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(s) (IDS) submitted on February 5, 2026 is/are in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement(s) is/are being considered by the examiner.
Amendments
This Office Action is in response to the amendment filed on February 27, 2026.
Claim(s) 1, 8, 15, 17, 18, 22, 24, and 25 have been amended.
No claims have been cancelled.
No new claims have been added.
The objections and rejections from the prior correspondence that are not restated herein are withdrawn.
Response to Arguments
Applicant's arguments filed on February 27, 2026 have been fully considered.
Applicant’s arguments regarding 35 USC 101 rejections have been fully considered, but are not persuasive. Applicant argues:
“Applicant submits that the claims are eligible under Step 2A, Prong 2 because various features of the claims, in fact, integrate any alleged abstract idea into a practical application – namely, improving computational efficiency of federated learning systems. In particular, Applicant submits that this practical application reflects an improvement to other technology or a technical field (namely the field of machine learning and, more specifically, to federated learning systems). Specifically, any alleged abstract idea has been integrated into the practical application of inducing parameter sparsity during training of federated machine learning models. Specification at para. [0080]. Applicant respectfully directs the Examiner's attention to Example 48, Claim 3. See July 2024 Subject Matter Eligibility Examples, p. 18. Example 48 describes an artificial-intelligence based method of analyzing speech signals and separating desired speech from extraneous or background speech. The USPTO explained that while Claim 3 of Example 48 recites abstract ideas, the Claim is directed to an improvement to existing speech-to-text technology. Specifically, the USPTO noted how the Claim recited that the deep neural network, which was trained on speech source separation, could be used to make "individual transcription of each separated speech signal possible." Id, p. 28 (citing M.P.E.P. § 2106.0S(a)). Accordingly, the USPTO said that this Claim was patent eligible under Step 2A, Prong 2.
Likewise, here, the features of the present claims reflect an improvement to a technology or technical field. For example, the practical application improves the technical field of machine learning (and more particularly, to federated learning). More specifically, inducing sparsity during training of a federated machine learning model may include a server generating a global model in a first state, where edges between the nodes in global model are "associated with parameters, including a weight w and a gate probability 0. " Specification at para. [0072]. The server may then sample the global model weights according to the associated gate probabilities in order to "generate various subsets of weights and gate probabilities for" a set of client devices. Id. at para. [0073]. The client devices may then generate updates with their associated subsets, where the updates may be used to update the global model at the server. Id. at paras. [0076]-[0077]. Certain techniques described therein may reduce the amount of computational resources used to perform federated learning. Id. Thus, particular solutions discussed therein improve upon conventional techniques, such as by enabling more efficient use of computational resources.
Thus, the pending claims are eligible because the claims as a whole improve federated learning technology and thus integrate the exception into a practical application of encouraging sparsity in federated learning and are therefore not "directed to" the judicial exception.
Further, Applicant respectfully directs the Examiner's attention to the Board decision in Ex parte Chari, in which claims directed to solutions rooted in computer technology to solve a problem that does not exist in the absence of computer technology were found to be eligible under Step 2A, Prong 2 of the Alice/Mayo test. Ex parte Chari, Appeal No. 2018-009014 (Sept. 10, 2020) at 12-13. In that case, the Board found that the recitation of a machine learning model that applies specific techniques to provide a specific improvement to a computer system-and allows for efficient and scalable use of a computing system-constituted integration into a practical application such that the claims were patent eligible under the Alice/Mayo test. Likewise, the present claims are not merely instructions to apply an exception using a generic computer component, but provide a specific solution ( of federated learning via encouragement of sparsity) to a problem that does not exist outside of actual and practical implementation of federated learning technology (e.g., the computational demands of conventional federated learning techniques). Accordingly, Applicant submits that the claims are directed to a specific improvement in machine learning technology, and are thus eligible subject matter under Step 2A, Prong 2 of the Alice/Mayo test.”
Examiner respectfully disagrees. Example 48 in the “July 2024 Subject Matter Eligibility Examples” integrates the judicial exception into a practical application because it discloses a technical improvement in speech-to-text conversion that makes possible the individual transcription of each separated speech signal. The present application induces sparsity for reducing the number of computational resources used to perform federated learning. However, an abstract idea cannot provide the integration into a practical application; the integration must be provided by the additional elements. The additional elements in the present application, when considered as a whole, do not integrate the judicial exception into a practical application. The alleged computational resource savings are the result of transmitting and training fewer model elements via sparsity. This is simply reducing the amount of data and computation, and does not provide a specific technological improvement to technology or how the computer, memory, or processor operates. Therefore, the claims as a whole do not provide an improvement to federated learning technology, as shown in the 101 rejections below.
Applicant further argues:
“In the Office Action, the Examiner argues that the claims do "not include additional elements that are sufficient to amount to significantly more than the judicial exception." Office Action, p. 4, 6, 8, 9, 10, 13, and 16. However, as discussed with respect to Step 2A, Prong 2 supra, the Specification provides sufficient details to show how the claimed features improve the technical field of machine learning (e.g., federated learning via induced sparsity). For example, the Specification describes how inducing sparsity during training of a federated machine learning model may include a server generating a global model in a first state, where edges between the nodes in global model are "associated with parameters, including a weight wand a gate probability 0. " Specification at para. [0072]. The server may then sample the global model weights according to the associated gate probabilities in order to "generate various subsets of weights and gate probabilities for" a set of client devices. Id. at para. [0073]. The client devices may then generate updates with their associated subsets, where the updates may be used to update the global model at the server. Id. at paras. [0076]-[0077]. Induced sparsity during federated learning can beneficially "reduce communications costs." Id. at para. [0081]. Id.
Applicant further respectfully directs the Examiner's attention to the Federal Circuit's decision in Cosmokey Solutions GMBH & Co. KG v. Duo Security LLC, in which a finding of ineligibility under Section 101 was reversed by the Federal Circuit. 15 F.4th 1091, 1097-98 (Fed. Cir. 2021 ). In that case, and in support of the holding that the claims were directed to eligible subject matter under Step 2B oftheAliceMayo test, the Federal Circuit noted that the specification "describes how the particular arrangement of steps in Claim 1 provides a technical improvement" and "emphasizes the inventive nature" of the steps recited in the claims. Id at 1099. Likewise, as discussed supra, the Specification discusses how the particular arrangement of steps recited in the claims provides a technical improvement. Thus, like the patents at issue in Cosmokey, the claims clearly provide a technical improvement to a technical field (e.g., a technical improvement to the field of federated learning).
Applicant further contends that independent Claims 1 and 8 are basically directed to (1) receiving model elements and a set of gate probabilities at a device from a server (another device), (2) generating a set of model updates based thereon at the device, and then (3) transmitting the set of model updates from the device to the server. Similarly, independent Claims 15 and 22 are basically directed to, for each client (device), (a) generating a subset of model elements at a server (another device), (b) transmitting the subset of model elements and a set of gate probabilities from the server to the respective client (device), and then ( c) receiving model updates at the server from each respective client (device). This transmission and reception between devices should reflect significantly more than any abstract idea. If not, then the claims of every wireless communication patent application (e.g., WiFi, WLAN, WAN, 3G, 4G, and 5G communications) should likely analogously be rejected for ineligible subject matter. However, this has not been the case from the USPTO. The claims of this application should be no different just because they are concerned with the subject matter of federated learning. The Examiner should be considering the claims as a whole.
For at least the reasons presented above, Applicant submits that the claims are eligible under Step 2B of the Alice/Mayo test and are therefore directed to eligible subject matter under § 101. Thus, Applicant respectfully requests that the rejection under 35 U.S.C. § 101 be withdrawn.”
Examiner respectfully disagrees. The additional elements do not amount to significantly more. It should be noted that generating/generate […] a set of model updates based on training the local machine learning model based on the set of model elements and the set of gate probabilities for inducing sparsity is a judicial exception, and thus cannot be analyzed under Alice/Mayo Step 2B. Furthermore, the additional elements recited in the claim describe generic components and steps for performing federated learning. The transmission and reception between is routine and conventional activity of receiving or transmitting data over a network (see MPEP § 2106.05(d)(ll)(i) - Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information)). Thus, the additional elements do not provide an unconventional or specific arrangement of elements or steps that would amount to significantly more than the judicial exception.
Applicant’s arguments regarding 35 USC 103 rejections have been fully considered, but are not persuasive. Applicant argues:
“The Examiner maps Louizos's description of"non-negative stochastic gates ... for Lo norm regularization for neural networks" to the claimed "set of gate probabilities." Office Action, p. 19 (citing Louizos at Abstract). Louizos describes "binary 'gates' that denote whether a parameter is present" (emphasis added). Louizos at p. 3, paragraph 1. However, because a binary indication whether a parameter is present does not teach "a likelihood that [an] associated model element will be included in a local machine learning model," Louizos does not teach "a set of gate probabilities, wherein each gate probability of the set of gate probabilities is associated with one model element of the subset of model elements and represents a likelihood that the associated model element will be included in a local machine learning model," as recited in Claim 1 and similar features recited in Claims 8, 15, and 22.
Caldas does not cure the deficiency of Louizos, nor does the Examiner rely on Caldas for teaching these features. Office Action, p. 19 and 26.
Accordingly, Applicant submits that Claims 1, 8, 15, and 22, as well as claims dependent thereon, are allowable over Caldas in view of Louizos and respectfully requests withdrawal of this rejection.”
Examiner respectfully disagrees. The newly amended limitation requires:
and represents a likelihood that the associated model element will be included in a local machine learning model;
CALDAS in combination with LOUIZOS teaches this limitation, as shown in the detailed 103 rejection below. Specifically, CALDAS teaches:
a local machine learning model (CALDAS [Pg. 2, section 1. Introduction] teaches: “Our approach enables each device to locally operate on a smaller sub-model (i.e., local machine learning model) (i.e. with smaller weight matrices) while still providing updates that can be applied to the larger global model on the server.”).
CALDAS is not relied upon for teaching, but LOUIZOS teaches: and represents a likelihood that the associated model element will be included in a […] machine learning model;
(LOUIZOS [Pg. 3, section 2.1 A General Recipe For Efficiently Minimizing
L
0
Norms] teaches: “Consider the
L
0
norm under a simple re-parametrization of
θ
:
PNG
media_image1.png
69
709
media_image1.png
Greyscale
where the
z
j
correspond to binary “gates” that denote whether a parameter is present and the
L
0
norm corresponds to the amount of gates being “on”.” By letting
q
z
j
|
π
j
=
B
e
r
n
π
j
be a Bernoulli distribution over each gate
z
j
we can reformulate the minimization of Eq. 1 as penalizing the number of parameters being used, on average, as follows:
PNG
media_image2.png
128
785
media_image2.png
Greyscale
where
⨀
corresponds to the elementwise product.” Examiner’s note: Under BRI, the associated model element can be interpreted as the gated parameter
θ
j
and represents a likelihood can be interpreted as
π
j
, which controls the probability that a gate
z
j
is “on” (i.e., included in a […] machine learning model).)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of CALDAS and LOUIZOS before them, to include LOUIZOS' collection of gates to determine which weights of a network to set to zero in CALDAS' federated learning method. Both CALDAS and LOUIZOS teach fully connected models/layers, and LOUIZOS teaches introducing a gate per input neuron for fully connected layers. One would have been motivated to make such a combination of including a gate per input neuron in order to improve the accuracy of the dropout equivalent network for practical computation savings while training, a benefit not possible with the commonly used independent dropout mask per spatial location (LOUIZOS [Pg. 6, section 2.4 Group Sparsity under an
L
0
Norm] and [Pg. 9, section 4.2 CIFAR Classification]).
Applicant further argues:
“Claims 7, 14, 21, and 28 stand rejected under 35 U.S.C. § 103 as being allegedly unpatentable over Caldas in view of Louizos and further in view of Jiang et al. ("Model Pruning Enables Efficient Federated Learning on Edge Devices," hereinafter "Jiang").
Applicant respectfully traverses this rejection.
Claims 7, 14, 21, and 28 each depend from one of independent Claims 1, 8, 15, and 22, which Applicant submits are allowable over Caldas in view of Louizos for at least the reasons discussed above. Further, the Examiner relies on Jiang as disclosing various elements of dependent Claims 7, 14, 21, and 28; however, Jiang fails to overcome the deficiencies of Caldas in view of Louizos with respect to independent Claims 1, 8, 15, and 22. Therefore, Claims 7, 14, 21, and 28 are believed to be allowable at least due to their dependence from an allowable base claim and for their additional distinguishing features recited therein. See M.P.E.P. § 2143.03; In re Fine, 837 F.2d at 1076. Withdrawal of this rejection is respectfully requested.”
Examiner respectfully disagrees. The independent claims 1, 8, 15, and 22 are not allowable over CALDAS in view of LOUIZOS, as shown in the detailed 103 rejections below.
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 1-28 are rejected under 35 U.S.C.101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1: Claims 1-7 and 15-21 are directed to a process. Claims 8-14 and 22-28 are directed to a machine or an article of manufacture.
With respect to claim(s) 1 and 8:
2A Prong 1: The claim(s) recite(s) an abstract idea. Specifically:
generating/generate […] a set of model updates based on training the local machine learning model based on the set of model elements and the set of gate probabilities; and (Mathematical concepts – generating model updates based on the set of model elements and the set of gate probabilities involves mathematical calculations (see [0050-0056])– see MPEP § 2106.04(a)(2)(I))
If claim limitations, under their broadest reasonable interpretation, cover performance of the limitations as a mental process, but for the recitation of generic computer components, then the claim limitations fall within the mathematical or mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
2A Prong 2: The additional elements recited in the claim(s) do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
(Claim 8) memory comprising computer-executable instructions; (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
(Claim 8) one or more processors configured to execute the computer-executable instructions and cause the processing system to: (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
by the device (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
receiving/receive at a device from a server managing federated learning of a global machine learning model: a subset of model elements from a set of model elements for the global machine learning model; and a set of gate probabilities, wherein each gate probability of the set of gate probabilities is associated with one model element of the subset of model elements and represents a likelihood that the associated model element will be included in a local machine learning model; (Mere data gathering – Adding insignificant extra-solution activity of mere data gathering to the judicial exception – see § MPEP2106.05(g).)
transmitting/transmit from the device to the server a set of model updates. (Adding insignificant extra-solution activity to the judicial exception – see § MPEP2106.05(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.
2B: The claim(s) do(es) not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
(Claim 8) memory comprising computer-executable instructions; and (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
(Claim 8) one or more processors configured to execute the computer-executable instructions and cause the processing system to: (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
by the device (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
receiving/receive at a device from a server managing federated learning of a global machine learning model: a subset of model elements from a set of model elements for the global machine learning model; and a set of gate probabilities, wherein each gate probability of the set of gate probabilities is associated with one model element of the subset of model elements and represents a likelihood that the associated model element will be included in a local machine learning model; (Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception (WURC)- see MPEP § 2106.05(d)(ll)(i) - Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information).)
transmitting/transmit from the device to the server a set of model updates. (Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception (WURC)- see MPEP § 2106.05(d)(ll)(i) - Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information).)
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.
With respect to claim(s) 2, 9, 16, and 23:
2A Prong 2: The additional elements recited in the claim(s) do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
wherein the subset of model elements comprises a subset of weights associated with edges connecting nodes in the global machine learning model. (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
2B: The claim(s) do(es) not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
wherein the subset of model elements comprises a subset of weights associated with edges connecting nodes in the global machine learning model. (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – 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. Therefore, the claim is not patent eligible.
With respect to claim(s) 3 and 10:
2A Prong 1: The claim(s) recite(s) an abstract idea. Specifically:
wherein the set of model updates comprises: a set of weight gradients associated with the local machine learning model; and a set of gate probability gradients associated with the local machine learning model. (Mathematical concepts – model updates comprising weight gradients and gate probability gradients recite mathematical concepts (See [0035-0038], and [0055-0065]) – see MPEP § 2106.04(a)(2)(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. Therefore, the claim is not patent eligible.
With respect to claim(s) 4 and 11:
2A Prong 1: The claim(s) recite(s) an abstract idea. Specifically:
wherein the set of model updates comprises: a set of weight gradients associated with the local machine learning model; and a binary gate variable value associated with each weight gradient of the set of weight gradients. (Mathematical concepts – model updates comprising weight gradients and binary gate variable value recite mathematical concepts (See [0052], [0035-0038], and [0055-0065]) – see MPEP § 2106.04(a)(2)(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. Therefore, the claim is not patent eligible.
With respect to claim(s) 5, 12, 19, and 26:
2A Prong 2: The additional elements recited in the claim(s) do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
wherein the subset of model elements comprises a subset of nodes in the global machine learning model. (Generally linking the use of a judicial exception to a particular technological environment or field of use – see MPEP § 2106.05(h).)
2B: The claim(s) do(es) not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
wherein the subset of model elements comprises a subset of nodes in the global machine learning model. (Generally linking the use of a judicial exception to a particular technological environment or field of use – 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. Therefore, the claim is not patent eligible.
With respect to claim(s) 6, 13, 20, and 27:
2A Prong 2: The additional elements recited in the claim(s) do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
wherein the subset of model elements comprises a subset of channels in a convolution filter of the global machine learning model. (Generally linking the use of a judicial exception to a particular technological environment or field of use – see MPEP § 2106.05(h).)
2B: The claim(s) do(es) not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
wherein the subset of model elements comprises a subset of channels in a convolution filter of the global machine learning model. (Generally linking the use of a judicial exception to a particular technological environment or field of use – 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. Therefore, the claim is not patent eligible.
With respect to claim 7:
2A Prong 1: The claim(s) recite(s) an abstract idea. Specifically:
wherein the final set of model elements corresponds to a pruned global machine learning model. (Mathematical concepts – a pruned global machine learning model involves mathematical calculations (see [0068]) – see MPEP § 2106.04(a)(2)(I))
2A Prong 2: The additional elements recited in the claim(s) do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
receiving at the device a final set of model elements from the server, (Mere data gathering – Adding insignificant extra-solution activity of mere data gathering to the judicial exception – see § MPEP2106.05(g).)
2B: The claim(s) do(es) not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
receiving at the device a final set of model elements from the server, (Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception (WURC)- see MPEP § 2106.05(d)(ll)(i) - Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information).)
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. Therefore, the claim is not patent eligible.
With respect to claim 14:
2A Prong 2: The additional elements recited in the claim(s) do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
receive a final set of model elements from the server, (Mere data gathering – Adding insignificant extra-solution activity of mere data gathering to the judicial exception – see § MPEP2106.05(g).)
wherein the final set of model elements corresponds to a pruned global machine learning model. (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
2B: The claim(s) do(es) not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
receive a final set of model elements from the server, (Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception (WURC)- see MPEP § 2106.05(d)(ll)(i) - Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information).)
wherein the final set of model elements corresponds to a pruned global machine learning model. (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – 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. Therefore, the claim is not patent eligible.
With respect to claim(s) 15 and 22:
2A Prong 1: The claim(s) recite(s) an abstract idea. Specifically:
(Claim 15) generating, […]
(Claim 22) generate […]
[…] a subset of model elements for the respective client based on sampling a gate probability distribution for each model element of a set of model elements for a global machine learning model; (Mathematical concepts – generating model elements based on a gate probability distribution involves mathematical calculations (See [0055-0065]) – see MPEP § 2106.04(a)(2)(I))
(Claim 15) updating […]
(Claim 22) update […]
[…] the global machine learning model based on the respective set of model updates from each respective client of the plurality of clients. (Mathematical concepts – updating the global machine learning model based on model updates involves mathematical calculations (see [0055-0065]) – see MPEP § 2106.04(a)(2)(I))
If claim limitations, under their broadest reasonable interpretation, cover performance of the limitations as a mental process, but for the recitation of generic computer components, then the claim limitations fall within the mathematical or mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
2A Prong 2: The additional elements recited in the claim(s) do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
(Claim 22) a memory comprising computer-executable instructions; and (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
(Claim 22) one or more processors configured to execute the computer-executable instructions and cause the processing system to: (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
(Claim 15) by a server, (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
for each respective client of a plurality of clients and for each training round in a plurality of training rounds: (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
(Claim 15) transmitting from the server […]
(Claim 22) transmitting […]
[…] to the respective client: the subset of model elements; and a set of gate probabilities based on the sampling, wherein each gate probability of the set of gate probabilities is associated with one model element of the subset of model elements and represents a likelihood that the associated model element will be included in a local machine learning model; (Adding insignificant extra-solution activity to the judicial exception – see § MPEP2106.05(g).)
(Claim 15) receiving at the server
(Claim 22) receiving […]
[…] from each respective client of the plurality of clients a respective set of model updates; and (Mere data gathering – Adding insignificant extra-solution activity of mere data gathering to the judicial exception – see § MPEP2106.05(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.
2B: The claim(s) do(es) not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
(Claim 22) a memory comprising computer-executable instructions; and (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
(Claim 22) one or more processors configured to execute the computer-executable instructions and cause the processing system to: (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
(Claim 15) by a server, (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
for each respective client of a plurality of clients and for each training round in a plurality of training rounds: (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
(Claim 15) transmitting from the server […]
(Claim 22) transmitting […]
[…] to the respective client: the subset of model elements; and a set of gate probabilities based on the sampling, wherein each gate probability of the set of gate probabilities is associated with one model element of the subset of model elements and represents a likelihood that the associated model element will be included in a local machine learning model; (Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception (WURC)- see MPEP § 2106.05(d)(ll)(i) - Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information).)
(Claim 15) receiving at the server
(Claim 22) receiving […]
[…] from each respective client of the plurality of clients a respective set of model updates; and (Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception (WURC)- see MPEP § 2106.05(d)(ll)(i) - Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information).)
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. Therefore, the claim is not patent eligible.
With respect to claim(s) 17 and 24:
2A Prong 1: The claim(s) recite(s) an abstract idea. Specifically:
wherein the respective set of model updates comprises: a set of weight gradients associated with the local machine learning model trained by the respective client; and a set of gate probability gradients associated with the local machine learning model trained by the respective client. (Mathematical concepts – model updates comprising weight gradients and gate probability gradients recite mathematical concepts (See [0035-0038], and [0055-0065]) – see MPEP § 2106.04(a)(2)(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. Therefore, the claim is not patent eligible.
With respect to claim(s) 18 and 25:
2A Prong 1: The claim(s) recite(s) an abstract idea. Specifically:
wherein the respective set of model updates comprises: a set of weight gradients associated with the local machine learning model trained by the respective client; and a binary gate variable value associated with each weight gradient of the set of weight gradients. (Mathematical concepts – model updates comprising weight gradients and binary gate variable value recite mathematical concepts (See [0052], [0035-0038], and [0055-0065]) – see MPEP § 2106.04(a)(2)(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. Therefore, the claim is not patent eligible.
With respect to claim 21:
2A Prong 1: The claim(s) recite(s) an abstract idea. Specifically:
wherein updating […] the global machine learning model based on the respective set of model updates from each respective client of the plurality of clients further comprises pruning the updated global machine learning model based on updated gate probabilities for the global machine learning model and a threshold gate probability value. (Mathematical concepts – pruning the updated global machine learning model based on updated gate probabilities and a threshold gate probability value involves mathematical calculations (see [0068]) – see MPEP § 2106.04(a)(2)(I))
2A Prong 2: The additional elements recited in the claim(s) do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
by the server (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
2B: The claim(s) do(es) not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
by the server (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – 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. Therefore, the claim is not patent eligible.
With respect to claim 28:
2A Prong 1: The claim(s) recite(s) an abstract idea. Specifically:
wherein in order to update the global machine learning model based on the respective set of model updates from each respective client of the plurality of clients, the one or more processors are further configured to prune the updated global machine learning model based on updated gate probabilities for the global machine learning model and a threshold gate probability value. (Mathematical concepts – pruning the updated global machine learning model based on updated gate probabilities and a threshold gate probability value involves mathematical calculations (see [0068]) – see MPEP § 2106.04(a)(2))
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. Therefore, the claim is not patent eligible.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-6, 8-13, 15-20, and 22-27 are rejected under 35 U.S.C. 103 as being unpatentable over CALDAS ("Expanding the Reach of Federated Learning by Reducing Client Resource Requirements") in view of LOUIZOS ("Learning Sparse Neural Networks Through
L
0
Regularization"), hereafter CALDAS and LOUIZOS.
Regarding Claim 1, CALDAS teaches:
receiving at a device from a server managing federated learning of a global machine learning model: (CALDAS [Pg. 2, Figure 1] teaches a server sending a compressed model constructed from a global model (i.e., global machine learning model) to a client’s device (i.e., receiving at a device) for federated learning (i.e., managing federated learning).)
receiving at a device from a server managing federated learning of a global machine learning model: […] a subset of model elements from a set of model elements for the global machine learning model; (CALDAS [Pg. 2, section I Introduction] and [Pg. 2, Figure 1] teaches a client’s device receiving a compressed model (e.g., sub-model) from a server by transmitting a reduced structure with only the necessary coefficients to the client for local training (see CALDAS [pg. 4, section 3.2 Federated Dropout]). Furthermore, CALDAS [pg. 4, section 3.2 Federated Dropout] teaches that the sub-models are subsets of the global model, and thus a subset of model elements from a set of model elements can be understood as the received compressed model from the global model (i.e., global machine learning model) sent by the server to the client’s device.)
generating by the device a set of model updates based on training the local machine learning model based on the set of model elements […]; (CALDAS [Pg. 4, section 3.2 Federated Dropout] teaches the client training (i.e., generating) the received sub-model based on the compressed global model (i.e., based on the set of model elements) and sending the sub-model updates (i.e., set of model updates based on training the local machine learning model) to the server.)
transmitting from the device to the server a set of model updates. (CALDAS [Pg. 2, Figure 1] teaches the client sending (i.e., transmitting from the device) the compressed final update (i.e., a set of model updates) back to the server.)
a local machine learning model; (CALDAS [Pg. 2, section 1. Introduction] teaches: “Our approach enables each device to locally operate on a smaller sub-model (i.e., local machine learning model) (i.e. with smaller weight matrices) while still providing updates that can be applied to the larger global model on the server.”)
CALDAS is not relied upon for teaching:
receiving […] a set of gate probabilities, wherein each gate probability of the set of gate probabilities is associated with one model element of the subset of model elements and represents a likelihood that the associated model element will be included in a […] machine learning model;
generating by the device a set of model updates […] the set of gate probabilities;
However, LOUIZOS teaches: receiving […] a set of gate probabilities (LOUIZOS [Abstract] teaches the inclusion of a collection of non-negative stochastic gates (i.e., a set of gate probabilities) for
L
0
norm regularization for neural networks and jointly optimizing the parameters (i.e., generating […] a set of model updates) of the distribution over the gates with the original network parameters.)
wherein each gate probability of the set of gate probabilities is associated with one model element of the subset of model elements (LOUIZOS [Pg. 6, section 2.4 Group Sparsity Under an
L
0
Norm] teaches introducing a gate per input neuron for fully connected layers (i.e., gate probabilities is associated with one model element of the subset of model elements).)
and represents a likelihood that the associated model element will be included in a […] machine learning model; (LOUIZOS [Pg. 3, section 2.1 A General Recipe For Efficiently Minimizing
L
0
Norms] teaches: “Consider the
L
0
norm under a simple re-parametrization of
θ
:
PNG
media_image1.png
69
709
media_image1.png
Greyscale
where the
z
j
correspond to binary “gates” that denote whether a parameter is present and the
L
0
norm corresponds to the amount of gates being “on”.” By letting
q
z
j
|
π
j
=
B
e
r
n
π
j
be a Bernoulli distribution over each gate
z
j
we can reformulate the minimization of Eq. 1 as penalizing the number of parameters being used, on average, as follows:
PNG
media_image2.png
128
785
media_image2.png
Greyscale
where
⨀
corresponds to the elementwise product.” Examiner’s note: Under BRI, the associated model element can be interpreted as the gated parameter
θ
j
and represents a likelihood can be interpreted as
π
j
, which controls the probability that a gate
z
j
is “on” (i.e., included in a […] machine learning model).)
generating by the device a set of model updates […] the set of gate probabilities; (LOUIZOS [Abstract] and [Pg. 6, section 2.4 Group Sparsity Under an
L
0
Norm] teaches jointly optimizing parameters over the distribution of gates and original network parameters, and thus a person of ordinary skill in the art could apply LOUIZOS’ gate probabilities to the compressed model structure sent by the server for conducting local training and generating updates, as taught in CALDAS [Pg. 4, section 3.2 Federated Dropout].)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of CALDAS and LOUIZOS before them, to include LOUIZOS' collection of gates to determine which weights of a network to set to zero in CALDAS' federated learning method. Both CALDAS and LOUIZOS teach fully connected models/layers, and LOUIZOS teaches introducing a gate per input neuron for fully connected layers. One would have been motivated to make such a combination of including a gate per input neuron in order to improve the accuracy of the dropout equivalent network for practical computation savings while training, a benefit not possible with the commonly used independent dropout mask per spatial location (LOUIZOS [Pg. 6, section 2.4 Group Sparsity under an
L
0
Norm] and [Pg. 9, section 4.2 CIFAR Classification]).
Regarding Claim 2, CALDAS in view of LOUIZOS teaches the elements of claim 1 as outlined above. Further, CALDAS teaches:
The method of Claim 1, wherein the subset of model elements comprises a subset of weights associated with edges connecting nodes in the global machine learning model. (CALDAS [Pg. 2, Introduction] teaches that their approach allows each device to locally operate on a smaller sub-model with smaller weight matrices (i.e., a subset of weights). Furthermore, CALDAS [Pg. 4, section 3.2 Federated Dropout] and [Pg. 2, Figure 1] teaches that these sub-models are reduced architectures constructed from a global model, and thus are subsets of the global model. Therefore, the subset of model elements can be understood as the sent compressed fully connected model (i.e., associated with edges connecting nodes in the global machine learning model) by the server to the client. A fully connected model (or network) is a type of neural network where every neuron (i.e., node) in a layer is connected (i.e., edges) to every neuron in preceding layers, and thus fully connected.)
Regarding Claim 3, CALDAS in view of LOUIZOS teaches the elements of claim 2 as outlined above. Further, CALDAS teaches:
The method of Claim 2, wherein the set of model updates comprises: a set of weight gradients associated with the local machine learning model; (CALDAS [Pg. 4, section 3.2 Federated Dropout] teaches the client training the received sub-model with smaller weight matrices and sending its updates (i.e., a set of model updates […] associated with the local machine learning model] to the server for mapping back to the global model. Furthermore, CALDAS [Pg. 5, section 4.2 Lossy compression] teaches client-to-server gradient updates (i.e., a set of weight gradients).)
Further, LOUIZOS teaches: and a set of gate probability gradients associated with the local machine learning model. (LOUIZOS [Abstract] and [Pg. 6, section 2.4 Group Sparsity Under an
L
0
Norm] teaches jointly optimizing parameters of the distribution over the gates (i.e., a set of gate probabilities) and the original network parameters, where each input neuron has a gate for a fully connected network. Furthermore, LOUIZOS [Pg. 4, section 2.1 A General Recipe for Efficiently Minimizing
L
0
Norms] teaches the re-parametrization of the network parameters
θ
as
θ
j
=
θ
j
~
z
j
, where
z
j
is the binary gate that denote whether a parameter (i.e., associated with the local machine learning model) is present and the
L
0
norm corresponds to the number of gates being “on”. LOUIZOS Eq. (4) and Eq. (5) teaches the giving the binary gates a hard-sigmoid rectification of
s
2
, and thus allowing for the gates to be a continuous random variable. Furthermore, LOUIZOS Eq. 9 teaches the total cost is differentiable w.r.t.
ϕ
, thus enabling for efficient stochastic gradient based optimization, and that due to rectifications, the gradient of log-likelihood w.r.t. the parameters
ϕ
of
q
s
is sparse.)
Regarding Claim 4, CALDAS in view of LOUIZOS teaches the elements of claim 2 as outlined above. Further, CALDAS teaches:
The method of Claim 2, wherein the set of model updates comprises: a set of weight gradients associated with the local machine learning model; (CALDAS [Pg. 4, section 3.2 Federated Dropout] teaches the client training the received sub-model with smaller weight matrices and sending its updates (i.e., a set of model updates […] associated with the local machine learning model] to the server for mapping back to the global model. Furthermore, CALDAS [Pg. 5, section 4.2 Lossy compression] teaches client-to-server gradient updates (i.e., a set of weight gradients).)
Further, LOUIZOS teaches: and a binary gate variable value associated with each weight gradient of the set of weight gradients. (LOUIZOS [Abstract] and [Pg. 6, section 2.4 Group Sparsity Under an
L
0
Norm] teaches jointly optimizing parameters of the distribution over the gates (i.e., a set of gate probabilities) and the original network parameters, where each input neuron has a gate for a fully connected network. Furthermore, LOUIZOS [Pg. 4, section 2.1 A General Recipe for Efficiently Minimizing
L
0
Norms] teaches the re-parametrization of the network parameters
θ
as
θ
j
=
θ
j
~
z
j
, where
z
j
is the binary gate (i.e., binary gate variable) that denote whether a parameter is present (i.e., associated with each weight gradient of the set of weight gradients) and the
L
0
norm corresponds to the number of gates being “on”.)
Regarding Claim 5, CALDAS in view of LOUIZOS teaches the elements of claim 1 as outlined above. Further, CALDAS teaches:
The method of Claim 1, wherein the subset of model elements comprises a subset of nodes in the global machine learning model. (CALDAS [Pg. 2, Introduction] teaches that their approach allows each device to locally operate on a smaller sub-model with smaller weight matrices. Furthermore, CALDAS [Pg. 4, section 3.2 Federated Dropout] and [Pg. 2, Figure 1] teaches that these sub-models are reduced architectures constructed from a global model, and thus are subsets of the global model. Therefore, the subset of model elements can be understood as the sent compressed fully connected model (i.e., a subset of nodes in the global machine learning model) by the server to the client.)
Regarding Claim 6, CALDAS in view of LOUIZOS teaches the elements of claim 1 as outlined above. Further, CALDAS teaches:
The method of Claim 1, wherein the subset of model elements comprises a subset of channels in a convolution filter of the global machine learning model. (CALDAS [Pg. 5, section 4.1 Experimental Setup] teaches various convolutional neural networks (CNN) with 32 and 64 channels for the convolution layers. Furthermore, CALDAS [Pg. 4, section 3.2 Federated Dropout] varying the percentage of filters (i.e., convolution filter) in convolutional layers that are kept on each layer of the model (i.e., of the global machine learning model).)
(CALDAS [Pg. 2, Introduction] teaches that their approach allows each device to locally operate on a smaller sub-model with smaller weight matrices. Furthermore, CALDAS [Pg. 4, section 3.2 Federated Dropout] and [Pg. 2, Figure 1] teaches that these sub-models are reduced architectures constructed from a global model, and thus are subsets of the global model. Therefore, the subset of model elements can be understood as the sent compressed fully connected model (i.e., a subset of nodes in the global machine learning model) by the server to the client.)
Regarding Claim 8, the claim recites similar limitations as corresponding claim 1 and is rejected for similar reasons as claim 1 using similar teachings and rationale. Additionally, CALDAS teaches:
memory comprising computer-executable instructions; and one or more processors configured to execute the computer-executable instructions and cause the processing system to: (CALDAS [Pg. 2, section 2 Related Work] teaches: "Federated Learning (FL) is a technique that aims to learn a global model over data distributed across multiple edge devices (usually mobile phones) […].")
Regarding Claim 9, CALDAS in view of LOUIZOS teaches the elements of claim 8 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 2 and is rejected for similar reasons as claim 2 using similar teachings and rationale.
Regarding Claim 10, CALDAS in view of LOUIZOS teaches the elements of claim 9 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 3 and is rejected for similar reasons as claim 3 using similar teachings and rationale.
Regarding Claim 11, CALDAS in view of LOUIZOS teaches the elements of claim 9 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 4 and is rejected for similar reasons as claim 4 using similar teachings and rationale.
Regarding Claim 12, CALDAS in view of LOUIZOS teaches the elements of claim 8 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 5 and is rejected for similar reasons as claim 5 using similar teachings and rationale.
Regarding Claim 13, CALDAS in view of LOUIZOS teaches the elements of claim 8 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 6 and is rejected for similar reasons as claim 6 using similar teachings and rationale.
Regarding Claim 15, CALDAS teaches:
A method for performing federated learning of a machine learning model, comprising: (CALDAS [Pg. 2, Figure 1] teaches a Federated Learning strategy.)
for each respective client of a plurality of clients and for each training round in a plurality of training rounds: generating, by a server, a subset of model elements for the respective client […] for each model element of a set of model elements for a global machine learning model; (CALDAS [Pg. 2, Figure 1], [Pg. 4, Section 3.2 Federated Dropout], and [Pg. 5, section 4.1 Experimental Setup, Hyperparameters] teaches a server constructing (i.e., generating) a sub-model by compressing a global model and sending the compressed model structure with smaller weight matrices (i.e., a subset of model elements) to each selected client (i.e., for the respective client of a plurality of clients) for training for one epoch per round (i.e., for each training round in a plurality of training rounds).)
transmitting from the server to the respective client: the subset of model elements; (CALDAS [Pg. 2, Figure 1] teaches the server transmitting the compressed global model (i.e., the subset of model elements) to various clients.)
receiving at the server from each respective client of the plurality of clients a respective set of model updates; and updating, by the server, the global machine learning model based on the respective set of model updates from each respective client of the plurality of clients. (CALDAS [Pg. 2, Figure 1] teaches: "This compressed model is then sent to the client, who (3) decompresses and trains it using local data, and (4) compresses the final update. This update is sent back to the server (i.e., receiving at the server), where it is (5) decompressed and finally, (6) aggregated into the global model." CALDAS [Pg. 4, Section 3.2 Federated Dropout] and [Pg. 5, section 4.1 Experimental Setup, Hyperparameters] teaches selected client (i.e., for the respective client) training their received models for one epoch per round and sending their updates (i.e., respective set of model updates) back to the server for aggregation into the global model.)
a local machine learning model; (CALDAS [Pg. 2, section 1. Introduction] teaches: “Our approach enables each device to locally operate on a smaller sub-model (i.e., local machine learning model) (i.e. with smaller weight matrices) while still providing updates that can be applied to the larger global model on the server.”)
CALDAS is not relied upon for teaching:
generating […] a subset of model elements […] based on sampling a gate probability distribution […]
a set of gate probabilities based on the sampling, wherein each gate probability of the set of gate probabilities is associated with one model element of the subset of model elements and represents a likelihood that the associated model element will be included in a […] machine learning model;
However, LOUIZOS teaches: generating […] a subset of model elements […] based on sampling a gate probability distribution […]; (LOUIZOS [Pg. 1-2, Introduction] teaches the hard concrete distribution, obtained by stretching a binary concrete random variable and then passing its samples through a hard-sigmoid of
s
. Furthermore, LOUIZOS [Pg. 4, section 2.1 A General Recipe for Efficiently Minimizing
L
0
Norms] eq. (4) teaches sampling
s
~
q
(
s
|
ϕ
)
(i.e., sampling a gate probability distribution), and thus
s
has a distribution of
q
(
s
|
ϕ
)
, which is sampled in eq. (5) to generate the binary mask gate probability
z
(i.e., a subset of model elements).)
[…] a set of gate probabilities based on the sampling, (LOUIZOS [Pg. 4, section 2.1 A General Recipe for Efficiently Minimizing
L
0
Norms] teaches associating distribution over the gates (i.e., gate probabilities) with parameters of the model (i.e., subset of model elements) by
θ
j
=
θ
j
~
z
j
, where
z
j
correspond to binary gates and
θ
j
~
correspond to the parameters of the network.)
wherein each gate probability of the set of gate probabilities is associated with one model element of the subset of model elements (LOUIZOS [Abstract] and [Pg. 6, section 2.4 Group Sparsity Under an
L
0
Norm] teaches jointly optimizing parameters of the distribution over the gates (i.e., gate probabilities) and the original network parameters, where each input neuron has a gate for a fully connected network (i.e., associated with one model element of the subset of model elements).
CALDAS teaches sending a compressed model to a client’s device, and thus a person of ordinary skill in the art could apply LOUIZOS’ gate probabilities to the compressed model structure sent by the server.)
and represents a likelihood that the associated model element will be included in a […] machine learning model; (LOUIZOS [Pg. 3, section 2.1 A General Recipe For Efficiently Minimizing
L
0
Norms] teaches: “Consider the
L
0
norm under a simple re-parametrization of
θ
:
PNG
media_image1.png
69
709
media_image1.png
Greyscale
where the
z
j
correspond to binary “gates” that denote whether a parameter is present and the
L
0
norm corresponds to the amount of gates being “on”.” By letting
q
z
j
|
π
j
=
B
e
r
n
π
j
be a Bernoulli distribution over each gate
z
j
we can reformulate the minimization of Eq. 1 as penalizing the number of parameters being used, on average, as follows:
PNG
media_image2.png
128
785
media_image2.png
Greyscale
where
⨀
corresponds to the elementwise product.” Examiner’s note: Under BRI, the associated model element can be interpreted as the gated parameter
θ
j
and represents a likelihood can be interpreted as
π
j
, which controls the probability that a gate
z
j
is “on” (i.e., included in a […] machine learning model).)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of CALDAS and LOUIZOS before them, to include LOUIZOS' collection of gates to determine which weights of a network to set to zero in CALDAS' federated learning method. Both CALDAS and LOUIZOS teach fully connected models/layers, and LOUIZOS teaches introducing a gate per input neuron for fully connected layers. One would have been motivated to make such a combination of including a gate per input neuron in order to improve the accuracy of the dropout equivalent network for practical computation savings while training, a benefit not possible with the commonly used independent dropout mask per spatial location (LOUIZOS [Pg. 6, section 2.4 Group Sparsity under an
L
0
Norm] and [Pg. 9, section 4.2 CIFAR Classification]).
Regarding Claim 16, CALDAS in view of LOUIZOS teaches the elements of claim 15 as outlined above. Additionally, the claim recites similar limitations as corresponding claims 2 and 9 and is rejected for similar reasons as claims 2 and 9 using similar teachings and rationale.
Regarding Claim 17, CALDAS in view of LOUIZOS teaches the elements of claim 16 as outlined above. Further, CALDAS teaches:
The method of Claim 16, wherein the respective set of model updates comprises: a set of weight gradients associated with the local machine learning model trained by the respective client; (CALDAS [Pg. 4, section 3.2 Federated Dropout] teaches the client training the received sub-model with smaller weight matrices and sending its updates (i.e., the respective set of model updates/a set of model updates […] associated with the local machine learning model] to the server for mapping back to the global model. Furthermore, CALDAS [Pg. 5, section 4.2 Lossy compression] teaches client-to-server gradient updates (i.e., a set of weight gradients). CALDAS [Pg. 4, Section 3.2 Federated Dropout] and [Pg. 5, section 4.1 Experimental Setup, Hyperparameters] teaches selected client training their received models (i.e., trained by the respective client) for one epoch per round and sending their updates back to the server for aggregation into the global model.)
Further, LOUIZOS teaches: and a set of gate probability gradients associated with the local machine learning model trained […]. (LOUIZOS [Abstract] and [Pg. 6, section 2.4 Group Sparsity Under an
L
0
Norm] teaches jointly optimizing parameters of the distribution over the gates (i.e., a set of gate probabilities) and the original network parameters, where each input neuron has a gate for a fully connected network. Furthermore, LOUIZOS [Pg. 4, section 2.1 A General Recipe for Efficiently Minimizing
L
0
Norms] teaches the re-parametrization of the network parameters
θ
as
θ
j
=
θ
j
~
z
j
, where
z
j
is the binary gate that denote whether a parameter (i.e., associated with the local machine learning model) is present and the
L
0
norm corresponds to the number of gates being “on”. LOUIZOS’ Eq. (4) and Eq. (5) teaches the giving the binary gates a hard-sigmoid rectification of
s
2
, and thus allowing for the gates to be a continuous random variable. Furthermore, LOUIZOS’ Eq. 9 teaches the total cost is differentiable w.r.t.
ϕ
, thus enabling for efficient stochastic gradient based optimization, and that due to rectifications, the gradient of log-likelihood w.r.t. the parameters
ϕ
of
q
s
is sparse. LOUIZOS’ gates denoting whether a parameter (i.e., weight) is present can be applied to the model architecture as disclosed in CALDAS [Pg. 4, section 3.2 Federated Dropout].
Regarding Claim 18, CALDAS in view of LOUIZOS teaches the elements of claim 16 as outlined above. Additionally, the claim recites similar limitations as corresponding claims 4 and 11 and is rejected for similar reasons as claims 4 and 11 using similar teachings and rationale.
Regarding Claim 19, CALDAS in view of LOUIZOS teaches the elements of claim 15 as outlined above. Additionally, the claim recites similar limitations as corresponding claims 5 and 12 and is rejected for similar reasons as claims 5 and 12 using similar teachings and rationale.
Regarding Claim 20, CALDAS in view of LOUIZOS teaches the elements of claim 15 as outlined above. Additionally, the claim recites similar limitations as corresponding claims 6 and 13 and is rejected for similar reasons as claims 6 and 13 using similar teachings and rationale.
Regarding Claim 22, the claim recites similar limitations as corresponding claim 15 and is rejected for similar reasons as claim 15 using similar teachings and rationale. Additionally, CALDAS teaches:
memory comprising computer-executable instructions; and one or more processors configured to execute the computer-executable instructions and cause the processing system to: (CALDAS [Pg. 2, section 2 Related Work] teaches: "Federated Learning (FL) is a technique that aims to learn a global model over data distributed across multiple edge devices (usually mobile phones) […].")
Regarding Claim 23, CALDAS in view of LOUIZOS teaches the elements of claim 22 as outlined above. Additionally, the claim recites similar limitations as corresponding claims 2, 9, and 16 and is rejected for similar reasons as claims 2, 9, and 16 using similar teachings and rationale.
Regarding Claim 24, CALDAS in view of LOUIZOS teaches the elements of claim 23 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 17 and is rejected for similar reasons as claim 17 using similar teachings and rationale.
Regarding Claim 25, CALDAS in view of LOUIZOS teaches the elements of claim 23 as outlined above. Additionally, the claim recites similar limitations as corresponding claims 4, 11, and 18 and is rejected for similar reasons as claims 4, 11, and 18 using similar teachings and rationale.
Regarding Claim 26, CALDAS in view of LOUIZOS teaches the elements of claim 22 as outlined above. Additionally, the claim recites similar limitations as corresponding claims 5, 12, and 19 and is rejected for similar reasons as claims 5, 12, and 19 using similar teachings and rationale.
Regarding Claim 27, CALDAS in view of LOUIZOS teaches the elements of claim 22 as outlined above. Additionally, the claim recites similar limitations as corresponding claims 6, 13, and 20 and is rejected for similar reasons as claims 6, 13, and 20 using similar teachings and rationale.
Claims 7, 14, 21, and 28 are rejected under 35 U.S.C. 103 as being unpatentable over CALDAS in view of LOUIZOS as applied respectively above to claims 1, 8, 15, and 22, and further in view of JIANG ("Model Pruning Enables Efficient Federated Learning on Edge Devices"), hereafter JIANG.
Regarding Claim 7, CALDAS in view of LOUIZOS teaches the elements of claim 1 as outlined above. Further, CALDAS teaches:
receiving at the device a […] set of model elements from the server, (CALDAS [Pg. 2, Figure 1] teaches the server sending a compressed model (i.e., set of model elements) to the client’s device. CALDAS [Pg. 6, Figure 3] teaches number of rounds for the local training.)
However, CALDAS in view of LOUIZOS is not relied for teaching, but JIANG teaches: receiving at the device a final set of model elements from the server, wherein the final set of model elements corresponds to a pruned global machine learning model. (JIANG [Abstract] teaches integrating model pruning with federated learning, which includes initial model pruning at the server. JIANG [Pg. 3, section 3.2.3 Overall Procedure] teaches the server distributing the initially pruned model to the clients.
CALDAS’ compression of the model can be applied to JIANG’s pruned model for client distribution.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of LOUIZOS, CALDAS, and JIANG before them, to include JIANG's pruning in LOUIZOS and CALDAS' federated learning method. One would have been motivated to make such a combination in order to effectively reduce the size of deep neural network models so that resource-limited clients can train them with their own data and contribute in the federated learning processes (JIANG [Pg. 10, section 7 Conclusion]).
Regarding Claim 14, CALDAS in view of LOUIZOS teaches the elements of claim 8 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 7 and is rejected for similar reasons as claim 7 using similar teachings and rationale. Additionally, CALDAS teaches:
wherein the one or more processors are further configured to […] (CALDAS [Pg. 2, section 2 Related Work] teaches: "Federated Learning (FL) is a technique that aims to learn a global model over data distributed across multiple edge devices (usually mobile phones) […].")
Regarding Claim 21, CALDAS in view of LOUIZOS teaches the elements of claim 15 as outlined above. Further, LOUIZOS teaches
[…] pruning the updated […] machine learning model based on updated gate probabilities for the […] machine learning model and a threshold gate probability value. (LOUIZOS [Pg. 1, Abstract] teaches pruning a neural network (i.e., machine learning model) during training by encouraging weights to become exactly zero. LOUIZOS [Pg. 4-5, section 2.2 The Hard Concrete Distribution] teaches using Eq. (13) to determine the final parameters under a hard concrete gate using the updated parameter
log
α
(i.e., based on updated gate probabilities) as:
z
^
=
m
i
n
(
1
,
max
0
,
S
i
g
m
o
i
d
log
α
ζ
-
γ
+
γ
)
Furthermore, LOUIZOS [Pg. 5, section 2.2 The Hard Concrete Distribution] teaches that this estimator produces a rounded version of the original binary concrete gate, where values smaller than
-
γ
ζ
-
γ
(i.e., a threshold gate probability value) are rounded to zero, and thus become pruned.)
However, LOUIZOS is not relied upon for teaching, but JIANG teaches: pruning the updated global machine learning model based on […] a threshold […] value. (JIANG [Abstract] teaches integrating model pruning with federated learning (i.e., pruning the updated global machine learning model), which includes initial model pruning at the server. Furthermore, JIANG [Pg. 3, section 3.2.3 Overall Procedure] teaches the server distributing the initially pruned model to the clients. Furthermore, JIANG [Pg. 5, section 5 Experimentation] teaches federated pruning by (i.e., based on) removing weights with small magnitudes, and thus the small magnitude acts as a threshold (i.e., a threshold […] value).)
LOUIZOS’ pruning can be applied to CALDAS’ federated learning method, and JIANG’s pruned model for client distribution can be applied to CALDA’s federated learning.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of LOUIZOS, CALDAS, and JIANG before them, to include JIANG's pruning in LOUIZOS and CALDAS' federated learning method. One would have been motivated to make such a combination in order to effectively reduce the size of deep neural network models so that resource-limited clients can train them with their own data and contribute in the federated learning processes (JIANG [Pg. 10, section 7 Conclusion]).
Regarding Claim 28, CALDAS in view of LOUIZOS teaches the elements of claim 22 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 21 and is rejected for similar reasons as claim 21 using similar teachings and rationale. Additionally, CALDAS teaches:
the one or more processors are further configured to […] (CALDAS [Pg. 2, section 2 Related Work] teaches: "Federated Learning (FL) is a technique that aims to learn a global model over data distributed across multiple edge devices (usually mobile phones) […].")
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Alvaro S Laham Bauzo whose telephone number is (571)272-5650. The examiner can normally be reached Mon-Fri 7:30 AM - 11:00 AM | 1:00 PM - 5:30 PM ET.
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 on (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.
/A.S.L./Examiner, Art Unit 2146
/USMAAN SAEED/Supervisory Patent Examiner, Art Unit 2146
Prosecution Insights