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 .
Amendments
This action is in response to amendments filed January 15th, 2026, in which Claims 1-3, 5, 9-12, 14, & 18-20 have been amended. No claims have been added or cancelled. The amendments have been entered, and Claims 1-20 are currently pending.
Response to Arguments
Regarding the applicant’s traversal of the 35 U.S.C. 101 rejections of the previous office action, the applicant’s arguments filed January 15th, 2026 have been fully considered, and are unpersuasive.
Applicant asserts that the present amendments overcome the rejection because they further demonstrate the direction toward a technical improvement, specifically that producing a global set of hyperparameters, that when used in a federated learning process, tunes the loss from the federated learning environment, further stating that this provides significantly more than the abstract idea because limitations claim improvements to the technical field of computing and improvements to the functioning of a computer itself.
The examiner respectfully submits that the amendments to claim 1, which states that the devices are part “of a federated learning (FL) environment”, which involves “using the global set of HPs to tune loss from the FL environment, wherein the using the global set of HPs comprises orchestrating the FL training with the global set of HPs”, merely describe how federated learning ordinarily functions. In a typical federated learning environment, global parameters are calculated or “orchestrated” using the various devices in the federated learning environment in an effort to minimize/tune loss for the models. Therefore, this does not recite a new technical improvement to the functioning of a computer, but a generic link to the field of federated learning.
Further, other independent claims recite similar limitations and amendments as claim 1 and are also maintained. Further, all dependent claims depend on one of these claims and their subsequent rejections have been maintained as well. Therefore, the rejections under 35 U.S.C. 101 of the previous office action are maintained.
Regarding the applicant’s traversal of the 35 U.S.C. 102 rejections of the previous office action, the applicant’s arguments filed January 15th, 2026 have been fully considered, and are unpersuasive.
Applicant asserts that MOZAFFARI does not teach the ranking of hyperparameters, and that for this reason alone, the rejection is improper. The examiner respectfully submits that the ranking of the hyperparameters is explicitly shown in:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received (“submitted by clients”) parameter rankings are aggregated) to generate the global ranking of the next training epoch.”)
Further, each limitation is rejected in detail, in the rejections, as shown below. Therefore, the rejections under 35 U.S.C 102 are maintained.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea (mathematical process) without significantly more.
Regarding claim 1, in Step 1 of the 101-analysis set forth in MPEP 2106, the claim recites “A computer-implemented method”. A method is one of the four statutory categories of invention.
In Step 2a Prong 1 of the 101-analysis set forth in the MPEP 2106, the examiner has determined that the following limitations recite a process that, under the broadest reasonable interpretation, covers a mathematical process but for recitation of generic computer components:
“computing, based on the set of HP/rank value pairs, a global set of HPs from the HPO results for FL training” (This is directed toward a mathematical process made up of mathematical calculations, in view of the specification [00104-00105, 00108-00118 and 0156-0157], and as such, is directed toward an abstract idea (MPEP 2106).)
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mathematical process but for the recitation of generic computer components, then it falls within the mathematical process grouping of abstract ideas. According, the claim “recites” an abstract idea.
In Step 2a Prong 2 of the 101-analysis set forth in MPEP 2106, the examiner has
determined that the following additional elements do not integrate this judicial exception into a
practical application:
“A computer-implemented method” (Uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f)).)
“issuing a hyperparameter optimization (HPO) query to a plurality of computing devices of a federated learning (FL) environment” (Mere instructions to apply the judicial exception (MPEP 2106.05(f)).)
“receiving, from the plurality of computing devices, HPO results, wherein the HPO results include a set of hyperparameter (HP)/rank value pairs” (Adding insignificant extra-solution activity (mere data gathering) to the judicial exception (MPEP 2106.05(g)).)
“using the global set of HPs to tune loss from the FL environment, wherein the using the global set of HPs comprises orchestrating the FL training with the global set of HPs” (Generally linking the use of the judicial exception to a particular technological environment or field of use (MPEP 2106.05(h)).)
Since the claim as a whole, looking at the additional elements individually and in combination, does not contain any other additional elements that are indicative of integration into a practical application, the claim is “directed” to an abstract idea.
In Step 2b of the 101-analysis set forth in the 2019 PEG, the examiner has determined that the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
As discussed above, additional element (ii) recites use of a computer as a tool to perform the abstract idea, which is not indicative of significantly more. Additional element (iii) recites mere instructions to apply the judicial exception, which is not indicative of significantly more. Additional element (iv) recites insignificant extra-solution activities. Further, this element recites steps of receiving/transmitting data via a network, which has been determined by the courts to recite a well-understood, routine, and conventional activity, which is not indicative of significantly more (Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362). Additional element (v) recites generally linking the use of the judicial exception to a particular technological environment or field of use, which is not indicative of significantly more. Considering the additional elements individually and in combination, and the claim as a whole, the additional elements do not provide significantly more than the abstract idea. Therefore, the claim is not patent eligible.
Regarding claim 2, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 2 recites “outputting an indication of the global set of HPs to the plurality of computing devices.” (In step 2A, prong 2, this recites insignificant extra-solution activity (mere data output) to the judicial exception (MPEP 2106.05(g).) In step 2B, the courts have found steps that present output of data to be a well-understood, routine, and conventional activity, which is not indicative of significantly more (Presenting offers and gathering statistics, OIP Techs., 788 F.3d at 1362-63, 115 USPQ2d at 1092-93).)
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Regarding claim 3, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 3 recites the following additional mathematical process:
“wherein computing, based on the set of HP/rank value pairs, the global set of HPs from the HPO results for FL training includes generating a unified loss surface using the HP/rank value pairs of the received HPO results, wherein a minimizer of a predetermined unified loss surface function is the global set of HPs” (This is directed toward a mathematical process made up of mathematical calculations, in view of the specification [00104-00105], and as such, is directed toward an abstract idea (MPEP 2106).)
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Regarding claim 4, it is dependent upon claim 3, and thereby incorporates the limitations of, and corresponding analysis applied to claim 3. Further, claim 4 recites the following additional mathematical process:
“wherein the unified loss surface is generated by training a predetermined machine learning model, wherein the HPs of the HP/rank value pairs are used as inputs of the predetermined machine learning model, wherein the ranks of the HP/rank value pairs are used as targets of the predetermined machine learning model” (This is directed toward a mathematical process made up of mathematical calculations, in view of the specification [00108-00118], and as such, is directed toward an abstract idea (MPEP 2106).)
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Regarding claim 5, it is dependent upon claim 4, and thereby incorporates the limitations of, and corresponding analysis applied to claim 4. Further, claim 5 recites the following additional mathematical process:
“wherein the trained predetermined machine learning model is a loss surface model that is used to compute the global set of HPs” (This is directed toward a mathematical process made up of mathematical calculations, in view of the specification [00108-00118], and as such, is directed toward an abstract idea (MPEP 2106).)
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Regarding claim 6, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 6 recites the following additional mathematical process:
“wherein each of the HP/rank value pairs are generated, by an associated one of the computing devices, using a local dataset and a current model to run a plurality of HPO operations” (This is directed toward a mathematical process made up of mathematical calculations, in view of the specification [00155], and as such, is directed toward an abstract idea (MPEP 2106).)
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Regarding claim 7, it is dependent upon claim 6, and thereby incorporates the limitations of, and corresponding analysis applied to claim 6. Further, claim 7 recites the following additional mathematical process:
“wherein each of the HP/rank value pairs are generated, by an associated one of the computing devices, using a predetermined mapping function to map each local loss value resulting from running the HPO operations to a rank value, wherein the predetermined mapping function is configured to assign ranks according to locally defined loss ranges” (This is directed toward a mathematical process made up of mathematical calculations, in view of the specification [00155], and as such, is directed toward an abstract idea (MPEP 2106).)
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Regarding claim 8, it is dependent upon claim 6, and thereby incorporates the limitations of, and corresponding analysis applied to claim 6. Further, claim 8 recites the following additional mathematical process:
“wherein each of the HP/rank value pairs are generated, by an associated one of the computing devices, using a predetermined mapping function to map each local loss value resulting from running the HPO operations to a rank value, wherein the predetermined mapping function is configured to assign ranks according to global pre-defined loss ranges” (This is directed toward a mathematical process made up of mathematical calculations, in view of the specification [00155-00156], and as such, is directed toward an abstract idea (MPEP 2106).)
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Regarding claim 9, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 9 recites “wherein each of the plurality of computing devices includes a party within the federated learning (FL) environment, wherein the computing and the use of the global set of HPs is performed by an aggregator that issues the HPO query and receives the HPO results, wherein the orchestrating includes instructing, on the computing devices, execution of single federated training using the global set of HPs.” (In step 2a, prong 2, this recites generally linking the use of the judicial exception to a particular technological environment or field of use (MPEP 2106.05(h).) In step 2B, generally linking the use of the judicial exception to a particular technological environment or field of use is not indicative of significantly more.)
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Regarding claim 10, in Step 1 of the 101-analysis set forth in MPEP 2106, the claim recites “A computer program product, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions readable and/or executable by an aggregator”. In view of the specification at [0026] a computer program product comprising a computer readable storage medium of the described configuration is within one of the four statutory categories of invention.
In Step 2a Prong 1 of the 101-analysis set forth in the MPEP 2106, the examiner has determined that the following limitations recite a process that, under the broadest reasonable interpretation, covers a mathematical process but for recitation of generic computer components:
“compute… based on the set of HP/rank value pairs, a global set of HPs from the HPO results for federated learning (FL) training” (This is directed toward a mathematical process made up of mathematical calculations, in view of the specification [00104-00105, 00108-00118 and 0156-0157], and as such, is directed toward an abstract idea (MPEP 2106).)
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mathematical process but for the recitation of generic computer components, then it falls within the mathematical process grouping of abstract ideas. According, the claim “recites” an abstract idea.
In Step 2a Prong 2 of the 101-analysis set forth in MPEP 2106, the examiner has
determined that the following additional elements do not integrate this judicial exception into a
practical application:
“A computer program product, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions readable and/or executable by a aggregator to cause the aggregator to…” (Uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f)).)
“issue, by the computer, a hyperparameter optimization (HPO) query to a plurality of computing devices of a federated learning (FL) environment” (Mere instructions to apply the judicial exception (MPEP 2106.05(f)).)
“receive, by the aggregator, from the plurality of computing devices, HPO results, wherein the HPO results include a set of hyperparameter (HP)/rank value pairs” (Adding insignificant extra-solution activity (mere data gathering) to the judicial exception (MPEP 2106.05(g)).)
“…by the aggregator…” (Uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f)).)
“use, by the aggregator, the global set of HPs to tune loss from the FL environment, wherein the using the global set of HPs comprises orchestrating the FL training with the global set of HPs” (Generally linking the use of the judicial exception to a particular technological environment or field of use (MPEP 2106.05(h)).)
Since the claim as a whole, looking at the additional elements individually and in combination, does not contain any other additional elements that are indicative of integration into a practical application, the claim is “directed” to an abstract idea.
In Step 2b of the 101-analysis set forth in the 2019 PEG, the examiner has determined that the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
As discussed above, additional elements (ii) and (v) recite use of a computer as a tool to perform the abstract idea, which is not indicative of significantly more. Additional element (iii) recites mere instructions to apply the judicial exception, which is not indicative of significantly more. Additional element (iv) recites insignificant extra-solution activities. Further, this element recites steps of receiving/transmitting data via a network, which has been determined by the courts to recite a well-understood, routine, and conventional activity, which is not indicative of significantly more (Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362). Additional element vi recites generally linking the use of the judicial exception to a particular technological environment or field of use, which is not indicative of significantly more. Considering the additional elements individually and in combination, and the claim as a whole, the additional elements do not provide significantly more than the abstract idea. Therefore, the claim is not patent eligible.
Regarding claims 11-18, they are dependent upon claim 10, and thereby incorporate the limitations of, and corresponding analysis applied to claim 10. Further, claims 11-18 recite similar additional limitations as claims 2-9, respectively, and are rejected under the same rationale.
Regarding claim 19, in Step 1 of the 101-analysis set forth in MPEP 2106, the claim recites “A system, comprising: a processor;”. A system of the described configuration is within one of the four statutory categories of invention.
In Step 2a Prong 1 of the 101-analysis set forth in the MPEP 2106, the examiner has determined that the following limitations recite a process that, under the broadest reasonable interpretation, covers a mathematical process but for recitation of generic computer components:
“compute, based on the set of HP/rank value pairs, a global set of HPs from the HPO results for FL training” (This is directed toward a mathematical process made up of mathematical calculations, in view of the specification [00104-00105, 00108-00118 and 0156-0157], and as such, is directed toward an abstract idea (MPEP 2106).)
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mathematical process but for the recitation of generic computer components, then it falls within the mathematical process grouping of abstract ideas. According, the claim “recites” an abstract idea.
In Step 2a Prong 2 of the 101-analysis set forth in MPEP 2106, the examiner has
determined that the following additional elements do not integrate this judicial exception into a
practical application:
“A system, comprising: a processor; and logic integrated with the processor, executable by the processor, or integrated with and executable by the processor, the logic being configured to…” (Uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f)).)
“issue a hyperparameter optimization (HPO) query to a plurality of computing devices of a federated learning (FL) environment” (Mere instructions to apply the judicial exception (MPEP 2106.05(f)).)
“receive, from the plurality of computing devices, HPO results, wherein the HPO results include a set of hyperparameter (HP)/rank value pairs” (Adding insignificant extra-solution activity (mere data gathering) to the judicial exception (MPEP 2106.05(g)).)
“use the global set of HPs to tune loss from the FL environment, wherein the using the global set of HPs comprises orchestrating the FL training with the global set of HPs” (Generally linking the use of the judicial exception to a particular technological environment or field of use (MPEP 2106.05(h)).)
Since the claim as a whole, looking at the additional elements individually and in combination, does not contain any other additional elements that are indicative of integration into a practical application, the claim is “directed” to an abstract idea.
In Step 2b of the 101-analysis set forth in the 2019 PEG, the examiner has determined that the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
As discussed above, additional element (ii) recites use of a computer as a tool to perform the abstract idea, which is not indicative of significantly more. Additional element (iii) recites mere instructions to apply the judicial exception, which is not indicative of significantly more. Additional element (iv) recites insignificant extra-solution activities. Further, this element recites steps of receiving/transmitting data via a network, which has been determined by the courts to recite a well-understood, routine, and conventional activity, which is not indicative of significantly more (Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362). Additional element (v) recites generally linking the use of the judicial exception with a particular technological environment or field of use, which is not indicative of significantly more. Considering the additional elements individually and in combination, and the claim as a whole, the additional elements do not provide significantly more than the abstract idea. Therefore, the claim is not patent eligible.
Regarding claim 20, it is dependent upon claim 19, and thereby incorporate the limitations of, and corresponding analysis applied to claim 19. Further, claim 20 recites similar additional limitations as claim 2, and is rejected under the same rationale.
Claim Rejections - 35 USC § 102
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 the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claims 1-20 are rejected under 35 U.S.C. 102(a)(1) as being clearly anticipated by Mozaffari, H. et al. “FRL: Federated Rank Learning.” Available on 8 October 2021 (hereafter, MOZAFFARI).
Regarding claim 1, MOZAFFARI teaches “A computer-implemented method, comprising: issuing a hyperparameter optimization (HPO) query to a plurality of computing devices of a federated learning (FL) environment”:
([Abstract] “…Federated learning (FL) allows mutually untrusted clients to collaboratively train a common machine learning model without sharing their private/proprietary training data among each other… we propose Federated Rank Learning (FRL). FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) …”) This citation shows that MOZAFFARI teaches a method to improve hyperparameter optimization (HPO) for “federated learning”, which is a known method in the art. Federated learning is, by definition, a setup that uses hyperparameter optimizations queried by the central aggregator to a plurality of computing devices connected to the “aggregator,” which then receives data from those devices in order to aggregate them into a single “global model,” which is also shown in the citation below.
And further:
([Introduction, Paragraph 1] “Federated Learning (FL) is an emerging learning paradigm, where mutually untrusted clients (e.g., Android devices) collaborate to train a shared model, called the global model, without explicitly sharing their local training data. FL training involves a server (e.g., a Google server) who repeatedly collects model updates that the clients compute using their local private data (a query is issued to optimize hyperparameters), aggregates the clients’ updates using an aggregation rule (AGR), and finally uses the aggregated updates to tune the jointly trained model (called the global model), which is broadcast to a subset of the clients at the end of each FL training round.”)
Further, MOZAFFARI teaches “receiving, from the plurality of computing devices, HPO results, wherein the HPO results include a set of hyperparameter (HP)/rank value pairs”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received (“submitted by clients”) parameter rankings are aggregated) to generate the global ranking of the next training epoch.”)
Further, MOZAFFARI teaches “computing, based on the set of HP/rank value pairs, a global set of HPs from the HPO results for FL training”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received parameter rankings are aggregated) to generate the global ranking of the next training epoch (the global set of hyperparameters is generated for further FL training).”)
Further, MOZAFFARI teaches “using the global set of HPs to tune loss from the FL environment, wherein the using the global set of HPS comprises orchestrating the FL training with the global set of HPs”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received parameter rankings are aggregated) to generate the global ranking of the next training epoch (the global set of hyperparameters is generated for further FL training, which is to say, that the global set will be dispersed to clients for further training, as is typical of federated learning systems.).”)
Regarding claim 2, MOZAFFARI teaches the limitations of claim 1. Further, MOZAFFARI teaches “outputting an indication of the global set of HPS to the plurality of computing devices”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received parameter rankings are aggregated) to generate the global ranking of the next training epoch (the global set of hyperparameters is generated for further FL training, which is to say, that the global set will be dispersed to clients for further training, as is typical of federated learning systems.).”)
And further:
([Page 5, Section 4.2, Paragraph 1] “In the tth round, FRL server randomly selects n clients among total N clients, and shares the global rankings Rtg with them. Each of the selected n clients locally reconstructs the weights 0w’s and scores 0s’s using SEED (Algorithm 2 line 9).”) Here, we explicitly see the global set of hyperparameter/rank value pairs being output to a plurality of devices.
Regarding claim 3, MOZAFFARI teaches the limitations of claim 1. Further, MOZAFFARI teaches “wherein computing, based on the set of HP/rank value pairs, the global set of HPs from the HPO results for FL training includes generating a unified loss surface using the HP/rank value pairs of the received HPO results, wherein a minimizer of a predetermined unified loss surface function is the global set of HPs”:
([Pages 4-5, Section 4 Federated Rank Learning: Design, Paragraphs 2-3, FRL objective] “The optimization problem of FRL (Federated Rank Learning) is to find a global ranking Rg which produces a global binary mask m that minimizes the average loss of all of clients with that subnetwork (the average loss of all clients is equivalent to a “unified loss surface” being generated) (0w
PNG
media_image1.png
12
14
media_image1.png
Greyscale
m). FRL aims to solve:
PNG
media_image2.png
77
344
media_image2.png
Greyscale
where N is the total number of participating clients and Li is the loss function for the ith client.
PNG
media_image3.png
18
16
media_image3.png
Greyscale
shows the importance of the ith client in empirical risk minimization;
PNG
media_image3.png
18
16
media_image3.png
Greyscale
=
PNG
media_image4.png
22
14
media_image4.png
Greyscale
gives same importance to all the participating clients. m is the final mask that contains the edges of top k ranks, i.e., edges in top k ranks (layer-wise) get ’1’ in the binary mask, and others get ’0’ in the mask. 0w
PNG
media_image1.png
12
14
media_image1.png
Greyscale
m shows the subnetwork inside the random and fixed weights 0w that all clients unanimously vote for.”) In this citation 0w is the weights/hyperparameters of the clients, k represents ranks, and the loss function of each clients is averaged to generate a unified loss surface using those hyperparameter/rank value pairs and the equation shows that the unified loss surface function is the global set of hyperparameters, as also shown in the previously cited abstract citation added below.
And further:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received parameter rankings are aggregated) to generate the global ranking of the next training epoch (the global set of hyperparameters is generated for further FL training, which is to say, that the global set will be dispersed to clients for further training, as is typical of federated learning systems.). We additionally see that the parameters that are ranked and aggregated by the above cited math, are used as the global set of hyperparameters.”)
Regarding claim 4, MOZAFFARI teaches the limitations of claim 3. Further, MOZAFFARI teaches “wherein the unified loss surface is generated by training a predetermined machine learning model”:
([Pages 4-5, Section 4 Federated Rank Learning: Design, Paragraphs 2-3, FRL objective] “The optimization problem of FRL (Federated Rank Learning) is to find a global ranking Rg which produces a global binary mask m that minimizes the average loss of all of clients with that subnetwork (the average loss of all clients is equivalent to a “unified loss surface” being generated) (0w
PNG
media_image1.png
12
14
media_image1.png
Greyscale
m).
And further:
([Abstract] “…Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) (a predetermined machine learning model) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch to generate the global ranking of the next training epoch…”)
And further:
([Page 2, Col. 2, Paragraph 1] “…FRL uses a novel learning paradigm called super-masks training [34, 44] to create edge rankings, which, as we will show, allows FRL to reduce communication costs while achieving significantly stronger robustness. Specifically, in FRL, clients collaborate to find a subnetwork within a randomly initialized neural network which we call the supernetwork (this is in contrast to conventional FL where clients collaborate to train a neural network). The goal of training in FRL is to collaboratively rank the supernetwork’s edges based on the importance of each edge and find a global ranking (Here, the neural network (predetermined machine learning model) is explicitly trained, in order to generate the unified loss surface). The global ranking can be converted to a supermask, which is a binary mask of 1’s and 0’s, that is superimposed on the random neural network (the supernetwork) to obtain the final subnetwork. For example, in our experiments, the final subnetwork is constructed using the top 50% of all edges. The subnetwork is then used for downstream tasks, e.g., image classification, hence it is equivalent to the global model in conventional FL. Note that in entire FRL training, weights of the supernetwork do not change.”) This citation describes using the neural network (predetermined machine learning model) by inputting the rankings (hyperparameter/rank value pairs) into it, to train it.
Further, MOZAFFARI teaches “wherein the HPs of the HP/rank value pairs are used as inputs of the predetermined machine learning model”:
([Page 2, Col. 2, Paragraph 1] “…FRL uses a novel learning paradigm called super-masks training [34, 44] to create edge rankings, which, as we will show, allows FRL to reduce communication costs while achieving significantly stronger robustness. Specifically, in FRL, clients collaborate to find a subnetwork within a randomly initialized neural network which we call the supernetwork (this is in contrast to conventional FL where clients collaborate to train a neural network). The goal of training in FRL is to collaboratively rank the supernetwork’s edges based on the importance of each edge and find a global ranking. The global ranking can be converted to a supermask, which is a binary mask of 1’s and 0’s, that is superimposed on the random neural network (the supernetwork) to obtain the final subnetwork. For example, in our experiments, the final subnetwork is constructed using the top 50% of all edges. The subnetwork is then used for downstream tasks, e.g., image classification, hence it is equivalent to the global model in conventional FL. Note that in entire FRL training, weights of the supernetwork do not change.”) This citation describes using the neural network (predetermined machine learning model) by inputting the rankings (hyperparameter/rank value pairs) into it.
Further, MOZAFFARI teaches “wherein the ranks of the HP/rank value pairs are used as targets of the predetermined machine learning model”:
([Page 2, Col. 2, Paragraph 1] “…FRL uses a novel learning paradigm called super-masks training [34, 44] to create edge rankings, which, as we will show, allows FRL to reduce communication costs while achieving significantly stronger robustness. Specifically, in FRL, clients collaborate to find a subnetwork within a randomly initialized neural network which we call the supernetwork (this is in contrast to conventional FL where clients collaborate to train a neural network). The goal of training in FRL is to collaboratively rank the supernetwork’s edges based on the importance of each edge and find a global ranking. The global ranking can be converted to a supermask, which is a binary mask of 1’s and 0’s, that is superimposed on the random neural network (the supernetwork) to obtain the final subnetwork. For example, in our experiments, the final subnetwork is constructed using the top 50% of all edges. The subnetwork is then used for downstream tasks, e.g., image classification, hence it is equivalent to the global model in conventional FL. Note that in entire FRL training, weights of the supernetwork do not change.”) This citation describes using the neural network (predetermined machine learning model) by inputting the rankings (hyperparameter/rank value pairs) into it. The hyperparameter/rank value pairs of the subnetwork is shown to be the target of the machine learning process.
Regarding claim 5, MOZAFFARI in view of X teaches the limitations of claim 4. Further, MOZAFFARI teaches “wherein the trained predetermined machine learning model is a loss surface model that is used to compute the global set of HPs”:
([Pages 4-5, Section 4 Federated Rank Learning: Design, Paragraphs 2-3, FRL objective] “The optimization problem of FRL (Federated Rank Learning) is to find a global ranking Rg which produces a global binary mask m that minimizes the average loss of all of clients with that subnetwork (the average loss of all clients is equivalent to a “unified loss surface” being generated) (0w
PNG
media_image1.png
12
14
media_image1.png
Greyscale
m).
And further:
([Abstract] “…Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) (a predetermined machine learning model) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch to generate the global ranking of the next training epoch…”) The global hyperparameters are generated from the loss surface model.
And further:
([Page 2, Col. 2, Paragraph 1] “…FRL uses a novel learning paradigm called super-masks training [34, 44] to create edge rankings, which, as we will show, allows FRL to reduce communication costs while achieving significantly stronger robustness. Specifically, in FRL, clients collaborate to find a subnetwork within a randomly initialized neural network which we call the supernetwork (this is in contrast to conventional FL where clients collaborate to train a neural network). The goal of training in FRL is to collaboratively rank the supernetwork’s edges based on the importance of each edge and find a global ranking (Here, the neural network (predetermined machine learning model) is explicitly trained, in order to generate the unified loss surface). The global ranking can be converted to a supermask, which is a binary mask of 1’s and 0’s, that is superimposed on the random neural network (the supernetwork) to obtain the final subnetwork. For example, in our experiments, the final subnetwork is constructed using the top 50% of all edges. The subnetwork is then used for downstream tasks, e.g., image classification, hence it is equivalent to the global model in conventional FL. Note that in entire FRL training, weights of the supernetwork do not change.”) This citation describes using the neural network (predetermined machine learning model) by inputting the rankings (hyperparameter/rank value pairs) into it, to train it.
Regarding claim 6, MOZAFFARI teaches the limitations of claim 1. Further, MOZAFFARI teaches “wherein each of the HP/rank value pairs are generated, by an associated one of the computing devices, using a local dataset and a current model to run a plurality of HPO operations”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL (Federated Learning) to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data (This explicitly shows that the clients of the federated learning system use local data). The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received parameter rankings are aggregated) to generate the global ranking of the next training epoch (the global set of hyperparameters is generated for further FL training).”)
And further:
([Page 2, Col. 2, Paragraph 1] “…FRL uses a novel learning paradigm called super-masks training [34, 44] to create edge rankings, which, as we will show, allows FRL to reduce communication costs while achieving significantly stronger robustness. Specifically, in FRL, clients collaborate to find a subnetwork within a randomly initialized neural network which we call the supernetwork (this is in contrast to conventional FL where clients collaborate to train a neural network). The goal of training in FRL is to collaboratively rank the supernetwork’s edges based on the importance of each edge and find a global ranking (Here, the neural network (predetermined machine learning model) is explicitly trained, in order to generate the unified loss surface). The global ranking can be converted to a supermask, which is a binary mask of 1’s and 0’s, that is superimposed on the random neural network (the supernetwork) to obtain the final subnetwork. For example, in our experiments, the final subnetwork is constructed using the top 50% of all edges. The subnetwork is then used for downstream tasks, e.g., image classification, hence it is equivalent to the global model in conventional FL. Note that in entire FRL training, weights of the supernetwork do not change.”) This citation describes using the neural network (predetermined machine learning model) by inputting the rankings (hyperparameter/rank value pairs) into it, to train it, which corresponds to running a plurality of HPO operations using the local dataset by an associated one of the plurality of devices..
Regarding claim 7, MOZAFFARI teaches the limitations of claim 6. Further, MOZAFFARI teaches “wherein each of the HP/rank value pairs are generated, by an associated one of the computing devices, using a predetermined mapping function to map each local loss value resulting from running the HPO operations to a rank value, wherein the predetermined mapping function is configured to assign ranks according to locally defined loss ranges”:
([Page 2, Col. 2, Paragraph 1] “…FRL uses a novel learning paradigm called super-masks training [34, 44] to create edge rankings (rank values), which, as we will show, allows FRL to reduce communication costs while achieving significantly stronger robustness. Specifically, in FRL, clients collaborate to find a subnetwork within a randomly initialized neural network which we call the supernetwork (this is in contrast to conventional FL where clients collaborate to train a neural network). The goal of training in FRL is to collaboratively rank the supernetwork’s edges based on the importance of each edge and find a global ranking. The global ranking can be converted to a supermask, which is a binary mask of 1’s and 0’s, that is superimposed on the random neural network (the supernetwork) to obtain the final subnetwork (a predetermined mapping function is applied to map each local loss value resulting from running the HPO operations to a rank value, based on locally defined loss ranges). For example, in our experiments, the final subnetwork is constructed using the top 50% of all edges. The subnetwork is then used for downstream tasks, e.g., image classification, hence it is equivalent to the global model in conventional FL. Note that in entire FRL training, weights of the supernetwork do not change.”) In the above citation, we see that FRL computes a ranking by running the “edge popup algorithm” …
([Page 2, Col. 2, Paragraph 2] “…Each FRL client will use the edge popup algorithm… and their data to compute their local rankings (the edge popup algorithm aims at learning which edges in a supernetwork are more important over the other edges by minimizing the loss of the subnetwork on their local data) …”)
…so that edges are ordered by importance, meaning they are determined by minimizing local subnetwork loss. Since FRL’s ranking is derived from minimizing local loss, making ranks a function of local loss-based importance, this aligns with the idea of mapping local loss to rank. This process and algorithm, therefore, is the predetermined mapping function, which uses the subnetwork for locally defined loss ranges.
Regarding claim 8, MOZAFFARI teaches the limitations of claim 6. Further, MOZAFFARI teaches “wherein each of the HP/rank value pairs are generated, by an associated one of the computing devices, using a predetermined mapping function to map each local loss value resulting from running the HPO operations to a rank value, wherein the predetermined mapping function is configured to assign ranks according to global pre-defined loss ranges”:
([Page 2, Col. 2, Paragraph 1] “…FRL uses a novel learning paradigm called super-masks training [34, 44] to create edge rankings, which, as we will show, allows FRL to reduce communication costs while achieving significantly stronger robustness. Specifically, in FRL, clients collaborate to find a subnetwork within a randomly initialized neural network which we call the supernetwork (this is in contrast to conventional FL where clients collaborate to train a neural network). The goal of training in FRL is to collaboratively rank the supernetwork’s edges based on the importance of each edge and find a global ranking. The global ranking can be converted to a supermask, which is a binary mask of 1’s and 0’s, that is superimposed on the random neural network (the supernetwork) to obtain the final subnetwork (a predetermined mapping function is applied to map each local loss value resulting from running the HPO operations to a rank value, based on locally defined loss ranges). For example, in our experiments, the final subnetwork is constructed using the top 50% of all edges. The subnetwork is then used for downstream tasks, e.g., image classification, hence it is equivalent to the global model in conventional FL. Note that in entire FRL training, weights of the supernetwork do not change.”) (a predetermined mapping function is applied to map each local loss value resulting from running the HPO operations to a rank value, based on locally defined loss ranges). For example, in our experiments, the final subnetwork is constructed using the top 50% of all edges. The subnetwork is then used for downstream tasks, e.g., image classification, hence it is equivalent to the global model in conventional FL. Note that in entire FRL training, weights of the supernetwork do not change.”) In the above citation, we see that FRL computes a ranking by running the “edge popup algorithm” …
([Page 2, Col. 2, Paragraph 2] “…Each FRL client will use the edge popup algorithm… and their data to compute their local rankings (the edge popup algorithm aims at learning which edges in a supernetwork are more important over the other edges by minimizing the loss of the subnetwork on their local data) …”)
…so that edges are ordered by importance, meaning they are determined by minimizing local subnetwork loss. Since FRL’s ranking is derived from minimizing local loss, making ranks a function of local loss-based importance, this aligns with the idea of mapping local loss to rank. This process and algorithm, therefore, is the predetermined mapping function, which uses the subnetwork for locally defined loss ranges. In the federated learning process, briefly described in [Page 1, Introduction, Paragraph 1], the values generated from local pre-defined loss ranges will then be used in the aggregator by way of the model updates of each client being sent to the aggregator to be combined, to send further updates to all the clients, where the updates will become global pre-defined loss ranges by way of aggregation.
Regarding claim 9, MOZAFFARI teaches the limitations of claim 1. Further, MOZAFFARI teaches “wherein each of the plurality of computing devices includes a party within the FL environment, wherin the computing and the use of the global set of HPs is performed by an aggregator that issues the HPO query and receives the HPO results, wherein the orchestrating includes instructing, on the computing devices, execution of single federated training using the global set of HPs”:
([Abstract] “…Federated learning (FL) allows mutually untrusted clients to collaboratively train a common machine learning model without sharing their private/proprietary training data among each other… we propose Federated Rank Learning (FRL). FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) …”) This citation shows that MOZAFFARI teaches a method to improve hyperparameter optimization (HPO) for “federated learning”. Federated Learning is a known method in the art. Federated learning is, by definition, a setup that uses hyperparameter optimizations queried by the central aggregator to a plurality of computing devices connected to the “aggregator,” which then receives data from those devices in order to aggregate them into a single “global model,” which is also shown in the citation below.
And further:
([Introduction, Paragraph 1] “Federated Learning (FL) is an emerging learning paradigm, where mutually untrusted clients (parties) (e.g., Android devices) collaborate to train a shared model, called the global model, without explicitly sharing their local training data. FL training involves a server (e.g., a Google server) who repeatedly collects model updates that the clients compute using their local private data (a query is issued to optimize hyperparameters), aggregates the clients’ updates using an aggregation rule (AGR), and finally uses the aggregated updates to tune the jointly trained model (called the global model), which is broadcast to a subset of the clients at the end of each FL training round.”)
Regarding claim 10, MOZAFFARI teaches “A computer program product, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions readable and/or executable by an aggregator to cause the aggregator to: issue, by the aggregator, a hyperparameter optimization (HPO) query to a plurality of computing devices of a federated learning (FL) environment”:
([Abstract] “…Federated learning (FL) allows mutually untrusted clients to collaboratively train a common machine learning model without sharing their private/proprietary training data among each other… we propose Federated Rank Learning (FRL) (A computer program product). FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) …”) This citation shows that MOZAFFARI teaches a method to improve hyperparameter optimization (HPO) for “federated learning”, which is a known method in the art. Federated learning is, by definition, a setup that uses hyperparameter optimizations queried by the central aggregator to a plurality of computing devices connected to the “aggregator,” which then receives data from those devices in order to aggregate them into a single “global model,” which is also shown in the citation below.
And further:
([Introduction, Paragraph 1] “Federated Learning (FL) is an emerging learning paradigm, where mutually untrusted clients (e.g., Android devices (computer readable storage medium having instructions embodied therewith)) collaborate to train a shared model, called the global model, without explicitly sharing their local training data. FL training involves a server (e.g., a Google server) who repeatedly collects model updates that the clients compute using their local private data (a query is issued to optimize hyperparameters), aggregates the clients’ updates using an aggregation rule (AGR), and finally uses the aggregated updates to tune the jointly trained model (called the global model), which is broadcast to a subset of the clients at the end of each FL training round.”)
Further, MOZAFFARI teaches “receive, by the aggregator, from the plurality of computing devices, HPO results, wherein the HPO results include a set of hyperparameter (HP)/rank value pairs”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received (“submitted by clients”) parameter rankings are aggregated) to generate the global ranking of the next training epoch.”)
Further, MOZAFFARI teaches “compute, by the aggregator, based on the set of HP/rank value pairs, a global set of HPs from the HPO results for FL training”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received parameter rankings are aggregated) to generate the global ranking of the next training epoch (the global set of hyperparameters is generated for further FL training).”)
Further, MOZAFFARI teaches “use, by the aggregator, the global set of HPs to tune loss from the FL environment, wherein the global set of HPs comprises orchestrating the FL training with the global set of HPs”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received parameter rankings are aggregated) to generate the global ranking of the next training epoch (the global set of hyperparameters is generated for further FL training, which is to say, that the global set will be dispersed to clients for further training, as is typical of federated learning systems.).”)
And further:
([Page 5, Section 4.2, Paragraph 1] “In the tth round, FRL server randomly selects n clients among total N clients, and shares the global rankings Rtg with them. Each of the selected n clients locally reconstructs the weights 0w’s and scores 0s’s using SEED (Algorithm 2 line 9).”) Here, we explicitly see the global set of hyperparameter/rank value pairs being output to a plurality of devices.
Regarding claims 11-18, MOZAFFARI teaches the limitations of claim 10. Further, claims 11-18 recite similar additional limitations as claims 2-9, respectively, and are rejected under the same rationale.
Regarding claim 19, MOZAFFARI teaches “A system, comprising: a processor; and logic integrated with the processor, executable by the processor, or integrated with and executable by the processor, the logic being configured to: issue a hyperparameter optimization (HPO) query to a plurality of computing devices of a federated learning (FL) environment”:
([Abstract] “…Federated learning (FL) allows mutually untrusted clients to collaboratively train a common machine learning model without sharing their private/proprietary training data among each other… we propose Federated Rank Learning (FRL). FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) …”) This citation shows that MOZAFFARI teaches a method to improve hyperparameter optimization (HPO) for “federated learning”, which is a known method in the art. Federated learning is, by definition, a setup that uses hyperparameter optimizations queried by the central aggregator to a plurality of computing devices connected to the “aggregator,” which then receives data from those devices in order to aggregate them into a single “global model,” which is also shown in the citation below.
And further:
([Introduction, Paragraph 1] “Federated Learning (FL) is an emerging learning paradigm, where mutually untrusted clients (e.g., Android devices (a system with a processor and logic integrated with the processor)) collaborate to train a shared model, called the global model, without explicitly sharing their local training data. FL training involves a server (e.g., a Google server) who repeatedly collects model updates that the clients compute using their local private data (a query is issued to optimize hyperparameters), aggregates the clients’ updates using an aggregation rule (AGR), and finally uses the aggregated updates to tune the jointly trained model (called the global model), which is broadcast to a subset of the clients at the end of each FL training round.”)
Further, MOZAFFARI teaches “receive, from the plurality of computing devices, HPO results, wherein the HPO results include a set of hyperparameter (HP)/rank value pairs”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received (“submitted by clients”) parameter rankings are aggregated) to generate the global ranking of the next training epoch.”)
Further, MOZAFFARI teaches “compute, based on the set of HP/rank value pairs, a global set of HPs from the HPO results for FL training”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received parameter rankings are aggregated) to generate the global ranking of the next training epoch (the global set of hyperparameters is generated for further FL training).”)
Further, MOZAFFARI teaches “use the global set of HPs to tune loss from the FL environment, wherein the using the global set of HPs comprises orchestrating the FL training with the global set of HPs”:
([Abstract] “FRL reduces the space of client updates from model parameter updates (a continuous space of float numbers) in standard FL to the space of parameter rankings (a discrete space of integer values) (This explicitly shows that instead of simply sending parameters from each client, MOZAFFARI sends rankings of parameters, which correlates to hyperparameter/rank value pairs) … FRL leverage ideas from recent super-masks training mechanisms. Specifically, FRL clients rank the parameters of a randomly initialized neural network (provided by the server) based on their local training data. The FRL server uses a voting mechanism to aggregate the parameter rankings submitted by clients in each training epoch (the received parameter rankings are aggregated) to generate the global ranking of the next training epoch (the global set of hyperparameters is generated for further FL training, which is to say, that the global set will be dispersed to clients for further training, as is typical of federated learning systems.).”)
And further:
([Page 5, Section 4.2, Paragraph 1] “In the tth round, FRL server randomly selects n clients among total N clients, and shares the global rankings Rtg with them. Each of the selected n clients locally reconstructs the weights 0w’s and scores 0s’s using SEED (Algorithm 2 line 9).”) Here, we explicitly see the global set of hyperparameter/rank value pairs being output to a plurality of devices.
Regarding claim 20, MOZAFFARI teaches the limitations of claim 19. Further, claim 20 recites similar additional limitations as claim 2, and is rejected under the same rationale.
Conclusion
THIS ACTION IS MADE FINAL. 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 MATTHEW LEE LEWIS whose telephone number is (571)272-1906. The examiner can normally be reached Monday: 12:00PM - 4:00PM and Tuesday - Friday: 12:00PM - 9PM.
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, Tamara Kyle can be reached at (571)272-4241. 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.
/Matthew Lee Lewis/Examiner, Art Unit 2144
/TAMARA T KYLE/Supervisory Patent Examiner, Art Unit 2144
Prosecution Insights