DETAILED ACTION
This final office action is responsive to the amendment filed on July 8, 2026. Claims 1-18 are pending. Claims 1, 7, and 13 are independent.
The objection to claim 14 is withdrawn in light of applicant’s amendment to the claim.
The claim rejections under 35 USC §101 of claims 3-4, 9-10, and 15-16 are withdrawn in light of applicant’s amendment and arguments, see section Response to Arguments below.
Claim rejections under 35 USC §103 of claims 1-18 are maintained. See sections Claim Rejections – 35 USC § 103 and Response to Arguments below.
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 .
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.
Claims 1-18 are rejected under 35 U.S.C. 103 as being unpatentable over Wink et al. (An Approach for Peer-to-Peer Federated Learning), hereinafter Wink, in view of Pothula et al. (US20210174257), hereinafter Pothula.
Regarding claim 1, Wink teaches:
A method of creating models with a plurality of entities, wherein the plurality of entities is at least three (Wink, page 152, column 1, paragraph 3: “Figure 1 shows an abstract overview of the iterative peer-to-peer model training process with three participants Peer1, Peer2 and Peer3 as an example.” – Wherein peers are analogous to entities.)
training, by the computing device, a first model for each of the plurality of entities based on the received data, wherein each first model is trained using recursive model stacking and each first model outputs a corresponding score; (Wink, page 152, column 1, paragraph 3: “In the first training round, each peer (including the initiator Peer1) trains its own respective copy of v0 locally. Local training steps are conducted by using local sample datasets only, which the given participant wants to keep secret.” And column 2, paragraph 3: “To partition a given weight value
w
of a given neural network model into
n
parts, the participant
P
e
e
r
i
(as owner of
w
) first generates
n
positive random numbers
r
n
1
,
⋯
,
r
n
n
. These values will be divided by their sum, such that
p
r
n
i
=
r
n
i
∑
k
=
1
n
r
n
k
,
i
∈
1
,
⋯
,
n
.
The resulting percentage distribution is then used to generate normalized
w
1
,
⋯
,
w
n
partitions of
w
, where
w
i
=
w
*
p
r
n
i
,
i
∈
1
,
⋯
,
n
.
For example, assume the weight
w
=
19
has been partitioned in three parts. The owner of
w
will first generate three random numbers, such as
r
n
1
=
9
,
r
n
2
=
18
,
and
r
n
3
=
30
leading to the percentage factors of
p
r
n
1
=
0.16
,
p
r
n
2
=
0.32
,
and
p
r
n
3
=
0.52
, respectively. The resulting parts of
w
are thus
w
1
=
3
,
w
2
=
6
,
and
w
3
=
10
(see also in Table I).” – v0 being trained locally is analogous to the first model for each of the entities being trained using recursive model stacking, see, e.g., fig. 1. The percentage factors for each peer is analogous to the score that’s being output by each model.)
training, by the computing device, a revised model for each of the plurality of entities based on the output of all of the first models, (Wink, page 152, column 2, paragraph 1: “In the next phase, participants collaboratively conduct Secure Average Computation (SAC) steps to determine the average of corresponding model weight values in v0 , v0 and v0 . During that process they do not exchange any model weight values among each other. As a result of SAC, all participants will possess identical updated model instances v1 at their local premises.” – The secure average computation is analogous to the revised model being based on the output of all of the first models, as seen in Fig. 1, after exchanging the secure average computation the model is further trained for each peer.) wherein the output of all of the first models comprises, for each record of a dataset, a plurality of scores generated by applying each of the first models to the dataset, and wherein the plurality of scores are used as input features for training the revised model using the recursive model stacking; (Wink, page 151, column 1, final paragraph: “A node runs the forward-pass using local training data. The resulting tensor is sent to the server, which uses it as input to train its model layers.” – The weights and secure average computation is contained in the tensor that is input into the model to train it, which is thus analogous to the scores being applied as input features for training the revised model, each peer has its own record of the dataset as the local training data.)
training, by the computing device, a second model for each of the plurality of entities based on the corresponding output of all of the revised models, (Wink, page 152, column 2, paragraph 1: “If it is not yet the case, they may start another training round and further re-train their local v1 instances while using again their respective local training datasets as input. It may take in total r rounds until the model vr will converge, i.e., hit the desired quality threshold, so that the training process can terminate.” – The local training repeating indicates that after each training round the secure average computation is performed and then the model at the peer is further trained. Thus, v1 is analogous to the revised model and v2 would be analogous to the second model.) wherein the output of all of the revised models comprises, for each record of the dataset, a plurality of scores generated by applying each of the revised models to the dataset, and wherein the plurality of scores of the output of all of the revised models are used as input features for training the second model using the recursive model stacking, (Wink, page 151, column 1, final paragraph: “A node runs the forward-pass using local training data. The resulting tensor is sent to the server, which uses it as input to train its model layers.” And algorithm III.2 – As noted above, the weights and secure average computation is contained in the tensor that is input into the model to train it, which is thus analogous to the scores being applied as input features for training the revised model, each peer has its own record of the dataset as the local training data.)
wherein the first models, the revised models, and the second models form a recursive hierarchical sequence of models in which each successive model is trained using outputs from a prior model as input features; and (Wink, Algorithm III.2 and Fig. 1 – The tensors being applied to the models at each iteration is analogous to the outputs from the prior model being used as input features, Fig. 1 shows the recursive hierarchical sequence of models.)
outputting one score, by the computing device, based on the output of all of the second models. (Wink, page 152, column 2, paragraph 1: “If it is not yet the case, they may start another training round and further re-train their local v1 instances while using again their respective local training datasets as input. It may take in total r rounds until the model vr will converge, i.e., hit the desired quality threshold, so that the training process can terminate.” – The retraining occurring again indicates that after v2, e.g., the second models, the process then conducts the secure average computation again and, therefore, outputs the one score based on the output of all the second models using the same process as noted above.)
Wink does not explicitly teach:
receiving, by a computing device, data from a plurality of entities
However, Pothula teaches:
receiving, by a computing device, data from a plurality of entities, wherein the plurality of entities is at least three; (Pothula, paragraph 0026: “In some embodiments, datasets 14 may be obtained from various entities (e.g. enterprises) and may include entity events (e.g., those involving the entity, like where the entity acted upon or was acted upon by a system being modeled, like a customer) and non-event attributes related to those events (e.g., exogenous events, like the weather, news events, holidays, etc.).”)
Pothula is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Wink, which already teaches training machine learning models for a plurality of entities but does not explicitly teach receiving data from the plurality of entities, to include the teachings of Pothula which does teach receiving data from the plurality of entities where the system may control the amount of data that is shared in order to provide the appropriate level of security and protections. (Pothula, paragraph 0027)
Regarding claim 2, Wink and Pothula teach the computerized-method of claim 1, as cited above.
Wink further teaches:
training a third model for each of the plurality of entities based on the one score that is output for all models, and wherein the third model for each of the plurality of entities is trained using the recursive model stacking. (Wink, page 152, column 2, paragraph 1: “If it is not yet the case, they may start another training round and further re-train their local v1 instances while using again their respective local training datasets as input. It may take in total r rounds until the model vr will converge, i.e., hit the desired quality threshold, so that the training process can terminate.” And Fig. 1 – The process repeating a total of r rounds indicates that after conducting the secure average computation using the output of v2, e.g., the second model, then v3 would be trained which is analogous to the third model for each peer based on the score output that is taught by Wink in claim 1 above.)
Regarding claim 3, Wink and Pothula teach the computerized-method of claim 1, as cited above.
Wink further teaches:
the one score is based on an aggregation statistic of a score output from all of the second models for each of the plurality of entities, (Wink, page 153, column 1, last paragraph: “In the next step, Peeri determines the subtotal psi by summing up wii and the wji values that it received from the other n-1 peers. The locally computed subtotal psi will than be sent to all other participants. Upon receiving the subtotals from all peers, Peeri is able to compute S as the sum of subtotals and can calculate the average Avg=S/n finally.” – The average is analogous to an aggregation statistic of a score output.) wherein the score output from each of the second models reflects a reduction in bias attributable to the corresponding entity's individual dataset achieved through multiple iterations of the recursive model stacking, and wherein the aggregation statistic is calculated across the score output from all of the second models to produce the one score independent of any single entity's data distribution. (Wink, page 151, column 1, paragraph 1: “Federated optimization (FO) is a method to create and continuously improve a neural-network model based on distributed training data….Unbalanced data: The amount of training data may vary. significantly between nodes. Some nodes might have a multiple of the data available compared to others.” – The process that is taught by Wink above is used to optimize and improve the models and specifically addresses unbalanced data which would is indicative of reducing bias in a data distribution.)
Regarding claim 4, Wink and Pothula teach the computerized-method of claim 3, as cited above.
Wink further teaches:
the aggregation statistic is an average. (Wink, page 153, column 2, first paragraph: “Upon receiving the subtotals from all peers, Peeri is able to compute S as the sum of subtotals and can calculate the average Avg=S/n finally.”)
Regarding claim 5, Wink and Pothula teach the computerized-method of claim 2, as cited above.
Wink further teaches:
training of the first model, the second model, and the third model is performed in a pipeline. (Wink, Fig. 1 – The figure shows that the training flows from the first model to the second to the third, etc., therefore the training is performed in a pipeline.)
Regarding claim 6, Wink and Pothula teach the computerized-method of claim 1, as cited above.
Wink further teaches:
the recursive model stacking is recursive federated learning. (Wink, page 152, column 1, paragraph 2: “Thus, the most basic requirement for a federated peer-to-peer ML system is that collaboratively trained models should have at least the same quality (prediction accuracy) as if a central server had been involved in the training.” And Fig. 1 – The federated peer-to-peer ML system is analogous to recursive federated learning. Fig. 1 shows that the learning method is recursive model stacking.)
Regarding claim 7, claim 7 has all the same limitations of claim 1 which are taught by Wink and Pothula, see claim 1 above.
Wink does not explicitly teach:
at least one processor configured to:
However, Pothula further teaches:
at least one processor configured to: (Pothula, paragraph 0008: “Some aspects include a system, including one or more processors, and memory storing instructions that, when executed by the processors, cause the processors to effectuate operations of the above-mentioned process.”)
Regarding claim 8, Wink and Pothula teach the system of claim 7, as cited above. Claim 8 additionally has the same limitations of claim 2 which are taught by Wink and Pothula – see claim 2 above.
Regarding claim 9, Wink and Pothula teach the system of claim 7, as cited above. Claim 9 additionally has the same limitations of claim 3 which are taught by Wink and Pothula – see claim 3 above.
Regarding claim 10, Wink and Pothula teach the system of claim 9, as cited above. Claim 10 additionally has the same limitations of claim 4 which are taught by Wink and Pothula – see claim 4 above.
Regarding claim 11, Wink and Pothula teach the system of claim 8, as cited above. Claim 11 additionally has the same limitations of claim 5 which are taught by Wink and Pothula – see claim 5 above.
Regarding claim 12, Wink and Pothula teach the system of claim 7, as cited above. Claim 12 additionally has the same limitations of claim 6 which are taught by Wink and Pothula – see claim 6 above.
Regarding claim 13, claim 13 has all the same limitations of claim 1 which are taught by Wink and Pothula, see claim 1 above.
Wink does not explicitly teach:
A non-transitory computer program product comprising instruction instructions which, when the program is executed by a processor cause the computer to:
However, Pothula further teaches:
A non-transitory computer program product comprising instruction instructions which, when the program is executed by a processor cause the computer to: (Pothula, paragraph 0115: “Instructions or other program code to provide the functionality described herein may be stored on a tangible, non-transitory computer readable media.”)
Regarding claim 14, Wink and Pothula teach the non-transitory computer program product of claim 13, as cited above. Claim 14 additionally has the same limitations of claim 2 which are taught by Wink and Pothula – see claim 2 above.
Regarding claim 15, Wink and Pothula teach the non-transitory computer program product of claim 13, as cited above. Claim 15 additionally has the same limitations of claim 3 which are taught by Wink and Pothula – see claim 3 above.
Regarding claim 16, Wink and Pothula teach the non-transitory computer program product of claim 15, as cited above. Claim 16 additionally has the same limitations of claim 4 which are taught by Wink and Pothula – see claim 4 above.
Regarding claim 17, Wink and Pothula teach the non-transitory computer program product of claim 14, as cited above. Claim 17 additionally has the same limitations of claim 5 which are taught by Wink and Pothula – see claim 5 above.
Regarding claim 18, Wink and Pothula teach the non-transitory computer program product of claim 13, as cited above. Claim 18 additionally has the same limitations of claim 6 which are taught by Wink and Pothula – see claim 6 above.
Response to Arguments
Applicant’s arguments regarding claim rejections under 35 USC §101 have been fully considered and are persuasive. In particular, page 10 of applicant’s remarks discusses how the amendment to claim 3 clarifies the technological nature of the claimed invention. The amendment to claims 3, 9, and 15 integrate the abstract idea previously noted into a practical application. Therefore, claim rejections under 35 USC §101 are withdrawn.
Applicant's arguments regarding claim rejections under 35 USC §103 have been fully considered but they are not persuasive. Applicant argues on page 16 that the framework of Wink and Pothula is different than that of the claimed invention, examiner disagrees. Wink discusses the issue of fraud detection in the Introduction section and discusses successive models that are built from outputs of prior models. Each peer trains its own version of the model using its dataset before the secure average computation, e.g. the scores of all of the previous models. This is similar to the present invention which has each entity training its own model version before using the scores from all the models to train successive models. The use of the secure average computation to train the model at each of the peers is analogous to the scores of each entity’s model being used to train the next model in the pipeline. Each peer is able to determine when the training is complete, see Wink, page 152, column 2, paragraph 1. Thus, each peer would converge into a local model that is optimized for its own data rather than simply having a single global shared model.
Applicant further argues on page 18 that the claimed architecture is different as it has “score-level feature generation and multi-layer model construction.” However, the weight tensors of Wink are being output from the model and used as inputs to determine the update weights of each model at the peer. See, e.g., Algorithm III.2. Additionally, neither the claims nor the specification discuss the alleged multi-layer model construction.
The arguments against Wink’s federated learning structure is also unpersuasive as claim 6 specifically states that “the recursive model stacking is recursive federated learning.”
The arguments against Pothula are moot as Pothula was not relied upon to teach any of the amended limitations, which are rather taught by Wink.
Therefore, claim rejections under 35 USC §103 are maintained.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Silberman (US11893520)
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JACQUELINE MEYER whose telephone number is (703)756-5676. The examiner can normally be reached M-F 8:00 am - 4:30 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, Tamara Kyle can be reached at 571-272-4241. 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.
/J.C.M./ Examiner, Art Unit 2144
/TAMARA T KYLE/ Supervisory Patent Examiner, Art Unit 2144
