CTFR 18/176,068 CTFR 98833 Notice of Pre-AIA or AIA Status 07-03-aia AIA 15-10-aia The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA. Response to Arguments In the previous Office Action issued January 12, 2026 (hereinafter “the previous Office Action”), claims 1-20 were pending. This action is in response to the amendment and remarks filed March 15, 2026. In the amendment, claims 1, 5, 10, 14, 16, 20 were amended, no claims were canceled, and claims 21-23 were added. Thus, claims 1-23 are pending. The rejections of claims 1-20 under 35 U.S.C. § 101, set forth in the previous Office Action, have been withdrawn in view of Applicant’s amendments and remarks. Claim Rejections - 35 USC § 103 07-103 AIA The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action. 07-21-aia AIA Claim s 1, 10, and 20-23 are rejected under 35 U.S.C. 103 as being unpatentable over Kim et al. (“Efficient Two-Stage Model Retraining for Machine Unlearning”), hereinafter Kim, in view of Wang et al. (US 20190347570), hereinafter Wang, and further in view of Fraboni et al. (US 20240086760), hereinafter Fraboni . Regarding Claim 1: Kim discloses: A computer program product for editing a machine learning model to forget data Kim, Abstract, “we propose an efficient machine unlearning architecture to be used for computer vision classification models. Our approach consists of two-stage models, where in the first stage we enables a deep learning model that loses information with contrastive labels in the requested dataset” Section 1.1, “we propose a efficient two-stage model retraining algorithm, which enables a deep learning model to efficiently erase the information of the requested dataset on deep learning model. First, we make the model forget the subset of the training dataset” calculating a loss function to measure a divergence of the reference output and the target output Kim, Section 2, “By using the KD algorithm, we measure of how one probability distribution is different from a second, reference probability distribution to mimic the teacher model” Section 3.2, “During training time, we apply the KL divergence loss with Teacher and Student model… In this training phase, the neutralized model is student model, and the original model is the teacher model [i.e., the K-L divergence loss measures the divergence between the outputs of the Teacher (reference) output and the Student (target) model output]. determining gradients that minimize an error of the loss function Kim, Section 3.2: “the final loss function is combined with soft label knowledge distillation loss 2 and cross entropy loss 3 with bias factor as follows: PNG media_image1.png 111 467 media_image1.png Greyscale , PNG media_image2.png 117 581 media_image2.png Greyscale , PNG media_image3.png 53 506 media_image3.png Greyscale [i.e., the determination of gradients are necessary in minimizing the error in the loss function calculations]. calculating optimized gradients from the determined gradients Kim, Section 3.2, “For training efficiency, we set alpha and beta value as 1.1, 0.9 each in 4 To validate our training efficiency that can merge faster, we stopped using knowledge distillation loss after 10th of epochs and used only cross entropy loss with no bias factor” [i.e., the L TOTAL is used in a training phase which inherently involves an optimizer (such as Adam optimizer) to optimize the gradients over many training iterations]. Section 4: “For preparing an original model, we train the model with Adam optimizer with the setting of learning rate as 0.001” [i.e., an Adam optimizer produces optimized gradients in the experiment]. applying the optimized gradients to update weights in the target model to produce an edited target model Kim, p. 4365, Section 3.2, “We retrain the neutralized model using the remaining data except the requested data…In this training phase, the neutralized model is student model” [i.e., retraining the neutralized model (applying the optimized gradients to update weights in the target model) to obtain a retrained neutralized model (produce an edited target model)]. deploying the edited target model to users as compliant with data privacy requirements that require removal of private data for users that request removal Kim, Section 1, “The General Data Protection Regulation (GDPR) provided regulatory guidance on this data management situation. The GDPR guarantees the ‘individual’s right to be forgotten’ that is individuals can limit the scope of commercial use on their personal information, and entities utilizing personal data must faithfully respond to individual data related requests…For compliance with the regulation, data managers must process the individual data according to individual’s requests.” Kim discloses that individuals can request [users that request removal] for their personal data to be forgotten [removal of private data] in compliance with regulatory guidance [compliant with data privacy requirements]. p. 4365, Section 3.2, “We retrain the neutralized model using the remaining data except the requested data…In this training phase, the neutralized model is student model” p. 4366, Section 4, “Thus, we leverage TTA to demonstrate that our retrained model does not remember the requested data points, and even the similar data points with maintaining the model performance.” p. 4366, Section 5, “we demonstrated that our method successfully retrained the neutralized model efficiently and effectively.” On p. 4365, Kim discloses that the student model [target model] is the neutralized model, and the neutralized model is retrained [the edited target model]. Further on p. 4366, in sections 4 and 5, Kim discloses that the retrained neutralized model’s implementation was able to efficiently and effectively forget the requested data points [deploying the edited target model to users]. Kim does not explicitly disclose: the computer program product comprising a computer readable storage medium having computer readable program code embodied therein that is executable to perform operations, the operations comprising: inputting forget data samples of data samples to forget into a reference model, trained on a non-private data set, to produce reference output inputting the forget data samples to a target model, trained on a total data set comprising the non-private data set and a private data set to produce target output wherein the private data set includes the forget data samples However, in the same field, analogous art Wang teaches: the computer program product comprising a computer readable storage medium having computer readable program code embodied therein that is executable to perform operations, the operations comprising Wang, [0055], “These computer programs (also known as programs, software, software applications or code) include machine instructions for a programmable processor, and can be implemented in a high-level procedural and/or object-oriented programming language, and/or in assembly/machine language. As used herein, the terms ‘machine-readable medium’ and ‘computer-readable medium’ refer to any computer program product, non-transitory computer readable medium, apparatus and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term ‘machine-readable signal’ refers to any signal used to provide machine instructions and/or data to a programmable processor” inputting forget data samples of data samples to forget into a reference model, trained on a non-private data set, to produce reference output Wang, [0027], “The base model generator 130 receives a first dataset 132 of sentence data pairs 133, 133a-n for training a sequence-to-sequence base model 134 until convergence” [i.e., the base model 134 receives the non-private dataset from the base model generator of sentence data pairs for training until convergence (producing a reference model output)]. [0031], “the adapted model 144, p(S i )=p(t i /s i ) and the base model 134, q(S i )=q(t i /s i ) may be used to determine a contrastive score 154 (or quality measure) of a sentence pair S i =(s i , t i ) as follows” [i.e., the base (reference) model produces a probability distribution (reference output) of the data input pair]. [0031], “Still referring to FIG. 1, the score determiner 150 is configured to receive a third dataset 152 with sentence pairs 153, 153a-n, the sequence-to-sequence base model 134, and the sequence-to-sequence adapted model 144 for determining a respective contrastive score” [i.e., the base model acts as a reference model that receives data samples (from a third dataset) analyzed for quality and potential removal (forget data)]. [0039], “After selection, a data pair sorter 340 sorts the selected data pairs 342 based on the respective contrastive score 154 for each selected data pair. A data pair remover 350 then removes, from the data batch 210, a removal ratio of the scored and sorted data pairs with the lowest contrastive scores 154. The removal ratio is equivalent to an inverse of the selection ratio, i.e., 1-r(t). For example, when r(t)=0.5, then 50% of the selected data pairs 342 will be removed from the data batch 210 (the 50% with the lowest contrastive scores 154)” [i.e., data samples with low contrastive scores are forgotten]. Wang, Kim, and the instant application are analogous art because they are all directed to Machine Learning methods (see, e.g., Kim, Section 1.1, Wang, paragraph [0009]). 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 Kim’s unlearning model to incorporate the non-private forget data sample dataset input teachings of Wang. Doing so would have allowed Kim to use Wang’s method for “training a convolutional network to classify good data or bad data, with a small amount of clean data (or in-domain data)”, as suggested by Wang (see, e.g., Wang, paragraph [0004]). Kim in view of Wang do not explicitly disclose: inputting the forget data samples to a target model, trained on a total data set comprising the non-private data set and a private data set to produce target output wherein the private data set includes the forget data samples However, in the same field, analogous art Fraboni teaches: inputting the forget data samples to a target model, trained on a total data set comprising the non-private data set and a private data set to produce target output Fraboni, [0046], “the noise sensitivity of a particular client bounds a difference between the machine learning model trained on the multiple datasets owned by all of the available clients 104 and the machine learning model trained on the multiple datasets owned by the available clients 104 excluding the dataset owned by the particular client” [i.e., the combined training sets (one owned by all available clients and owned by only available clients excluding the particular client's dataset) corresponding to a non-private data set and a private data set within a total data set]. [0055]: “To remove the specificities of a dataset owned by client i from a global machine learning model with learned machine learning model parameters θn, Gaussian noise vi(n) with 0 mean and standard deviation σi(n) can be added to the machine learning model parameters θn to obtain new machine learning model parameters” [i.e., the final output of the unlearning process is the new global model parameters which functions as the target]. [0059]-[0060]: “training iterations since retraining began. Therefore, θr 0 represents the initial retraining model forgetting every client in request Wr and previous requests {Ws}s=1 r−1… In these implementations the noise sensitivity can be given by Equation (9) below. Ψ(i,r,n):=∑s=0n-1 Δ(θrs,Dr)-Δ(θrs,Dr\Di)2 (9)… In Equation (9), θr s represents the global machine learning model parameters where r represents the request to forget and s represents the amount of training iterations since retraining began, Dr represents a dataset that includes the remaining client datasets after the set of clients Wr in request r have been forgotten” [i.e., input for forgetting data samples is inherently expressed in Equation 9, and input into the global machine learning model (target model)]. wherein the private data set includes the forget data samples Fraboni, [0061]: “To forget a first client W 1 ={m 1 }, client m 1 is forgotten on the global model with coordinates (ζ 1 =0, T 1 ) such that the privacy criteria described above is satisfied for client m 1 . To forget a new client W 2 ={m 2 } after forgetting W 1 ={m 1 } with m 1 ≠m 2 , there are two cases to consider based on the noise sensitivity of client m 2 during forgetting index r=0 and r=1” [i.e., the data requested for removal is used to satisfy the privacy criteria and the forgotten data samples are the specific subset of the larger private data set whose removal is being requested]. Kim, Wang, Fraboni, and the instant application are analogous art because they are all directed to Machine Learning methods (see, e.g., Kim, Section 1.1, Wang, paragraph [0009], Fraboni, paragraph [0007]). 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 Kim in view of Wang to incorporate the private and non-private data sets of forget data samples of Fraboni. Doing so would have allowed Kim in view of Wang to use Fraboni’s method for “adding differential privacy to the trained machine learning model”, as suggested by Fraboni (see, e.g., Fraboni, paragraph [0005]). Regarding Claim 10: Claim 10 corresponds to claim 1 and is rejected for at least the same reasons as given in the rejection of claim 1, with the exception of the following limitations. Kim teaches: A system for editing a machine learning model to forget data Kim, Abstract, “we propose an efficient machine unlearning architecture to be used for computer vision classification models. Our approach consists of two-stage models, where in the first stage we enables a deep learning model that loses information with contrastive labels in the requested dataset” Section 1.1, “we propose a efficient two-stage model retraining algorithm, which enables a deep learning model to efficiently erase the information of the requested dataset on deep learning model. First, we make the model forget the subset of the training dataset” Wang teaches: comprising: at least one processor; and a computer readable storage medium having computer readable program code embodied therein that when executed by the at least one processor performs operations, the operations comprising Wang, [0055], “These computer programs (also known as programs, software, software applications or code) include machine instructions for a programmable processor, and can be implemented in a high-level procedural and/or object-oriented programming language, and/or in assembly/machine language. As used herein, the terms ‘machine-readable medium’ and ‘computer-readable medium’ refer to any computer program product, non-transitory computer readable medium, apparatus and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term ‘machine-readable signal’ refers to any signal used to provide machine instructions and/or data to a programmable processor” 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 Kim’s unlearning model to incorporate the non-private forget data sample dataset input teachings of Wang. Doing so would have allowed Kim to use Wang’s method for “training a convolutional network to classify good data or bad data, with a small amount of clean data (or in-domain data)”, as suggested by Wang (see, e.g., Wang, paragraph [0004]). Regarding Claim 16: Claim 16 corresponds to claim 1 and is rejected for at least the same reasons as given in the rejection of claim 1. Regarding Claim 21: As discussed above, Kim, Wang, and Fabroni teach [ the ] computer program product of claim 1 , and Kim further discloses: wherein the optimized gradients are optimized to mimic a probability distribution of output of the edited target model on remainder data comprising the data samples with the forget data samples removed Kim, Section 2, “By using the KD algorithm, we measure of how one probability distribution is different from a second, reference probability distribution to mimic the teacher model” p. 4365, Section 3.2, “We retrain the neutralized model using the remaining data except the requested data…In this training phase, the neutralized model is student model” p. 4366, Section 4, “Thus, we leverage TTA to demonstrate that our retrained model does not remember the requested data points, and even the similar data points with maintaining the model performance.” p. 4366, Section 5, “we demonstrated that our method successfully retrained the neutralized model efficiently and effectively.” On p. 4365, Kim discloses that the student model [target model] is the neutralized model, and the neutralized model is retrained [the edited target model] using the remaining data except the requested data [remainder data comprising the data samples with the forget data samples removed]. Further on p. 4366, Kim discloses that the retrained neutralized model was able to maintain performance while forgetting the requested data points [the optimized gradients are optimized to mimic a probability distribution of the output of the edited target model]. Regarding Claims 22: Claim 22 corresponds to claim 1 and are rejected for at least the same reasons as given in the rejection of claim 1. Regarding Claims 23: Claim 23 corresponds to claim 1 and are rejected for at least the same reasons as given in the rejection of claim 1 . 07-21-aia AIA Claim s 2-9, 11-15, and 17-20 are rejected under 35 U.S.C. 103 as being unpatentable over the combination of Kim, Wang and Fraboni in view of Mitchell et al. (“Fast Model Editing at Scale”), hereinafter Mitchell . Regarding Claim 2: As discussed above Kim, Wang, and Fraboni teach [ the ] computer program product of claim 1 , but do not explicitly disclose: wherein the calculating the optimized gradients comprises: inputting the gradients into an edit model to produce optimized gradients However, in the same field, analogous art Mitchell teaches: wherein the calculating the optimized gradients comprises: inputting the gradients into an edit model to produce optimized gradients Mitchell, Abstract, “we propose Model Editor Networks with Gradient Decomposition (MEND), a collection of small auxiliary editing networks that use a single desired input-output pair to make fast, local edits… MEND learns to transform the gradient obtained by standard fine-tuning” [i.e., the MEND network ('edit model') receives standard fine-tuning gradients as input]. Section 3.2: “MEND uses an editing training set Dtr edit to learn parameters φ` for each of the MEND networks g`… We optimize LMEND with respect to the MEND parameters at each time step using the Adam optimizer” [i.e., the MEND networks (edit models) and are trained and use an optimizer to optimize gradients]. “We then compute the parameter update for each layer W˜ = W` − α` ∇ ˜ W` ( ∇ ˜ W` from Eq. 2)… L e = − log p θW˜ (y' e |x ' e ), L loc = KL(pθW (·|x loc )kp θW˜ (·|x loc )). (4a,b) Intuitively, Le is small if the model has successfully updated its output for the edit example’s equivalence neighborhood, while L loc is small if the edit did not affect the model’s behavior on unrelated inputs” [i.e., the training objective L MEND forces the edit model (the MEND networks) to learn parameters that transform the optimized gradients ( ∇ ˜ W`)]. Kim, Wang, Fraboni, Mitchell and the instant application are analogous art because they are all directed to Machine Learning methods (see, e.g., Kim, Section 1.1, Wang, paragraph [0009], Fraboni, paragraph [0007], and Mitchell Section 1. Introduction). 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 combination of Kim, Wang and Fraboni to incorporate the edit model production of optimized gradients teaching of Mitchell. Doing so would have allowed the combination of Kim, Wang and Fraboni to use Mitchell’s method in order to “produce edits to a pre-trained model’s weights when provided with the standard fine-tuning gradient of a given correction as input”, as suggested by Mitchell (see, e.g., Mitchell, Section 1). Regarding Claim 3: As discussed above Kim, Wang, Fraboni, and Mitchell teach [ the ] computer program product of claim 2 , and Mitchell further teaches: wherein the operations further comprise: training the edit model to output the optimized gradients Mitchell, Section 3.2, “MEND uses an editing training set Dtr edit to learn parameters φ` for each of the MEND networks g`… We optimize LMEND with respect to the MEND parameters at each time step using the Adam optimizer” [i.e., the MEND networks (edit models) and are trained and use an optimizer to optimize gradients]. “We then compute the parameter update for each layer W˜ = W` − α` ∇ ˜ W` ( ∇ ˜ W` from Eq. 2)… Le = − log pθW˜ (y' e |x ' e ), Lloc = KL(pθW (·|xloc)kpθW˜ (·|xloc)). (4a,b) Intuitively, Le is small if the model has successfully updated its output for the edit example’s equivalence neighborhood, while Lloc is small if the edit did not affect the model’s behavior on unrelated inputs” [i.e., the training objective LMEND forces the edit model (the MEND networks) to learn parameters that transform the optimized gradients ( ∇ ˜ W`)]. 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 combination of Kim, Wang and Fraboni to incorporate training the edit model outputting optimized gradients teaching of Mitchell. Doing so would have allowed the combination of Kim, Wang and Fraboni to use Mitchell’s method in order to produce “a scalable algorithm for fast model editing that can edit very large pre-trained language models by leveraging the low-rank structure of fine-tuning gradients”, as suggested by Mitchell (see, e.g., Mitchell, Section 1). Regarding Claim 4: As discussed above Kim, Wang, Fraboni, and Mitchell teach [ the ] computer program product of claim 3 , and Mitchell further teaches: wherein the edit model is trained to minimize an error of a combination of a first loss function and a second loss function Mitchell, Section 3.2, “The total training loss for a MEND network is computed as LMEND = ceLe(θW˜ ) + Lloc(θW, θW˜ ). We optimize LMEND with respect to the MEND parameters)” [i.e., MEND (L MEND is the combination of the edit success loss L e and the locality loss L loc , meeting the functionality of minimizing a combination of a first loss function and a second loss function]. wherein the first loss function measures a divergence of the reference output and the target output Mitchell, Section 2, “By using the KD algorithm, we measure of how one probability distribution is different from a second, reference probability distribution to mimic the teacher model” Section 3.2, “During training time, we apply the KL divergence loss with Teacher and Student model… In this training phase, the neutralized model is student model, and the original model is the teacher model” [i.e., the K-L divergence loss measures the divergence between the outputs of the Original model/Teacher model (reference output) and the Student model (target output)]. 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 combination of Kim, Wang and Fraboni to incorporate the edit model’s minimization of an error between two loss functions measuring a divergence of a reference output and a target output teaching of Mitchell. Doing so would have allowed the combination of Kim, Wang and Fraboni to use Mitchell’s method in order to produce “a scalable algorithm for fast model editing that can edit very large pre-trained language models by leveraging the low-rank structure of fine-tuning gradients”, as suggested by Mitchell (see, e.g., Mitchell, Section 1). Regarding Claim 5: As discussed above Kim, Wang, Fraboni, and Mitchell teach [ the ] computer program product of claim 4 , and Kim further teaches: wherein the operations further comprise: applying a first weight to the first loss function to calculate a weighted first loss function; and applying a second weight to the second loss function to calculate a weighted second loss function Kim, Section 3.2, “the final loss function is combined with soft label knowledge distillation loss 2 and cross entropy loss 3 with bias factor as follows. For training efficiency, we set alpha and beta value as 1.1, 0.9 each”… PNG media_image1.png 111 467 media_image1.png Greyscale , PNG media_image2.png 117 581 media_image2.png Greyscale , PNG media_image3.png 53 506 media_image3.png Greyscale [i.e., the alpha and beta terms are the first and second weights (bias factors) applied to the two loss components (Cross entropy loss and knowledge distillation losses)]. wherein the combination of the first loss function and the second loss function comprises a sum of the weighted first loss function and the weighted second loss function Kim, Section 3.2, PNG media_image3.png 53 506 media_image3.png Greyscale [i.e., equation 4 is the sum of the two weighted loss functions forming the total objective loss function L TOTAL ]. wherein the calculating the optimized gradients comprises solving partial derivatives of the combination of the first loss function and the second loss function to determine the optimized gradients Kim, 1.1, “We proposed practical measurement method of forgetness in deep learning model and validated the forgetness of requested dataset by using augmentation…” and Section 3.2: “the final loss function is combined with soft label knowledge distillation loss2 and crossentropy loss3 with bias factor as follows PNG media_image3.png 53 506 media_image3.png Greyscale …we are retraining the neutralized model with dataset D \ P” [i.e., the operation of retraining a model in Deep Learning inherently involves minimizing its objective function, L TOTAL which is minimized by updating the model weights by computing partial derivatives multiplied by bias factors alpha and beta, in the gradient calculation of the loss functions]. Regarding Claim 6: As discussed above Kim, Wang, Fraboni, and Mitchell teach [ the ] computer program product of claim 3 , and Mitchell further teaches: wherein the optimized gradients minimize an error of a combination of a first loss function and a second loss function Mitchell, Section 3.2, “ultimately producing ∇ ˜ W` (Eq. 2). The final edited weights are W˜ = W` − α ∇ ˜ W` The total training loss for a MEND network is computed as LMEND = ceLe(θW˜ ) + Lloc(θW, θW˜ ). We optimize LMEND with respect to the MEND parameters” [i.e., MEND (L MEND ) is the combination of the edit success loss (Le and the locality loss L loc using optimized gradients ( ∇ ˜ W`), meeting the functionality of ensuring optimized gradients minimize the combined error]. wherein the first loss function measures a divergence of the reference output and the target output Mitchell, Section 3.2, “The training losses are Le, which measures edit success and L loc , which measures edit locality (the KL divergence between the pre-edit and post-edit model conditioned on the locality input x loc ” [i.e., the loss is produced by the KL divergence loss between pre-edit (reference) model output and post-edit (target) output]. wherein the operations further comprise: determining interim gradients that minimize an error of the first loss function Mitchell, Section 1, “Our approach trains lightweight model editor networks to produce edits to a pre-trained model’s weights when provided with the standard fine-tuning gradient of a given correction as input, leveraging the gradient as an information-rich starting point for editing” [i.e., The standard fine-tuning gradient is the interim gradient, a type of optimization]. wherein the edit model is trained to output the optimized gradients from input comprising the interim gradients Mitchell, Section 3.1, PNG media_image4.png 202 368 media_image4.png Greyscale …“MEND learns functions g l , with parameters φ`, which map u l i and δ i l+1 to pseudoactivations u˜ l i and pseudodelta ˜δ i l+1 . The model edit for weight matrix W` is then ∇ ˜ W` = Σ B i=1 ˜δ i l+1 u˜ i l T ” [i.e., the MEND network (edit model) takes the components of the fine-tuning gradient (interim gradients δ l+1 and u l ) as input and outputs the pseudo-gradients (˜δ i l+1 and u˜ i l T ), which are combined into the optimized gradient ∇ ˜ W` in equation 2]. 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 combination of Kim, Wang and Fraboni to incorporate the interim gradients that minimize an error of a loss function that an edit model outputs optimized gradients using the interim gradients teaching of Mitchell. Doing so would have allowed the combination of Kim, Wang and Fraboni to use Mitchell’s method in order to produce “a scalable algorithm for fast model editing that can edit very large pre-trained language models by leveraging the low-rank structure of fine-tuning gradients”, as suggested by Mitchell (see, e.g., Mitchell, Section 1). Regarding Claim 7: As discussed above Kim, Wang, Fraboni, and Mitchell teach [ the ] computer program product of claim 6 , and Fraboni further teaches: from input comprising the forget data samples Fraboni, [0061]: “To forget a first client W 1 ={m 1 }, client m 1 is forgotten on the global model with coordinates (ζ 1 =0, T 1 ) such that the privacy criteria described above is satisfied for client m 1 . To forget a new client W 2 ={m 2 } after forgetting W 1 ={m 1 } with m 1 ≠m 2 , there are two cases to consider based on the noise sensitivity of client m 2 during forgetting index r=0 and r=1” [i.e., the data requested for removal]. 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 Kim in view of Wang to incorporate the private and non-private data sets of forget data samples of Fraboni. Doing so would have allowed Kim in view of Wang to use Fraboni’s method for “adding differential privacy to the trained machine learning model”, as suggested by Fraboni (see, e.g., Fraboni, paragraph [0005]). Mitchell further teaches: wherein the reference output and the target output in the first loss function result Mitchell, Section 3.2, “The training losses are Le, which measures edit success and L loc , which measures edit locality (the KL divergence between the pre-edit and post-edit model conditioned on the locality input x loc ” [i.e., the loss is produced by the KL divergence loss between pre-edit (reference) model output and post-edit (target) output]. 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 combination of Kim, Wang and Fraboni to incorporate reference output and target output in a first loss function resulting from input comprising forget data samples teaching of Mitchell. Doing so would have allowed the combination of Kim, Wang and Fraboni to use Mitchell’s method in order to produce “a procedure that yields reliable, local, and general edits, while easily scaling to models with over 10 billion parameters”, as suggested by Mitchell (see, e.g., Mitchell, Section 1). Regarding Claim 8: As discussed above Kim, Wang, Fraboni, and Mitchell teach [ the ] computer program product of claim 3 , and Kim further teaches: applying the interim gradients to weights in the target model to yield a temporary target model Kim, Section 3.2, “The concept of neutralization is to make our model loss the critical information about the specific subset of dataset P. In this stage, we train the model M D … PNG media_image5.png 288 748 media_image5.png Greyscale , PNG media_image6.png 118 581 media_image6.png Greyscale …To validate our training efficiency that can merge faster, we stopped using knowledge distillation loss after 10th of epochs and used only cross entropy loss” [i.e., the interim gradients (determined by minimizing L CE on P) are applied to the weights of the model M' D (target model) during the train operation of the neutralization phase (see, e.g., Figure 1), the resulting neutralized model serves as the temporary target model that is the intermediate model between stage 1 and 2]. wherein the second loss function measures a divergence of the target model and the temporary target model Kim, Section 2, “By using the KD algorithm, we measure of how one probability distribution is different from a second, reference probability distribution to mimic the teacher model” and Section 3.2: “During training time, we apply the KL divergence loss with Teacher and Student model” [i.e., the second loss function (knowledge distillation loss) is defined as a measure of how the Student/Neutralized model's output (temporary target model output) diverge from the Teacher model output (target model)]. Mitchell further teaches: wherein the optimized gradients minimize an error of a combination of a first loss function and a second loss function Mitchell, Section 3.2, “ultimately producing ∇ ˜ W` (Eq. 2). The final edited weights are W˜ = W` − α ∇ ˜ W` The total training loss for a MEND network is computed as LMEND = ceLe(θW˜ ) + Lloc(θW, θW˜ ). We optimize LMEND with respect to the MEND parameters” [i.e., MEND (LMEND) is the combination of the edit success loss L e and the locality loss L loc using optimized gradients ( ∇ ˜ W`), meeting the functionality of ensuring optimized gradients minimize the combined error]) wherein the first loss function measures a divergence of the reference output and the target output Mitchell, Section 3.2, “The training losses are L e , which measures edit success and L loc , which measures edit locality (the KL divergence between the pre-edit and post-edit model conditioned on the locality input x loc ” [i.e., the loss is produced by the KL divergence loss between pre-edit (reference) model output and post-edit (target) output]. wherein the operations further comprise: determining interim gradients that minimize an error of the first loss function Mitchell, Section 1, “Our approach trains lightweight model editor networks to produce edits to a pre-trained model’s weights when provided with the standard fine-tuning gradient of a given correction as input, leveraging the gradient as an information-rich starting point for editing” [i.e., The standard fine-tuning gradient functions as the interim gradient] and Section 3.1: “The input to a MEND network g` is the fine-tuning gradient ∇ W l l e (x e , y e , θ) at layer l… PNG media_image7.png 397 726 media_image7.png Greyscale ” [i.e., interim gradient ~W l is calculated with respect to the loss operation that minimizes the error of the NLL loss function (the first loss function)] 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 combination of Kim, Wang and Fraboni to incorporate the minimizing of an error of a combination of loss functions, divergence measure from a first loss function of the reference output and target output, and determination of interim gradients that minimize an error of a first loss function, taught by Mitchell. Doing so would have allowed the combination of Kim, Wang and Fraboni to use Mitchell’s method in order to “enable[s] editability by training a collection of MLPs to modify model gradients to produce local model edits that do not damage model performance on unrelated inputs”, as suggested by Mitchell (see, e.g., Mitchell, Section 1 Introduction). Regarding Claim 9: As discussed above Kim, Wang, Fraboni, and Mitchell teach [ the ] computer program product of claim 8 , and Kim further teaches: wherein the calculating the second loss function comprises: forming a remainder data set comprising the total data set excluding the forget data set removed Kim, Section 3.1, “D = {(x1, y1),(x2, y2),(x3, y3)..,(xn, yn)} denotes the original dataset…Requested dataset should be removed in D will be notated as P(P ⊂ D)” and Section 3.2: “we are retraining the neutralized model with dataset D \ P” [i.e., The remainder data set is defined as the total dataset D excluding the forget data set (P) resulting in D \ P (the remaining data)]. inputting the remainder data set to the target model to produce target output Kim Figure 1: PNG media_image8.png 432 1017 media_image8.png Greyscale [i.e., the remain data set is input into the Teacher model and neutralized model (target and temporary target models) to produce the target retrained model output]. inputting the remainder data set to the temporary target model to produce temporary target output Kim, Section 3.2, “we retrain the neutralized deep learning model with remaining dataset using kl-divergence loss only with remaining dataset…We retrain the neutralized model using the remaining data except the requested data” [i.e., the retraining stage is used to produce a target model output using the remaining data as input into the neutralized model (temporary target model)]. wherein the divergence of the target model and the temporary target model comprises a divergence of the target output and the temporary target output Kim, Section 2, “In this work, we use KD to efficiently retrain the model that forgets the requested data. By using the KD algorithm, we measure of how one probability distribution is different from a second, reference probability distribution to mimic the teacher model” Section 3.2, “During training time, we apply the KL divergence loss with Teacher and Student model…to train the model faster and ensure the stability of training. In this training phase, the neutralized model is student model, and the original model is the teacher model” [i.e., the KL divergence loss L KD measures the divergence between the outputs (probability distribution/soft labels) of the Teacher model (target output) and the Student model (temporary target output)]. Regarding Claim 11-15: Claims 11-15 correspond to claims 2-5 and 8 and are rejected for at least the same reasons as given in the rejections of claims 2-5 and 8. In particular, 11:2, 12:3, 13:4, 14:5, 15:8. Regarding Claim 17-20: Claims 17-20 correspond to claims 2-5 and are rejected for at least the same reasons as given in the rejections of claims 2-5. In particular, 17:2, 18:3, 19:4, 20:5 . Response to Arguments 35 U.S.C. § 101: Applicant’s arguments, see “Remarks” pp. 8-11, filed March 15, 2026, with respect to claims 1-20 in view of 35 U.S.C. § 101 have been fully considered and are persuasive. The rejections of claims 1-20 under section 35 U.S.C. § 101 have been withdrawn. 35 U.S.C. § 103: 07-37 AIA Applicant's arguments filed March 15, 2026 (“Remarks”) have been fully considered but they are not persuasive. Remarks, pp. 12. Applicant argues with respect to claims 1, 10, and 16, that Wang does not teach the claimed limitation. Examiner respectfully disagrees. Although the vocabulary differs, the functionality is the same. In other words, Wang teaches the removal of the data which is functionally the same as data to be forgotten because in both instances, data is to be removed. Remarks, pp. 12-13. Applicant argues with respect to claims 1, 10, and 16 that Fabroni does not teach the claimed limitation. Examiner respectfully disagrees. Similar to above, although the vocabulary differs, the functionality is the same. In other words, Fabroni teaches the removal of data client data from a global client set (private data and non-private data) which is functionally the same as data to be forgotten because in both instances, data is to be removed. Remarks, pp. 13-14. Applicant argues with respect to claims 1, 10, and 16 that Kim does not teach divergence of outputs from a reference model having non-private data and a target model having non-private and private data. Examiner respectfully disagrees. As discussed under section 103, Kim discloses a KL (Kullback-Leibler) divergence between a teacher model (reference) model) and the student model (target model). Conclusion 07-40 AIA Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL . See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a). A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action. Any inquiry concerning this communication or earlier communications from the examiner should be directed to STEVEN PHUNG whose telephone number is (703) 756-1499. The examiner can normally be reached Monday-Thursday: 9:00AM-4:00PM 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, KAMRAN AFSHAR can be reached at (571) 272-7796. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300. Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000. /S.H.P./Examiner, Art Unit 2125 /KAMRAN AFSHAR/Supervisory Patent Examiner, Art Unit 2125 Application/Control Number: 18/176,068 Page 2 Art Unit: 2125 Application/Control Number: 18/176,068 Page 3 Art Unit: 2125 Application/Control Number: 18/176,068 Page 4 Art Unit: 2125 Application/Control Number: 18/176,068 Page 5 Art Unit: 2125 Application/Control Number: 18/176,068 Page 6 Art Unit: 2125 Application/Control Number: 18/176,068 Page 7 Art Unit: 2125 Application/Control Number: 18/176,068 Page 8 Art Unit: 2125 Application/Control Number: 18/176,068 Page 9 Art Unit: 2125 Application/Control Number: 18/176,068 Page 10 Art Unit: 2125 Application/Control Number: 18/176,068 Page 11 Art Unit: 2125 Application/Control Number: 18/176,068 Page 12 Art Unit: 2125 Application/Control Number: 18/176,068 Page 13 Art Unit: 2125 Application/Control Number: 18/176,068 Page 14 Art Unit: 2125 Application/Control Number: 18/176,068 Page 15 Art Unit: 2125 Application/Control Number: 18/176,068 Page 16 Art Unit: 2125 Application/Control Number: 18/176,068 Page 17 Art Unit: 2125 Application/Control Number: 18/176,068 Page 18 Art Unit: 2125 Application/Control Number: 18/176,068 Page 19 Art Unit: 2125 Application/Control Number: 18/176,068 Page 20 Art Unit: 2125 Application/Control Number: 18/176,068 Page 21 Art Unit: 2125 Application/Control Number: 18/176,068 Page 22 Art Unit: 2125 Application/Control Number: 18/176,068 Page 23 Art Unit: 2125 Application/Control Number: 18/176,068 Page 24 Art Unit: 2125 Application/Control Number: 18/176,068 Page 25 Art Unit: 2125 Application/Control Number: 18/176,068 Page 26 Art Unit: 2125 Application/Control Number: 18/176,068 Page 27 Art Unit: 2125 Application/Control Number: 18/176,068 Page 28 Art Unit: 2125 Application/Control Number: 18/176,068 Page 29 Art Unit: 2125 Application/Control Number: 18/176,068 Page 30 Art Unit: 2125
Prosecution Insights