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 .
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.
Claim(s) 1, 4-5, 7-8, 11, 14-15, 17-19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Jianwu Tang et al. (hereinafter Tang) (“FedRAD: Heterogeneous Federated Learning via Relational Adaptive Distillation,” 2023) in view of Dianbo Sui et al. (hereinafter Sui) (“FedED: Federated Learning via Ensemble Distillation for Medical Relation Extraction,” 2020).
Regarding claim 1, Tang teaches;
A system, comprising: a processor that executes computer-executable components stored in a non-transitory computer-readable memory, wherein the computer-executable components comprise (Tang teaches a computer implemented process): a gathering component that aggregates at least one trained machine learning model ([pg. 5] each selected client performs their own local model training … After the local training phase … The server collects the selected local models and aggregates them as a new round global model) … and a training component that updates parameters of a central machine learning model based on an optimization of a federated loss function ([pg. 9] global model … updated by minimizing … [Equation] (3) [pg. 7, equation 3] Lglobal = αLCE + (1−α) [DKL(pn(x) || pg(x))+ LRKD]) computed from the at least one trained machine learning model ([pg. 9] pn(x): prediction of … local model … DKL(pn(x) || pg(x)), wherein the federated loss function ([pg. 7, equation 3] Lglobal = αLCE + (1−α) [DKL(pn(x) || pg(x))+ LRKD]) comprises a federated error ([pg. 7] DKL is the Kullback–Leibler (KL) Divergence, which is used to measure the similarity of the distributions between the global model output and the local model output) and a base loss ([pg. 7] The cross-entropy loss is calculated using the predicted value and the ground truth labels … LCE).
Tang fails to explicitly teach but Sui teaches;
trained machine learning model in connection with a respective healthcare institution ([pg. 4] local model is only trained on its own data … Local models are deployed in local platforms, such as hospitals);
OBVIOUSNESS TO COMBINE SUI WITH TANG:
Tang teaches a federated learning framework having a plurality of trained local models, while Sui teaches a federated learning framework where each local model can be deployed at a respective healthcare institution and trained using private medical text data maintained at the institution. Sui further explains that to “[pg. 2] prevent private information leakage, we leverage federated learning without sharing raw privacy sensitive medical texts.” Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement Tang’s federated learning framework at such healthcare institutions to obtain the known privacy benefits of federated learning for medical text data.
Regarding claim 4, Tang teaches;
further comprising a network component that shares the central machine learning model after updating the parameters based on the optimization of the federated loss function with the at least one trained machine learning model ([pg. 9-10] In each round t, we first sample a subset based on fraction C of clients, then each client receives the global model … global model … updated by minimizing [Equation] (3) … The above process continues cyclically until the communication round reaches the preset T [pg. 7, equation 3] Lglobal = αLCE + (1−α) [DKL(pn(x) || pg(x))+ LRKD]) … [pg. 9] pn(x): prediction of … local model)
Regarding claim 5, Tang fails to explicitly teach but Sui teaches;
wherein training datasets utilized to train the at least one trained machine learning model ([pg. 4] Local models are deployed in local platforms, such as hospitals, and can access their respective private local data … The local model in each medical platform is trained on its own private data) comprises protected health information (PHI) reports, patient records, clinical notes, lab results, imaging reports, medication histories, or relevant clinical textual information from the respective healthcare institution ([pg. 5] 2010 i2b2/VA challenge dataset … consists of discharge-summary and progress notes of the patients, and is manually annotated by medical practitioners [pg. 6] For 2010 i2b2/VA challenge dataset … The training data is randomly shuffled and then partitioned into K local platforms each receiving 1/K of the training data).
OBVIOUSNESS: Using the same reasoning from claim 1.
Regarding claim 7, Tang teaches;
wherein the at least one trained machine learning model and the central machine learning model receives an input (the dataset of local model n is Dn = {x, y},y ∈ {1,2,...,K} … predictions p(x) = pg(x) for the global model and p(x) = pn(x) for the local model) … , wherein the at least one trained machine learning model outputs one or more encodings of a prediction based on the input ([pg. 7] predictions ... p(x) = pn(x) for the local model) … and the central machine learning model outputs an encoding of another prediction based on the input ([pg. 7] predictions p(x) = pg(x) for the global model) … , and wherein the federated error is an error between the encoding generated from the central machine learning model and the one or more encodings generated from the at least one trained machine learning model ([pg. 7-8] DKL is the Kullback–Leibler (KL) Divergence, which is used to measure the similarity of the distributions between the global model output and the local model output … DKL (pn(x) || pg(x))).
Tang fails to explicitly teach but Sui teaches;
the at least one trained machine learning model and the central machine learning model receives an input sequence ([pg. 3] Given the prepared sequence ŝ as input, the output of BERT is expressed as H [pg. 4] the local model and the central model, which share the same network structure)
OBVIOUSNESS: Using the same reasoning from claim 1, it would have been obvious to implement Tang’s federated learning framework for Sui’s medical text data, such that Tang’s input x comprises Sui’s prepared input sequence ŝ, to obtain the privacy benefits taught by Sui while processing the private medical text data maintained at the respective healthcare institutions.
Regarding claim 8, Tang teaches;
wherein the base loss is an error between the encoding generated from the central machine learning model ([pg. 7] predictions p(x) = pg(x) for the global model) and a ground-truth ([pg. 7] The cross-entropy loss is calculated using the predicted value and the ground truth labels … LCE = ∑x∈Dn −ylog[p(x)]).
Regarding claims 11, 14-15, 17-18
Claims 11, 14-15, 17-18 are substantially similar to claims 1, 4-5, 7-8, 11, respectively, and are rejected using the same reasoning.
Regarding claim 19
Claim 19 is substantially similar to claim 1 and is rejected using the same reasoning.
Claim(s) 2, 9, 12, 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Tang (“FedRAD: Heterogeneous Federated Learning via Relational Adaptive Distillation,” 2023) in view of Sui (“FedED: Federated Learning via Ensemble Distillation for Medical Relation Extraction,” 2020) as applied to claim 1 above, further in view of Le Sun et al. (hereinafter Sun) (CN 117788949 A, 2024-03-29).
Regarding claim 2, Tang and Sui fail to explicitly teach but Sun teaches;
a scaling component that weights the federated error (Lkl , see below) based on volume of training data ([pg. 5] | X i | represents the sample size of the i-th client) of each of the at least one trained machine learning model ([Abstract] obtaining an updated local classification model by minimizing the classification loss … counting the data size of the local data set and uploading the data size to the cloud server … local classification model of each client).
PNG
media_image1.png
237
783
media_image1.png
Greyscale
NOTE: | X i | represents the sample size of the local classification model training data of the i-th client, and | X i | is used in the weighting factor | Xi | / (∑nj=1 | Xj |) that controls the respective client contributions in Pc (z, y) and Ps (z, y), which are used to determine Lkl.
OBVIOUSNESS TO COMBINE SUN:
Sun teaches weighting federated error based on each client’s local data amount and explains that “[pg. 6] clients with unbalanced data may result in poor performance.” Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to modify the federated knowledge distillation training of Tang as modified by Sui to weight the respective local model contributions used in determining the federated error, as taught by Sun, to account for client data imbalance, predictably improving the resulting model’s performance and accuracy.
Regarding claim 9, Tang and Sui fail to explicitly teach but Sun teaches;
wherein weights of the federated error ([0054-0055] | Xi | / (∑nj=1 | Xj |)) are directly proportional to the volume of training data ([pg. 5] | X i | represents the sample size of the i-th client).
OBVIOUSNESS: Using the same reasoning from claim 2.
Regarding claim 12
Claim 12 is substantially similar to claim 2 and is rejected using the same reasoning.
Regarding claim 20
Claim 20 is substantially similar to claim 2 and is rejected using the same reasoning.
Claim(s) 3, 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Tang (“FedRAD: Heterogeneous Federated Learning via Relational Adaptive Distillation,” 2023) in view of Sui (“FedED: Federated Learning via Ensemble Distillation for Medical Relation Extraction,” 2020), further in view of Sun (CN 117788949 A, 2024-03-29) as applied to claim 2 above, further in view of Grzegorz Andrzej Toporek et al. (hereinafter Toporek) (US 20230316141 A1, 2023-10-05) further in view of Michael Zhang et al. (hereinafter Zhang) (“Personalized Federated Learning with First Order Model Optimization,” 2021).
Regarding claim 3, Tang and Sui fail to explicitly teach but Sun teaches;
wherein the scaling component weights, … the federated error
Using the same reasoning and obviousness rational from claim 2
Tang, Sui, and Sun fail to teach but Toporek teaches;
weights, … based on geographic location of the respective healthcare institution ([0057] local participant 150 as a medical institute [0058] local model parameter updates weighted according to … institute criteria based on organizational characteristics of the local participants [0059] attributes in the institute criteria … include, … a geographical location of the local participant).
OBVIOUSNESS TO COMBINE TOPOREK:
Sui teaches local models associated with respective healthcare institutes, and Sun teaches applying specific weights to control the contribution of respective local models used in determining the federated error of the combined system of Tang, Sui, and Sun. Toporek teaches weighting respective local model contributions based on geographic locations of the respective healthcare institutions. Toporek further teaches that “[0056] sizes, types, and qualities of training datasets respective to participants can be biased according to … respective status of public health and demography in locations and/or regions in which the respective participants operate” and that “[0056] Such statistical heterogeneity and biases in datasets respective to the participants may result in degraded performance in a resulting federated model.” Accordingly, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to use geographic location as a weighting criterion for the respective healthcare institution contributions used in determining the federated error to account for geographically related heterogeneity and bias, thereby mitigating degradation in the resulting federated model caused by such heterogeneity and bias.
Tang, Sui, Sun, and Toporek fail to teach but Zhang teaches;
weights, according to user preference (Note: The client’s specified interests constitute user preference - [pg. 1] client-specific target tasks and distributions … for each client we compute a weighted combination of the available models to best align with that client’s interests [pg. 2] client-specified local objectives)
OBVIOUSNESS TO COMBINE ZHANG:
Sui teaches local models associated with respective healthcare institutes, Sun teaches applying specific weights to control the contribution of respective local models used in determining the federated error, and Toporek teaches weighting respective local model contributions based on geographic locations of the respective healthcare institutions. Zhang teaches computing client specific model weights based on a client specific target task/distribution to “[pg. 1] best align with that client’s interests.” Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to apply Zhang’s client specific weighting approach to Toporek’s geographic weighting criterion such that the relative weights associated with the respective healthcare institution locations are configured according to a user/client specific target or interest (i.e., preference), thereby tailoring the resulting federated model toward the geographic location of interest to the user.
Regarding claim 13
Claim 13 is substantially similar to claim 3 and is rejected using the same reasoning.
Claim(s) 6, 16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Tang (“FedRAD: Heterogeneous Federated Learning via Relational Adaptive Distillation,” 2023) in view of Sui (“FedED: Federated Learning via Ensemble Distillation for Medical Relation Extraction,” 2020) as applied to claim 1 above, further in view of Aref Jafari et al. (hereinafter Jafari) (US 20210383238 A1, 2021-12-09).
Regarding claim 6, Tang teaches;
wherein the training component updates the parameters of the central machine learning model, via [computed gradients], based on the optimization of the federated loss function (Note: ∇ is the gradient operator - [pg. 9] global model ... updated by minimizing [Lglobal] ... wtg,i ← wtg,i − η∇Lglobal).
Tang and Sui fail to explicitly teach but Jafari teaches;
backpropagation ([0005] The KD loss function is used in addition to the standard loss function for backpropagation during training of the student DNN model)
OBVIOUSNESS TO COMBINE JAFARI:
Tang already teaches updating the central/global model parameters according to the gradient of the federated loss function, which includes a standard cross entropy loss and knowledge distillation losses. Jafari teaches that a knowledge distillation loss is used with a standard loss for backpropagation during training of a student DNN. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement Tang’s disclosed gradient based optimization using backpropagation, as taught by Jafari, as a known technique for minimizing the federated loss function and updating the central/global model parameters.
Regarding claim 16
Claim 16 is substantially similar to claim 6 and is rejected using the same reasoning.
Claim(s) 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Tang (“FedRAD: Heterogeneous Federated Learning via Relational Adaptive Distillation,” 2023) in view of Sui (“FedED: Federated Learning via Ensemble Distillation for Medical Relation Extraction,” 2020) as applied to claim 1 above, further in view of Peyman Passban et al. (hereinafter Passban) (US 20220076136 A1, 2022-03-10).
Regarding claim 10, Tang teaches;
a sum of the [federated error] ([pg. 7] DKL) and base loss ([pg. 7] LCE) is used to compute the federated loss function ([pg. 7, equation 3] Lglobal = αLCE + (1−α) [DKL(pn(x) || pg(x))+ LRKD])
Tang fails to teach but Sui teaches
respective healthcare institutions ([pg. 4] Local models are deployed in local platforms, such as hospitals)
OBVIOUSNESS: Using the same reasoning from claim 1
Tang and Sui fail to explicitly teach but Passban teaches;
wherein the training component computes a weighted average of the federated error if the [teacher/local models] comprise more than one [teacher/local model] (Note: Passban computes a weighted average of the respective KD errors LKD associated with the individual teachers, wherein the respective teacher weights ∈pq sum to one [see below])
[0144] the attention weights of all K teachers …, where ∑i=1K ∈iq = 1) are computed. … [0142] a weighted KD loss based on a loss computed for each of the teacher's 204 is computed as per Equation (12):
PNG
media_image2.png
171
748
media_image2.png
Greyscale
OBVIOUSNESS TO COMBINE PASSBAN:
Tang teaches a sum of a federated error (knowledge distillation) and a base loss being used to compute a federated loss, and Sui teaches a plurality of local/teacher models each associated with a respective healthcare institution. Passban teaches an attention-based knowledge distillation approach with multiple teachers, where weights for respective teachers are determined and used to compute a weighted average of the individual KD losses. Passban further teaches optimizing attention weights “[0116] in an attempt to achieve the optimal knowledge distillation.” Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to modify the knowledge distillation training of Tang as modified by Sui to weight the respective distillation errors associated with the plurality of local/teacher models according to Passban’s attention-based weighting scheme, predictably allowing the respective local/teacher contributions to be weighted in an attempt to achieve optimal knowledge distillation, as taught by Passban.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Matthew Alan Cady whose telephone number is (571) 272-7229. The examiner can normally be reached Monday - Friday, 7:30 am - 5:00 pm ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Cesar Paula can be reached on (571)272-4128. 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.
/MATTHEW ALAN CADY/ Examiner, Art Unit 2145
/CESAR B PAULA/ Supervisory Patent Examiner, Art Unit 2145
Prosecution Insights