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 .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on October 27, 2023, and March 12, 2026, and May 21, 2026 have been considered by the examiner.
Claim Interpretation
The following is a quotation of 35 U.S.C. 112(f):
(f) Element in Claim for a Combination. – An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The following is a quotation of pre-AIA 35 U.S.C. 112, sixth paragraph:
An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The claims in this application are given their broadest reasonable interpretation using the plain meaning of the claim language in light of the specification as it would be understood by one of ordinary skill in the art. The broadest reasonable interpretation of a claim element (also commonly referred to as a claim limitation) is limited by the description in the specification when 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is invoked.
As explained in MPEP § 2181, subsection I, claim limitations that meet the following three-prong test will be interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph:
(A) the claim limitation uses the term “means” or “step” or a term used as a substitute for “means” that is a generic placeholder (also called a nonce term or a non-structural term having no specific structural meaning) for performing the claimed function;
(B) the term “means” or “step” or the generic placeholder is modified by functional language, typically, but not always linked by the transition word “for” (e.g., “means for”) or another linking word or phrase, such as “configured to” or “so that”; and
(C) the term “means” or “step” or the generic placeholder is not modified by sufficient structure, material, or acts for performing the claimed function.
Use of the word “means” (or “step”) in a claim with functional language creates a rebuttable presumption that the claim limitation is to be treated in accordance with 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph. The presumption that the claim limitation is interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is rebutted when the claim limitation recites sufficient structure, material, or acts to entirely perform the recited function.
Absence of the word “means” (or “step”) in a claim creates a rebuttable presumption that the claim limitation is not to be treated in accordance with 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph. The presumption that the claim limitation is not interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is rebutted when the claim limitation recites function without reciting sufficient structure, material or acts to entirely perform the recited function.
Claim limitations in this application that use the word “means” (or “step”) are being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action. Conversely, claim limitations in this application that do not use the word “means” (or “step”) are not being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action.
This application includes one or more claim limitations that do not use the word “means,” but are nonetheless being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, because the claim limitations use a generic placeholder that is coupled with functional language without reciting sufficient structure to perform the recited function and the generic placeholder is not preceded by a structural modifier.
Such claim limitations are found in claim 31 for a receiver configured to receive, from each of a plurality of worker computing devices, … a determiner configured to determine a common portion of the parts of trained ML models … a generator configured to generate an updated common portion of the ML model… and a transmitter configured to initiate transmission… , Because these claim limitations are being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, they are being interpreted to cover the corresponding structure described in the specification in [0032 , 0038] from US PG Pub. US20240370737A1 as performing the claimed function, and equivalents thereof.
If applicant does not intend to have these limitations interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, applicant may: (1) amend the claim limitations to avoid them being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph (e.g., by reciting sufficient structure to perform the claimed function); or (2) present a sufficient showing that the claim limitations recite sufficient structure to perform the claimed function so as to avoid them being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
The term “maximum” in ‘a maximum common portion of the ML model…’ in claims 2 and 17, is a relative term which renders the claim indefinite. The term “maximum” is not defined by the claim, the specification does not provide a standard for ascertaining the requisite degree, and one of ordinary skill in the art would not be reasonably apprised of the scope of the invention. Appropriate correction is required.
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 1-16, 17-23, 28, 31, and 32 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea (mental process) without significantly more.
Claim 1:
Regarding claim 1, in step 1 of the 101-analysis set forth in MPEP 2106, the claim recites
A method for developing a Machine Learning, ML, model, the method comprising:
receiving, at a leader computing device from each of a plurality of worker computing devices, weights and model architecture information for part of a trained ML model;
determining, at the leader computing device, a common portion of the parts of trained ML models that is useable by all of the plurality of worker computing devices;
generating, at the leader computing device, an updated common portion of the ML model using the common portion of the parts of trained ML models and the weights and model architecture information received from each of the plurality of worker computing devices; and initiating transmission of the generated updated common portion of the ML model from the leader computing device to the worker computing devices, and 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 mental process but for recitation of generic computer components:
determining, …, a common portion of the parts of trained ML models that is useable by all of the plurality of worker computing devices; (mental process, a person can mentally evaluate and determine a common portion of the trained models used by worker devices, see MPEP 2106.04(a)(2)(III)),
generating, …, an updated common portion of the ML model, (mental process, a person can mentally evaluate and create an updated common portion of a model, see MPEP 2106.04(a)(2)(III))
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, 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:
receiving, at a leader computing device from each of a plurality of worker computing devices, weights and model architecture information for part of a trained ML model; (In step 2A, prong 2, this recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g)),
A method for developing a Machine Learning, ML, model, the method comprising …using the common portion of the parts of trained ML models and the weights and model architecture information received from each of the plurality of worker computing devices; (This recites mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
at the leader computing device, (In step 2A, prong 2, this recites a generic computer being used as a tool – see MPEP 2106.05(f)),
and initiating transmission of the generated updated common portion of the ML model from the leader computing device to the worker computing devices, (In step 2A, prong 2, this recites mere data sending and transmitting, which is considered insignificant extra-solution activity – see MPEP 2106.05(g)),
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 iv and v recite mere instructions to apply the judicial exception using generic computer components or a generic computer being used as a tool, which are not indicative of significantly more. The additional elements iii and vi recite mere data gathering, and are considered insignificant extra-solution activities. In step 2B, these insignificant extra-solution activities are well understood routine and conventional activities, which include receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016), – see MPEP 2106.05(d) (II)(i)),
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.
Claim 2:
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 the following abstract idea:
determining, …, a maximum common portion of the ML model that is useable by all of the plurality of worker computing devices, using the ML model architecture privacy information; (This recites a mental process, a person can mentally evaluate and create a maximum common portion of a model, see MPEP 2106.04(a)(2)(III)),
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
Further, claim 2 recites the following additional elements:
The method of claim 1, further comprising, prior to receiving the weights and model architecture information for the parts of trained ML models from the worker computing devices: receiving, at the leader computing device from the plurality of worker computing devices, ML model architecture privacy information; (In step 2A, prong 2, this recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016) – see MPEP 2106.05(d) (II)(i),
and initiating transmission of initialization information for the maximum common portion of the ML model to all of the plurality of worker computing devices. (In step 2A, prong 2, this recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016) – see MPEP 2106.05(d) (II)(i),
at the leader computing device, (In step 2A, prong 2, this recites a generic computer being used as a tool – see MPEP 2106.05(f)), In step 2B, this also recites a generic computer being used as a tool – see MPEP 2106.05(f)),
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.
Claim 3:
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 abstract idea:
The method of claim 1, wherein the step of determining the common portion of the ML models comprises detecting a variation in a model architecture from among the model architectures of the trained ML models, (This recites a mental process, a person can mentally evaluate and detect a variation in model architecture, see MPEP 2106.04(a)(2)(III)),
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
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.
Claim 4:
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 element:
the method of claim 3, wherein, if the variation is detected, the updated common portion of the ML model distributed to the worker computing devices comprises weights and model architecture information, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
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.
Claim 5:
Regarding claim 5, it is dependent upon claim 3, and thereby incorporates the limitations of, and corresponding analysis applied to claim 3. Further, claim 5 recites the following additional element:
The method of claim 3, wherein, if the variation is not detected, the updated common portion of the ML model distributed to the worker computing devices comprises weights, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
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.
Claim 6:
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 element:
The method of claim 1, further comprising receiving, at the leader computing device from each of the plurality of worker computing devices, metadata. (In step 2A, prong 2, this recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016) – see MPEP 2106.05(d) (II)(i),
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.
Claim 7:
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 element:
The method of claim 6, wherein the metadata is used by the leader computing device when determining the common portion of the parts of trained ML models that may be used by all of the plurality of worker computing devices, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
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.
Claim 8:
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 elements:
The method of claim 6, wherein the metadata comprises one or more of: resource availability information; validation information; updated model architecture privacy information; (In step 2A, prong 2, this recites an indication to a field of use or technological environment – see MPEP 2106.05(h)), (In step 2B, this also recites a field of use or technological environment – see MPEP 2106.05(h)),
and notification of a variation in a model architecture from among the model architectures of the trained ML models, (In step 2A, prong 2, this recites an indication to a field of use or technological environment – see MPEP 2106.05(h)), (In step 2B, this also recites a field of use or technological environment – see MPEP 2106.05(h)),
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.
Claim 9:
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 the following additional element:
The method of claim 1, wherein the updated common portion of the ML model is generated using federated averaging, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
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.
Claim 10:
Regarding claim 10, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 10 recites the following additional element:
The method of claim 1, wherein the weights and model architecture information the leader computing device receives from each of the plurality of worker computing devices is weights and model architecture information of part of a trained ML model that has been trained by a given worker computing device using data private to the given worker computing device, (In step 2A, prong 2, this recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016) – see MPEP 2106.05(d) (II)(i),
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.
Claim 11:
Regarding claim 11, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 11 recites the following additional element:
The method of claim 1, wherein the updated common portion of the ML model comprises the output layer of the ML model, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
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.
Claim 12:
Regarding claim 12, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 12 recites the following additional element:
The method of claim 1, further comprising, by each of the worker computing devices, using the updated common portion of the ML model as part of a worker specific ML model, wherein each worker specific ML model is used to provide suggested actions for an environment, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
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.
Claim 13:
Regarding claim 13, it is dependent upon claim 12, and thereby incorporates the limitations of, and corresponding analysis applied to claim 12. Further, claim 13 recites the following additional element:
The method of claim 12, wherein the environment is one or more base stations in a communications network, or wherein the environment is one or more servers in a data center, (In step 2A, prong 2, this recites an indication to a field of use or technological environment – see MPEP 2106.05(h)), (In step 2B, this also recites a field of use or technological environment – see MPEP 2106.05(h)),
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.
Claim 14:
Regarding claim 14, it is dependent upon claim 12, and thereby incorporates the limitations of, and corresponding analysis applied to claim 12. Further, claim 14 recites the following additional element:
The method of claim 12, further comprising modifying the environment based on the suggested actions, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
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.
Claim 15:
Regarding claim 15, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 15 recites the following additional element:
The method of claim 1, wherein the trained ML models are trained local ML models and the updated common portion is an updated common global portion, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
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.
Claim 16:
Regarding claim 16, in step 1 of the 101-analysis set forth in MPEP 2106, the claim recites “A leader computing device configured to develop a Machine Learning, ML, model, the leader computing device comprising processing circuitry and a memory containing instructions executable by the processing circuitry, whereby the leader computing device is operable to: receive, from each of a plurality of worker computing devices, weights and model architecture information for part of a trained ML model; ... “ , and a device or machine 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 mental process but for recitation of generic computer components:
determine a common portion of the parts of trained ML models that is useable by all of the plurality of worker computing devices; (mental process, a person can mentally evaluate and determine a common portion of the trained models used by worker devices, see MPEP 2106.04(a)(2)(III)),
generate an updated common portion of the ML model, (mental process, a person can mentally evaluate and create an updated common portion of a model, see MPEP 2106.04(a)(2)(III)),
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, 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:
using the common portion of the parts of trained ML models the weights and model architecture information from each of the plurality of worker computing devices, (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
A leader computing device configured to develop a Machine Learning, ML, model, the leader computing device comprising processing circuitry and a memory containing instructions executable by the processing circuitry, (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
whereby the leader computing device is operable to: receive, from each of a plurality of worker computing devices, weights and model architecture information for part of a trained ML model; (In step 2A, prong 2, this recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g)),
and initiate transmission of the updated common portion of the ML model to the worker computing devices. (In step 2A, prong 2, this recites mere data receiving, which is considered insignificant extra-solution activity – see MPEP 2106.05(g)),
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 iii and iv recite mere instructions to apply the judicial exception using generic computer components, which are not indicative of significantly more. The additional elements v and vi recite mere data gathering, and are considered insignificant extra-solution activities. In step 2B, these insignificant extra-solution activities are well understood routine and conventional activities, which include receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016), – see MPEP 2106.05(d) (II)(i)),
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.
Claims 17-23, 28:
Regarding claim 16, all of claim 16’s dependent claims follow the deficiencies of their parent claim. Since claims 17-23, and 28 recite similar limitations as corresponding claims 2-8, and 12, respectively, listed above, and are rejected for similar reasons under 35 U.S.C. 101.
Claim 31,
Regarding claim 31, in step 1 of the 101-analysis set forth in MPEP 2106, the claim recites
“A leader computing device configured to develop a Machine Learning, ML, model, the leader computing device comprising: a receiver configured to receive, from each of a plurality of worker computing devices, weights and model architecture information for part of a trained ML model …, and a device or machine 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 mental process but for recitation of generic computer components:
…configured to determine a common portion of the parts of trained ML models that is useable by all of the plurality of worker computing devices; (mental process, a person can mentally evaluate and determine a common portion of the trained models used by worker devices, see MPEP 2106.04(a)(2)(III)),
…configured to generate an updated common portion of the ML model, (mental process, a person can mentally evaluate and create an updated common portion of a model, see MPEP 2106.04(a)(2)(III)),
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, 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:
using the common portion of the parts of trained ML models the weights and model architecture information from each of the plurality of worker computing devices, (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
A leader computing device configured to develop a Machine Learning, ML, model, (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
the leader computing device comprising a receiver …a determiner … a generator … a transmitter… (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
…a receiver configured to receive, from each of a plurality of worker computing devices, weights and model architecture information for part of a trained ML model; (In step 2A, prong 2, this recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g)),
and a transmitter configured to initiate transmission of the updated common portion of the ML model to the worker computing devices. (In step 2A, prong 2, this recites mere data receiving, which is considered insignificant extra-solution activity – see MPEP 2106.05(g)),
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 iii, iv, and v recite mere instructions to apply the judicial exception using generic computer components, which are not indicative of significantly more. The additional elements vi and vii recite mere data gathering, and are considered insignificant extra-solution activities. In step 2B, these insignificant extra-solution activities are well understood routine and conventional activities which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016), – see MPEP 2106.05(d) (II)(i)),
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.
Claim 32:
Regarding claim 32, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 32 recites the following additional element:
A computer program product comprising a non-transitory computer-readable medium storing a computer program comprising instructions which, when executed on processing circuitry, cause the processing circuitry to perform a method in accordance with claim 1, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
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.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1, 15, 16, and 31 are rejected under 35 U.S.C. 103 over Savazzi, S. et al., in “Federated Learning With Cooperating Devices: A Consensus Approach for Massive IoT Networks”, published on January 6th, 2020, available in https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8950073 , (hereafter, SAVAZZI), in view of Langer, M. et al., in “Distributed Training of Deep Learning Models: A Taxonomic Perspective,” published on December 1st, 2020, available at https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9120226 , (hereafter, LANGER).
Claim 1:
Regarding claim 1, SAVAZZI teaches “A method for developing a Machine Learning, ML, model, the method comprising: receiving, at a leader computing device from each of a plurality of worker computing devices, weights and model architecture information for part of a trained ML model;
See SAVAZZI in page 4644, section A. Consensus-Based Federated Averaging, describe “After initialization3 of W0,k at time t = 0, on each communication round t > 0, device k sends its model updates Wt,k (once per round) and receives weights from neighbors Wt,i, i ∈ N¯ k. Based on received data, the device updates its model Wt,k sequentially to obtain the aggregated model,” where the foot note 3 mentions “Each device hosts a model W of the same architecture and initialized similarly.” SAVAZZI mentions that each device receives model updates including weights, and each device hosts a model W of the same architecture, which means also receiving model architecture information.
Further, see SAVAZZI in page 4641 describe "The goal of FL systems is to train a shared global model, i.e., a neural network (NN) from a federation of participating devices acting as local learners under the coordination of a central server for model aggregation." Here, SAVAZZI shows a central server (i.e. viewed as a leader computing device) coordinates from each of the participating devices (i.e. worker devices) to train a shared global model (i.e. a common portion of the parts of trained ML models).
Further, SAVAZZI teaches “determining, at the leader computing device, a common portion of the parts of trained ML models that is useable by all of the plurality of worker computing devices;”
See SAVAZZI in page 4641 describe "The goal of FL systems is to train a shared global model, i.e., a neural network (NN) from a federation of participating devices acting as local learners under the coordination of a central server for model aggregation." Here, SAVAZZI shows a central server (i.e. leader computing device) coordinates the participating devices (i.e. worker devices) to train a shared global model (i.e. a common portion of the parts of trained ML models).
Further, SAVAZZI teaches “generating, at the leader computing device, an updated common portion of the ML model using the common portion of the parts of trained ML models and the weights and model architecture information received from each of the plurality of worker computing devices;”
See SAVAZZI in page 4644, section A. Consensus-Based Federated Averaging, describe “After initialization3 of W0,k at time t = 0, on each communication round t > 0, device k sends its model updates Wt,k (once per round) and receives weights from neighbors Wt,i, i ∈ N¯ k. Based on received data, the device updates its model Wt,k sequentially to obtain the aggregated model,” where the foot note 3 mentions “Each device hosts a model W of the same architecture and initialized similarly.”
PNG
media_image1.png
98
728
media_image1.png
Greyscale
, Here, SAVAZZI shows a device send model updates (viewed as updated common portion of the ML model) of model W which has weights and architecture, among each of the worker devices.
Further, see SAVAZZI in section D. Communication Overhead and Complexity Analysis, page 4646, in 4th paragraph describe "During initialization, i.e., at time t=t0 , devices might use the last available global model received from the server, Wt0,k=WNs , after Ns communication rounds of the previous FA phase, and obtain a local update via SGD: Wt+1,k=WNs−μt∇Lt,k(WNs) . This is fed back to neighbors to start CFA or CFA-GE iterations."
Further, see SAVAZZI in page 4649, section V. part A. Data collection and processing describe "As before, we further assume that, during the initial stage, each device has knowledge of the ML global model structure (see layers and dimensions in Table I). At each new communication round, model parameters for each layer are multiplexed and propagated simultaneously by using a time-division multiple access (TDMA) scheme." Here, SAVAZZI shows that each device is aware of the updates in the ML global model, and model parameters are propagated simultaneously, which means multiple parameters are updated, shared, or passed through the system at the same time for a common part or global model.
Further, see SAVAZZI in page 4643 states “In conventional centralized ML (i.e., learning without federation), used here as benchmark, the server collects all local training data from the devices and obtains the optimization of model parameters by applying an incremental gradient method over a number of batches from the training data set. For iteration t, the model parameters are thus updated by the server” and “Updates ∇Lt,k (6), or local models Wt+1,k,are sent back to the server, after quantization, anonymization [7], and compression stages, modeled here by the operator P . A global model update is obtained by the server through aggregation according to Wt+1 = Wt −μs 1 nt k=1 Ek E P ∇Lt,k(Wt)”. Here, SAVAZZI shows the global model is obtained by the server (i.e. leader computing device) on the updates from the local models on local devices.
Further, SAVAZZI teaches “initiating transmission of the generated updated common portion of the ML model from the leader computing device to the worker computing devices.”
See SAVAZZI in page 4644, section A. Consensus-Based Federated Averaging, describe “After initialization3 of W0,k at time t = 0, on each communication round t > 0, device k sends its model updates Wt,k (once per round) and receives weights from neighbors Wt,i, i ∈ N¯ k. Based on received data, the device updates its model Wt,k sequentially to obtain the aggregated model,” where the foot note 3 mentions “Each device hosts a model W of the same architecture and initialized similarly.”
PNG
media_image1.png
98
728
media_image1.png
Greyscale
, Here, SAVAZZI shows that the initialization of the model updates Wt, k , that is part of the generated updated part of the model.
Further, see SAVAZZI on page 4641 describe “The server obtains a global model by fusion of local models and then feeds back such model to the devices. Multiple rounds are repeated until convergence is reached. The objective of FL is thus to build a global model y =ˆ y(W; x) by the cooperation of a number of devices. Model is characterized by parameters W (i.e., NN weights and biases for each layer) for the output quantity y and the observed (input) data x. Since it decouples the ML stages from the need to send data to the server, FL provides strong privacy advantages compared to conventional centralized learning methods.” Here, SAVAZZI teaches a server (i.e. leader computing device) sending back the fusion of local models back to the device.
However, SAVAZZI did not teach “initiating transmission of the generated updated common portion of the ML model from the leader computing device to the worker computing devices.”
In an analogous art, LANGER teaches “initiating transmission of the generated updated common portion of the ML model from the leader computing device to the worker computing devices.”
See LANGER in Page 2805, section 3.2.2 Decentralized optimization, second paragraph of this section, describe " To achieve collaborative training, the workers have to exchange model parameters with each other. In this example, we assume the existence of a dedicated master node, which processes the individual parameter adjustments suggested by the workers
and comes up with a new global model state that is then shared with them ". See figure 4 in LANGER for details. Here, LANGER illustrates a master node or leader device that comes up with a new global model state or shared common updated model for the worker devices.
PNG
media_image2.png
522
1078
media_image2.png
Greyscale
Further, see LANGER in page 2807, section 3.3.2 Decentralized synchronous systems, second paragraph within this section, mention "First, the initial model parameters (w~0) are distributed among the workers to initialize the local models (wi). Then, the workers randomly sample mini-batches from their locally available partition of the training dataset, determine per-parameter gradients and adjust their model to minimize the loss function (L)... Due to the different properties of the mini-batches, each worker eventually arrives at a slightly better (w.r.t. L), but different model. The master node acts as a synchronization conduit. After each exploration phase, the worker models (wiτ) are merged to form a new joint model (w~t+τ)." Here, LANGER describes the initializing step where the master node (i.e. leader computing device) synchronize the new joint model (i.e. generated updated common portion of the ML model) and send to the workers.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of SAVAZZI, and incorporate with the teachings of LANGER by using the teachings of SAVAZZI of determining a common portion of a trained model useable by worker devices, with LANGER’s teaching of initiating transmission of the generated updated common portion of the ML model from a leader device to worker devices.
One of ordinary skill in the art would be motivated to do so because by integrating LANGER’s framework into the methods of SAVAZZI, achieve a method that “improve upon previous approaches by introducing new concepts, such as defining backup workers (cf. Section 3.3.1), optimizing model partitioning using self-tuning heuristic models, and improving scalability by allowing hierarchical parameter servers to be configured” (See LANGER in page 2814, section 4.1, part of first half paragraph).
Claim 15:
Regarding claim 15, SAVAZZI in view of LANGER, teach the limitations in claim 1.
Further, LANGER teaches “The method of claim 1, wherein the trained ML models are trained local ML models and the updated common portion is an updated common global portion.”
See LANGER in Page 2805, section 3.2.2 Decentralized optimization, second paragraph of this section, describe “To achieve collaborative training, the workers have to exchange model parameters with each other. In this example, we assume the existence of a dedicated master node, which processes the individual parameter adjustments suggested by the workers and comes up with a new global model state that is then shared with them.” See figure 4 in LANGER for details. Here, LANGER illustrates a master node or leader device that comes up with a new global model state or shared common updated model for the worker devices.
Further, see LANGER in page 2807, section 3.3.2 Decentralized synchronous systems, second paragraph within this section, mention "First, the initial model parameters (w~0) are distributed among the workers to initialize the local models (wi). Then, the workers randomly sample mini-batches from their locally available partition of the training dataset, determine per-parameter gradients and adjust their model to minimize the loss function (L)... Due to the different properties of the mini-batches, each worker eventually arrives at a slightly better (w.r.t. L), but different model. The master node acts as a synchronization conduit. After each exploration phase, the worker models (wiτ) are merged to form a new joint model (w~t+τ)." Here, LANGER describes the initializing step where the master node (i.e. leader device) synchronize the new joint model (i.e. generated updated common portion of the ML model) and send to the workers. See the specification describe in [0005] “The ML models may be trained at the worker nodes … resulting in trained (local) ML models that typically vary between worker nodes (for example, the weights assigned to connections between neurons differ between the different trained local ML models). The trained ML models from the worker nodes may then be sent back to the leader node and combined to produce a collaboratively trained (global) ML model. This collaboratively trained ML model may then be used,” where the global model is construed by the examiner to mean the same as a common shared part of the model. LANGER from pages 2805 and 2807 describes the trained models, which have the local models from each worker device, includes a new global model state with a new joint model (i.e. updated common part of the model).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of SAVAZZI, and incorporate with the teachings of LANGER by using the teachings of SAVAZZI of determining a common portion of a trained model useable by worker devices, with LANGER’s teaching of where the trained local models have an updated common portion that is an updated common global portion.
One of ordinary skill in the art would be motivated to do so because by integrating LANGER’s framework into the methods of SAVAZZI, helps to achieve a method that “improve upon previous approaches by introducing new concepts, such as defining backup workers (cf. Section 3.3.1), optimizing model partitioning using self-tuning heuristic models, and improving scalability by allowing hierarchical parameter servers to be configured” (See LANGER in page 2814, section 4.1, part of first half paragraph).
Claim 16:
Regarding claim 16, further, SAVAZZI teaches “A leader computing device configured to develop a Machine Learning, ML, model, the leader computing device comprising processing circuitry and a memory containing instructions executable by the processing circuitry, whereby the leader computing device is operable to…”
See SAVAZZI in page 4651 describe “a device equipped with a 1.5-GHz quad core ARM Cortex-A72 processor with 4-GB internal RAM.6 Focusing on a realistic IIoT environment, the device has thus limited computational capabilities, compared with the server. Execution time depends in general on the specific CPU or tensor processing unit (TPU) performances;” SAVAZZI mentions a processor called 1.5-GHz quad core ARM Cortex-A72 processor and a memory called 4-GB internal RAM.6, and these systems are used to perform the methods such as running the models and algorithms within the study.
Regarding claim 16, the claim recites similar limitations as corresponding independent claim 1 and is rejected for similar reasons as claim 1 using similar teachings and rationale.
Claim 31:
Regarding claim 31, the terms “receiver”, “determiner”, “generator”, “transmitter”,
SAVAZZI further teaches “a receiver configured to receive, …”
See SAVAZZI in page 4646, section III part D describe “devices might use the last available global model received from the server … ” SAVAZZI mentions using a server, which relates to a device, to receive information. A receiver is construed to mean a computer system or machine that can perform the act of gathering or receiving data or information for a model.
Further, SAVAZZI teaches “a determiner configured to determine…”
See SAVAZZI in page 4651 describe “a device equipped with a 1.5-GHz quad core ARM Cortex-A72 processor with 4-GB internal RAM.6 Focusing on a realistic IIoT environment, the device has thus limited computational capabilities, compared with the server. Execution time depends in general on the specific CPU or tensor processing unit (TPU) performances;” SAVAZZI mentions a processor called 1.5-GHz quad core ARM Cortex-A72 processor and a memory called 4-GB internal RAM.6, and these systems are used to perform the methods such as running the models and algorithms within the study.
Further, see SAVAZZI in page 4642, section II. FL for Model Optimization, describe “FL approach defines an incremental algorithm for model optimization over a large population of devices ... The goal is to learn a global model ˆ y(W;x) for inference problems.” Here, SAVAZZI shows using a system that incorporates federated learning to define a model, and to learn a global model (i.e. common portion of a ML model) from devices. A determiner is construed to mean a computer system or machine that can perform the act of defining, identifying or determining information for a model. See SAVAZZI in page 4648, section IV part A. Data collection and processing for more information.
Further, SAVAZZI teaches “a generator configured to generate…”
See SAVAZZI in page 4651 describe “a device equipped with a 1.5-GHz quad core ARM Cortex-A72 processor with 4-GB internal RAM.6 Focusing on a realistic IIoT environment, the device has thus limited computational capabilities, compared with the server. Execution time depends in general on the specific CPU or tensor processing unit (TPU) performances;” SAVAZZI mentions a processor called 1.5-GHz quad core ARM Cortex-A72 processor and a memory called 4-GB internal RAM.6, and these systems are used to perform the methods such as running the models and algorithms within the study. Here, SAVAZZI shows that using the computational system, to generate information. A generator is construed to mean a computer system or machine that can perform the act of generating data or information for a model.
Further, see SAVAZZI in page 4644, section A. Consensus-Based Federated Averaging, describe “After initialization3 of W0,k at time t = 0, on each communication round t > 0, device k sends its model updates Wt,k (once per round) and receives weights from neighbors Wt,i, i ∈ N¯ k. Based on received data, the device updates its model Wt,k sequentially to obtain the aggregated model,” where the foot note 3 mentions “Each device hosts a model W of the same architecture and initialized similarly.” SAVAZZI mentions that each device receives model updates including weights, and each device hosts a model W of the same architecture, and later obtain the aggregated model relates to generating the model.
Further, see SAVAZZI in page 4641 describe "The goal of FL systems is to train a shared global model, i.e., a neural network (NN) from a federation of participating devices acting as local learners under the coordination of a central server for model aggregation." Here, SAVAZZI shows a central server (i.e. leader computing device) coordinates the participating devices (i.e. worker devices) to train a shared global model (i.e. a common portion of the parts of trained ML models).
Further, SAVAZZI teaches “and a transmitter configured to initiate transmission…”
See SAVAZZI in page 4651 describe “a device equipped with a 1.5-GHz quad core ARM Cortex-A72 processor with 4-GB internal RAM.6 Focusing on a realistic IIoT environment, the device has thus limited computational capabilities, compared with the server. Execution time depends in general on the specific CPU or tensor processing unit (TPU) performances;” SAVAZZI mentions a processor called 1.5-GHz quad core ARM Cortex-A72 processor and a memory called 4-GB internal RAM.6, and these systems are used to perform the methods such as running the models and algorithms within the study.
Further, see SAVAZZI in page 4641, last paragraph on page of I. Introduction section, describe “Model is characterized by parameters W (i.e., NN weights and biases for each layer) for the output quantity y and the observed (input) data x. Since it decouples the ML stages from the need to send data to the server, FL provides strong privacy advantages compared to conventional centralized learning methods.” Here, SAVAZZI teaches using a system to send information such as data to the server, which is similar to transmitting data. A transmitter is construed to mean a computer system or machine that can perform the act of sending data or information for a model.
Regarding claim 31, the claim recites similar limitations as corresponding independent claim 16 and is rejected for similar reasons as claim 16 using similar teachings and rationale.
Claims 6, 7, 21, 22, and 32 are rejected under 35 U.S.C. 103 over SAVAZZI, in view of LANGER, and further in view of Hall, J. et al., (Pub. No. WO2021056043A1), published on April 1st, 2021, (hereafter, HALL).
Claim 6:
Regarding claim 6, SAVAZZI in view of LANGER, teach the limitations in claim 1.
However, SAVAZZI in view of LANGER, did not teach “The method of claim 1, further comprising receiving, at the leader computing device from each of the plurality of worker computing devices, metadata.”
In an analogous art, HALL teaches “The method of claim 1, further comprising receiving, at the leader computing device from each of the plurality of worker computing devices, metadata.”
See HALL in [0099] note " In some embodiments additional model metadata may be exported/saved and sent along with the network weights, such as model accuracy, number of epochs, etc., that may further characterise the model, or otherwise assist in constructing another model (e.g. a Student model) on another node/server."
Further, see HALL in [00122] describe “In the case of the decentralized training stage, where all model updates are sent to the master node, the hyper-parameter optimization problem must be treated collectively.” Here, HALL mentions all updates are sent to the master node, which is viewed as a leader device.
Later, see HALL in [0010] describe “when performing distributed training, a distribution strategy needs to be chosen. This defines how the workload will be divided among the different worker nodes. The two methods for this are Model Parallelism and Data Parallelism. Model Parallelism splits the workload by segmenting the model weights into N partitions, where N is the number of workers amongst which to split the work.” HALL mentions the nodes can be viewed as worker nodes, which when processing the model, also receive information such as metadata about the model. This information, along with model updates, is sent to the master node or leader device.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of SAVAZZI along with the secondary reference of LANGER, and incorporate with the teachings of HALL by using the teachings of SAVAZZI and LANGER, with HALL’s teaching of a leader computing device receive metadata.
One of ordinary skill in the art would be motivated to do so because by integrating HALL’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve the goal of providing “can be used to improve runtime inference efficiency (reducing costs) for a loss in accuracy (compared with the ensemble),” (See HALL in [00121]).
Claim 7:
Regarding claim 7, SAVAZZI in view of LANGER, further in view of HALL, teach the limitations in claim 6.
Further, HALL teaches “The method of claim 6, wherein the metadata is used by the leader computing device when determining the common portion of the parts of trained ML models that may be used by all of the plurality of worker computing devices.”
See HALL at [00105] note " where sharing of the dataset itself is not permitted, but sharing of the neural network weights and non-confidential metadata is permitted (e.g. model architectures, de-identified file numbers, number of total data points in each region, etc.)... The training code and the code for receiving and sending network's weight are assumed to be available at each location/machine. For simplicity, S and T, denote the model names and are also the model's weight that can be sent to other nodes. A cloud-based master or administrative server node controls the training procedure and collects the final trained model for production." HALL here describes a master server node (i.e. leader computing device), collects information about metadata pertaining to the models, so model information can be sent to the other nodes, which other nodes in this case is viewed to be worker devices. Later, see HALL in [00207] describe “The AI model may be deployed in distributed systems with separate computer systems at the user and administrator (central node) sides.”
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of SAVAZZI along with the secondary reference of LANGER, and incorporate with the teachings of HALL by using the teachings of SAVAZZI and LANGER, with HALL’s teaching of a leader computing device receive metadata.
One of ordinary skill in the art would be motivated to do so because by integrating HALL’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve the goal of providing “can be used to improve runtime inference efficiency (reducing costs) for a loss in accuracy (compared with the ensemble),” (See HALL in [00121]).
Claim 21:
Regarding claim 21, SAVAZZI in view of LANGER, teach the limitations in claim 16. Regarding claim 21, the claim recites similar limitations as corresponding claim 6 and is rejected for similar reasons as claim 6 using similar teachings and rationale.
Claim 22:
Regarding claim 22, SAVAZZI in view of LANGER, further in view of HALL, teaches the limitations in claim 21. Regarding claim 22, the claim recites similar limitations as corresponding claim 7 and is rejected for similar reasons as claim 7 using similar teachings and rationale.
Claim 32:
Regarding claim 32, SAVAZZI in view of LANGER, teach the limitations in claim 1.
Further, SAVAZZI teaches “a computer program product comprising a non-transitory computer-readable medium storing a computer program comprising instructions which, when executed on processing circuitry, cause the processing circuitry to perform a method in accordance with claim 1.”
See SAVAZZI describe in page 4651 “Local execution time considers here the ML stages only, while data preprocessing and acquisition steps are not included, being negligible compared to the learning steps. The execution time shown in Table IV is measured using the timeit Python module, on a device equipped with a 1.5-GHz quad core ARM Cortex-A72 processor with 4-GB internal RAM.6 Focusing on a realistic IIoT environment, the device has thus limited computational capabilities, compared with the server. Execution time depends in general on the specific CPU or tensor processing unit (TPU) performances; nevertheless, the analysis of the results in Table IV is useful to highlight the scaling performance of the proposed methods compared to the plain FA algorithm. Considering FA, the execution time on the device is ruled by local SGD rounds: it is 140 and 145 ms for CNN and 2NN, respectively …”, where SAVAZZI mentions a processor called 1.5-GHz quad core ARM Cortex-A72 processor and a memory called 4-GB internal RAM.6, and these systems are used to perform the methods such as running the models and algorithms within the study.
However, SAVAZZI in view of LANGER did not teach “a computer program product comprising a non-transitory computer-readable medium…”
In an analogous art, HALL teaches “a computer program product comprising a non-transitory computer-readable medium storing a computer program comprising instructions which, when executed on processing circuitry, cause the processing circuitry to perform a method in accordance with claim 1.”
See HALL mention in [00217] " In some aspects the computer-readable media may comprise non-transitory computer-readable media (e.g., tangible media)... In another aspect, the computer readable medium may be integral to the processor. The processor and the computer readable medium may reside in an ASIC or related device." Further, see HALL describe in [0024] “In most cases, the distributed model and the training process across all the separated localities can be managed by an off-site server, such as one administered by a third party that is intending to provide the resultant machine learning model as a product or service.” Here, HALL mentions a non-transitory computer-readable medium connected to a processor and providing a machine learning model, generated from computer system, as a product.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI and LANGER, and incorporate with the teachings of HALL by using the teachings of SAVAZZI and LANGER, with HALL’s teaching of a non-transitory computer readable medium connected with a processor.
One of ordinary skill in the art would be motivated to do so because by integrating HALL’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve the goal of providing “can be used to improve runtime inference efficiency (reducing costs) for a loss in accuracy (compared with the ensemble),” (See Hall in [00121]).
Claims 2 and 17 are rejected under 35 U.S.C. 103 over SAVAZZI in view of LANGER, and further in view of Froelicher, D., et al., in “Scalable Privacy-Preserving Distributed Learning,” published on July 2, 2020, available on https://arxiv.org/pdf/2005.09532v2 , (hereafter, FROELICHER).
Claim 2:
Regarding claim 2, SAVAZZI in view of LANGER, teach the limitations of claim 1.
Further, SAVAZZI teaches “and initiating transmission of initialization information for the maximum common portion of the ML model to all of the plurality of worker computing devices,”
See SAVAZZI in page 4644, section A. Consensus-Based Federated Averaging, describe “After initialization3 of W0,k at time t = 0, on each communication round t > 0, device k sends its model updates Wt,k (once per round) and receives weights from neighbors Wt,i, i ∈ N¯ k. Based on received data, the device updates its model Wt,k sequentially to obtain the aggregated model,” where the foot note 3 mentions “Each device hosts a model W of the same architecture and initialized similarly.”
PNG
media_image1.png
98
728
media_image1.png
Greyscale
, Here, SAVAZZI shows that the initialization of the model updates Wt, k , that is part of the generated updated part of the model.
Further, SAVAZZI teaches “determining, at the leader computing device, a maximum common portion of the ML model that is useable by all of the plurality of worker computing devices,.. ;”
See SAVAZZI in page 4641 describe "The goal of FL systems is to train a shared global model, i.e., a neural network (NN) from a federation of participating devices acting as local learners under the coordination of a central server for model aggregation." Here, SAVAZZI shows a central server (i.e. leader computing device) coordinates the participating devices (i.e. worker devices) to train a shared global model (i.e. a common portion of the parts of trained ML models) usable by the participating devices. Since the term maximum is a relative term and not defined clearly what is considered minimum or maximum relative to a value, the examiner construes the claim to mean similar meaning as a common portion of the machine learning model.
However, SAVAZZI in view of LANGER, did not teach
“The method of claim 1, further comprising, prior to receiving the weights and model architecture information for the parts of trained ML models from the worker computing devices: receiving, at the leader computing device from the plurality of worker computing devices, ML model architecture privacy information,”
“determining, at the leader computing device, a maximum common portion of the ML model that is useable by all of the plurality of worker computing devices, using the ML model architecture privacy information;”
In an analogous field, FROELICHER teaches “The method of claim 1, further comprising, prior to receiving the weights and model architecture information for the parts of trained ML models from the worker computing devices: receiving, at the leader computing device from the plurality of worker computing devices, ML model architecture privacy information,”
See FROELICHER in page 3, section II. Related work, describe "Differential privacy has also been envisioned in distributed settings, where to collectively train a model, multiple parties exchange or send differentially private model parameters to a central server. " Here, FROELICHER mentions sending differentially private model parameters to a central server (i.e. leader device). The model architecture privacy information is construed to include any information related to the model, including model parameters. Differentially private is interpreted by examiner to include privacy information about the model.
Further, FROELICHER teaches “determining, at the leader computing device, a maximum common portion of the ML model that is useable by all of the plurality of worker computing devices, using the ML model architecture privacy information;”
See FROELICHER in page 3, section II. Related work, describe " Differential privacy has also been envisioned in distributed settings, where to collectively train a model, multiple parties exchange or send differentially private model parameters to a central server." Here, FROELICHER mentions sending differentially private model parameters (model privacy information) to a central server (i.e. leader device). Later, see FROELICHER in page 6, section B. SPINDLE Protocols, subsection Prepare describe “The data providers (DPs) collectively agree on the training parameters: the maximum number of global g and local m iterations, and the learning parameters lp={α,ρ,b}, where α is the learning rate, ρ the elastic rate, and b the batch size. The DPs also collectively initialize the cryptographic keys for the distributed CKKS scheme by executing DKeyGen(·) (see Section IV-A). Then, the DPs initialize their local weights and pre-compute operations that involve only their input data.” Further, see FROELICHER in page 6, section B. SPINDLE Protocols, subsection Map mention “As depicted in Protocol 2, the DPs execute m iterations of the cooperative gradient-descent local update (Section IV-A). The local weights of DPi (i.e., w(i,j,l−1) ) are updated at a global iteration j and a local iteration l by computing the gradient (Protocol 2, lines 4, 5, and 6) that is then combined with the current global weights.” Here, FROELICHER shows determining model training parameters including the greatest number of global and local iterations, and other model information, to be used by data providers (viewed as worker computing devices).
Later, see FROELICHER in page 5, section A. Background, second to last paragraph describe “this enables the data providers (DPs) to train a collectively encrypted model,” Here, FROELICHER mentions using information from page 6, to train a collectively encrypted model, using information such as encrypted keys (from page 6), and other model parameters (i.e. model architecture privacy information). See FROELICHER in page 5, section A. Cooperative gradient descent mention “The data providers (DPs), each of which owns a part of the dataset, locally perform multiple iterations of the SGD before aggregating their model weights into the global model weights.” Here, the data providers are viewed as worker devices, locally performing model updates before updating into the global model.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI and LANGER, and incorporate with the teachings of FROELICHER by using the teachings of SAVAZZI and LANGER, with FROELICHER’s teaching of receiving model architecture privacy information.
One of ordinary skill in the art would be motivated to do so because by integrating FROELICHER’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve a method where “the total amount of data remains constant, [the method] SPINDLE’s execution time decreases linearly, as the workload is efficiently distributed among the DPs [data providers],” (see FROELICHER in page 13, second to last paragraph), and “the parallelization is efficient up to c=28, where the maximum thread-utilisation is reached,” (see FROELICHER in page 13, scalability section).
Claim 17:
Regarding claim 17, SAVAZZI in view of LANGER, teach the limitations in claim 16. Regarding claim 17, the claim recites similar limitations as corresponding claim 2 and is rejected for similar reasons as claim 2 using similar teachings and rationale.
Claims 3, 4, 5, 18, 19 and 20 are rejected under 35 U.S.C. 103 over SAVAZZI in view of LANGER, and further in view of Lim, W. et al. in "Federated learning in mobile edge networks: A comprehensive survey," published on April 8th, 2020, available at: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9060868 , (hereafter, LIM).
Claim 3:
Regarding claim 3, SAVAZZI in view of LANGER, teach the limitations in claim 1.
However, SAVAZZI in view of LANGER, did not teach “The method of claim 1, wherein the step of determining the common portion of the ML models comprises detecting a variation in a model architecture from among the model architectures of the trained ML models.”
In an analogous field, LIM teaches “wherein the step of determining the common portion of the ML models comprises detecting a variation in a model architecture from among the model architectures of the trained ML models.”
See LIM in page 2042, section C. Importance-based updating, describe “ In each iteration, a participant’s local update is first compared with the global update to identify if the update is relevant. A relevance score is computed where the score equates to percentage of same-sign parameters in the local and global update.” LIM here shows comparing the model of a participant (i.e. worker device) with a global or common model update if update is relevant (i.e. identify a variation). Later, see LIM in page 2052, section 3). Free-riding attacks, describe “blockchain-based FL architecture, called BlockFL, in which the participants’ local learning model updates are exchanged and verified by leveraging the blockchain technology. In particular, each participant trains and sends the trained global model to its associated miner in the blockchain network and then receives a reward that is proportional to the number of trained data samples”. LIM shows by using a method to verify the participants’ model updates from trained models (trained ML models), this shows the verification process is similar to identifying if there is a change.
Further, see LIM in page 2035, section A. Deep learning, describe "The training iterations are then repeated over many epochs, ... there also exists several DNN networks and architectures tailored to process the varying natures of input data, e.g., Multilayer Perceptron (MLP) [51], Convolutional Neural Network (CNN) [52] typically for CV tasks, and Recurrent Neural Network (RNN)". Here, this shows LIM describes within each round of model training, each model contains its own type of architecture such as MLP, CNN or RNN. Any comparison or verification will also detect all aspects of the model, including a model's architecture.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI and LANGER, and incorporate with the teachings of LIM by using the teachings of SAVAZZI and LANGER, with LIM’s teaching of detecting a variation among worker models.
One of ordinary skill in the art would be motivated to do so because by integrating LIM’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve “simulation results show that the proposed scheme can significantly improve the accuracy of the FL model since unreliable workers are detected and not selected for FL training,” (see LIM in section D. Incentive mechanism, page 2047).
Claim 4:
Regarding claim 4, SAVAZZI in view of LANGER, and further in view of LIM, teach the limitations in claim 3.
Further, SAVAZZI teaches “The method of claim 3, wherein, if the variation is detected, the updated common portion of the ML model distributed to the worker computing devices comprises weights and model architecture information.”
See SAVAZZI in page 4649, section A. Data collection and processing, describe “each device has knowledge of the ML global model structure (see layers and dimensions in Table I). At each new communication round, model parameters for each layer are multiplexed and propagated simultaneously by using a time-division multiple access (TDMA) scheme [14].” Model architecture is construed by examiner as any part of the model, including its structure or content, such as including the type or number of model parameters. Here, SAVAZZI describes that at each round of updating into a global shared model, the model parameters for each layer are propagated simultaneously, which relates to updating the model architecture.
Further, SAVAZZI in page 4644, section A. Consensus-Based Federated Averaging, describe “After initialization3 of W0,k at time t = 0, on each communication round t > 0, device k sends its model updates Wt,k (once per round) and receives weights from neighbors Wt,i, i ∈ N¯ k. Based on received data, the device updates its model Wt,k sequentially to obtain the aggregated model,” where the foot note 3 mentions “Each device hosts a model W of the same architecture and initialized similarly.” SAVAZZI mentions that each device receives model updates including weights, and each device hosts a model W of the same architecture, which means also receiving model architecture information to update the model back to the worker device. See SAVAZZI in page 4643 in first paragraph for details.
However, SAVAZZI did not explicitly teach “The method of claim 3, wherein, if the variation is detected, ...”
Further, LIM teaches “The method of claim 3, wherein, if the variation is detected, ...”
See LIM in page 2042, section C. Importance-based updating, describe “ In each iteration, a participant’s local update is first compared with the global update to identify if the update is relevant. A relevance score is computed where the score equates to percentage of same-sign parameters in the local and global update.” LIM here shows comparing the model of a participant (i.e. worker device) with a global or common model update if update is relevant (i.e. identify a variation). Later, see LIM in page 2052, section 3). Free-riding attacks, describe “blockchain-based FL architecture, called BlockFL, in which the participants’ local learning model updates are exchanged and verified by leveraging the blockchain technology. In particular, each participant trains and sends the trained global model to its associated miner in the blockchain network …” LIM shows by using a method to verify the participants’ model updates from trained models (trained ML models), this shows the verification process is similar to identifying if there is a change.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI and LANGER, and incorporate with the teachings of LIM by using the teachings of SAVAZZI and LANGER, with LIM’s teaching of detecting a variation among worker models.
One of ordinary skill in the art would be motivated to do so because by integrating LIM’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve “simulation results show that the proposed scheme can significantly improve the accuracy of the FL model since unreliable workers are detected and not selected for FL training,” (see LIM in section D. Incentive mechanism, page 2047).
Claim 5:
Regarding claim 5, SAVAZZI in view of LANGER, and further in view of LIM, teach the limitations in claim 3.
Further, LIM teaches “The method of claim 3, wherein, if the variation is not detected, the updated common portion of the ML model distributed to the worker computing devices comprises weights.”
See LIM in introduction section, page 2032, second full paragraph describes “In FL, mobile devices use their local data to cooperatively train an ML model required by an FL server. They then send the model updates, i.e., the model’s weights, to the FL server for aggregation. The steps are repeated in multiple rounds until a desirable accuracy is achieved.” LIM here shows updating the model using only weights by default.
Further, see LIM in page 2049. Section 3) Collaborative training solutions, mention “The key idea of this technique is that instead of uploading the whole set of trained parameters to the server and updating the whole global parameters to its local model, each participant wisely selects the number of gradients to upload and the number of parameters from the global model to update as illustrated in Fig. 11.” Later, see LIM in page 2035, section B. Federated Learning, mention “In general, there are two main entities in the FL system, i.e., the data owners (viz. participants) and the model owner (viz. FL server). Let N={1,…,N} denote the set of N data owners, each of which has a private dataset Di∈N . Each data owner i uses its dataset Di to train a local model wi and send only the local model parameters to the FL server. Then, all collected local models are aggregated w=∪i∈Nwi to generate a global model wG .” See LIM in page 2036, section B. Federated Learning describe “Step 3 (Global model aggregation and update): The server aggregates the local models from participants and then sends the updated global model parameters wt+1 G back to the data owners.” Here, LIM shows selective update of model information, which includes model weights, after identifying an update, and this updating this information to the data owners (i.e. worker devices) from aggregating the local models to generate a global model (i.e. updated common portion of model).
Additionally, LIM in page 2039, describe "Importance-based Updating: This strategy involves selective communication such that only the important or relevant updates [95] are transmitted in each communication round. " Here, LIM describes that if an update is not considered relevant, this relates to situations where a variation among model structures is not detected.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI and LANGER, and incorporate with the teachings of LIM by using the teachings of SAVAZZI and LANGER, with LIM’s teaching of updating weights to worker computing devices.
One of ordinary skill in the art would be motivated to do so because by integrating LIM’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve “simulation results show that the proposed scheme can significantly improve the accuracy of the FL model since unreliable workers are detected and not selected for FL training,” (see LIM in section D. Incentive mechanism, page 2047).
Claim 18:
Regarding claim 18, SAVAZZI in view of LANGER, teach the limitations in claim 16. Regarding claim 18, the claim recites similar limitations as corresponding claim 3 and is rejected for similar reasons as claim 3 using similar teachings and rationale.
Claim 19:
Regarding claim 19, SAVAZZI in view of LANGER, and further in view of LIM, teach the limitations in claim 18. Regarding claim 19, the claim recites similar limitations as corresponding claim 4 and is rejected for similar reasons as claim 4 using similar teachings and rationale.
Claim 20:
Regarding claim 20, SAVAZZI in view of LANGER, and further in view of LIM, teach the limitations in claim 18. Regarding claim 20, the claim recites similar limitations as corresponding claim 5 and is rejected for similar reasons as claim 5 using similar teachings and rationale.
Claim 9 rejected under 35 U.S.C. 103 over SAVAZZI in view of LANGER, and further in view of Wahab, O. et al., in “Federated Machine Learning: Survey, Multi-Level Classification, Desirable Criteria and Future Directions in Communication and Networking Systems,” published on February 10th, 2021,
available at https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9352033 , (hereafter, WAHAB).
Claim 9:
Regarding claim 9, SAVAZZI in view of LANGER, teach the limitations in claim 1.
However, SAVAZZI in view of LANGER, did not teach wherein the updated common portion of the ML model is generated using federated averaging.”
In an analogous field, WAHAB teaches “The method of claim 1, wherein the updated common portion of the ML model is generated using federated averaging.”
See WAHAB in page 1347, section A. Federated Learning, describe "The concept of federated learning was first introduced in [8] as a distributed training model that is executed by a set of mobile devices that share local model updates with a central server whose role is to aggregate these updates to build a global machine learning model. An aggregation model called Federated Averaging (explained in Section II-A1) is also introduced to allow the server to combine local stochastic gradients from the different devices using iterative model averaging. A federated learning scenario consists of one central server called parameter server and a set of N clients, each having its own local dataset…” Here, WAHAB teaches the method of using federated averaging to let devices (i.e. worker devices) share local model updates with a central server whose role is to aggregate these updates to build a global machine learning model (i.e. updated common portion of the ML model).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI and LANGER, and incorporate with the teachings of WAHAB by using the teachings of SAVAZZI and LANGER, with WAHAB’s teaching of using federated averaging.
One of ordinary skill in the art would be motivated to do so because by integrating WAHAB’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve a method that "help researchers in the domain better understand the field which would enable them to design more detailed and efficient solutions," (see WAHAB in page 1344, Introduction, section A. Related Work), and "the clients learn to add changes of limited magnitudes into the data, which helps improve the performance of the federated model in the presence of noisy data," (see WAHAB in page 1384, section VII Privacy Concerns (challenge 5), A. Differential Privacy, subsection Machine learning).
Claim 10 is rejected under 35 U.S.C. 103 over SAVAZZI in view of LANGER, and further in view of Amannejad, Y., in “Building and evaluating federated models for edge computing.” published on November 30, 2020, available at https://dl.ifip.org/db/conf/cnsm/cnsm2020/1570662810.pdf, (hereafter, AMANNEJAD).
Claim 10:
Regarding claim 10, SAVAZZI in view of LANGER, teach the limitations in claim 1. However, SAVAZZI in view of LANGER, did not teach “The method of claim 1, wherein the weights and model architecture information the leader computing device receives from each of the plurality of worker computing devices is weights and model architecture information of part of a trained ML model that has been trained by a given worker computing device using data private to the given worker computing device.”
In an analogous field, AMANNEJAD teaches “… wherein the weights and model architecture information the leader computing device receives from each of the plurality of worker computing devices is weights and model architecture information of part of a trained ML model that has been trained by a given worker computing device using data private to the given worker computing device.”
See Amannejad in page 1, Introduction, mention " each device trains a local model using its own private data and shares the model parameters, e.g., model weights, with an FL server that aggregates them into a global model. " Here, AMANNEJAD describes each device (i.e. worker computing device) has model information including model weights and parameters (construed by examiner to also mean model architecture) and trains its model by using its own private data. The FL server relates to a leader device, and since model parameter information is shared from the devices to the FL server, this relates to the leader receives information from each worker device. The ‘using data private to the given’ worker device is construed to mean using data that is available only to a specific worker device. In this case, a worker device’s own private data is data that is available only to a specific worker device.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI, and LANGER, and incorporate with the teachings of AMANNEJAD by using the teachings of SAVAZZI, and LANGER, with AMANNEJAD’s teaching of a model trained by a worker device using data private to the given worker device.
One of ordinary skill in the art would be motivated to do so because by integrating AMANNEJAD’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve “results show[ing] that the federated model built using the tool is able to recognize images with high accuracy comparable with the central solution, and the tool provides insights on the selection of federated parameters,” (see AMANNEJAD in page 1, Introduction, last paragraph).
Claim 11 is rejected under 35 U.S.C. 103 over SAVAZZI in view of LANGER, and further in view of Hu, Y. et al., in “FDML: A collaborative machine learning framework for distributed features,” published on July 25, 2019, available at https://dl.acm.org/doi/pdf/10.1145/3292500.3330765 , (hereafter, HU).
Claim 11:
Regarding claim 11, SAVAZZI in view of LANGER, teach the limitations in claim 1.
However, SAVAZZI in view of LANGER, did not teach “The method of claim 1, wherein the updated common portion of the ML model comprises the output layer of the ML model.”
In an analogous method, HU teaches “…wherein the updated common portion of the ML model comprises the output layer of the ML model.”
See HU in page 2237, section 7. Experiments describe “We run both a logistic regression (LR) and a two layered fully connected neural network (NN) under three different training schemes for both data sets ... FDML: use FDML system to train a joint model for app recommendation based on all 8,700 features distributed in three apps or train the a9a classification model on all 124 features from two different parties,” and see HU in figure 1 on page 2234, illustrating an output of the FDML model. Here, HU mentions the results are aggregated into an output, and HU mentions using a model with two layered fully connected neural network (NN), and since this model can include layers that relate to an output (i.e. output layer).
PNG
media_image3.png
544
703
media_image3.png
Greyscale
See HU in page 2233, Introduction, describe “A highlight of our system is that during each training iteration, every party is solely responsible for updating its own local model parameters (local net) using its own mini-batch of local feature sets, and for each record, only needs to share its local prediction to the central server (or to other parties directly in a fully decentralized scenario)... , the FDML system preserves data locality and is much less vulnerable to model inversion attacks [ 12 ] targeting other collaborative learning algorithms [ 27, 31] that share model parameters between parties.” HU describes here of each party or worker device sharing a local prediction from the model to the central server with the FDML system preserving data locality (i.e. updated common part of the model). See HU in page 2234, section 3. Problem formulation, last paragraph, describe “in this model, both the local features ξ j and the sub-model parameters x j are stored and processed locally within party j, while only the local predictions α j (x j , ξ j ) need be shared to produce the final prediction.” HU mentions the local prediction will be shared to produce a final prediction is updating a common shared part of the model since the results from each party or worker device is collected and shared to get a final prediction for the model.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI and LANGER, and incorporate with the teachings of HU by using the teachings of SAVAZZI and LANGER, with HU’s teaching of an updated common portion of the ML model comprises the output layer of the ML model.
One of ordinary skill in the art would be motivated to do so because by integrating HU’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve “a performance improvement over learning only based on local data, while achieving stronger privacy guarantee due to the perturbations introduced to the shared local prediction results,” (see HU in page 2238, 7. Experiments, last paragraph).
Claims 12 and 28 are rejected under 35 U.S.C. 103 over SAVAZZI in view of LANGER, and further in view of Gu, R. et al., in “From Server-Based to Client-Based Machine Learning: A Comprehensive Survey,” published on January 2nd, 2021, available at https://dl.acm.org/doi/10.1145/3424660 , (hereafter, GU).
Claim 12:
Regarding claim 12, SAVAZZI in view of LANGER, teach the limitations in claim 1.
However, SAVAZZI in view of LANGER, did not teach "the method of claim 1, further comprising, by each of the worker computing devices, using the updated common portion of the ML model as part of a worker specific ML model, wherein each worker specific ML model is used to provide suggested actions for an environment."
In an analogous field, GU teaches "the method of claim 1, further comprising, by each of the worker computing devices, using the updated common portion of the ML model as part of a worker specific ML model, wherein each worker specific ML model is used to provide suggested actions for an environment."
See GU in page 6:21, section 5.2, subsection 5.2.1 Workflow where GU mentions “as shown in Figure 6, a general workflow of client-based training consists of five phases: (1) Initialization: The workers {C1,C2,...,Cm} run on mobile devices download hyperparameters from the server S and get themselves initialized. (2) Distribution: The server S arranges the tasks and sends them to the workers {C1,C2,...,Cm}. The model to be trained should also be distributed to the workers in this step.(3)Local Training: The workers {C1,C2,...,Cm} train the local models based on the local datasets {D1,D2,...,Dm}. Updates will be generated after a certain number of local training iterations. (4) Uploading: The workers {C1,C2,...,Cm} upload the updates to the server S. (5)Aggregation: The server S aggregates all updates to generate a final update and applies it to the global model W. A single round of learning ends here. Go back to step 2 to start a new round of learning. A similar workflow has been adopted in the federated learning protocol [12].” Here, GU mentions that the worker device trains data and upload updates back to the server, indicating transfer of data to server (i.e. provides a suggested action for an environment). With the workers each uploading their local updates to a server, and the server performs the process of aggregating updates to create a final update and apply to a global model, this method shows using the updated common portion of the model as a part of a worker model.
Later, GU in page 6:22, second paragraph describes “The Communication module takes part in step1, step2, and step4. It is implemented on both the server and the worker, which means the server and the worker have to cooperate with each other for communication. The communication module is responsible for two jobs: task arrangement and data transfer.”
The specification in lines 3-15, of page 13, (or paragraph [0040] from US PG Pub.), notes that “Taking the example wherein the environment is a communications network wherein each worker node forms part of a base station, the suggested actions may comprise, for example, rerouting or dropping traffic or reprioritizing certain traffic. In the further example wherein the environment is a data center and each worker node forms part of a server, the suggested actions may comprise transferring data between servers, duplicating or deleting data, activating backup servers, and so on.” The suggested action can be transferring data, copying or removing data, activating backup servers. Here in page 6:22, GU mentions a specific suggested action of data transfer where the worker device communicates with the server for this task.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI and LANGER, and incorporate with the teachings of GU by using the teachings of SAVAZZI and LANGER, with GU’s teaching of a worker device provide suggested actions for an environment.
One of ordinary skill in the art would be motivated to do so because by integrating GU’s framework into the methods of SAVAZZI and LANGER, one with ordinary skill in the art would achieve “this design is useful in client-based training since it improves the algorithm’s compatibility for heterogeneous mobile devices. Thus, regarding the local complexity, CoCoA and CoCoA+ outperform the others,” (see GU in page 6:17, section Local complexity).
Claim 28:
Regarding claim 28, SAVAZZI in view of LANGER, teach the limitations in claim 16. Regarding claim 28 , the claim recites similar limitations as corresponding claim 12 and is rejected for similar reasons as claim 12 using similar teachings and rationale.
Claim 13 is rejected under 35 U.S.C. 103 over SAVAZZI in view of LANGER, and further in view of GU, and further in view of WAHAB.
Claim 13:
Regarding claim 13, SAVAZZI in view of LANGER, and further in view of GU, teach the limitations of claim 12.
However, did not teach “The method of claim 12, wherein the environment is one or more base stations in a communications network, or wherein the environment is one or more servers in a data center.”
In an analogous field, WAHAB teaches “… wherein the environment is one or more base stations in a communications network, or wherein the environment is one or more servers in a data center.”
See WAHAB in page 1350, section 2. Variants of federated averaging describe “HierFAVG is a communication-oriented aggregation technique that aims to foster partial model aggregation at the level of edge servers. Specifically, HierFAVG is based on a hierarchical client-edge-cloud architecture whereby each edge server is allowed to aggregate the model updates of its own clients. Subsequently, after a fixed number of model aggregations, the edge-level aggregate models are forwarded to a cloud server for a global aggregation. Such a multi-level structure enables a more efficient model exchange over the existing edge-cloud architecture.” Here, WAHAB shows using edge servers to perform model aggregation.
Later, see WAHAB in page 1354, section reliability, part 3, Edge computing, describe “Having a multitude of edge data centers and edge computing devices makes it more unlikely for a single failure to entirely disrupt the whole service. Thus, several pathways are available to reroute the exchanged model updates to in case of any failure to maintain users’ access to their federated learning services.” Here, WAHAB describes the environment includes having data centers for edge devices.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI, LANGER, and GU, and incorporate with the teachings of WAHAB by using the teachings of SAVAZZI, LANGER, and GU, with WAHAB’s teaching of the type of environment.
One of ordinary skill in the art would be motivated to do so because by integrating WAHAB’s framework into the methods of SAVAZZI, LANGER, and GU, one with ordinary skill in the art would achieve a method that "help researchers in the domain better understand the field which would enable them to design more detailed and efficient solutions," (see WAHAB in page 1344, Introduction, section A. Related Work), and "the clients learn to add changes of limited magnitudes into the data, which helps improve the performance of the federated model in the presence of noisy data," (see WAHAB in page 1384, section VII Privacy Concerns (challenge 5), A. Differential Privacy, subsection Machine learning).
Claim 14 is rejected under 35 U.S.C. 103 over SAVAZZI in view of LANGER, and further in view of GU, and further in view of Abad, M. et al., in “Hierarchical Federated Learning ACROSS Heterogeneous Cellular Networks,” published on April 9th, 2020, available at https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9054634 , (hereafter, ABAD).
Claim 14:
Regarding claim 14, SAVAZZI in view of LANGER, and further in view of GU, teach the limitations of claim 12.
However, SAVAZZI in view of LANGER, and further in view of GU, did not teach “the method of claim 12, further comprising modifying the environment based on the suggested actions.”
In an analogous art, ABAD teaches “…further comprising modifying the environment based on the suggested actions.”
See ABAD in page 8866 in abstract describe "We consider federated edge learning (FEEL), where mobile users (MUs) collaboratively learn a global model by sharing local updates on the model parameters rather than their datasets, with the help of a mobile base station (MBS). We optimize the resource allocation among MUs to reduce the communication latency in learning iterations. Observing that the performance in this centralized setting is limited due to the distance of the cell-edge users to the MBS, we introduce small cell base stations (SBSs) orchestrating FEEL among MUs within their cells, and periodically exchanging model updates with the MBS for global consensus. We show that this hierarchical federated learning (HFL) scheme significantly reduces the communication latency without sacrificing the accuracy." Here, ABAD shows a change in the environment from a mobile base station to small cell base stations (i.e. modify environment) based on reducing communication latency in learning iterations (i.e. suggested action).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI, LANGER, and GU, and incorporate with the teachings of ABAD by using the teachings of SAVAZZI, LANGER, and GU, with ABAD’s teaching of a leader computing device.
One of ordinary skill in the art would be motivated to do so because by integrating ABAD’s framework into the methods of SAVAZZI, LANGER, and GU, one with ordinary skill in the
art would achieve “this hierarchical federated learning (HFL) scheme significantly reduces the communication latency without sacrificing the accuracy,” (see ABAD in page 8866, abstract).
Claims 8 and 23 are rejected under 35 U.S.C. 103 over SAVAZZI in view of LANGER, further in view of HALL, and further in view of Badmus, A. et al., in “Compliance-aware DevOps for generative AI: Integrating legal risk management, data controls, and model governance to mitigate deepfake and data privacy risks in synthetic media deployment”, published on December 29, 2020, available at
https://saudijournals.com/media/articles/SIJLCJ_312_468-478_Replace.pdf , (hereafter, BADMUS).
Claim 8:
Regarding claim 8, SAVAZZI in view of LANGER, further in view of HALL, teach the limitations in claim 6.
However, SAVAZZI in view of LANGER, further in view of HALL, did not teach “the method of claim 6, wherein the metadata comprises one or more of: resource availability information; validation information; updated model architecture privacy information; and notification of a variation in a model architecture from among the model architectures of the trained ML models,”
In an analogous field, BADMUS teaches “metadata comprises one or more of: resource availability information;
See BADMUS discuss in page 471, “Tools like Kubernetes and Docker, which once served purely operational roles, are now evolving to support policy-based orchestration and runtime risk evaluation. For example, Kubernetes clusters can be equipped with AI agents that dynamically allocate resources based on usage patterns, while also enforcing guardrails to prevent data leakage or misuse.” Here, BADMUS describes tracking usage patterns (i.e. resource availability information) to prevent data misuse or leakage while evaluating runtime operations. For details, see BADMUS in page 469, section 2. Literature review, second paragraph.
Further, BADMUS teaches “validation information;”
See BADMUS in page 472, section 5. Regulatory landscape and legal risk management, fourth paragraph, mention “To meet these regulatory obligations, legal risk management must be integrated directly into AI development workflows. This includes embedding compliance checks at each phase of the DevOps pipeline, such as validating data provenance during ingestion, performing privacy risk assessments before model training, and scanning outputs for sensitive or legally problematic content before release... An emerging concept in legal risk mitigation is the “Model Card,” which documents the intended use, limitations, data sources, and ethical considerations of an AI model. While originally proposed as a transparency mechanism, model cards now play a vital legal role by providing a compliance record and defining boundaries for safe deployment.” Here, BADMUS mentions performing compliance checks such as validating data provenance, where data provenance is viewed as a documented history of a dataset and is part of metadata. Later, BADMUS mentions using a model card (i.e. metadata) to document usage, data sources and information regarding an AI model, and includes information such as validating data provenance (i.e. validation information).
Further, BADMUS teaches “updated model architecture privacy information; and notification of a variation in a model architecture from among the model architectures of the trained ML models”
See BADMUS in page 473, section 6, second paragraph, describe “Organizations must be able to document not just what the model generates, but how it was trained, what data was used, and under what conditions it was deployed. Model cards, audit logs, and deployment metadata are all tools that help ensure traceability, which is particularly important in regulatory environments that require disclosure of AI decision-making processes.” Later, see BADMUS in page 473, section 7, Data Privacy Controls and Ethical AI Practices, mention “another critical practice is implementing privacy-preserving architectures within the AI pipeline. This includes deploying techniques such as homomorphic encryption, secure multi-party computation, and local differential privacy. These methods ensure that data remains protected both in transit and at rest, without compromising model performance.” Here, BADMUS describes the model cards, audit logs, and deployment metadata as part of metadata that helps organize information such as local differential privacy and privacy-preserving architectures for models. Further, BADMUS in page 474, section 8. Analysis and evaluation, second to last paragraph, describes “when regulatory environments change, such as the introduction of the EU AI Act or regional digital rights laws, these systems enable dynamic policy injection without requiring the restructuring of the entire pipeline. This modularity ensures sustained compliance without productivity loss.” Here, BADMUS mentions that the systems that enable dynamic policy injection shows that the metadata is dynamically updated to reflect any changes.
See BADMUS in page 476, section 10, describe “develop unified APIs and open compliance frameworks that allow these tools to share context, propagate alerts,” where BADMUS shows that alerts is construed to mean the same as notifications. Further, see BADMUS in page 474, section 8. Analysis and Evaluation, describe “time compliance checks and policy gates throughout the CI/CD lifecycle has been shown to significantly reduce instances of unauthorized data use and leakage, particularly in organizations that utilize federated data architectures,” and see BADMUS mention in page 474, section 7 “these tools evaluate datasets and outputs for disparities and recommend mitigation actions such as data rebalancing or fairness-aware model retraining (Khandelwal & Rao, 2018)... Documentation mechanisms, such as data sheets for datasets and model cards for AI systems, help operationalize this transparency. These documents provide metadata, usage policies, and known limitations,” BADMUS here in section 7 describes a system that detects any disparities (i.e. variation) of the models later used for retraining (construed to mean trained models), from the federated data architectures (i.e. federated learning model architectures).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of SAVAZZI, LANGER, and HALL, and incorporate with the teachings of BADMUS by using the teachings of SAVAZZI, LANGER, and HALL, with the teaching of BADMUS of the types of metadata.
One of ordinary skill in the art would be motivated to do so because by integrating BADMUS’s framework into the methods of SAVAZZI, LANGER, and HALL, one with ordinary skill in the art would achieve a method "compliance-aware DevOps … represents a cultural and strategic shift toward responsible innovation in the age of generative AI. It aligns development efficiency with societal accountability, ensuring that as synthetic content becomes more influential, it remains governed by trust, transparency, and law. .. requires adaptive governance models that can respond to the rapidly evolving dynamics of AI regulation and risk”, (see BADMUS in page 477, last paragraph).
Claim 23:
Regarding claim 23, SAVAZZI in view of LANGER, further in view of HALL, teaches the limitations in claim 21. Regarding claim 23, the claim recites similar limitations as corresponding claim 8 and is rejected for similar reasons as claim 8 using similar teachings and rationale.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to WENWEI ZENG whose telephone number is (571)272-7111. The examiner can normally be reached Monday-Friday, 8am-5pm.
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, Usmaan Saeed can be reached at (571) 272-4046. 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.
/WenWei Zeng/Examiner, Art Unit 2146
/USMAAN SAEED/Supervisory Patent Examiner, Art Unit 2146
Prosecution Insights