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 .
Claims 1, 3-9, 11-17, and 19-24 are presented for examination.
Response to Amendment
The prior objections to the claims have been obviated by the amendment, and therefore are withdrawn. Most, but not all, of the prior objections to the specification have been obviated by the amendment. To the extent that an objection appears in both this Office Action and the previous Office Action it is maintained, otherwise, it is withdrawn.
Specification
The disclosure is objected to because of the following informalities:
[00106]: “input into the I/O interface 112” (last sentence of paragraph) should read “input into the I/O interface 212”
Appropriate correction is required.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1, 3-9, 11-17, and 19-24 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. The analysis of the claims will follow the 2019 Revised Patent Subject Matter Eligibility Guidance (“2019 PEG”).
Claim 1
Step 1: The claim is directed to a model training method, and therefore is directed to the statutory category of processes.
Step 2A Prong 1: The claim recites:
“obtaining… a first loss value based on the first prediction result”: This limitation encompasses mentally obtaining a first loss value by mentally determining a difference between the first prediction result and an expected output
“obtaining… a second loss value based on the second prediction result”: This limitation encompasses mentally obtaining a second loss value by mentally determining a difference between the second prediction result and an expected output
“performing… combination processing on the first loss value and the second loss value to obtain a third loss value, for updating the first private model”: This limitation encompasses mentally combining the first and second loss values to obtain a third loss value, and using the third loss value to mentally update parameters of the first private model
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “receiving, by the first client side device, a first shared model from the server,” “obtaining, by the first client side device, a data set comprising a first data set requiring privacy protection and a second data set not requiring privacy protection,” “outputting, by the first client side device, a first prediction result for the data set through the first shared model,” and “outputting, by the first client side device, a second prediction result for the data set through the first private model of the first client side device.” However, these limitations recite insignificant extra-solution activity, as they amount to mere data gathering and outputting (MPEP § 2106.05(g)). The claim additionally recites that the obtaining a first loss value, obtaining a second loss value, and performing combination processing steps are performed “by the first client side device,” and “wherein the first shared model and a first private model of the first client side device share a feature extraction model,” however, these limitations amount to mere instructions to apply a judicial exception on a generic computer programmed with a generic class of computer algorithms (MPEP § 2106.05(f)).
Step 2B: The claim does not contain significantly more than the judicial exception. The receiving a first shared model and obtaining a data set limitations, in addition to being insignificant extra-solution activity, is also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP § 2106.05(d)(II); buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network). The outputting limitations, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of storing and retrieving information in memory (MPEP § 2106.05(d)(II); Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93). Otherwise, the analysis at this step mirrors that of Step 2A Prong 2. As an ordered whole, the claim is directed to a mentally performable process of obtaining loss values for updating a model. Nothing in the claim provides significantly more than this. As such, the claim is not patent eligible.
Claim 3
Step 1: A process, as above.
Step 2A Prong 1: The claim recites:
“updating… a feature extraction model of the second private model based on the feature extraction model of the first shared model, to obtain the first private model”: This limitation encompasses mentally updating parameters of a feature extraction model of the second private model to match the feature extraction model of the first shared model, to obtain the first private model
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “wherein before the first shared model is received, the first client side device stores a second private model.” However, this limitation is directed to insignificant extra-solution activity as it amounts to mere data gathering (MPEP § 2106.05(g)), since the first client side device is storing data corresponding to a second private model, for use in performing the judicial exception. The claim additionally recites that the updating a feature extraction model is performed “by the first client side device,” however, this limitation amounts to mere instructions to apply a judicial exception on a generic computer (MPEP § 2106.05(f))
Step 2B: The claim does not contain significantly more than the judicial exception. The “stores a second private model” limitation, in addition to being insignificant extra-solution activity, is also directed to the well-understood, routine, and conventional activity of storing and retrieving information in memory (MPEP § 2106.05(d)(II); Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93). Otherwise, the analysis at this step mirrors that of Step 2A Prong 2.
Claim 4
Step 1: A process, as above.
Step 2A Prong 1: The claim recites:
“wherein the third loss value is used to update the first shared model to obtain a second shared model”: This limitation encompasses mentally updating parameters of the first shared model using the third loss value, to obtain a second shared model
“wherein the second shared model is used…to update the first shared model”: This limitation encompasses mentally updating parameters of the first shared model to match the second shared model
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “sending, by the first client side device, the second shared model to the server.” However, this limitation is directed to insignificant extra-solution activity, as it amounts to mere data gathering and outputting (MPEP § 2106.05(g)). The claim also recites that the “server” performs the updating of the first shared model. However, this limitation amounts to mere instructions to apply an exception using a generic computer (MPEP § 2106.05(f)).
Step 2B: The claim does not contain significantly more than the judicial exception. The sending limitation, in addition to being insignificant extra-solution activity, is also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP § 2106.05(d)(II); buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network). The server limitation is mere instructions to apply an exception using a generic computer (MPEP § 2106.05(f)) as stated above.
Claim 5
Step 1: A process, as above.
Step 2A Prong 1: The claim recites:
“wherein the first loss value comprises at least one of a cross-entropy loss value and/or a mutual distillation loss value, and the second loss value comprises a cross-entropy loss value and/or a mutual distillation loss value”: This limitation further limits the obtained first and second loss values to be at least one of a cross-entropy loss value and/or mutual distillation loss value, which are both mathematical calculations
Step 2A Prong 2: This judicial exception is not integrated into a practical application. See analysis of claim 1.
Step 2B: The claim does not contain significantly more than the judicial exception. See analysis of claim 1.
Claim 6
Step 1: A process, as above.
Step 2A Prong 1: The claim recites the same judicial exception as claim 1.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “outputting the first prediction result for the second data set through the first shared model,” and “outputting the second prediction result for the first data set and the second data set through the first private model.” However, these limitations are directed to insignificant extra-solution activity, as they amount to mere data gathering and outputting (MPEP § 2106.05(g)).
Step 2B: The claim does not contain significantly more than the judicial exception. The outputting limitations, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of storing and retrieving information in memory (MPEP § 2106.05(d)(II); Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93).
Claim 7
Step 1: A process, as above.
Step 2A Prong 1: The claim recites:
“performing weighted combination processing on the first loss value and the second loss value to obtain the third loss value”: This limitation encompasses the mathematical concept of calculating a weighted combination of two values to obtain a third value, and is also mentally performable
Step 2A Prong 2: This judicial exception is not integrated into a practical application. See analysis of claim 1.
Step 2B: The claim does not contain significantly more than the judicial exception. See analysis of claim 1.
Claim 8
Step 1: A process, as above.
Step 2A Prong 1: The claim recites:
“obtaining an average value of the first loss value and the second loss value as the third loss value”: This limitation encompasses the mathematical concept of averaging two values to obtain a third value, and is also mentally performable
Step 2A Prong 2: This judicial exception is not integrated into a practical application. See analysis of claim 1.
Step 2B: The claim does not contain significantly more than the judicial exception. See analysis of claim 1.
Claim 9
Step 1: The claim is directed to a data processing method, and therefore is directed to the statutory category of processes.
Step 2A Prong 1: The claim recites:
“wherein the target model is obtained by updating a first private model based on a third loss value”: this limitation encompasses mentally updating parameters of a first private model based on a third loss value to obtain a target model
“the third loss value is obtained by performing combination processing on a first loss value and a second loss value”: this limitation encompasses mentally combining a first loss value and a second loss value to obtain the third loss value
“the first loss value is obtained based on a first prediction result”: this limitation encompasses mentally obtaining a first loss value by determining a difference between a first prediction result and an expected outcome
“the second loss value is obtained based on a second prediction result”: this limitation encompasses mentally obtaining a second loss value by determining a difference between a second prediction result and an expected outcome
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “obtaining to-be-processed data,” “the first prediction result is output by a first shared model for a data set,” “the data set comprises a first data set requiring privacy protection and a second data set not requiring privacy protection,” “the first shared model is obtained from the server,” and “the second prediction result is output by the first private model for the data set.” However, these limitations are directed to insignificant extra-solution activity, as they amount to mere data gathering and outputting (MPEP § 2106.05(g)). The claim additionally further recites “the first shared model and the first private model share a feature extraction model,” however, this limitation amounts to mere instructions to apply an exception on a generic computer programmed with a generic class of computer algorithms (MPEP § 2106.05(f)). The claim also further recites “processing the to-be-processed data based on a target model stored in the first client side device, to obtain a prediction result.” However, this limitation amounts to mere instructions to apply an exception on a generic computer (MPEP § 2106.05(f)), as it is merely applying the model created by the judicial exception on a generically recited client side device.
Step 2B: The claim does not contain significantly more than the judicial exception. The “obtaining to-be processed data” and “the first shared model is obtained from the server” limitations, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP § 2106.05(d)(II); buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network). The output limitations, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of storing and retrieving information in memory (MPEP § 2106.05(d)(II); Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93). The processing data, and the first shared model and first private model sharing a feature extraction model limitations amount to mere instructions to apply an exception on a generic computer (MPEP § 2106.05(f)) for the same reasons stated above. As an ordered whole, the claim is directed to a mentally performable process of obtaining a target model for processing data by updating a model using obtained loss values. Nothing in the claim provides significantly more than this. As such, the claim is not patent eligible.
Claims 11-16
Step 1: A process, as above.
Step 2A Prong 1: Claims 10-16 recite the same judicial exception as claims 2-8, respectively, except insofar as they inherit the abstract ideas from claim 9 rather than claim 1.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. Claims 10-16 recite the same additional elements as claims 2-8, respectively, except insofar as they inherit the additional elements recited in claim 9 rather than claim 1.
Step 2B: The claim does not contain significantly more than the judicial exception. The analysis at this step mirrors that of claims 2-8, respectively, except insofar as the claims inherit the additional elements recited in claim 9 rather than claim 1.
Claim 17
Step 1: The claim is directed to a client side device, and is therefore directed to the statutory category of machines.
Step 2A Prong 1: The claim recites:
“obtain a first loss value based on the first prediction result”: This limitation encompasses mentally obtaining a first loss value by mentally determining a difference between the first prediction result and an expected output
“obtain a second loss value based on the second prediction result”: This limitation encompasses mentally obtaining a second loss value by mentally determining a difference between the second prediction result and an expected output
“perform combination processing on the first loss value and the second loss value to obtain a third loss value for updating the first private model”: This limitation encompasses mentally combining the first and second loss values to obtain a third loss value, and using the third loss value to mentally update parameters of the first private model
Step 2A Prong 2: This judicial exception is not integrated into a practical application. The claim further recites “receive a first shared model sent by a server,” “obtain a data set comprising a first data set requiring privacy protection and a second data set not requiring privacy protection,” “output a first prediction result for the data set through the first shared model,” and “output a second prediction result for the data set through the first private model of the client side device.” However, these limitations recite insignificant extra-solution activity, as they amount to mere data gathering and outputting (MPEP § 2106.05(g)). The claim additionally recites a client side device that “comprises a transceiver and a trainer,” and “wherein the first shared model and a first private model of the client side device share a feature extraction model,” however, these limitations amount to mere instructions to apply a judicial exception using a generic computer programmed with a generic class of computer algorithms (MPEP § 2106.05(f)).
Step 2B: The claims do not contain significantly more than the judicial exception. The receive a first shared model and obtain a data set limitations, in addition to being insignificant extra-solution activity, is also directed to the well-understood, routine, and conventional activity of receiving or transmitting data over a network (MPEP § 2106.05(d)(II); buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network). The output limitations, in addition to being insignificant extra-solution activity, are also directed to the well-understood, routine, and conventional activity of storing and retrieving information in memory (MPEP § 2106.05(d)(II); Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93). Otherwise, the analysis at this step mirrors that of Step 2A Prong 2. As an ordered whole, the claim is directed to a mentally performable process of obtaining loss values for updating a model. Nothing in the claim provides significantly more than this. As such, the claim is not patent eligible.
Claims 19-24
Step 1: A machine, as above.
Step 2A Prong 1: Claims 19-24 recite the same judicial exception as claims 3-8, respectively.
Step 2A Prong 2: This judicial exception is not integrated into a practical application. Claims 17-24 recite the same additional elements as claims 1-8, respectively, except insofar as they are directed to a client side device that “comprises a transceiver and a trainer.” However, the client side device limitation amounts to mere instructions to apply a judicial exception using a generic computer (MPEP § 2106.05(f)).
Step 2B: The claims do not contain significantly more than the judicial exception. The analysis at this step mirrors that of claims 1-8, respectively, except insofar as the claims are directed to a client side device, which amounts to mere instructions to apply a judicial exception using a generic computer (MPEP § 2106.05(f)) as stated above.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
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, 3-9, and 11-16 are rejected under 35 U.S.C. 103 as being unpatentable over Lin et al. (NPL: “LINDT: Tackling Negative Federated Learning with Local Adaptation”) (“Lin”) in view of Yang et al. (“Heterogeneous Data-Aware Federated Learning”) (“Yang”) and Zhao et al. (“Federated Learning with Non-IID Data”) (“Zhao”).
Regarding claim 1, Lin discloses “A model training method, wherein the method is applicable to a machine learning system, the machine learning system comprises a server (Figure 2: box labeled “Server”) and at least two client side devices (3.1: “Given a system of N clients”); including a first client side device (Figure 2: box labeled “Client”), and the method comprises:
receiving, by the first client side device, a first shared model from the server… (Figure 2, (1): “broadcast global model” with arrow pointing from “Server” box to “Client” box);
obtaining, by the first client side device, a data set… (Figure 2: “Local Data” in “Client” box)
outputting a first prediction result for a data set through the first shared model (Figure 2: “prediction” output from “Layer M” of “Layers in Global Model” in “Client” box, and “Local Data” input into “Layer 1”);
obtaining a first loss value based on the first prediction result (4.2: “For updating the dual-model, we denote by
G
M
x
,
w
i
r
a
n
d
L
i
M
x
,
v
i
,
the respective pre-softmax logit outputs from the last layer of the dual-model (one global and one local), and let
l
c
r
o
s
s
be cross entropy”; Equation (10):
l
c
r
o
s
s
G
M
x
,
w
i
r
,
y
; the examiner notes that
l
c
r
o
s
s
G
M
x
,
w
i
r
,
y
corresponds to a first loss value based on the first prediction result, as it is the cross entropy loss of the output from the global model);
outputting a second prediction result for the data set through a first private model of the first client side device (Figure 2: “prediction” output from “Layer M” of “Layers in Local Model” in “Client” box, and “Local Data” input into “Layer 1”);
obtaining a second loss value based on the second prediction result (4.2: “For updating the dual-model, we denote by
G
M
x
,
w
i
r
a
n
d
L
i
M
x
,
v
i
,
the respective pre-softmax logit outputs from the last layer of the dual-model (one global and one local), and let
l
c
r
o
s
s
be cross entropy”; Equation (10):
l
c
r
o
s
s
L
i
M
x
,
v
i
,
y
; the examiner notes that
l
c
r
o
s
s
L
i
M
x
,
v
i
,
y
corresponds to a second loss value based on the second prediction result, as it is the cross entropy loss of the output from the local model);
and performing combination processing on the first loss value and the second loss value to obtain a third loss value, for updating the first private model (4.2: “For updating the dual-model… Each client i objects to minimize
l
(
w
i
r
,
v
i
) = …”; see equation (10); the examiner notes that a third loss value
l
(
w
i
r
,
v
i
) is obtained by finding the expected value
E
of the combination of the first and second loss values, and that updating the dual-model requires updating the local model).”
Lin does not appear to explicitly disclose the further limitations of the claim.
However, Yang discloses “wherein…[a] first shared model and…[a] first private model share a
feature extraction model” (Yang, Section 4, paragraph 2: “Following this split of generic and specific
layers, we propose to share the generic feature extraction, like the convolutional layers, between the
servers and the clients and more precisely to keep the specific layer as the classification layer only
on the clients side”; paragraph 4: “The server then broadcasts the initialized parameters to all the
clients. At each round t of communications, the clients update the local parameters by copying the generic parameters wbt. For the first round of communication, the specific parameters are copied as well”; the examiner notes that the generic feature extraction convolutional layers that are shared between the servers and the clients correspond to “a feature extraction model”, the “initialized
parameters” that are broadcast to all the clients corresponds to “a first shared model” and the client’s “local parameters” correspond to “a first private model”).
Yang and the instant application both relate to federated learning and are analogous. It would
have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the first shared model and the first private model disclosed by Lin to share a
feature extraction model as disclosed by Yang, and one would have been motivated to do so for the
purposes of allowing the feature extraction part of the model to benefit from all the data and therefore
increase its robustness (Yang, Section 4, paragraph 2).
Neither Lin nor Yang appear to explicitly disclose the further limitations of the claim.
However, Zhao discloses “obtaining, by… [a] first client side device, a data set comprising a first data set requiring privacy protection and a second data set not requiring privacy protection” (Zhao, Section 4.1, Figure 6: “Private Data” and “Shared Data”).
Zhao and the instant application both relate to federated learning and are analogous. It would
have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed
invention, to have modified the data set disclosed by the combination of Lin and Yang to comprise a first data set requiring privacy protection and a second data set not requiring privacy protection as disclosed by Zhao, and one would have been motivated to do so for the purpose of improving training on non-IID data by creating a subset of data which is globally shared, decreasing weight divergence and leading to improved accuracy of federated learning (see Zhao, Abstract).
Regarding claim 3, the rejection of claim 1 is incorporated. Lin as modified by Yang and Zhao further discloses “wherein before the first shared model is received, the first client side device stores a second private model (Yang, Section 4, paragraph 4: “The server then broadcasts the initialized parameters to all the clients. At each round t of communications, the clients update the local parameters by copying the generic parameters wbt”; the examiner notes that the client’s “local parameters” prior to the first round of communications correspond to a second private model); and the method further comprises:
updating, by the client side device, a feature extraction model of the second private model based on the feature extraction model of the first shared model, to obtain the first private model (Yang, Section 4, paragraph 4: “The server then broadcasts the initialized parameters to all the clients. At each round t of communications, the clients update the local parameters by copying the generic parameters wbt. For the first round of communication, the specific parameters are copied as well. The clients update the received model on their local data with fixed epochs E and send back the updated generic parameters
w
b
t
k
to server; the examiner notes that updating the generic local parameters by copying the generic shared parameters corresponds to “updating a feature extraction model of the second private model based on the feature extraction model of the first shared model,” and the updated local parameters correspond to “the first private model”).
Yang and the instant application both relate to federated learning and are analogous. It would
have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the combination of Lin and Zhao to include storing a second private model before the first shared model is received, and updating a feature extraction model of the second private model based on the feature extraction model of the first shared model to obtain the first private model, as disclosed by Yang, and one would have been motivated to do so for the purposes of allowing the feature extraction part of the model to benefit from all the data and therefore increase its robustness (Yang, Section 4, paragraph 2).
Regarding claim 4, the rejection of claim 1 is incorporated. Lin as modified by Yang and Zhao further discloses “wherein the third loss value is used to update the first shared model (4.2: “For updating the dual-model… Each client i objects to minimize
l
(
w
i
r
,
v
i
) = …”; see equation (10); the examiner notes that updating the dual-model requires updating the global model) to obtain a second shared model (Figure 2: “new global model”); and the method further comprises:
sending, by the first client side device, the second shared model to the server (Figure 2, (3) “upload new global model” with arrow pointing from “Client” box to “Server” box),
wherein the second shared model is used by the server to update the first shared model” (Figure 2, (4) “aggregation” with arrow pointing from “Server” where the “new global model” was received to the original global model; 4.2: "At the end of each round r of dual-model training, clients are required to return the updated
w
i
r
to the server for aggregation (same as line c5 in Algorithm 1). The server, upon receiving the first K updates from the active clients (Cr), aggregates the parameters in the same way as conventional FL.”).
Regarding claim 5, the rejection of claim 1 is incorporated. Lin as modified by Yang and Zhao further discloses “wherein the first loss value comprises at least one of a cross-entropy loss value and/or a mutual distillation loss value, and the second loss value comprises at least one of a cross-entropy loss value and/or a mutual distillation loss value (4.2, equation (10); the examiner notes that the first loss value
l
c
r
o
s
s
G
M
x
,
w
i
r
,
y
and the second loss value
l
c
r
o
s
s
L
i
M
x
,
v
i
,
y
are both cross-entropy loss values).
Regarding claim 6, the rejection of claim 1 is incorporated. Lin as modified by Yang and Zhao further discloses “wherein the outputting a first prediction result for the data set through the first shared model comprises:
outputting the first prediction result for the second data set through the first shared model (Zhao, Section 4.2, Paragraph 1: “Herein, we propose a data-sharing strategy in the federated learning setting as illustrated in Figure 6. A globally shared dataset G that consists of a uniform distribution over classes is centralized in the cloud. At the initialization stage of FedAvg, the warm-up model trained on G and a random α portion of G are distributed to each client”; the examiner notes that the “warm-up model” that is shared to each client corresponds to “a first shared model,” and that training the warm-up model on the shared dataset (corresponding to the second data set) inherently involves outputting a first prediction result for the second data set, as the output is needed to train the model);
and the outputting a second prediction result for the data set through the first private model of the first client side device comprises:
outputting the second prediction result for the first data set and the second data set through the first private model.” (Zhao, Section 4.2, Paragraph 1: “The local model of each client is trained on the shared data from G together with the private data from each client”; the examiner notes that the “local model” of one of the clients corresponds to “a first private model of a first client side device,” and that training the local model on the shared data from G (the random portion of G that was distributed to each client can be the whole set) and the private data from each client inherently involves outputting a second prediction result for the first data set and the second data set, as the output is needed to train the model)
Zhao and the instant application both relate to federated learning and are analogous. It would
have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed
invention, to have modified the outputting prediction result steps disclosed by the combination of Lin and Yang, to include using the second data set not requiring privacy protection with the first shared model, and both the first data set requiring privacy protection and the second data set not requiring privacy protection with the first private model as disclosed by Zhao. One would have been motivated to do so for the purpose of improving training on non-IID data by creating a subset of data which is globally shared, decreasing weight divergence and leading to improved accuracy of federated learning (see Zhao, Abstract).
Regarding claim 7, the rejection of claim 1 is incorporated. Lin as modified by Yang and Zhao further discloses “wherein the performing combination processing on the first loss value and the second loss value to obtain a third loss value comprises:
performing weighted combination processing on the first loss value and the second loss value to obtain the third loss value” (4.2, equation (10); the examiner notes that the third loss value is obtained by finding the expected value
E
of the combination of the first and second loss values, and expected value is a weighted average calculation).
Regarding claim 8, the rejection of claim 7 is incorporated. Lin as modified by Yang and Zhao further discloses “wherein the performing weighted combination processing on the first loss value and the second loss value to obtain the third loss value comprises:
obtaining an average value of the first loss value and the second loss value as the third loss value” (4.2, equation (10); the examiner notes that the third loss value is obtained by finding the expected value
E
of the combination of the first and second loss values, and expected value is a weighted average calculation).
Regarding claim 9, Lin discloses “A data processing method, wherein the method is applicable to a machine learning system, the machine learning system comprises a server (Figure 2: box labeled “Server”) and at least two client side devices (3.1: “Given a system of N clients); including a first client side device (Figure 2: box labeled “Client”), and the method comprises:
obtaining to-be-processed data (Figure 2: “Local Data”);
and processing the to-be-processed data based on a target model stored in the first client side device, to obtain a prediction result (Figure 2 and Figure 2 Caption: “But when testing, each client only considers the prediction from its local model”; the examiner notes that the “Prediction” output from “Layers in Local Model” during testing (as opposed to training) corresponds to a “prediction result” and the “Layers in Local Model” during testing (which were updated during training) correspond to “a target model”),
wherein the target model is obtained by updating a first private model based on a third loss value (4.2: “For updating the dual-model… Each client i objects to minimize
l
(
w
i
r
,
v
i
) = …”; see equation (10); the examiner notes that updating the dual-model by minimizing a third loss value
l
(
w
i
r
,
v
i
) requires updating the local model, and the updated local model corresponds to the target model),
the third loss value is obtained by performing combination processing on a first loss value and a second loss value (4.2: “For updating the dual-model… Each client i objects to minimize
l
(
w
i
r
,
v
i
) = …”; see equation (10); the examiner notes that the third loss value
l
(
w
i
r
,
v
i
) is obtained by finding the expected value
E
of the combination of a first loss value
l
c
r
o
s
s
G
M
x
,
w
i
r
,
y
and a second loss value
l
c
r
o
s
s
L
i
M
x
,
v
i
,
y
),
the first loss value is obtained based on a first prediction result (4.2: “For updating the dual-model, we denote by
G
M
x
,
w
i
r
a
n
d
L
i
M
x
,
v
i
,
the respective pre-softmax logit outputs from the last layer of the dual-model (one global and one local), and let
l
c
r
o
s
s
be cross entropy”; Equation (10):
l
c
r
o
s
s
G
M
x
,
w
i
r
,
y
; the examiner notes that
l
c
r
o
s
s
G
M
x
,
w
i
r
,
y
corresponds to a first loss value obtained based on a first prediction result, as it is the cross entropy loss of the output prediction from the global model),
the first prediction result is output by a first shared model for a data set… (Figure 2: “prediction” output from “Layer M” of “Layers in Global Model” in “Client” box, and “Local Data” input into “Layer 1”)
the first shared model is obtained from the server (Figure 2, (1): “broadcast global model” with arrow pointing from “Server” box to “Client” box),
the second loss value is obtained based on a second prediction result (4.2: “For updating the dual-model, we denote by
G
M
x
,
w
i
r
a
n
d
L
i
M
x
,
v
i
,
the respective pre-softmax logit outputs from the last layer of the dual-model (one global and one local), and let
l
c
r
o
s
s
be cross entropy”; Equation (10):
l
c
r
o
s
s
L
i
M
x
,
v
i
,
y
; the examiner notes that
l
c
r
o
s
s
L
i
M
x
,
v
i
,
y
corresponds to a second loss value based on a second prediction result, as it is the cross entropy loss of the output prediction from the local model),
and the second prediction result is output by the first private model for the data set” (Figure 2: “prediction” output from “Layer M” of “Layers in Local Model” in “Client” box, and “Local Data” input into “Layer 1”).
Lin does not appear to explicitly disclose the further limitations of the claim.
However, Zhao discloses “…[a] data set comprises a first data set requiring privacy protection and a second data set not requiring privacy protection” (Zhao, Section 4.1, Figure 6: “Private Data” and “Shared Data”).
Zhao and the instant application both relate to federated learning and are analogous. It would
have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed
invention, to have modified the data set disclosed Lin to comprise a first data set requiring privacy protection and the second data set not requiring privacy protection as disclosed by Zhao, and one would have been motivated to do so for the purpose of improving training on non-IID data by creating a subset of data which is globally shared, decreasing weight divergence and leading to improved accuracy of federated learning (see Zhao, Abstract).
Lin as modified by Zhao does not appear to explicitly disclose the further limitations of the claim.
However, Yang discloses “…[a] first shared model and…[a] first private model share a
feature extraction model” (Yang, Section 4, paragraph 2: “Following this split of generic and specific
layers, we propose to share the generic feature extraction, like the convolutional layers, between the
servers and the clients and more precisely to keep the specific layer as the classification layer only
on the clients side”; paragraph 4: “The server then broadcasts the initialized parameters to all the
clients. At each round t of communications, the clients update the local parameters by copying the generic parameters wbt. For the first round of communication, the specific parameters are copied as well”; the examiner notes that the generic feature extraction convolutional layers that are shared between the servers and the clients correspond to “a feature extraction model”, the “initialized
parameters” that are broadcast to all the clients corresponds to “a first shared model” and the client’s “local parameters” correspond to “a first private model”).
Yang and the instant application both relate to federated learning and are analogous. It would
have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the first shared model and the first private model disclosed by the combination of Lin and Zhao to share a feature extraction model as disclosed by Yang, and one would have been motivated to do so for the purposes of allowing the feature extraction part of the model to benefit from all the data and therefore increase its robustness (Yang, Section 4, paragraph 2).
Regarding claim 11, the rejection of claim 9 is incorporated. The further limitations of the claim correspond to those of claim 3, and the remainder of the rejection follows the same rationale as the rejection of claim 3 above.
Regarding claim 12, the rejection of claim 9 is incorporated. The further limitations of the claim correspond to those of claim 4, and the remainder of the rejection follows the same rationale as the rejection of claim 4 above.
Regarding claim 13, the rejection of claim 9 is incorporated. The further limitations of the claim correspond to those of claim 5, and the remainder of the rejection follows the same rationale as the rejection of claim 5 above.
Regarding claim 14, the rejection of claim 9 is incorporated. The further limitations of the claim correspond to those of claim 6, and the remainder of the rejection follows the same rationale as the rejection of claim 6 above.
Regarding claim 15, the rejection of claim 9 is incorporated. The further limitations of the claim correspond to those of claim 7, and the remainder of the rejection follows the same rationale as the rejection of claim 7 above.
Regarding claim 16, the rejection of claim 15 is incorporated. The further limitations of the claim correspond to those of claim 8, and the remainder of the rejection follows the same rationale as the rejection of claim 8 above.
Claims 17 and 19-24 are rejected under 35 U.S.C. 103 as being unpatentable over Lin in view of Yang, Zhao, and Alabbasi et al. (WO2021121585) (“Alabbasi”).
Regarding claim 17, Lin discloses “A client side device (Figure 2: box labeled “Client”), wherein the client side device is applicable to a machine learning system, the machine learning system comprises at least two client side devices (3.1: “Given a system of N clients); including the client side device… wherein the… [client side device] is configured to receive a first shared model sent by a server… (Figure 2, (1): “broadcast global model” with arrow pointing from “Server” box to “Client” box); and the… [client side device] is configured to:
obtain a data set… (Figure 2: “Local Data” in “Client” box)
output a first prediction result for the data set through the first shared model (Figure 2: “prediction” output from “Layer M” of “Layers in Global Model” in “Client” box, and “Local Data” input into “Layer 1”);
obtain a first loss value based on the first prediction result (4.2: “For updating the dual-model, we denote by
G
M
x
,
w
i
r
a
n
d
L
i
M
x
,
v
i
,
the respective pre-softmax logit outputs from the last layer of the dual-model (one global and one local), and let
l
c
r
o
s
s
be cross entropy”; Equation (10):
l
c
r
o
s
s
G
M
x
,
w
i
r
,
y
; the examiner notes that
l
c
r
o
s
s
G
M
x
,
w
i
r
,
y
corresponds to a first loss value based on the first prediction result, as it is the cross entropy loss of the output from the global model);
output a second prediction result for the data set through the first private model of the client side device (Figure 2: “prediction” output from “Layer M” of “Layers in Local Model” in “Client” box, and “Local Data” input into “Layer 1”);
obtain a second loss value based on the second prediction result (4.2: “For updating the dual-model, we denote by
G
M
x
,
w
i
r
a
n
d
L
i
M
x
,
v
i
,
the respective pre-softmax logit outputs from the last layer of the dual-model (one global and one local), and let
l
c
r
o
s
s
be cross entropy”; Equation (10):
l
c
r
o
s
s
L
i
M
x
,
v
i
,
y
, the examiner notes that
l
c
r
o
s
s
L
i
M
x
,
v
i
,
y
corresponds to a second loss value based on the second prediction result, as it is the cross entropy loss of the output from the local model);
and perform combination processing on the first loss value and the second loss value to obtain a third loss value for updating the first private model” (4.2: “For updating the dual-model… Each client i objects to minimize
l
(
w
i
r
,
v
i
) = …”; see equation (10); the examiner notes that a third loss value
l
(
w
i
r
,
v
i
) is obtained by finding the expected value
E
of the combination of the first and second loss values, and that updating the dual-model requires updating the local model).
Lin does not appear to explicitly disclose the further limitations of the claim.
However, Yang discloses “wherein…[a] first shared model and…[a] first private model of the client side device share a feature extraction model” (Yang, Section 4, paragraph 2: “Following this split of generic and specific layers, we propose to share the generic feature extraction, like the convolutional layers, between the servers and the clients and more precisely to keep the specific layer as the classification layer only on the clients side”; paragraph 4: “The server then broadcasts the initialized parameters to all the clients. At each round t of communications, the clients update the local parameters by copying the generic parameters wbt. For the first round of communication, the specific parameters are copied as well”; the examiner notes that the generic feature extraction convolutional layers that are shared between the servers and the clients correspond to “a feature extraction model”, the “initialized parameters” that are broadcast to all the clients corresponds to “a first shared model” and the client’s “local parameters” correspond to “a first private model”).
Yang and the instant application both relate to federated learning and are analogous. It would
have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the first shared model and the first private model disclosed by Lin to share a
feature extraction model as disclosed by Yang, and one would have been motivated to do so for the
purposes of allowing the feature extraction part of the model to benefit from all the data and therefore
increase its robustness (Yang, Section 4, paragraph 2).
Neither Lin nor Yang appear to explicitly disclose the further limitations of the claim.
However, Zhao discloses “obtain a data set comprising a first data set requiring privacy protection and a second data set not requiring privacy protection” (Zhao, Section 4.1, Figure 6: “Private Data” and “Shared Data”).
Zhao and the instant application both relate to federated learning and are analogous. It would
have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed
invention, to have modified the data set disclosed by the combination of Lin and Yang to comprise a first data set requiring privacy protection and a second data set not requiring privacy protection as disclosed by Zhao, and one would have been motivated to do so for the purpose of improving training on non-IID data by creating a subset of data which is globally shared, decreasing weight divergence and leading to improved accuracy of federated learning (see Zhao, Abstract).
Neither Lin, Yang, nor Zhao appear to explicitly disclose that the client side device “comprises a transceiver and a trainer.”
However, Alabbasi discloses “…[a] client side device comprises a transceiver and a trainer” (Alabbasi, Figure 7 and [00101]: Client Computing Device 700, Transceiver 701, and Processing Circuit 732; the examiner notes that the processing circuit corresponds to a “trainer” because it “performs respective operations discussed herein” of the client computing device, including training, see [0008]: “The client computing device can perform further operations training the aggregated machine learning model in iterations with inputs”).
Alabbasi and the instant application both relate to federated learning and are analogous. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to have modified the client side device disclosed the combination of Lin/Yang/Zhao to comprise a transceiver and a trainer as disclosed by Alabbasi, and one would have been motivated to do so for the purpose of providing the hardware necessary to allow the client device to send and receive communications to/from a central server, and train a model with the client’s own data to generate model updates for an aggregate model (see Alabbasi, [0008] and [00101]).
Regarding claim 19, the rejection of claim 17 is incorporated. The further limitations of the claim correspond to those of claim 3, and the remainder of the rejection follows the same rationale as the rejection of claim 3 above, except insofar as the combination of Lin/Zhao/Alabbasi (as opposed to only Lin and Zhao) is modified to include the teachings of Yang.
Regarding claim 20, the rejection of claim 17 is incorporated. The further limitations of the claim correspond to those of claim 4, and the remainder of the rejection follows the same rationale as the rejection of claim 4 above.
Regarding claim 21, the rejection of claim 17 is incorporated. The further limitations of the claim correspond to those of claim 5, and the remainder of the rejection follows the same rationale as the rejection of claim 5 above.
Regarding claim 22, the rejection of claim 17 is incorporated. The further limitations of the claim correspond to those of claim 6, and the remainder of the rejection follows the same rationale as the rejection of claim 6 above, except insofar as the combination of Lin /Yang/Alabbasi (as opposed to only Lin and Yang) is modified to include the teachings of Zhao.
Regarding claim 23, the rejection of claim 17 is incorporated. The further limitations of the claim correspond to those of claim 7, and the remainder of the rejection follows the same rationale as the rejection of claim 7 above.
Regarding claim 24, the rejection of claim 23 is incorporated. The further limitations of the claim correspond to those of claim 8, and the remainder of the rejection follows the same rationale as the rejection of claim 8 above.
Response to Arguments
Applicant's arguments filed April 20, 2026 have been fully considered but they are not persuasive.
Regarding the rejection under 35 USC § 101, Applicant argues on pages 18-19 that amended claim 1 falls outside the mental process grouping because “such coordinated federated learning operations involving multiple models, shared feature extraction, privacy-partitioned training data, and loss-based model updating cannot practically be performed in the human mind.” Examiner submits that while the entire claimed process cannot be performed mentally, the analysis at Step 2A Prong 1 is whether the claim recites a judicial exception. Examiner maintains that obtaining a first loss value based on the first prediction result, obtaining a second loss value based on the second prediction result, and combining the first loss value and second loss value to obtain a third loss value are steps that can be practically performed in the human mind.
Applicant further argues on page 19 that the “receiving” limitation is not insignificant extra solution activity because claim 1 requires that the received shared model is one of the operative models used in the claimed training process. Examiner respectfully disagrees. MPEP 2106.05(g) states “[a]n example of pre-solution activity is a step of gathering data for use in a claimed process,” therefore, receiving a first shared model for use in the claimed training process falls under insignificant extra-solution activity. Applicant likewise argues that the “outputting” limitations are not mere post-solution reporting, because claim 1 requires that the first and second prediction results are the specific model outputs from which the first and second loss values are obtained, and those loss values are then used in the claimed combination processing to obtain the third loss value for updating the private model. MPEP 2106.05(g) states that one of the considerations for determining whether an additional element is insignificant extra-solution activity is “[w]hether the limitation amounts to necessary data gathering and outputting (i.e., all uses of the recited judicial exception require such data gathering or data output).” The outputting limitations are necessary for gathering the first and second prediction results for performing the judicial exceptions of obtaining first and second loss values, and then performing combination processing on the first and second loss values to obtain the third loss value, and thus amount to insignificant extra-solution activity.
Applicant further argues on pages 19-20 that the additional elements integrate any such abstract idea into a practical application by providing a technological improvement to federated-learning model training, alleging that the claimed process improves performance of the private model. Examiner fails to see a clear nexus between the claim language and this alleged improvement. Amended claim 1 (and similarly 9 and 17) does not reflect different portions of the training data being handled differently, the shared model participating in federated learning and the private model not directly participating, and does not positively recite the training/updating of the models. Rather, the claim is directed to outputting prediction results from each of the models to obtain a first and second loss value, and then combining the loss values to obtain a third loss value, with updating the private model nominally recited as an intended use of the third loss value. Examiner does not see a nexus between any additional elements recited in the claim and a technological improvement, and the improvement cannot come from the judicial exception itself of performing combination processing on the first loss value and second loss value to obtain a third loss value (MPEP 2106.05(a): “the judicial exception alone cannot provide the improvement).
Regarding the prior art rejections, Applicant argues on page 20, regarding amended claim 1, that Lin fails to disclose that “the first shared model and a first private model of the first client side device share a feature extraction model” and “obtaining, by the first client side device, a data set comprising a first data set requiring privacy protection and a second data set not requiring privacy protection.” The new ground of rejection under 35 U.S.C § 103 relies on the combination of Lin, Yang, and Zhao to teach these limitations, with Yang teaching a shared feature extraction model and Zhao teaching a data set comprising a first data set requiring privacy protection and a second data set not requiring privacy protection.
Applicant further argues that Lin fails to disclose the claimed operations of obtaining a first loss value based on a first prediction result output by the first shared model, obtaining a second loss value based on a second prediction result output by the first private model, and performing combination processing on the first and second loss values to obtain a third loss value for updating the first private model. Examiner respectfully disagrees. As mapped in the rejection, Lin discloses calculating a loss value (third loss value) that combines a cross-entropy loss based on a prediction from the global model layers (first loss value) and a cross-entropy loss based on a prediction from the local model layers (second loss value) (see Lin, eq 10), and updating the dual-model involves updating the local model (first private model), as the local model is part of the dual-model.
Applicant further argues on page 21 that Yang fails to rectify the deficiencies of Lin, because Yang describes a single split model having a shared generic portion and a client-specific portion, not a first shared model and a first private model that coexist on the client and share a feature extraction model, and the Office Action's mapping of Yang's broadcast initialized parameters to a "first shared model" and Yang's local parameters to a "first private model" allegedly improperly recharacterizes parameter subsets of Yang's single split model as two separate models. Examiner respectfully disagrees. The server contains generic parameters wbt which can be characterized as a first shared model, because they are parameters corresponding to generic feature extraction layers of a model and are broadcast (shared) to the clients, and the client contains local parameters
w
b
t
k
and
w
l
t
k
which can be characterized as a first private model because they are parameters corresponding to generic feature extraction layers and specific layers that make up a client model. The client then copies the generic parameters sent by the server, thus the server model and client model share a feature extraction model. The client model is different from the server model because it contains the client’s own specific layers as well as the shared feature extraction layers. There is no requirement for Yang to disclose that the first shared model and first private model coexist on the client, as this is disclosed by Lin. It would have been obvious to one of ordinary skill in the art to have modified the first private model and first shared model disclosed by Lin to share a feature extraction model as taught by Yang, for the motivation given in the Office Action.
Applicant further argues on page 22 that Zhao does not teach a client-side training framework having both a first shared model and a first private model operating together, much less such models sharing a feature extraction model. Examiner submits that Lin and Yang are used to teach these limitations, rather than Zhao. Applicant additionally argues that Zhao does not teach obtaining a first loss value based on a first prediction result output by a first shared model, obtaining a second loss value based on a second prediction result output by a first private model, and performing combination processing on the first and second loss values to obtain a third loss value for updating the first private model. Examiner submits that Lin discloses these limitations.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GWYNEVERE A DETERDING whose telephone number is (571)272-7657. The examiner can normally be reached Mon-Fri. 9am-5pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Kamran Afshar can be reached at (571) 272-7796. 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.
/G.A.D./Examiner, Art Unit 2125
/KAMRAN AFSHAR/Supervisory Patent Examiner, Art Unit 2125
Prosecution Insights