DETAILED ACTION
This action is in response to the application filed 02/28/2024. Claims 1-17 are pending and have been examined.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 02/28/2024 is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Interpretation
The following is a quotation of MPEP 2111.04 II:
The broadest reasonable interpretation of a method (or process) claim having contingent limitations requires only those steps that must be performed and does not include steps that are not required to be performed because the condition(s) precedent are not met. For example, assume a method claim requires step A if a first condition happens and step B if a second condition happens. If the claimed invention may be practiced without either the first or second condition happening, then neither step A or B is required by the broadest reasonable interpretation of the claim. If the claimed invention requires the first condition to occur, then the broadest reasonable interpretation of the claim requires step A. If the claimed invention requires both the first and second conditions to occur, then the broadest reasonable interpretation of the claim requires both steps A and B.
The broadest reasonable interpretation of a system (or apparatus or product) claim having structure that performs a function, which only needs to occur if a condition precedent is met, requires structure for performing the function should the condition occur. The system claim interpretation differs from a method claim interpretation because the claimed structure must be present in the system regardless of whether the condition is met and the function is actually performed.
Claim 10 recites a step of generating information regarding the learning data based on histogramming of the data label, only performed if the type of learning data is supervised. Thus, this limitation is found to be contingent, and is consequently interpreted as not being a required component of the claimed method under broadest reasonable interpretation. To become a required limitation of the method, it must be rewritten as a positively recited element.
Claims 11-15 recite further steps of generating information regarding the learning data, only performed if the type of learning data is unsupervised. Thus, these limitations are found to be contingent, and are consequently interpreted as not being required components of the claimed method under broadest reasonable interpretation. To become required limitations of the method, they must be rewritten as positively recited elements.
Specification
The disclosure is objected to because of the following informalities:
[0176]: “the server performs normalization for the received and the calculated” is improper grammar.
Appropriate correction is required.
Claim Objections
Claim 11 is objected to because of the following informalities: “transmitting, to the base station, the information acquired by the information acquired by histogramming the learning data” is improper grammar. Appropriate correction is required.
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 1-17 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.
Claim 1 recites the limitation “receiving, from a base station, a first downlink signal for requesting information regarding learning data for the federated learning, which is used by the one terminal" in its first limitation. There is insufficient antecedent basis “the one terminal” for this limitation in the claim. Thus, the scope of the claim is rendered indefinite. This deficiency is present in substantially similar independent claims 16-17, and inherited by all dependent claims. “the one terminal” is interpreted as one of the plurality of terminals declared in the preamble.
Additionally, claim 1 recites “receiving, from the base station, a second downlink signal including parameter information regarding a parameter related to a configuration for performing the federated learning, which is determined based on the learning data” in its fourth limitation. It’s unclear whether “which is determined based on the learning data” is meant to apply to the second downlink signal, the parameter information, the configuration for performing the federated learning, or some combination thereof. Thus, the scope of the claim is rendered indefinite. This deficiency is present in substantially similar independent claims 16-17, and inherited by all dependent claims. This limitation is interpreted as stating that the second downlink signal and / or the parameter information within it is determined based on the learning data.
Claim 11 recites “the data label” in its preamble. There’s a lack of antecedent for this term. Thus, the scope of the claim is rendered indefinite. This deficiency is inherited by dependent claims 12-15.
Additionally, claim 11 recites “transmitting, to the base station, the information acquired by the information acquired by histogramming the learning data”. It’s unclear which information “the information” is meant to refer to, be it the learning data, histogram information, or something else. Thus, the scope of the claim is rendered indefinite. This deficiency is inherited by dependent claims 12-15. “the information” is interpreted as referring to information acquired by histogramming the learning data.
Additionally, claim 11 recites “wherein when the type of learning data is unsupervised learning data in which the data label is not assigned to the learning data” in its preamble, as well as “receiving, from the base station, label information for assigning the data label for the learning data” in its fourth limitation. These limitations seem to contradict one another, being directed to a system that both does and does not assign data labels to the learning data. Thus, the scope of the claim is rendered indefinite. This deficiency is inherited by dependent claims 12-15. Given the preamble stating this is a form of “unsupervised learning”, and most claim limitations being directed toward clustering, a known type of unlabeled data organization, the Examiner is assuming that limitation 4, “receiving, from the base station, label information for assigning the data label for the learning data”, was written in error, and finds it to have no patentable weight.
Claim 13 recites “the number of at least one or more clusters is determined based on the number of clusters”. It’s unclear whether the “number of at least one or more clusters” is meant to be synonymous with the “number of clusters”. If the terms are not synonymous, there’s a lack of antecedent for “the number of clusters”. Thus, the scope of the claim is rendered indefinite. This deficiency is inherited by dependent claims 14-15. “the number of at least one or more clusters” and “the number of clusters” are interpreted as being synonymous.
Claim 14 recites “wherein the number of at least one or more clusters is equal to the number of clusters generated for the global data obtained based on the learning data of each of the plurality of terminals”. There’s a lack of antecedent basis for “clusters generated for the global data”. Thus, the scope of the claim is rendered indefinite. This deficiency is inherited by dependent claim 15.
Claims 14 recites “the global data” There is lack of antecedent basis for this term. Thus, the scope of the claim is rendered indefinite. This deficiency is inherited by dependent claim 15.
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claims 1, 10, and 17 are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Zhu (Federated Learning on Non-IID Data: A Survey, published 6/12/2021, arXiv:2106.06843v1).
Regarding claim 1, Zhu discloses [a] method for performing, by a plurality of terminals, federated learning in a wireless communication system, the method comprising:
receiving, from a base station, a first downlink signal for requesting information regarding learning data for the federated learning, which is used by the one terminal; transmitting, to the base station, the information regarding the learning data based on a type of learning data:
“Local fine-tuning, as the most classic and powerful personalization method, aims to fine-tune the local models after receiving the global model from the server (base station) using local data” (Zhu, page 15, paragraph 7)
PNG
media_image1.png
440
927
media_image1.png
Greyscale
(Zhu, page 3, Algorithm 1). A server (base station) sends a global update (learning data) to a plurality of clients (terminals). Each client updates its model using the global update, and sends a local update (information regarding learning data) back to the server.
the information regarding the learning data being information related to the distribution of the learning data used by the one terminal:
“That is, local models having the same initial parameters will convergence to different models because of the heterogeneity in local data distributions” (Zhu, page 1, paragraph 4)
“The training data on each client in FL heavily depends on the usage of particular local devices, and therefore, the data distribution of connected clients (terminal[s]) may be totally different with each other. This phenomenon is known as Non-IID [95], which may cause severe model divergence” (Zhu, page 6, paragraph 2)
receiving, from the base station, a second downlink signal including parameter information regarding a parameter related to a configuration for performing the federated learning, which is determined based on the learning data; and performing the federated learning based on the information regarding the parameter:
PNG
media_image1.png
440
927
media_image1.png
Greyscale
”An illustrative example of data partition in horizontal federated learning” (Zhu, page 3, Algorithm 1). A second downlink signal is set at 2nd round t = 2, containing parameter information in the form of a global model update, determined based on the learning data, as discussed above.
Zhu relates to federated learning based on global-local divergence and is analogous to the claimed invention.
Regarding claim 10, the rejection of claim 1 is incorporated. Zhu further discloses a method, wherein when the type of learning data is supervised learning data in which a data label is assigned to the learning data, the information regarding the learning data is generated based on histogramming of the data label:
“Decision trees are typical non-parametric models and the gradient boosting decision tree (GBDT) [63] becomes very popular at present in solving both regression and classification (supervised learning) problems because of its good performance. Among gradient boosting tree models, xgboost [16] is the most powerful one, which has already been applied in both horizontal FL [149] and vertical FL [135, 19]. The following modifications should be made when the FedAvg algorithm is applied to xgboost:
1. Set binning points (thresholds) for all data features on all connected clients and the server
2. Each client computes the local histograms for all binning points and send all histograms to the server
3. The server sums up the received local histograms, and calculates the corresponding impurity values for all binning points. Then, the server will find the best split binning points (information regarding the learning data) and send them back to the clients” (Zhu, page 11, paragraph 1)
Regarding claim 17, Zhu discloses [a] method for performing, by a base station, federated learning with a plurality of terminals in a wireless communication system, the method which the base station performs with one terminal of the plurality of terminals, comprising:
transmitting, to the one terminal, a first downlink signal for requesting information regarding learning data for the federated learning, which is used by the one terminal; receiving, from the one terminal, the information regarding the learning data based on a type of learning data:
“Local fine-tuning, as the most classic and powerful personalization method, aims to fine-tune the local models after receiving the global model from the server (base station) using local data” (Zhu, page 15, paragraph 7)
PNG
media_image1.png
440
927
media_image1.png
Greyscale
(Zhu, page 3, Algorithm 1). A server (base station) sends a global update (learning data) to a plurality of clients (terminals). Each client updates its model using the global update, and sends a local update (information regarding learning data) back to the server.
the information regarding the learning data being information related to the distribution of the learning data used by the one terminal:
“That is, local models having the same initial parameters will convergence to different models because of the heterogeneity in local data distributions” (Zhu, page 1, paragraph 4)
“The training data on each client in FL heavily depends on the usage of particular local devices, and therefore, the data distribution of connected clients (terminal[s]) may be totally different with each other. This phenomenon is known as Non-IID [95], which may cause severe model divergence” (Zhu, page 6, paragraph 2)
transmitting, to the one terminal, a second downlink signal including parameter information regarding a parameter related to a configuration for performing the federated learning, which is determined based on the learning data; and performing the federated learning based on the information regarding the parameter:
PNG
media_image1.png
440
927
media_image1.png
Greyscale
”An illustrative example of data partition in horizontal federated learning” (Zhu, page 3, Algorithm 1). A second downlink signal is set at 2nd round t = 2, containing parameter information in the form of a global model update, determined based on the learning data, as discussed above.
Zhu relates to federated learning based on global-local divergence and is analogous to the claimed invention.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
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.
Claim 2 is rejected under 35 U.S.C. 103 as being unpatentable over Zhu (Federated Learning on Non-IID Data: A Survey, published 6/12/2021, arXiv:2106.06843v1) in view of Kamp (Efficient Decentralized Deep Learning by Dynamic Model Averaging, published 11/13/2018, arXiv:1807.03210v2).
Regarding claim 2, the rejection of claim 1 is incorporated. Zhu further discloses a method, wherein the parameter information includes transmission period information regarding a transmission period of a local parameter of the one terminal and grouping information regarding whether terminal grouping is performed for the plurality of terminals:
PNG
media_image1.png
440
927
media_image1.png
Greyscale
(Zhu, page 3, Algorithm 1). Each client receives a global update from the global server, aggregating information from clients across the group connected to the same server / aggregator.
While the aforementioned references fail to disclose the further limitations of the claim, Kamp discloses a method, wherein the parameter information includes transmission period information regarding a transmission period of a local parameter of the one terminal:
“Using this, we define the dynamic averaging operator that allows to omit synchronization in cases where the divergence of a model configuration is low” (Kamp, page 5, paragraph 1)
“An operator adhering to this definition does not generally put all nodes into sync (albeit we still refer to it as synchronization operator). In particular it allows to leave all models untouched as long as the divergence remains below
∆
or to only average a subset of models in order to satisfy the divergence constraint” (Kamp, page 5, paragraph 1)
“The dynamic averaging protocol D = (
φ
,
σ
∆
,
b
) synchronizes the local learners using the dynamic averaging operator
σ
∆
,
b
. This operator only communicates (transmission period of local parameter[s]) when the model divergence exceeds a divergence threshold
∆
. In order to decide when to communicate locally, at round
t
∈
N
, each local learner
i
∈
[
m
]
monitors the local condition
|
|
f
t
i
-
r
|
|
2
for a reference model
r
∈
F
[36] (transmission period information) that is common among all learners. Thus, the first choice for the reference model is the average model from the last synchronization step.” (Kamp, page 5, paragraph 4)
Examiner’s note: The time at which a global update happens (local updates are aggregated and redistributed to local models) is dependent on how long local models adhere to the divergence threshold, based on a reference model passed to them.
Kamp relates to federal learning with variable synchronicity and is analogous to the claimed invention. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the primary reference to vary the rate of global updates in a client cluster based on similarity metrics, as disclosed by Kamp. Kamp’s method achieves a high predictive performance with less data transmission needed than contemporary methods, and is adaptive to concept drifts due to federated performed with non-iid data. See Kamp, page 14, paragraph 2 & page 22, paragraph 1.
Claims 3 and 5-8 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu (Federated Learning on Non-IID Data: A Survey, published 6/12/2021, arXiv:2106.06843v1) in view of Kamp (Efficient Decentralized Deep Learning by Dynamic Model Averaging, published 11/13/2018, arXiv:1807.03210v2), and further in view of Wang (Local Averaging Helps: Hierarchical Federated Learning and Convergence Analysis, published 3/4/2021, arXiv:2010.12998v2).
Regarding claim 3, Kamp further discloses a method, wherein the transmission period information and the grouping information are determined based on distances calculated based on (i) the distribution of learning data for each of the plurality of terminals, and (ii) the distribution of global data obtained based on the learning data for each of the plurality of terminals:
“We consider a decentralized learning setting with
m
∈
N
local learners (plurality of terminals) , where each learner
i
∈
[
m
]
runs the same learning algorithm
φ
:
F
×
2
X
×
2
Y
→
F
that trains a local model
f
i
from a model space F using local samples (distribution[s] of learning data) from an input space X and output space Y” (Kamp, page 3, paragraph 5).
“In each round
t
∈
N
, local learners use a synchronization operator
σ
:
F
m
→
F
m
that transfers the current set of local models, called the current model configuration
f
t
=
{
f
t
1
,
…
,
f
t
m
}
into a single stronger global model” (Kamp, page 4, paragraph 2)
“A simple measure to quantify the effect of synchronizations is given by the divergence (mean distance) of the current model configuration, i.e.,
PNG
media_image2.png
96
317
media_image2.png
Greyscale
” (Kamp, page 5, paragraph 1)
“The dynamic averaging protocol D = (
φ
,
σ
∆
,
b
) synchronizes the local learners using the dynamic averaging operator
σ
∆
,
b
. This operator only communicates when the model divergence (mean distance) exceeds a divergence threshold
∆
. In order to decide when to communicate locally (transmission period information), at round
t
∈
N
, each local learner
i
∈
[
m
]
(plurality of terminals) monitors the local condition
|
|
f
t
i
-
r
|
|
2
for a reference model
r
∈
F
[36] (global data model) that is common among all learners” (Kamp, page 5, paragraph 4)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the existing combination to vary the rate of global updates in a client cluster based on similarity metrics, as disclosed by Kamp. Kamp’s method achieves a high predictive performance with less data transmission needed than contemporary methods, and is adaptive to concept drifts due to federated performed with non-iid data. See Kamp, page 14, paragraph 2 & page 22, paragraph 1.
Wang discloses a method, wherein the transmission period information and the grouping information are determined based on distances calculated based on (i) the distribution of learning data for each of the plurality of terminals, and (ii) the distribution of global data obtained based on the learning data for each of the plurality of terminals:
“In our two-level hierarchical FL system (see Figure 1), all clients are partitioned into N groups” (Wang, page 3, paragraph 6)
“In this paper, we assume the following …
PNG
media_image3.png
118
947
media_image3.png
Greyscale
” (Wang, page 5, paragraph 1)
“the global divergence (variance of distance) corresponds to the relationship between the gradients of intra-group objectives (defined on clients’ (plurality of terminals) data within each group) (distribution[s] of learning data) and the gradient of the overall objective (defined on all clients’ data) (distribution of global data). A larger global divergence means a more heterogeneous data distribution between groups. Similarly, the local divergence i characterizes the data heterogeneity of clients within each group i” (Wang, page 5, paragraph 4)
“Different grouping strategies are considered to best utilize the divergence (variance of distance) measures to reduce communication costs while accelerating learning convergence” (Wang, page 13, paragraph 1)
Wang relates to grouping local clients based on local-global divergence metrics and is analogous to the claimed invention. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the existing combination to form aggregation groups of clients based on divergence between local and global distributions, as disclosed by Wang. Doing so would select a grouping that reduces communication costs while accelerating learning convergence. See Wang, page 13, paragraph 1.
Regarding claim 5, the rejection of claim 3 is incorporated. Kamp further discloses a method, wherein a transmission period value included in the transmission period information is determined based on a mean value of the distances: “A simple measure to quantify the effect of synchronizations is given by the divergence (mean distance) of the current model configuration, i.e.,
PNG
media_image2.png
96
317
media_image2.png
Greyscale
” (Kamp, page 5, paragraph 1)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the existing combination to vary the rate of global updates in a client cluster based on mean divergence of local and global model distributions, as disclosed by Kamp. Kamp’s method achieves a high predictive performance with less data transmission needed than contemporary methods, and is adaptive to concept drifts due to federated performed with non-iid data. See Kamp, page 14, paragraph 2 & page 22, paragraph 1.
Regarding claim 6, the rejection of claim 5 is incorporated. Kamp further discloses a method, wherein the transmission period value is determined in proportion to a size of the mean value of the distances: “The dynamic averaging protocol D = (
φ
,
σ
∆
,
b
) synchronizes the local learners using the dynamic averaging operator
σ
∆
,
b
. This operator only communicates when the model divergence (mean distance) exceeds a divergence threshold
∆
. In order to decide when to communicate locally (transmission period information), at round
t
∈
N
, each local learner
i
∈
[
m
]
(plurality of terminals) monitors the local condition
|
|
f
t
i
-
r
|
|
2
for a reference model
r
∈
F
[36] (global data model) that is common among all learners” (Kamp, page 5, paragraph 4)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the existing combination to vary the rate of global updates in a client cluster based on mean divergence of local and global model distributions, as disclosed by Kamp. Kamp’s method achieves a high predictive performance with less data transmission needed than contemporary methods, and is adaptive to concept drifts due to federated performed with non-iid data. See Kamp, page 14, paragraph 2 & page 22, paragraph 1.
Regarding claim 7, the rejection of claim 3 is incorporated. Wang further discloses a method, wherein whether the terminal grouping being performed included in the grouping information is determined based on a variance value of the distances: “In this paper, we assume the following …
PNG
media_image3.png
118
947
media_image3.png
Greyscale
” (Wang, page 5, paragraph 1). As one of ordinary skill in the art would know, Wang’s global divergence calculation is a variance value of the distances.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the existing combination to form aggregation groups of clients based on divergence (in the form of variances) between local and global distributions, as disclosed by Wang. Doing so would select a grouping that reduces communication costs while accelerating learning convergence. See Wang, page 13, paragraph 1.
Regarding claim 8, the rejection of claim 7 is incorporated. Wang further discloses a method, wherein the terminal grouping is performed in a scheme in which the overall distribution of learning data of terminals grouped into one group is similar to the distribution of the global data: “If the knowledge of data distribution (global distribution) of all clients is available, we can adjust the grouping strategy to achieve a smaller global divergence. For example, we can make the data from each group distributed like a subset sampled uniformly from the union of all clients’ datasets, which is referred to as “group-IID”. In this way, we can increase G, and hence, reduce the global communication cost without affecting the convergence performance of HF-SGD. 6 Extension to Multi-level” (Wang, page 8, paragraph 5)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the existing combination to form aggregation groups of clients based on divergence between local and global distributions, as disclosed by Wang. Doing so would select a grouping that reduces communication costs while accelerating learning convergence. See Wang, page 13, paragraph 1.
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Zhu (Federated Learning on Non-IID Data: A Survey, published 6/12/2021, arXiv:2106.06843v1) in view of Kamp (Efficient Decentralized Deep Learning by Dynamic Model Averaging, published 11/13/2018, arXiv:1807.03210v2), and further in view of Wang (Local Averaging Helps: Hierarchical Federated Learning and Convergence Analysis, published 3/4/2021, arXiv:2010.12998v2), and Desai (TRANSFER LEARNING WITHOUT LOCAL DATA EXPORT IN MULTI-NODE MACHINE LEARNING, published 10/24/2019, US 2019/0325350 A1).
Regarding claim 4, the rejection of claim 3 is incorporated. Kamp further discloses a method, wherein each of the distances is a difference value between (i) a normalized value of the distribution of the learning data of each of the plurality of terminals, and (ii) a normalized value of the distribution of the global data:
“We consider a decentralized learning setting with
m
∈
N
local learners … that trains a local model
f
i
… We assume a streaming setting, where in each round
t
∈
N
” (Kamp, page 3, paragraph 5)
“A simple measure to quantify the effect of synchronizations is given by the divergence (mean distance) of the current model configuration, i.e.,
PNG
media_image2.png
96
317
media_image2.png
Greyscale
” (Kamp, page 5, paragraph 1).
“
f
t
-
=
σ
∆
,
b
(
f
t
)
-
” (Kamp, page 5, paragraph 2).
f
-
is the globally aggregated model, formed from averaging local models results.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the existing combination to vary the rate of global updates in a client cluster based on divergence of local and global model distributions, as disclosed by Kamp. Kamp’s method achieves a high predictive performance with less data transmission needed than contemporary methods, and is adaptive to concept drifts due to federated performed with non-iid data. See Kamp, page 14, paragraph 2 & page 22, paragraph 1.
While Kamp fails to disclose the further limitations of the claim, Desai discloses (i) a normalized value of the distribution of the learning data of each of the plurality of terminals, and (ii) a normalized value of the distribution of the global data: “The application distributes the normalized model parameters to the nodes (plurality of local terminals) in the cluster (block 622). The application optionally adjusts the base model (global model) based on the normalized model parameters” (Desai, [0130])
Desai relates to normalizing model values in distributed systems and is analogous to the claimed invention. The existing combination teaches a method of calculating divergence by measuring the difference between local and global models. Desai teaches a method of normalizing local and global model parameters. It would have been obvious to one of ordinary skill in the art to combine the existing combination and Desai by normalizing the model values before calculating divergences. This would achieve the predictable result of storing and performing operations on smaller, normalized model values in memory, with the existing combination’s method and Desai’s normalization performing the same together as they did separately. (MPEP 2143 I. (A) Combining prior art elements according to known methods to yield predictable results).
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Zhu (Federated Learning on Non-IID Data: A Survey, published 6/12/2021, arXiv:2106.06843v1) in view of Kamp (Efficient Decentralized Deep Learning by Dynamic Model Averaging, published 11/13/2018, arXiv:1807.03210v2), and further in view of Wang (Local Averaging Helps: Hierarchical Federated Learning and Convergence Analysis, published 3/4/2021, arXiv:2010.12998v2), and Takasaki (PARALLEL CROSS VALIDATION IN COLLABORATIVE MACHINE LEARNING, filed 4/27/2021, US 20220343219 A1).
Regarding claim 9, the rejection of claim 8 is incorporated. While the aforementioned references fail to disclose the further limitations of the claim, Takasaki discloses a method, wherein the terminal grouping for the plurality of terminals is performed when the variance value of the distances is equal to or larger than a specific value: “The computer-implemented method includes grouping, by a server, local models on respective ones of local devices (terminals) into groups … The computer-implemented method further includes selecting, by the server, groups whose variances of the accuracies are not below a predetermined variance threshold” (Takasaki, [0007])
Takasaki relates to grouping devices based on variance in a distributed learning system and is analogous to the claimed invention. The existing combination teaches a method of grouping devices in a federated learning system based on variance. Takasaki teaches a method of grouping devices by comparing device variance to a threshold. It would have been obvious to one of ordinary skill in the art to combine the existing combination and Takasaki by determining groups in the existing combination’s system by comparing individual devices to a variance threshold. This would achieve the predictable result of grouping high-variance devices together, with the existing combination’s system and Takasaki’s variance thresholds performing the same together as they did separately. (MPEP 2143 I. (A) Combining prior art elements according to known methods to yield predictable results).
Claims 11-15 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu (Federated Learning on Non-IID Data: A Survey, published 6/12/2021, arXiv:2106.06843v1) in view of Kamp (Efficient Decentralized Deep Learning by Dynamic Model Averaging, published 11/13/2018, arXiv:1807.03210v2), and further in view of Hammouda (Hierarchically Distributed Peer-to-Peer Document Clustering and Cluster Summarization, published May 2009, IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 21, NO. 5, pp. 681-697).
Regarding claim 11, the rejection of claim 1 is incorporated. While the aforementioned references fail to disclose the further limitations of the claim, Hammouda discloses a method, wherein when the type of learning data is unsupervised learning data in which the data label is not assigned to the learning data transmitting the information regarding the learning data further includes:
generating at least one or more clusters based on clustering data constituting the learning data: “HP2PC is a hierarchically distributed P2P architecture for scalable distributed clustering of horizontally partitioned data” (Hammouda, page 683, right column, paragraph 5)
mapping the data constituting the learning data to a centroid of each of the at least one or more clusters: “The HP2PC algorithm is a distributed iterative clustering process. It is a centroid-based clustering algorithm, where a set of cluster centroids is generated to describe the clustering solution. In HP2PC, each neighborhood converges to a set of centroids that describe the data set in that neighborhood.” (Hammouda, page 686, left column, paragraph 2)
transmitting, to the base station, centroid information for each of at least one or more clusters:
“Each neighborhood has a supernode (base station). Communication between neighborhoods is achieved through their respective supernodes” (Hammouda, page 683, right column, paragraph 7)
“The notion of a neighborhood accompanied by a supernode (base station) can be applied recursively to construct a multilevel overlay hierarchy of peers; i.e., a group of supernodes can form a higher level neighborhood, which can communicate with other neighborhoods on the same level of the hierarchy through their respective (higher level) supernodes. This type of hierarchy is illustrated in Fig. 2” (Hammouda, page 683, right column, paragraph 8)
PNG
media_image4.png
337
748
media_image4.png
Greyscale
(Hammouda, Figure 2)
“Once a neighborhood converges to a set of centroids, those centroids are acquired by the supernode (base station) of that neighborhood.” (Hammouda, page 686, left column, paragraph 2)
receiving, from the base station, label information for assigning the data label for the learning data: This limitation is in direct contradiction with the preamble of claim 11, and is thus found to have no patentable weight. See rejections under 35 U.S.C. 112(b) for more information.
transmitting, to the base station, the information acquired by the information acquired by histogramming the learning data:
“the tightness of a cluster,
φ
(
c
k
)
, calculated as the skew of its histogram
H
k
, is
PNG
media_image5.png
87
570
media_image5.png
Greyscale
” (Hammouda, page 686, right column, paragraph 3)
“The final set of centroids for each iteration is calculated from all peer centroids. Unlike K-means (or P2P K-means), those final centroids are weighed by the skew and size of clusters at individual peers. The weight of a cluster
c
k
at peer
p
j
is defined as
PNG
media_image6.png
44
235
media_image6.png
Greyscale
” (Hammouda, page 687, right column, paragraph 1)
Examiner’s note: Histograms are computed for clusters, which are then used to compute cluster skews. Cluster skews are used to ultimately determine final centroids, which are passed to the base station / supernode on the next level of the hierarchy.
Hammouda relates to unsupervised federated clustering and is analogous to the claimed invention. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the primary reference to perform federated clustering with Hammouda’s method. Zhu discloses a generic federated learning system, broadly applicable to a wide variety of tasks. Hammouda discloses a particular usage of such a system, namely for organization of unlabeled data. Hammouda’s method in particular achieves comparable quality to contemporary methods with a significant speedup by comparison. It’s scalable with respect to network size, and is compatible both with standard peer to peer and hierarchical distributed networks. And, compared to non-federated clustering systems, Hammouda’s system maintains data privacy much better than a centralized system. See Hammouda, page 697, right column, paragraphs 1-2.
Regarding claim 12, the rejection of claim 11 is incorporated. Hammouda further discloses a method, further comprising: receiving, from the base station, information on the number of clusters generated based on the clustering by the one terminal:
“Once a neighborhood converges to a set of centroids, those centroids are acquired by the supernode (base station) of that neighborhood.” (Hammouda, page 686, left column, paragraph 2). Base stations receive information about the clusters generated by their child nodes, including terminal nodes at the lowest level of the tree.
“Peer hierarchy formation is bottom-up, so the lowest level of the hierarchy is h = 0. The supernodes of level 0 neighborhoods form the overlay network at level h = 1. Recursively, at level h = 2 are the supernodes of level 1 neighborhoods (groups of level 1 supernodes).” (Hammouda, page 685, left column, paragraph 4). Supernodes of one level in the hierarchy pass their clusters up to their own supernodes.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the primary reference to perform federated clustering with Hammouda’s method. Zhu discloses a generic federated learning system, broadly applicable to a wide variety of tasks. Hammouda discloses a particular usage of such a system, namely for organization of unlabeled data. Hammouda’s method in particular achieves comparable quality to contemporary methods with a significant speedup by comparison. It’s scalable with respect to network size, and is compatible both with standard peer to peer and hierarchical distributed networks. And, compared to non-federated clustering systems, Hammouda’s system maintains data privacy much better than a centralized system. See Hammouda, page 697, right column, paragraphs 1-2.
Regarding claim 13, the rejection of claim 12 is incorporated. Hammouda further discloses a method, wherein the number of at least one or more clusters is determined based on the number of clusters: “For each iteration, each node computes a new set of K (number of clusters) centroids by averaging all neighborhood centroids” (Hammouda, page 688, right column, paragraph 3)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the primary reference to perform federated clustering with Hammouda’s method. Zhu discloses a generic federated learning system, broadly applicable to a wide variety of tasks. Hammouda discloses a particular usage of such a system, namely for organization of unlabeled data. Hammouda’s method in particular achieves comparable quality to contemporary methods with a significant speedup by comparison. It’s scalable with respect to network size, and is compatible both with standard peer to peer and hierarchical distributed networks. And, compared to non-federated clustering systems, Hammouda’s system maintains data privacy much better than a centralized system. See Hammouda, page 697, right column, paragraphs 1-2.
Regarding claim 14, the rejection of claim 13 is incorporated. Hammouda further discloses a method, wherein the number of at least one or more clusters is equal to the number of clusters generated for the global data obtained based on the learning data of each of the plurality of terminals
“A neighborhood at level h consists of a set of peers, each having a set of K centroids. To merge those clusters, the centroids are collected and clustered at the supernode of this neighborhood, using K-means clustering” (Hammouda, page 688, left column, paragraph 2)
“For levels above 0, each neighborhood is responsible for metaclustering a set of KSQ centroids into K centroids using K-means. Then, for each neighborhood at level h, the required computation is
PNG
media_image7.png
68
267
media_image7.png
Greyscale
” (Hammouda, page 688, right column, paragraph 6)
Examiner’s note: Every node, be it a terminal node or an aggregator supernode, calculates a final set of K clusters / centroids.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the primary reference to perform federated clustering with Hammouda’s method. Zhu discloses a generic federated learning system, broadly applicable to a wide variety of tasks. Hammouda discloses a particular usage of such a system, namely for organization of unlabeled data. Hammouda’s method in particular achieves comparable quality to contemporary methods with a significant speedup by comparison. It’s scalable with respect to network size, and is compatible both with standard peer to peer and hierarchical distributed networks. And, compared to non-federated clustering systems, Hammouda’s system maintains data privacy much better than a centralized system. See Hammouda, page 697, right column, paragraphs 1-2.
Regarding claim 15, the rejection of claim 14 is incorporated. Hammouda further discloses a method, wherein the cluster generated for the global data is generated based on clustering for centroids of the clusters generated by the plurality of terminals, respectively:
“For levels above 0, each neighborhood is responsible for metaclustering a set of KSQ centroids into K centroids using K-means. Then, for each neighborhood at level h, the required computation is
PNG
media_image7.png
68
267
media_image7.png
Greyscale
” (Hammouda, page 688, right column, paragraph 6). Each supernode (global to its neighborhood of child nodes) calculates new centroids based on the centroids of its children, including terminal children at the lowest level of the tree.“Peer hierarchy formation is bottom-up, so the lowest level of the hierarchy is h = 0. The supernodes of level 0 neighborhoods form the overlay network at level h = 1. Recursively, at level h = 2 are the supernodes of level 1 neighborhoods (groups of level 1 supernodes). The root supernode is at level H, the height of the hierarchy; i.e., there exists exactly one
p
(
H
)
in the system.” (Hammouda, page 685, left column, paragraph 4)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the primary reference to perform federated clustering with Hammouda’s method. Zhu discloses a generic federated learning system, broadly applicable to a wide variety of tasks. Hammouda discloses a particular usage of such a system, namely for organization of unlabeled data. Hammouda’s method in particular achieves comparable quality to contemporary methods with a significant speedup by comparison. It’s scalable with respect to network size, and is compatible both with standard peer to peer and hierarchical distributed networks. And, compared to non-federated clustering systems, Hammouda’s system maintains data privacy much better than a centralized system. See Hammouda, page 697, right column, paragraphs 1-2.
Claim 16 is rejected under 35 U.S.C. 103 as being unpatentable over Zhu (Federated Learning on Non-IID Data: A Survey, published 6/12/2021, arXiv:2106.06843v1) in view of Kim (METHOD AND DEVICE FOR PERFORMING FEDERATED LEARNING IN WIRELESS COMMUNICATION SYSTEM, filed 5/21/2021, US 20240223407 A1).
Regarding claim 16, Zhu discloses operations includ[ing]:
receiving, from a base station, a first downlink signal for requesting information regarding learning data for the federated learning, which is used by the one terminal; transmitting, to the base station, the information regarding the learning data based on a type of learning data:
“Local fine-tuning, as the most classic and powerful personalization method, aims to fine-tune the local models after receiving the global model from the server (base station) using local data” (Zhu, page 15, paragraph 7)
PNG
media_image1.png
440
927
media_image1.png
Greyscale
(Zhu, page 3, Algorithm 1). A server (base station) sends a global update (learning data) to a plurality of clients (terminals). Each client updates its model using the global update, and sends a local update (information regarding learning data) back to the server.
the information regarding the learning data being information related to the distribution of the learning data used by the one terminal:
“That is, local models having the same initial parameters will convergence to different models because of the heterogeneity in local data distributions” (Zhu, page 1, paragraph 4)
“The training data on each client in FL heavily depends on the usage of particular local devices, and therefore, the data distribution of connected clients (terminal[s]) may be totally different with each other. This phenomenon is known as Non-IID [95], which may cause severe model divergence” (Zhu, page 6, paragraph 2)
receiving, from the base station, a second downlink signal including parameter information regarding a parameter related to a configuration for performing the federated learning, which is determined based on the learning data; and performing the federated learning based on the information regarding the parameter:
PNG
media_image1.png
440
927
media_image1.png
Greyscale
”An illustrative example of data partition in horizontal federated learning” (Zhu, page 3, Algorithm 1). A second downlink signal is set at 2nd round t = 2, containing parameter information in the form of a global model update, determined based on the learning data, as discussed above.
Zhu relates to federated learning based on global-local divergence and is analogous to the claimed invention.
While Zhu fails to disclose the further limitations of the claim, Kim discloses
A terminal for performing federated learning with a plurality of terminals in a wireless communication system, the terminal comprising: a transmitter for transmitting a radio signal; a receiver for receiving the radio signal; at least one processor:
“The present disclosure relates to a method of performing federated learning, and more particularly to a method of performing, by a plurality of user equipments [sic] (UEs), federated learning in a wireless communication system and a device therefor” (Kim, [0001])
“The wireless devices (plurality of terminals) and the BSs/the wireless devices may transmit/receive radio signals to/from each other through the wireless communication/connections” (Kim, [0347])
at least one computer memory operably connectable to the at least one processor, and storing instructions of performing operations when executed by the at least one processor: “The present disclosure provides one user equipment (UE) of a plurality of UEs performing a federated learning in a wireless communication system, the one UE comprising a transmitter configured to transmit a radio signal, a receiver configured to receive the radio signal, at least one processor, and at least one computer memory operably connectable to the at least one processor, wherein the at least one computer memory is configured to store instructions that allow the at least one processor to perform operations based on being executed by the at least one processor” (Kim, [0024])
Kim relates to wireless federated learning and is analogous to the claimed invention. Zhu teaches a method of performing federated learning. The claimed invention improves upon this method by storing it in the form of instructions on wireless computer hardware. Kim teaches wireless computer hardware, applicable to Zhu’s federated learning system. A person of ordinary skill in the art would have recognized that storing Zhu’s method as computer instructions on Kim’s hardware would lead to the predictable result of the method being executable by a wireless computing system, and would improve the known device by allowing it to be performed with real data (MPEP 2143 I. (D) Applying a known technique to a known device (method, or product) ready for improvement to yield predictable results).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
Ferreira (EDGE DATA DISTRIBUTION CLIQUES, filed 5/28/2021, US 20220383184 A1) discloses a method of grouping edge devices in a federated learning network based on a divergence threshold
Gardner (ROBUST MODEL PERFORMANCE ACROSS DISPARATE SUB-GROUPS WITHIN A SAME GROUP, filed 9/30/2020, US 20230222377 A1) discloses a method of incorporating divergence minimization into a loss function for model group selection
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Aaron P Gormley whose telephone number is (571)272-1372. The examiner can normally be reached Monday - Friday 12:00 PM - 8:00 PM EST.
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, Michelle T Bechtold can be reached at (571) 431-0762. 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.
/AG/Examiner, Art Unit 2148 /MICHELLE T BECHTOLD/Supervisory Patent Examiner, Art Unit 2148
Prosecution Insights