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 .
Status of Claims
Claims 3, 8, 9, 12, 18 are amended. Claims 1 – 18 are pending and examined herein.
Claims 9, 18 are rejected under 35 U.S.C. 112(b).
Claims 1 – 18 are rejected under 35 U.S.C. 101.
Claims 1 – 18 are rejected under 35 U.S.C. 103.
Response to Amendment
The amendment filed February 12th, 2026 has been entered. Claims 3, 8, 9, 12, 18 are amended. Claims 1 – 18 are pending and examined herein. Applicant’s amendments to the claims have overcome most of the objection and 112(b) rejection previously set forth in the Non-Final Rejection Office Action mailed October 01th, 2025. However, the 112(b) rejection regarding claims 9 and 18 remains.
Response to Arguments
Applicant's arguments filed February 12th, 2026 regarding the 35 U.S.C. 101 rejection of claims 1-15 have been fully considered but they are not persuasive.
Applicant argues, on pages 10 – 11, that the claims are not directed to mathematical calculations and data transmission. Applicant argues that the claims address a concrete technical problem, like insufficient bandwidth for devices receiving large global model updates, and that the claimed solution of quantizing the global model and performing quantization aware training on reconstructed feature maps is a technical improvement rather than an abstract idea. This argument is not persuasive because the claims do not recite a specific improvement to how the server, client devices, transceiver, processor, or network itself operates. Instead, the claims recite mathematical processing of machine learning model data in a federated learning environment. For example, the claims recite updating a global model based on local updates, quantizing the updated global model, reconstructing feature maps based on the received local updates, and refining the quantized, updated global model based on the reconstructed feature maps. These steps involve mathematical calculations and transformations of model parameters, local updates, gradients, and feature map data.
Applicant’s reliance on reduced downlink communication cost and compensation for compression loss is not persuasive because those benefits result from applying the recited mathematical model processing steps to the model data. The claims do not recite a particular transmission protocol, packet structure, channel allocation technique, hardware configuration, or other specific network improvement. The server receives local updates, performs the claimed model update, quantization, reconstruction, refinement steps, and transmits the resulting refined, quantized model.
The additional elements, including the server, federated network, client devices, transceiver, and processor, merely provide the computer and network environment in which the mathematical processing is performed. These elements do not integrate the mathematical concepts into a practical application because they do not impose any meaningful limitation beyond receiving data, performing the recited calculations, and transmitting the result.
Accordingly, the claims recite mathematical concepts, including mathematical relationships, calculations, and model optimization steps, applied in a federated learning environment. The additional elements, considered individually and as an ordered combination, do not integrate the abstract idea into a practical application and do not amount to significantly more than the abstract idea itself. The rejection under 35 U.S.C. 101 is maintained.
Applicant’s arguments regarding the rejection of claims 9 and 18 under 35 U.S.C. 112(b) have been fully considered but are not persuasive.
Claims 9 and 18 recite an equation for reconstructing the feature maps. The claims define fl as an input feature map of layer l. However, the equation also separately recites fl*. Applicant states, on page 10, that fl* refers to a feature map, but the claim language does not define what type of feature map fl* represents. Because the equation uses both and fl, fl* it is unclear whether fl* refers to a reconstructed feature maps, target feature maps, original feature map, optimized feature map, or some other feature map. Applicant’s remarks in the argument does not make the claim language clear because the metes and bounds of the claimed equation still depend on an undefined variable.
Additionally, the claims recite TV(fl) in the equation, while the claim language states that TV(x) is a total variation of x. While it appears as applicant intends TV to be a total variation operator applied to fl, the claim should clearly state the relationship with respect to the claimed equation.
Accordingly, the rejection of claims 9 and 18 under 35 U.S.C. 112(b) is maintained.
Applicant’s arguments regarding the rejection of claims 1, 8, 10, and 17 under 35 U.S.C. 103 have been fully considered but do not overcome the rejection.
Applicant argues, on pages 11 – 15, that Boroujeni does not teach a server quantizing the updated global model because Boroujeni’s cited quantization examples describe client side quantization of local updates. This argument has been considered. However, Boroujeni is relied upon for a federated learning system in which the base station receives local updates from UEs, updates the global machine learning component, and retransmits the updated machine learning component to the UEs for further federated learning. Boroujeni also teaches a compressor/expander quantization technique in the same federated learning system for reducing communication overhead. As discussed during the interview and written on the examiner interview summary record, Boroujeni’s paragraphs [0095]-[0096] further support this understanding. Paragraph [0095] describes the UE transmitting a quantized value, the base station receiving the quantized value, and the base station applying an expander to decode the feedback from the quantized value. Paragraph [0096] further describes the base station updating the machine learning component based on the update, determining a global update for the machine learning component, and recursively retransmitting an updated machine learning component for additional training by the UEs.
Therefore, even assuming Boroujeni’s express quantization example describes client side quantization of feedback, Boroujeni as a whole teaches both the known quantization mechanism and the repeated base station transmission of updated machine learning information in the same federated learning system. It would have been obvious to one of ordinary skill in the art to apply Boroujeni’s transmission of the updated global model because that repeated transmission presents the same communication overhead concern addressed by Boroujeni’s quantization technique. The modification would have been the predictable use of a known technique for its known purpose of reducing communication overhead in the same federated learning system.
The claims do not recite a particular downlink specific quantization algorithm, compressor structure, or other technical feature that would distinguish the claimed server side quantization from the predictable application of Boroujeni’s known quantization technique to the base station’s repeated transmission of the updated global model. Accordingly, Applicant’s arguments do not overcome the rejection and the rejection of claims 1 – 18 under 35 U.S.C. 103 is maintained.
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.
Claims 9, 18 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claims 9 and 18 recite an equation for reconstructing feature maps. The claims define fl as an input feature map of layer l. However, the equation also separately recites fl*, and the claims do not define what fl* represents. Because the equation uses both fl and fl*, it is unclear whether fl* refers to a reconstructed feature map, target feature map, original feature map, optimized feature map, or some other feature map. The scope of the claimed equation is unclear. No clear interpretation can be given to fl* based on the present claim language. Claims 9 and 18 further recites TV(fl), while the claim language states that TV(x) is a total variation of x. Because the claim defines fl as an input feature map of layer l, and only states that f1 = x for the first layer, it is unclear whether TV(fl) is intended to mean a total variation of the feature map fl, a total variation of the input image x, or something else.
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 - 18 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
MPEP § 2109(III) sets out steps for evaluating whether a claim is drawn to patent-eligible subject matter. The analysis of claims 1 – 18, in accordance with these steps, follows.
Step 1 Analysis:
Step 1 is to determine whether the claim is directed to a statutory category (process, machine, manufacture, or composition of matter.
Claims 1 – 9 are directed to a method, meaning that it is directed to the statutory category of process. Claims 10 - 18 are directed to a server with transceiver and processor for a federated network, which is directed to the statutory category of apparatus.
Step 2A Prong One, Step 2A Prong Two, and Step 2B Analysis:
Step 2A Prong One asks if the claim recites a judicial exception (abstract idea, law of nature, or natural phenomenon). If the claim recites a judicial exception, analysis proceeds to Step 2A Prong Two, which asks if the claim recites additional elements that integrate the abstract idea into a practical application. If the claim does not integrate the judicial exception, analysis proceeds to Step 2B, which asks if the claim amounts to significantly more than the judicial exception. If the claim does not amount to significantly more than the judicial exception, the claim is not eligible subject matter under 35 U.S.C. 101.
Regarding claim 1, the following claim elements are abstract ideas:
updating a global model based on the received local updates; (Updating global model based on the local updates is merely applying algebraic computation, which is mathematical concept.)
quantizing the updated global model; (Quantizing the global model recites the mathematical calculation, which is mathematical concept.)
reconstructing feature maps based on the received local updates; (Feature maps are matrix of numbers and reconstructing them with updates is merely mathematical calculations, which is mathematical concept.)
refining the quantized, updated global model based on the reconstructed feature maps; (Refining model based on reconstructed feature map recites the mathematical calculation, which is mathematical concept.)
The following claim elements are additional elements which, taken alone or in combination with the other additional elements, do not integrate the judicial exception into a practical application nor amount to significantly more than the judicial exception:
performed by a server in a federated network (This is mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
receiving local updates from client devices; (This is mere data gathering, an insignificant extra solution activity, which does not integrate the judicial exception into a practical application. The broadest reasonable interpretation of this claim is storing information in memory, which is a well-understood, routine conventional activity. See MPEP § 2106.05(d)(II)(iv). Therefore, this does not amount to significantly more than the judicial exception.)
and transmitting the refined, quantized, updated global model to the client devices. (This is mere transmitting data, which is a well-understood, routine conventional activity. It does not integrate the judicial exception into a practical application. See MPEP § 2106.05(d). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 2, the rejection of claim 1 is incorporated herein. Further, claim 2 recites the following abstract ideas:
mixing-up the reconstructed feature maps; (Mixing up the reconstructed feature map is merely mathematical calculations, which is mathematical concept.)
performing quantization-aware training, (Performing quantization aware training (QAT) is applying mathematical calculations, which is mathematical concept. )
performing quantization bias compensation, (Performing quantization bias compensation is applying mathematical calculations, which is mathematical concept.)
Claim 2 further recites the following additional element
based on the mixed-up reconstructed feature maps. (This falls under mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 3, the rejection of claim 2 is incorporated herein. Further, claim 3 recites the following abstract idea:
mixing-up the reconstructed feature maps is performed using:
PNG
media_image1.png
18
142
media_image1.png
Greyscale
wherein f1 and f2 represent a pair of reconstructed feature maps, and λ represents a mixing weight. (The limitation merely recites mathematical equation, which is mathematical concept.)
Claim 3 does not recite additional elements.
Claims 4 – 7 recite substantially similar subject matter to claim 2 respectively and are rejected with the same rationale, mutatis mutandis.
Regarding claim 8, the rejection of claim 1 is incorporated herein. Further, claim 8 recites the following abstract idea:
, and wherein reconstructing the feature maps based on the received local updates comprises inverting the received local gradients. (Inverting the local gradient is merely mathematical calculation, which is mathematical concept.)
Claim 8 further recites the following additional element:
wherein the received local updates include received local gradients (This falls under mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 9, the rejection of claim 1 is incorporated herein. Further, claim 9 recites the following abstract idea:
wherein reconstructing the feature maps based on the received local updates is performed using:
PNG
media_image2.png
35
289
media_image2.png
Greyscale
wherein θ represents parameters of the federated network, L represents a loss function, fl is an input feature map of layer 1, where for a first layer (1= 1), f1= x is an input image and y is a ground-truth label for input x, α is a hyper parameter that is greater than zero, and TV(x) is a total variation of x. (The limitation merely recites mathematical equation, which is mathematical concept.)
Regarding claim 10, the following claim elements are additional elements:
via the transceiver, (This falls under mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
The rest of claim 10 and claims 11 – 18 recite substantially similar subject matter to claims 1 – 9 respectively and are rejected with the same rationale, mutatis mutandis.
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.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1, 8, 10, 17 are rejected under 35 U.S.C. 103 as being unpatentable over Taherzadeh Boroujeni et al. (U.S. Pub. 2022/0103221 A1) in view of Geiping et al. (NPL: ” Inverting Gradients - How easy is it to break privacy in federated learning?”), further in view of Oki et al. (NPL: “Mixup of Feature Maps in a Hidden Layer for Training of Convolutional Neural Network”).
Regarding Claim 1, Taherzadeh Boroujeni teaches
receiving local updates from client devices; ([0051] of Taherzadeh Boroujeni states “server device … means for receiving a quantized value based at least in part on feedback from the client device.” And [0065] of Taherzadeh Boroujeni states “the UEs 120 may each transmit their respective local updates … a local update may include a compressed version of a local update. For example, in some aspects, a UE 120 may transmit a compressed set of gradients”)
updating a global model based on the received local updates; ([0066] of Taherzadeh Boroujeni states “the base station (e.g., using the second communication manager 330) may aggregate the updates received from the UEs…. As shown by reference number 380, the second communication manager 330 may update the global machine learning component based on the aggregated updates”)
quantizing the updated global model; ([0050] of Taherzadeh Boroujeni states “means for transmitting a quantized value... wherein the quantized value is determined based at least in part on distances between the feedback and a non-uniform set of quantized digits.” [0095] of Boroujeni states ”As shown in connection with reference number 515, the UE 120 may transmit, and the base station 110 may receive, the quantized value. The base station 110 may determine an update based at least in part on the quantized value. In some aspects, the base station 110 may determine the update based at least in part on aggregating quantized values from a plurality of UEs (e.g., as described in connection with FIG. 3). When the UE 120 uses nonlinear companding, the base station 110 may apply an expander (e.g., an inverse of the compressor) to decode the feedback from the quantized value.” [0096] of Boroujeni states “Accordingly, the base station 110 may update the machine learning component based at least in part on the update. In some aspects, the base station 110 may determine a plurality of updates (e.g., based at least in part on aggregating quantized values from multiple sets of UEs, where each set includes one or more UEs) and aggregate the plurality of updates to determine a global update for the machine learning component. In some aspects, example 500 may be recursive, where the base station 110 re-transmits an updated machine learning component for additional training by the UE 120 and/or other UEs in the federated learning.”)
… the quantized, updated global model … ([0050] of Taherzadeh Boroujeni states “means for transmitting a quantized value... wherein the quantized value is determined based at least in part on distances between the feedback and a non-uniform set of quantized digits.”)
and transmitting the refined, quantized, updated global model to the client devices. ([0066] of Taherzadeh Boroujeni states “the base station… may transmit an update associated with the updated global machine learning component to the UEs 120.”)
However, Taherzadeh Boroujeni doesn’t explicitly teach
reconstructing feature maps based on the received local updates;
refining … based on the reconstructed feature maps;
Geiping and Oki teaches that
reconstructing feature maps based on the received local updates; (Pg. 3 I. Introduction section of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images.”)
refining … based on the reconstructed feature maps; (Pg. 1 Abstract of Oki states “In this paper, we propose to apply mix-up to the feature maps in a hidden layer. From the experimental comparison, it is observed that the mix-up of the feature maps obtained from the first convolution layer is more effective than the original image mix-up.” And Pg. 3 I. Introduction section of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images.”)
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings from Taherzadeh Boroujeni with Geiping and Oki. Taherzadeh Boroujeni teaches the standard federated learning method that received client updates (including compressed local gradients as local updates) and aggregate those updates to update a global model. Geiping teaches that gradients can be inverted to reconstruct feature maps. Oki teaches that mixing feature maps is an effective technique to improve robustness of the model. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Geiping and Oki into Taherzadeh Boroujeni to apply mixup technique to reconstructed feature maps available on the server to improve the global model before transmitting the updated models to clients. It is predictable combination to apply a known data augmentation and refine technique in a federated network environment.
Regarding claim 8, the rejection of claim 1 is incorporated herein. The combination of Taherzadeh Boroujeni, Geiping, and Oki teach
wherein the received local updates include received local gradients ([0065] of Taherzadeh Boroujeni states “As shown by reference number 360, the UEs 120 may each transmit their respective local updates (shown as “local update 1, . . . , local update k, . . . , local update K”). In some aspects, a local update may include a compressed version of a local update. For example, in some aspects, a UE 120 may transmit a compressed set of gradients (e.g., represented by {tilde over (g)}.sub.k.sup.(n) =q(g.sub.k.sup.(n)), where q represents a compression scheme applied to the set of gradients g.sub.k.sup.(n)).”)
, and wherein reconstructing the feature maps based on the received local updates comprises inverting the received local gradients. (Pg.2 I. Introduction of Geiping states “In this paper we discuss privacy limitations of federated learning first in an academic setting, honing in on the case of gradient inversion from one image and showing that • Reconstruction of input data from gradient information is possible for realistic deep and non-smooth architectures with both, trained and untrained parameters.”)
Claims 10, 17 recite substantially similar subject matter as claims 1, 8 respectively, and are rejected with the same rationale, mutatis mutandis.
Claims 5, 14 are rejected under 35 U.S.C. 103 as being unpatentable over Taherzadeh Boroujeni et al. (U.S. Pub. 2022/0103221 A1) in view of Geiping et al. (NPL: ” Inverting Gradients - How easy is it to break privacy in federated learning?”), Oki et al. (NPL: “Mixup of Feature Maps in a Hidden Layer for Training of Convolutional Neural Network”), further in view of Sriram et al. (U.S. Pub. 2022/0044114 A1).
Regarding claim 5, the rejection of claim 1 is incorporated herein. The combination of Taherzadeh Boroujeni, Geiping, and Oki teach
Wherein refining the quantized, updated global model based on the reconstructed feature maps comprises: mixing-up the reconstructed feature maps; (Pg. 1 Abstract of Oki states “In this paper, we propose to apply mix-up to the feature maps in a hidden layer” and Pg. 3 I. Introduction of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images” Pg. 1 Abstract of Geiping also states “we show that it is actually possible to faithfully reconstruct images at high resolution from the knowledge of their parameter gradients, and demonstrate that such a break of privacy is possible even for trained deep networks.”)
, based on the mixed-up reconstructed feature maps. (Pg. 1 Abstract of Oki states “In this paper, we propose to apply mix-up to the feature maps in a hidden layer” and Pg. 3 I. Introduction of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images” Pg. 1 Abstract of Geiping also states “we show that it is actually possible to faithfully reconstruct images at high resolution from the knowledge of their parameter gradients, and demonstrate that such a break of privacy is possible even for trained deep networks.”)
However, the combination of Taherzadeh Boroujeni, Geiping, and Oki doesn’t explicitly teach that and performing quantization-aware training
Sriram teaches that
and performing quantization-aware training ([0057] of Sriram states “In some instances, quantizing models such as rounding the weights and activations to lower-bit integers (e.g., 8-bit integers) is performed during training (e.g., quantization-aware training (QAT))” [0112] of Sriram states “In at least one embodiment, a model is a neural network model that is part of one or more federated learning processes. In at least one embodiment, QAT is applied, during training of a neural network”)
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings from the combination of Taherzadeh Boroujeni, Geiping, Oki with Sriram to further stabilize the step of refining the quantized, updated global model in federated network. Taherzadeh Boroujeni teaches receiving and aggregating local updates and transmitting quantized global models. Geiping teaches reconstructing feature maps through gradient inversion and Oki teaches mixing feature maps. Sriram teaches that quantization-aware training could be applied when quantizing the global model in federated network to further improve accuracy. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Sriram into the combination of Taherzadeh Boroujeni, Geiping, Oki so that the refined, quantized, updated model transmitted by the server preserve the accuracy on client devices.
Regarding claim 14, the rejection of claim 10 is incorporated herein. Claim 14 recites substantially similar subject matter as claim 5 respectively, and is rejected with the same rationale, mutatis mutandis.
Claims 6, 15 are rejected under 35 U.S.C. 103 as being unpatentable over Taherzadeh Boroujeni et al. (U.S. Pub. 2022/0103221 A1) in view of Geiping et al. (NPL: ” Inverting Gradients - How easy is it to break privacy in federated learning?”), Oki et al. (NPL: “Mixup of Feature Maps in a Hidden Layer for Training of Convolutional Neural Network”), further in view of Tang et al. (NPL: “DOUBLESQUEEZE: Parallel Stochastic Gradient Descent with Double-pass Error-Compensated Compression”).
Regarding claim 6, the rejection of claim 1 is incorporated herein. The combination of Taherzadeh Boroujeni, Geiping, and Oki teach
Wherein refining the quantized, updated global model based on the reconstructed feature maps comprises: mixing-up the reconstructed feature maps; (Pg. 1 Abstract of Oki states “In this paper, we propose to apply mix-up to the feature maps in a hidden layer” and Pg. 3 I. Introduction of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images” Pg. 1 Abstract of Geiping also states “we show that it is actually possible to faithfully reconstruct images at high resolution from the knowledge of their parameter gradients, and demonstrate that such a break of privacy is possible even for trained deep networks.”)
, based on the mixed-up reconstructed feature maps. (Pg. 1 Abstract of Oki states “In this paper, we propose to apply mix-up to the feature maps in a hidden layer” and Pg. 3 I. Introduction of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images” Pg. 1 Abstract of Geiping also states “we show that it is actually possible to faithfully reconstruct images at high resolution from the knowledge of their parameter gradients, and demonstrate that such a break of privacy is possible even for trained deep networks.”)
However, the combination of Taherzadeh Boroujeni, Geiping, and Oki doesn’t explicitly teach that and performing quantization bias compensation
Tang teaches that and performing quantization bias compensation (Pg.1 Abstract of Tang states “In this work, we
provide a detailed analysis on this two-pass communication model, with error-compensated compression both on the worker nodes and on the parameter server. We show that the error-compensated stochastic gradient algorithm admits three very nice properties: 1) it is compatible with an arbitrary compression technique; 2) it admits an improved convergence rate than the non error-compensated stochastic gradient methods such as QSGD and sparse SGD; 3) it admits linear speedup with respect to the number of workers. An empirical study is also conducted to validate our theoretical results.”)
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings from the combination of Taherzadeh Boroujeni, Geiping, Oki with Tang to compensate the bias/error that occurs during quantization for the quantized, updated global model in federated network. Taherzadeh Boroujeni teaches receiving and aggregating local updates and transmitting quantized global models. Geiping teaches reconstructing feature maps through gradient inversion and Oki teaches mixing feature maps. Tang teaches that error compensated compression could eliminate the bias introduced by compression and reduce the noise associated with quantized updates. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Tang into the combination of Taherzadeh Boroujeni, Geiping, Oki so that the refined, quantized, updated model transmitted by the server preserve the accuracy on client devices.
Regarding claim 15, the rejection of claim 10 is incorporated herein. Claim 15 recites substantially similar subject matter as claim 6 respectively, and is rejected with the same rationale, mutatis mutandis.
Claims 9, 18 are rejected under 35 U.S.C. 103 as being unpatentable over Taherzadeh Boroujeni et al. (U.S. Pub. 2022/0103221 A1) in view of Geiping et al. (NPL: ” Inverting Gradients - How easy is it to break privacy in federated learning?”), Oki et al. (NPL: “Mixup of Feature Maps in a Hidden Layer for Training of Convolutional Neural Network”), further in view of Jeon et al. (NPL: “Gradient Inversion with Generative Image Prior”).
Regarding claim 9, the rejection of claim 1 is incorporated herein. The combination of Taherzadeh Boroujeni, Geiping, and Oki teach
wherein reconstructing the feature maps based on the received local updates is performed using:
PNG
media_image2.png
35
289
media_image2.png
Greyscale
wherein θ represents parameters of the federated network, L represents a loss function, fl is an input feature map of layer 1, where for a first layer (1= 1), f1= x is an input image and y is a ground-truth label for input x, α is a hyper parameter that is greater than zero, and TV(x) is a total variation of x. (Pg. 4 4 A numerical Reconstruction Method section of Geiping shows Equation 4. We further constrain our search space to images within [0; 1] and add only total variation [30] as a simple image prior to the overall problem,
PNG
media_image3.png
67
471
media_image3.png
Greyscale
)
However, the combination doesn’t explicitly teach that
[Apply for every layer]Jeon teaches
[Apply for every layer] (Pg. 4 4. Methods section of Jeon states “When inverting images, Geiping et al. [9] propose to add the total variation regularization RTV(x) to the cost function in (3) since neighboring pixels of natural images are likely to have similar values. To be specific, Yin et al. [29] propose to employ the regularizer RBN(x1; :::; xB; ) which quantifies the discrepancy between the BN statistics of estimated xj ’s and those of true xj’s on each layer of the learning model.)
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings from the combination of Taherzadeh Boroujeni, Geiping, Oki with Jeon to further apply the reconstructing feature maps on each layers of feature maps. Taherzadeh Boroujeni teaches receiving and aggregating local updates and transmitting quantized global models. Geiping teaches reconstructing feature maps through gradient inversion and Oki teaches mixing feature maps. Jeon teaches that gradient inversion can be improved by applying regularizers across layers to reconstruct more accurate feature maps. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Jeon into the combination of Taherzadeh Boroujeni, Geiping, Oki so that the reconstructed feature maps used in refining the quantized, updated model would be more accuracy and robust across different layers.
Regarding claim 18, the rejection of claim 10 is incorporated herein. Claim 18 recites substantially similar subject matter as claim 6 respectively, and is rejected with the same rationale, mutatis mutandis.
Claims 2 – 4, 7, 11 – 13, 16 are rejected under 35 U.S.C. 103 as being unpatentable over Taherzadeh Boroujeni et al. (U.S. Pub. 2022/0103221 A1) in view of Geiping et al. (NPL: ” Inverting Gradients - How easy is it to break privacy in federated learning?”), Oki et al. (NPL: “Mixup of Feature Maps in a Hidden Layer for Training of Convolutional Neural Network”), Sriram et al. (U.S. Pub. 2022/0044114 A1), further in view of Tang et al. (NPL: “DOUBLESQUEEZE: Parallel Stochastic Gradient Descent with Double-pass Error-Compensated Compression”).
Regarding claim 2, the rejection of claim 1 is incorporated herein. The combination of Taherzadeh Boroujeni, Geiping, and Oki teach
Wherein refining the quantized, updated global model based on the reconstructed feature maps comprises: mixing-up the reconstructed feature maps; (Pg. 1 Abstract of Oki states “In this paper, we propose to apply mix-up to the feature maps in a hidden layer” and Pg. 3 I. Introduction of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images” Pg. 1 Abstract of Geiping also states “we show that it is actually possible to faithfully reconstruct images at high resolution from the knowledge of their parameter gradients, and demonstrate that such a break of privacy is possible even for trained deep networks.”)
, based on the mixed-up reconstructed feature maps; (Pg. 3 I. Introduction section of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images”.)
However, the combination of Taherzadeh Boroujeni, Geiping, and Oki doesn’t teach
performing quantization-aware training
performing quantization bias compensation
Sriram and Tang teach
performing quantization-aware training ([0057] of Sriram states “In some instances, quantizing models such as rounding the weights and activations to lower-bit integers (e.g., 8-bit integers) is performed during training (e.g., quantization-aware training (QAT))” [0112] of Sriram states “In at least one embodiment, a model is a neural network model that is part of one or more federated learning processes. In at least one embodiment, QAT is applied, during training of a neural network”)
performing quantization bias compensation (Pg.1 Abstract of Tang states “In this work, we
provide a detailed analysis on this two-pass communication model, with error-compensated compression both on the worker nodes and on the parameter server. We show that the error-compensated stochastic gradient algorithm admits three very nice properties: 1) it is compatible with an arbitrary compression technique; 2) it admits an improved convergence rate than the non error-compensated stochastic gradient methods such as QSGD and sparse SGD; 3) it admits linear speedup with respect to the number of workers. An empirical study is also conducted to validate our theoretical results.”)
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings from the combination of Taherzadeh Boroujeni, Geiping, Oki with Sriram and Tang to further enhance the accuracy and stability of the quantized, updated global model in federated network. Taherzadeh Boroujeni teaches receiving and aggregating local updates and transmitting quantized global models. Geiping teaches reconstructing feature maps through gradient inversion and Oki teaches mixing feature maps. Sriram teaches that quantization-aware training could be applied when quantizing the global model in federated network to further improve accuracy. Tang teaches that error compensated compression could eliminate the bias introduced by compression and reduce the noise associated with quantized updates. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Sriram and Tang into the combination of Taherzadeh Boroujeni, Geiping, Oki so that the refined, quantized, updated model transmitted by the server maintains quality while minimizing error/bias caused by quantization.
Regarding claim 3, the rejection of claim 2 is incorporated herein. The combination of Taherzadeh Boroujeni, Geiping, Oki, Sriram, and Tang teach
mixing-up the reconstructed feature maps is performed using:
PNG
media_image1.png
18
142
media_image1.png
Greyscale
wherein f1 and f2 represent a pair of reconstructed feature maps, and λ represents a mixing weight. (Pg. 3 2.2 Mixup section of Oki states “Let x1 and x2 be the images randomly extracted from the training samples and t1 and t2 are the corresponding teacher vectors. Then the new image x and the corresponding teacher vector t is generated by Equation 2
PNG
media_image4.png
16
118
media_image4.png
Greyscale
where λ is random number generated from the beta distribution”)
Regarding claim 4, the rejection of claim 1 is incorporated herein. The combination of Taherzadeh Boroujeni, Geiping, Oki, Sriram, and Tang teach
Wherein refining the quantized, updated global model based on the reconstructed feature maps comprises: performing quantization-aware training, based on the reconstructed feature maps; ([0057] of Sriram states “In some instances, quantizing models such as rounding the weights and activations to lower-bit integers (e.g., 8-bit integers) is performed during training (e.g., quantization-aware training (QAT))” [0112] of Sriram states “In at least one embodiment, a model is a neural network model that is part of one or more federated learning processes. In at least one embodiment, QAT is applied, during training of a neural network” Pg. 3 I. Introduction section of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images”)
and performing quantization bias compensation, based on the reconstructed feature maps. (Pg.1 Abstract of Tang states “In this work, we provide a detailed analysis on this two-pass communication model, with error-compensated compression both on the worker nodes and on the parameter server. We show that the error-compensated stochastic gradient algorithm admits three very nice properties: 1) it is compatible with an arbitrary compression technique; 2) it admits an improved convergence rate than the non error-compensated stochastic gradient methods such as QSGD and sparse SGD; 3) it admits linear speedup with respect to the number of workers. An empirical study is also conducted to validate our theoretical results.” Pg. 3 I. Introduction section of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images”.)
Regarding claim 7, the rejection of claim 1 is incorporated herein. The combination of Taherzadeh Boroujeni, Geiping, Oki, Sriram, and Tang teach
wherein refining the quantized, updated global model based on the reconstructed feature maps comprises one of: mixing-up the reconstructed feature maps; (Pg. 1 Abstract of Oki states “In this paper, we propose to apply mix-up to the feature maps in a hidden layer” and Pg. 3 I. Introduction of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images” Pg. 1 Abstract of Geiping also states “we show that it is actually possible to faithfully reconstruct images at high resolution from the knowledge of their parameter gradients, and demonstrate that such a break of privacy is possible even for trained deep networks.”)
performing quantization-aware training, based on the reconstructed feature maps; ([0057] of Sriram states “In some instances, quantizing models such as rounding the weights and activations to lower-bit integers (e.g., 8-bit integers) is performed during training (e.g., quantization-aware training (QAT))” [0112] of Sriram states “In at least one embodiment, a model is a neural network model that is part of one or more federated learning processes. In at least one embodiment, QAT is applied, during training of a neural network” Pg. 3 I. Introduction section of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images”)
and performing quantization bias compensation, based on the reconstructed feature maps. (Pg.1 Abstract of Tang states “In this work, we provide a detailed analysis on this two-pass communication model, with error-compensated compression both on the worker nodes and on the parameter server. We show that the error-compensated stochastic gradient algorithm admits three very nice properties: 1) it is compatible with an arbitrary compression technique; 2) it admits an improved convergence rate than the non error-compensated stochastic gradient methods such as QSGD and sparse SGD; 3) it admits linear speedup with respect to the number of workers. An empirical study is also conducted to validate our theoretical results.” Pg. 3 I. Introduction section of Geiping states “Reconstruction of multiple, separate input images from their averaged gradient is possible in practice, over multiple epochs, using local mini-batches, or even for a local gradient averaging of up to 100 images”.)
Regarding claims 11 – 13, and 16, the rejection of claim 10 is incorporated herein. Claims 11 – 13, and 16 recite substantially similar subject matter as claims 2 – 4, and 7 respectively, and are rejected with the same rationale, mutatis mutandis.
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 BYUNGKWON HAN whose telephone number is (571)272-5294. The examiner can normally be reached M-F: 9:00AM-6PM PST.
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, Li B Zhen can be reached at (571)272-3768. 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.
/BYUNGKWON HAN/Examiner, Art Unit 2121
/Li B. Zhen/Supervisory Patent Examiner, Art Unit 2121
Prosecution Insights