turow teaches:
evaluating, by each source node, a loss of the respective pruned candidate model (Turow discloses that performance results may include accuracy level of the candidate model among other criteria [see Turow, Col. 8, lines 13-22], and providing performance results for the candidate models from the ML reduction service [see Turow, Col. 15, lines 64-66]. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on. Turow further discloses comparing the efficiency results of candidate models against each other with an efficiency curve that models the processing time with the accuracy results of the candidate models, where the efficiency curve is generated based on a curve fitting technique to the efficiencies of the plurality of candidate models [see Turow, Col. 15, lines 20-39; FIG. 3-4]. Thus, the loss of candidate models are evaluated against each other by comparing the accuracy efficiencies against the efficiency curve fit to the plurality of candidate models);
discarding, by each source node, the respective pruned candidate models whose loss is less than or equal to the threshold (Turow discloses that candidates are validated to obtain performance results and evaluate performance criteria for the models, such as accuracy, wherein the models that do not satisfy the threshold criteria for accuracy are removed [see Turow, Col. 16, lines 31-44]. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on. Thus, for a threshold criteria where accuracy must be above a certain threshold, when the loss of a candidate model is above a threshold loss that puts the accuracy below the accuracy threshold, the candidate would be removed)
transmitting, at least one of the candidate models whose loss is less than or equal to the threshold (Turow discloses that candidates are validated to obtain performance results and evaluate performance criteria for the models, such as accuracy, wherein the models that do not satisfy the threshold criteria for accuracy are removed [see Turow, Col. 16, lines 31-44]. Further, it would have been obvious that the accuracy of the model would reflect loss values for data in which it is validated on. Thus, for a threshold criteria where accuracy must be above a certain threshold, when the loss of a candidate model is above a threshold loss that puts the accuracy below the accuracy threshold, the candidate would be removed).
It would have been obvious to one of ordinary skill, in the art at the time before the effective filing date of the invention to incorporate evaluating, by each source node, a loss of the respective pruned candidate model; discarding, by each source node, the respective pruned candidate models whose loss is greater than a threshold; and transmitting, at least one of the candidate models whose loss is less than or equal to the threshold because both methods perform machine learning (see Xu, Abstract; see Turow, Abstract). Incorporating the teaching of Turow into the combination of Xu and Balevi would efficient candidate ML models (in terms of the performance criteria) that satisfy the relative threshold, while inefficient candidate ML models that do not satisfy the relative threshold may be removed (see Turow, Col. 3, lines 55-58).
However, the combination of Xu, Balevi, and Turow fails to teach
sampling, by source nodes, the model parameters to obtain a respective initial model for each source node;
training the respective initial model with a respective distilled dataset, wherein the respective distilled dataset of a first source node is different from a respective distilled dataset of a second source node different from the first source node.
In the same field of endeavor, Wang teaches:
sampling, the model parameters to obtain a respective initial model for each source node (Wang discloses initializing the neural network with random samples from a specific distribution [see Wang, Sect. 3.2, para. 1]);
training the respective initial model with a respective distilled dataset (Wang discloses distilling the training dataset and then training a model with the distilled dataset [see Wang, Sect. 1, para. 1]).
It would have been obvious to one of ordinary skill, in the art at the time before the effective filing date of the invention to incorporate sampling, the model parameters to obtain a respective initial model for each source node and training the respective initial model with a respective distilled dataset as suggested in Wang into the combination of Xu, Balevi, and Turow because both methods perform machine learning (see Xu, Abstract; see Wang, Abstract). Incorporating the teaching of Wang into the combination of Xu, Balevi, and Turow would provide fast model training in only a few gradient steps (see Wang, Sect. 1, para. 3).
However, the combination of Xu, Balevi, Turow, and Wang fails to teach wherein the respective distilled dataset of a first source node is different from a respective distilled dataset of a second source node different from the first source node.
In the same field of endeavor Pushkin teaches:
wherein the respective dataset of a first source node is different from a respective dataset of a second source node different from the first source node (Pushkin discloses training a ML model by a plurality of clients where each client trains the model using local data [see Pushkin, Col. 9, ln 8-15]).
It would have been obvious to one of ordinary skill, in the art at the time before the effective filing date of the invention to incorporate wherein the respective dataset of a first source node is different from a respective dataset of a second source node different from the first source node as suggested in Pushkin into the combination of Xu, Balevi, Turow, and Wang because methods are directed to distributed learning (see Xu, Abstract; see Pushkin, Abstract). Incorporating the teaching of Pushkin into the combination of Xu, Balevi, Turow, and Wang would improve data privacy by not sharing the local training data with the central server (see Pushkin, Col. 1, ln 58-61 and Col. 6, ln 11-26).
Response to Amendment
Applicant’s amendments, filed 01/30/2026, to the specification are accepted and the objections to the specification and drawings are respectfully withdrawn.
Response to Arguments
Applicant’s arguments, filed 01/30/2026, traversing the rejection of claims 6-9 and 16-19 under 35 U.S.C. 112(b), on p. 10, have been fully considered and are persuasive, the rejection of claims 6-9 and 16-19 under 35 U.S.C. 112(b) are respectfully withdrawn.
Applicant’s arguments, filed 01/30/2026, traversing the rejection of claim 20 under 35 U.S.C. 101, on p. 10-11, have been fully considered and are persuasive, the rejection of claim 20 under 35 U.S.C. 101 is respectfully withdrawn.
Applicant’s arguments, filed 01/30/2026, traversing the rejection of claims 1, 6-11, 16-19 under 35 U.S.C. 103, on p. 11-14, have been fully considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
M. N. H. Nguyen et al. ("Self-Organizing Democratized Learning: Toward Large-Scale Distributed Learning Systems," in IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 12, pp. 10698-10710, Dec. 2023, doi: 10.1109/TNNLS.2022.3170872) discloses evaluating learning models in a distributed learning system on local test data only.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JAKE BREEN whose telephone number is (571)272-0456. The examiner can normally be reached Monday - Friday, 7:00 AM - 3:00 PM EST.
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) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Jennifer Welch can be reached at (571) 272-7212. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/J.T.B./Examiner, Art Unit 2143
/JENNIFER N WELCH/Supervisory Patent Examiner, Art Unit 2143
Prosecution Insights