DETAILED ACTION
Claims 1-11 are presented for examination.
This office action is in response to submission of application on 10-NOVEMBER-2025.
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 14-JANUARY-2022 is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Response to Amendment
The amendment filed 10-NOVEMBER-2025 in response to the non-final office action mailed 11-AUGUST-2025 has been entered. Claims 1-11 remain pending in the application.
With regards to the non-final office action’s rejections under 103, the amendments to the claims necessitated a new consideration of the art. After this consideration, the examiner respectfully disagrees with the applicant’s arguments that the art referenced in the previous office action does not teach the amendment claim limitations. However, in order to address the amendment limitations, previously presented art Dempsey has been integrated into claim 1. A new 103 rejection over the prior art has been provided.
The applicant argues that neither Hu nor Buesser teaches classifying, by a client, training data into two categories comprising a forgettable sample and an unforgettable sample; generating, by the client, a learning mini-batch by selectively including forgettable and unforgettable samples so that a ratio of forgettable to unforgettable samples in the mini-batch corresponds to a preset ratio. However, examiner believes that Hu along with newly incorporated into claim 1’s rejection Dempsey disclose all aspects of these limitations:
The applicant argues that while Hu does teach separation of samples into two categories, it does not specifically teach that those two categories are a forgettable sample and an unforgettable sample. The examiner agrees that Hu does not fully teach the amended limitation. However, Dempsey resolves the deficiencies of Hu as Dempsey teaches providing a first item of training data and comparing it to a second item of training data already used to train a model in order to determine what data the model should be retrained with in order to produce a consistent model (Page 4, lines 15-26). Determining if the first data should be used to retrain the model or not would be analogous to classifying it as a forgettable or unforgettable sample as it determines if it is key to the model’s performance or not.
Therefore, Hu in view of Dempsey would teach classifying, by a client, training data into two categories comprising a forgettable sample and an unforgettable sample. This would have provided the advantage of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5)
Furthermore, the applicant argues that Hu nor Buesser does not teach generating, by the client, a learning mini-batch by selectively including forgettable and unforgettable samples so that a ratio of forgettable to unforgettable samples in the mini-batch corresponds to a preset ratio. However, the examiner believes that Dempsey resolves the deficiencies of this art:
The examiner has previously argued that Buesser in the same field of endeavor of machine learning teaches generating a learning mini-batch by selectively including samples such that a ratio of adversarial samples to non-adversarial samples (Paragraph 65) such that a ratio of samples in the batch corresponds with preset ratio, wherein the preset ratio is the initial value of the ratio before its increase. The applicant argues that this ratio is not preset, as it changes over time. However, at minimum, the initial mini-batch of Buesser would have a preset ratio that the other ratios are then modified from. Therefore, Buesser does teach a preset ratio in at least one case.
Buesser does not disclose a forgettable sample and an unforgettable sample. However, as discussed above, this aspect of the limitation has been previously taught by Dempsey. The classification of samples of Dempsey may be applied to the methodology of Buesser, for purposes of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5) as provided by Dempsey’s teachings.
Furthermore, regarding the applicant’s arguments on claim 11 that none of the above art discloses wherein the ratio adjustment mitigates catastrophic forgetting and enhances model performance by repeatedly exposing the forgettable samples during training while maintaining a proportion of the forgettable samples that avoids performance degradation:
As discussed above, Buesser and Dempsey are believed by the examiner to disclose a ratio adjustment by repeatedly exposing the forgettable samples during training while maintaining a proportion of the forgettable samples, as this describes the process of the selective inclusion of forgettable samples across multiple rounds. Furthermore, the ratio of Buesser is for purposes of avoiding performance degradation, as it is to avoid the model producing an error or mistake (Paragraph 65).
Furthermore, regarding mitigation of catastrophic forgetting and enhancing model performance, Buesser enhances model performance through this method through preventing mistakes by machine learning model (Buesser, Paragraph 65). The applicant claims that Dempsey does not teach catastrophic forgetting, as it removes forgettable samples before catastrophic forgetting could occur (Page 4, lines 15-26). However, this would be a form of mitigating catastrophic forgetting.
Regarding the combination of these teachings, the methodology of Buesser that improves performance by repeatedly exposes a category of samples while maintaining a proportion of this category of samples to avoid performance degradation could have the forgettable samples and mitigation of catastrophic forgetting incorporated into it by Dempsey, with forgettable samples as the category of samples that are exposed and maintained, and mitigation of catastrophic forgetting a consequence of the changing number of said samples.
Furthermore, regarding the applicant’s amendments of claims 2-3:
Regarding claim 2, the amendment states that the previous limitation is based on a retraining comparison process, which is disclosed by Buesser:
Buesser teaches retraining a model in order to improve its performance (Paragraph 65) which would be in comparison to the previous model.
Regarding claim 3, the amendment states that its process is based on catastrophic forgetting analysis performed across multiple training rounds:
As previously argued, Dempsey teaches catastrophic forgetting analysis as it removes samples that would lead to catastrophic forgetting (Page 4, lines 15-26). Furthermore, Buesser teaches training across multiple rounds (Paragraph 65). In combination, Buesser and Dempsey would address this limitation as it would be obvious to combine the catastrophic forgetting analysis of Dempsey with Buesser’s training rounds for the purposes of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5).
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.
Claims 1-7 are rejected under 35 U.S.C. 103 as being unpatentable over Hu et al. (Pub. No. US20230106985 A1, filed October 19th 2019, hereinafter Hu) in view of Buesser et al. (Pub. No. US 20210110045 A1, filed October 14th 2019, hereinafter Buesser) further in view of Dempsey et al. (Pub. No. WO 02063558 A2, filed January 31st 2002, hereinafter Dempsey).
Regarding claim 1:
Claim 1 recites:
A processor-implemented federated learning framework local model training method, comprising: classifying, by a client, training data into two categories comprising a forgettable sample and an unforgettable sample; generating, by the client, a learning mini-batch by selectively including forgettable and unforgettable samples so that a ratio of forgettable to unforgettable samples in the mini-batch corresponds to a preset ratio; and training, by the client, a learning model by implementing the mini-batch with the adjusted sample ratio.
Hu discloses a processor-implemented federated learning framework local model training method, comprising: [classifying], by a client, training data into two categories [comprising a forgettable sample and an unforgettable sample];
Hu teaches the division of training samples into two categories, positive and negative (Paragraph 53) which would consist of separating the samples by a client as the data set and its respective ratio are associated with the client itself. Furthermore, Hu teaches a federated learning framework machine-learning model (Paragraph 59).
Hu does not disclose classifying data as a forgettable sample and an unforgettable sample specifically. This aspect of the limitation is taught further below.
Buesser discloses generating, by the client, a learning mini-batch by selectively including [forgettable and unforgettable] samples so that a ratio [of forgettable to unforgettable samples] in mini-batch corresponds to a preset ratio:
Buesser in the same field of endeavor of machine learning teaches generating a learning mini-batch by selectively including samples such that a ratio of adversarial samples to non-adversarial samples (Paragraph 65) such that a ratio of samples in the batch corresponds with preset ratio, wherein the preset ratio is the initial value of the ratio before its increase. While this ratio does change over time, at a minimum the initial mini-batch of Buesser would have a preset ratio that the other ratios are then modified from. Therefore, Buesser does teach a preset ratio in at least one case.
Hu, Buesser, and the present application are all analogous art because they are all in the same field of endeavor of machine learning
Buesser does not disclose a forgettable sample and an unforgettable sample. This aspect of the limitation is taught further below by Dempsey.
Buesser discloses training, by the client, a learning model by implementing the mini-batch with the adjusted sample ratio:
Buesser teaches retraining the machine learning model using the adjusted sample ratio mini-batch (Paragraph 65).
Dempsey discloses classifying data as a forgettable sample and an unforgettable sample:
Dempsey in same field of endeavor of machine learning teaches providing a first item of training data and comparing it to a second item of training data already used to train a model in order to determine what data the model should be retrained with in order to produce a consistent model (Page 4, lines 15-26). Determining if the first data should be used to retrain the model or not would be analogous to classifying it as a forgettable or unforgettable sample as it determines if it is key to the model’s performance or not.
Hu, Buesser, Dempsey, and the present application are all analogous art because they are all in the same field of endeavor of machine learning.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu, the teachings of Buesser, and the teachings of Dempsey. This would have granted the advantage of improving the error of machine learning models when given novel or malicious inputs (Buesser, Paragraph 65) as well as avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5).
Regarding claim 2, which depends upon claim 1:
Claim 2 recites:
The method of claim 1, wherein in the classifying the training data into the two categories, the client is configured to classify the training data into a forgettable sample and an unforgettable sample based on a retraining comparison process.
Hu in view of Buesser further in view of Dempsey discloses claim 1 upon which claim 2 depends. Furthermore, Dempsey discloses wherein in the classifying the training data into the two categories, the client is configured to classify the training data into a forgettable sample and an unforgettable sample:
Dempsey teaches providing a first item of training data and comparing it to a second item of training data already used to train a model in order to determine what data the model should be retrained with in order to produce a consistent model (Page 4, lines 15-26). Determining if the first data should be used to retrain the model or not would be analogous to classifying it as a forgettable or unforgettable sample as it determines if it is key to the model’s performance or not.
Buesser discloses based on a retraining comparison process:
Buesser teaches retraining a model in order to improve its performance (Paragraph 65) which would be in comparison to the previous model.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu in view of Buesser, and the teachings of Dempsey. This would have provided the advantage of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5)
Regarding claim 3, which depends upon claim 2:
Claim 3 recites:
The method of claim 2, wherein the client is configured to classify the training data into the forgettable sample and the unforgettable sample based on catastrophic forgetting analysis performed across multiple training rounds.
Hu in view of Buesser further in view of Dempsey discloses claim 2 upon which claim 3 depends. Furthermore, Dempsey discloses wherein the client is configured to classify the training data into the forgettable sample and the unforgettable sample based on catastrophic forgetting analysis:
Dempsey teaches providing a first item of training data and comparing it to a second item of training data already used to train a model in order to determine what data the model should be retrained with in order to produce a consistent model (Page 4, lines 15-26). The training data would include local training data stored in memory and the determination of which data to use when retraining the model for best consistency would be analogous catastrophic forgetting analysis as catastrophic forgetting refers to inconsistent results across model versions.
Buesser discloses multiple training rounds:
Buesser teaches retraining a model in order to improve its performance (Paragraph 65) which would provide multiple training rounds.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu in view of Buesser, and the teachings of Dempsey. This would have provided the advantage of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5)
Regarding claim 4, which depends upon claim 2:
Claim 4 recites:
The method of claim 2, wherein the client is configured to classify the training data into the forgettable sample and the unforgettable sample by comparing a result obtained by training the learning model with the training data, and a result obtained by retraining, with the training data, the learning model trained with the training data.
Hu in view of Buesser further in view of Dempsey discloses claim 2 upon which claim 4 depends. Furthermore, Dempsey discloses the limitation of claim 4:
Dempsey teaches determining that a sample is conflicting with another if it would produce different output elements than its identical pair. (Page 9, lines 25-30). Furthermore, this conflict arises from the use of old data previously used to train the model and new data used to retrain the model (Page 8, lines 20-25). Therefore, one of the identical pair of samples is the result obtained by training the learning model with the training data, and the other is the result obtained by retraining with the training data, where the two results are compared in order to determine if a sample is conflicting, i.e., forgettable.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu in view of Buesser and the teachings of Dempsey. This would have provided the advantage of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5)
Regarding claim 5, which depends upon claim 4:
Claim 5 recites:
The method of claim 4, wherein the client is configured to classify a sample in which a result obtained by training the learning model with the training data is different from a result obtained by retraining the learning model with the training data as the forgettable sample.
Hu in view of Buesser further in view of Dempsey discloses claim 4 upon which claim 5 depends. Furthermore, Dempsey discloses the limitation of claim 5:
Dempsey in the same field of endeavor of reinforcement learning teaches determining that a sample is conflicting with another if it would produce different output elements than its identical pair. (Page 9, lines 25-30). Furthermore, this conflict arises from the use of old data previously used to train the model and new data used to retrain the model (Page 8, lines 20-25). As discussed previously, this pair is a result of the original trained model and a result of the retrained model that are compared. If they are in conflict, the sample is considered for removal from the knowledge base (Page 18, lines 15-25). This would make it a forgettable sample as it may not be included in a future model.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu in view of Buesser and the teachings of Dempsey. This would have provided the advantage of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5)
Regarding claim 6, which depends upon claim 4:
Claim 6 recites:
The method of claim 4, wherein the client is configured to classify a sample in which a result obtained by retraining the learning model with the training data is an incorrect answer, among samples in which a result obtained by training the learning model with the training data is a correct answer, as the forgettable sample.
Hu in view of Buesser further in view of Dempsey discloses claim 4 upon which claim 6 depends. Furthermore, Dempsey discloses the limitation of claim 6:
Dempsey teaches that results may be erroneous in a retrained model if conflicts are included in the training data (Page 4, lines 1-5). As discussed previously, the conflict is an example of a forgettable sample, and as such determining a conflict through an erroneous classification would be classifying a sample based on an incorrect answer in the retrained learning model.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu in view of Buesser and the teachings of Dempsey. This would have provided the advantage of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5)
Regarding claim 7, which depends upon claim 1:
Claim 7 recites:
The method of claim 1, wherein the client is configured to receive the preset ratio from a server.
Hu in view of Buesser further in view of Dempsey disclose claim 1 upon which claim 7 depends. Furthermore, Buesser discloses the limitation of claim 7:
Buesser teaches a preset ratio (Buesser, Paragraph 65) as discussed above in claim 1. A preset ratio would be an example of a hyperparameter. Hu teaches clients receiving hyperparameters from the server (Hu, Paragraph 110). Therefore in combination, Hu in view of Buesser would teach a client configured to receive the preset ratio from the server.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu and the teachings of Buesser. This would have granted the advantage of improving the error of machine learning models when given novel or malicious inputs (Buesser, Paragraph 65).
Claims 8-11 are rejected under 35 U.S.C. 103 as being unpatentable over Hu in view of Buesser further in view of Dempsey further in view of Injung et al. (Pub. No. KR 20190103090 A, published September 4th 2019, hereinafter Injung).
Regarding claim 8, which depends upon claim 1:
Claim 8 recites:
The method of claim 1, wherein the client is configured to receive information on the learning model from a server before the client classifies the training data into the two categories.
Hu in view of Buesser further in view of Dempsey discloses the method of claim 1 upon which claim 8 depends. However, Injung discloses the limitation of claim 8:
Injung recites: “Hyper-parameters (Hyper Parameters) used to learn the mutual location learning model can be received from the server every round and used” (Federated Learning).
Injung in the same field of endeavor of reinforcement learning teaches receiving hyper-parameters from the server, which would be an example of information on the learning model.
Hu, Buesser, Injung, and the present application would be analogous art because they are all in the same field of endeavor of reinforcement learning.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu in view of Buesser further in view of Dempsey and the teachings of Injung. This would have provided the advantage of multiple clients learning jointly through a common predictive model (Injung, “Using federated learning, since the terminal can manage the predictive data, and learn jointly using a common predictive model, there is no need to store the data in a separate server or cloud” (Federated Learning).)
Regarding claim 9, which depends upon claim 8:
Claim 9 recites:
The method of claim 8, wherein the client is configured to transmit information on the trained learning model to the server after the client trains the learning model.
Hu in view of Buesser further in view of Dempsey further in view of Injung discloses the method of claim 8 upon which claim 9 depends. Furthermore, Injung discloses the limitation of claim 9:
Injung recites: “the weight parameter of the mutual location learning model of the terminal. When the learning is completed, the weight parameter of the mutual location learning model is uploaded to the server” (Federating Learning).
Injung teaches that weight parameters are uploaded to the server when the learning is complete, which is analogous to transmitting information on the trained learning model after the client trains the learning model.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu in view of Buesser further in view of Dempsey and the teachings of Injung. This would have provided the advantage of multiple clients learning jointly through a common predictive model (Injung, “Using federated learning, since the terminal can manage the predictive data, and learn jointly using a common predictive model, there is no need to store the data in a separate server or cloud” (Federated Learning).)
Regarding claim 10, which depends upon claim 9:
Claim 10 recites:
The method of claim 9, wherein the information on the learning model and the information on the trained learning model are weights for the learning model.
Hu in view of Buesser further in view of Dempsey further in view of Injung discloses the method of claim 9 upon which claim 10 depends. Furthermore, Injung discloses the limitation of claim 10:
Injung recites: “the weight parameter of the mutual location learning model of the terminal. When the learning is completed, the weight parameter of the mutual location learning model is uploaded to the server” (Federating Learning).
Injung teaches that weight parameters are uploaded to the server when the learning is complete, so the information transmitted of Injung are weights.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu in view of Buesser and the teachings of Injung. This would have provided the advantage of multiple clients learning jointly through a common predictive model (Injung, “Using federated learning, since the terminal can manage the predictive data, and learn jointly using a common predictive model, there is no need to store the data in a separate server or cloud” (Federated Learning))
Regarding claim 11:
Claim 11 recites:
A processor-implemented method for improving training stability of a local machine learning model in a federated learning framework, the method comprising: receiving, by a processor of a client device, from a server device, an initial set of model parameters and a preset sample ratio; storing, by the processor in a memory, local training data; performing, by the processor, catastrophic forgetting analysis on the local training data, wherein the catastrophic forgetting analysis comprises: training the local model using the local training data to generate first output results, retraining the local model using the same training data to generate second output results, and identifying, by the processor, as forgettable samples, those samples whose output results differ between the first and second output results, and identifying other samples as unforgettable samples; generating, by the processor, a mini-batch for training the local model by adjusting the ratio of forgettable samples to unforgettable samples within the mini-batch according to the preset sample ratio, wherein the ratio adjustment mitigates catastrophic forgetting and enhances model performance; training, by the processor, the local model using the generated mini-batch with the adjusted sample ratio; and transmitting, by the processor, updated model parameters derived from the training to the server device for aggregation into a global model of the federated learning framework, wherein the global model is configured to improve convergence and accuracy across client devices.
Hu in view of Buesser discloses improving training stability of a local machine learning model in a federated learning framework, the method comprising: receiving, by a processor of a client device, from a server device, an initial set of model parameters and a preset sample ratio:
Hu teaches receiving a seed version of a machine learning model meant to represent the model for transmittal to a client in a federated learning framework (Paragraph 11), which would include the model parameters.
Furthermore, Buesser teaches a preset ratio (Buesser, Paragraph 65) as discussed above in claim 1. A preset ratio would be an example of a hyperparameter. Hu teaches clients receiving hyperparameters from the server (Hu, Paragraph 110). Therefore in combination, Hu in view of Buesser would teach a client configured to receive the preset ratio from the server.
Furthermore, Buesser discloses generating, by the processor, a mini-batch for training the local model by adjusting the ratio of [forgettable samples to unforgettable] samples within the mini-batch according to the preset sample ratio, wherein the ratio [adjustment mitigates catastrophic forgetting and] enhances model performance by repeatedly exposing forgettable samples during training while maintaining proportion of the forgettable samples that avoids performance degradation:
Buesser teaches generating a learning mini-batch by adjusting the ratio samples such that a ratio of adversarial samples to non-adversarial samples (Paragraph 65) such that a ratio of samples in the batch corresponds with preset ratio, wherein the preset ratio is the initial value of the ratio before its increase. While this ratio does change over time, at a minimum the initial mini-batch of Buesser would have a preset ratio that the other ratios are then modified from. Therefore, Buesser does teach a preset ratio in at least one case.
Furthermore, Buesser discloses a ratio adjustment by repeatedly exposing the forgettable samples during training while maintaining a proportion of the forgettable samples, as this describes the process of the adjusting of a ratio of samples across multiple rounds. Furthermore, the ratio of Buesser is for purposes of avoiding performance degradation, as it is to avoid the model producing an error or mistake (Paragraph 65).
Furthermore, regarding enhancing model performance, Buesser enhances model performance through this method through preventing mistakes by machine learning model (Buesser, Paragraph 65).
Buesser does not disclose a forgettable sample and an unforgettable sample or mitigation of catastrophic forgetting. This aspect of the limitation is taught further below by Dempsey.
Buesser discloses training, by the processor, the local model using the generated mini-batch with the adjusted sample ratio:
Buesser teaches retraining the machine learning model using the adjusted sample ratio mini-batch (Paragraph 65).
Dempsey discloses storing, by the processor in a memory, local training data; performing, by the processor, catastrophic forgetting analysis on the local training data, wherein the catastrophic forgetting analysis comprises:
Dempsey teaches providing a first item of training data and comparing it to a second item of training data already used to train a model in order to determine what data the model should be retrained with in order to produce a consistent model (Page 4, lines 15-26). The training data would include local training data stored in memory and the determination of which data to use when retraining the model for best consistency would be analogous catastrophic forgetting analysis as catastrophic forgetting refers to inconsistent results across model versions.
Dempsey discloses training the local model using the local training data to generate first output results, retraining the local model using the same training data to generate second output results, and identifying, by the processor, as forgettable samples, those samples whose output results differ between the first and second output results, and identifying other samples as unforgettable samples:
Dempsey teaches determining that a sample is conflicting with another if it would produce different output elements than its identical pair. (Page 9, lines 25-30). Furthermore, this conflict arises from the use of old data previously used to train the model and new data used to retrain the model (Page 8, lines 20-25). Therefore, one of the identical pair of samples is the result obtained by training the learning model with the training data, and the other is the result obtained by retraining with the training data, where the two results are compared in order to determine if a sample is conflicting for the purpose of retraining the model, i.e., forgettable.
Dempsey further discloses the mitigation of catastrophic forgetting:
Dempsey teaches removing forgettable samples before catastrophic forgetting could occur (Page 4, lines 15-26). However, this would be a form of mitigating catastrophic forgetting even if the catastrophic forgetting does not itself occur.
Regarding the combination of these teachings, the methodology of Buesser that improves performance by repeatedly exposes a category of samples while maintaining a proportion of this category of samples to avoid performance degradation could have the forgettable samples and mitigation of catastrophic forgetting incorporated into it by Dempsey, with forgettable samples as the category of samples that are exposed and maintained, and mitigation of catastrophic forgetting a consequence of the changing number of said samples. This would provide the advantages described below.
Injung discloses transmitting, by the processor, updated model parameters derived from the training to the server device for aggregation into a global model of the federated learning framework, wherein the global model is configured to improve convergence and accuracy across client devices:
Injung recites: “the weight parameter of the mutual location learning model of the terminal. When the learning is completed, the weight parameter of the mutual location learning model is uploaded to the server” (Federating Learning).
Injung teaches that weight parameters are uploaded to the server when the learning is complete, which is analogous to transmitting information on the trained learning model after the client trains the learning model which would include updated model parameters derived from the training. Furthermore, in a federated learning system this would be to the server device for aggregation into a global model of the federated learning framework, wherein the global model is configured to improve convergence and accuracy across client devices
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Hu, the teachings of Buesser, the teachings of Dempsey, and the teachings of Injung. This would have provided the advantage of improving the error of machine learning models when given novel or malicious inputs (Buesser, Paragraph 65) as well as avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5) as well as multiple clients learning jointly through a common predictive model (Injung, “Using federated learning, since the terminal can manage the predictive data, and learn jointly using a common predictive model, there is no need to store the data in a separate server or cloud” (Federated Learning)).
Response to Arguments
Applicant’s arguments filed 10-NOVEMBER-2025 have been fully considered, but the examiner believes that not all are fully persuasive.
Regarding the applicant’s remarks on the non-final office action’s 103 rejection of the claims, the applicant argues that Hu, Buesser, Dempsey, and Injung do not teach the amended limitations of these claims. As such, the applicant argues that all claims dependent on the above would additionally not be obvious under 103. However, the examiner believes that Hu, Buesser, Dempsey, and Injung does teach the amended and previous limitations and respectfully requests applicant’s consideration of the following:
The applicant argues that neither Hu nor Buesser teaches classifying, by a client, training data into two categories comprising a forgettable sample and an unforgettable sample; generating, by the client, a learning mini-batch by selectively including forgettable and unforgettable samples so that a ratio of forgettable to unforgettable samples in the mini-batch corresponds to a preset ratio. However, examiner believes that Hu along with newly incorporated into claim 1’s rejection Dempsey disclose all aspects of these limitations:
The applicant argues that while Hu does teach separation of samples into two categories, it does not specifically teach that those two categories are a forgettable sample and an unforgettable sample. The examiner agrees that Hu does not fully teach the amended limitation. However, Dempsey resolves the deficiencies of Hu as Dempsey teaches providing a first item of training data and comparing it to a second item of training data already used to train a model in order to determine what data the model should be retrained with in order to produce a consistent model (Page 4, lines 15-26). Determining if the first data should be used to retrain the model or not would be analogous to classifying it as a forgettable or unforgettable sample as it determines if it is key to the model’s performance or not.
Therefore, Hu in view of Dempsey would teach classifying, by a client, training data into two categories comprising a forgettable sample and an unforgettable sample. This would have provided the advantage of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5)
Furthermore, the applicant argues that Hu nor Buesser does not teach generating, by the client, a learning mini-batch by selectively including forgettable and unforgettable samples so that a ratio of forgettable to unforgettable samples in the mini-batch corresponds to a preset ratio. However, the examiner believes that Dempsey resolves the deficiencies of this art:
The examiner has previously argued that Buesser in the same field of endeavor of machine learning teaches generating a learning mini-batch by selectively including samples such that a ratio of adversarial samples to non-adversarial samples (Paragraph 65) such that a ratio of samples in the batch corresponds with preset ratio, wherein the preset ratio is the initial value of the ratio before its increase. The applicant argues that this ratio is not preset, as it changes over time. However, at minimum, the initial mini-batch of Buesser would have a preset ratio that the other ratios are then modified from. Therefore, Buesser does teach a preset ratio in at least one case.
Buesser does not disclose a forgettable sample and an unforgettable sample. However, as discussed above, this aspect of the limitation has been previously taught by Dempsey. The classification of samples of Dempsey may be applied to the methodology of Buesser, for purposes of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5) as provided by Dempsey’s teachings.
Furthermore, regarding the applicant’s arguments on claim 11 that none of the above art discloses wherein the ratio adjustment mitigates catastrophic forgetting and enhances model performance by repeatedly exposing the forgettable samples during training while maintaining a proportion of the forgettable samples that avoids performance degradation:
As discussed above, Buesser and Dempsey are believed by the examiner to disclose a ratio adjustment by repeatedly exposing the forgettable samples during training while maintaining a proportion of the forgettable samples, as this describes the process of the selective inclusion of forgettable samples across multiple rounds. Furthermore, the ratio of Buesser is for purposes of avoiding performance degradation, as it is to avoid the model producing an error or mistake (Paragraph 65).
Furthermore, regarding mitigation of catastrophic forgetting and enhancing model performance, Buesser enhances model performance through this method through preventing mistakes by machine learning model (Buesser, Paragraph 65). The applicant claims that Dempsey does not teach catastrophic forgetting, as it removes forgettable samples before catastrophic forgetting could occur (Page 4, lines 15-26). However, this would be a form of mitigating catastrophic forgetting.
Regarding the combination of these teachings, the methodology of Buesser that improves performance by repeatedly exposes a category of samples while maintaining a proportion of this category of samples to avoid performance degradation could have the forgettable samples and mitigation of catastrophic forgetting incorporated into it by Dempsey, with forgettable samples as the category of samples that are exposed and maintained, and mitigation of catastrophic forgetting a consequence of the changing number of said samples.
Furthermore, regarding the applicant’s arguments for claims 2-6:
Regarding claim 2, the amendment states that the previous limitation is based on a retraining comparison process, which is disclosed by Buesser:
Buesser teaches retraining a model in order to improve its performance (Paragraph 65) which would be in comparison to the previous model.
Regarding claim 3, the amendment states that its process is based on catastrophic forgetting analysis performed across multiple training rounds:
As previously argued, Dempsey teaches catastrophic forgetting analysis as it removes samples that would lead to catastrophic forgetting (Page 4, lines 15-26). Furthermore, Buesser teaches training across multiple rounds (Paragraph 65). In combination, Buesser and Dempsey would address this limitation as it would be obvious to combine the catastrophic forgetting analysis of Dempsey with Buesser’s training rounds for the purposes of avoiding erroneous results in later version of the model by resolving conflicting results (Dempsey, page 4, lines 1-5).
Regarding claim 4, the applicant claims that Dempsey does not disclose comparing results of training versus retraining. However, the examiner respectfully disagrees and believes that the previous 103 rejection has disclosed this limitation:
Dempsey teaches determining that a sample is conflicting with another if it would produce different output elements than its identical pair. (Page 9, lines 25-30). Furthermore, this conflict arises from the use of old data previously used to train the model and new data used to retrain the model (Page 8, lines 20-25). Therefore, one of the identical pair of samples is the result obtained by training the learning model with the training data, and the other is the result obtained by retraining with the training data, where the two results are compared in order to determine if a sample is conflicting, i.e., forgettable.
Regarding claims and 5 and 6, the applicant claims that Dempsey does not disclose specific criteria for identifying forgettable samples, based on differences in outputs (claim 5) or retraining producing incorrect answers. The examiner respectfully disagrees and believes that the previous 103 rejection has disclosed this limitation:
Dempsey discloses identifying forgettable samples based on differences in outputs as it teaches determining that a sample is conflicting with another if it would produce different output elements than its identical pair. (Page 9, lines 25-30).
Dempsey discloses identifying forgettable samples based on retraining producing incorrect answers as it teaches that results may be erroneous in a retrained model if conflicts are included in the training data (Page 4, lines 1-5). The conflict is an example of a forgettable sample, and as such determining a conflict through an erroneous classification would be classifying a sample based on an incorrect answer in the retrained learning model.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ALEXANDRIA JOSEPHINE MILLER whose telephone number is (703)756-5684. The examiner can normally be reached Monday-Thursday: 7:30 - 5:00 pm, every other Friday 7:30 - 4:00.
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, Mariela Reyes can be reached on (571) 270-1006. 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.
/A.J.M./Examiner, Art Unit 2142
/Mariela Reyes/Supervisory Patent Examiner, Art Unit 2142