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 .
Response to Amendment
This action is in response to amendments filed 08/25/2026. In which claims 1, 6, 8, 13, 15, and 20 are amended and claims 3-5, 10-12, 17-19 are canceled.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1-2, 6-9, 13-16, 20 are rejected under 35 U.S.C. 103 as being unpatentable over Wu (PUMA: Performance Unchanged Model Augmentation for Training Data Removal) in view of Bourtoule (Machine Unlearning).
Regarding claim 1, Wu teaches a computer-implemented method comprising building an optimal model by minimizing a prediction error over the original training data set (Pg. 4, Let us assume the model parameter θorg is the optimal solution of the (original)training objective Jorg, Equation (3) goes on to show that the optimal model minimizes the loss over the data set); determining a loss function that measures a difference between a first prediction of the optimal model when trained with the original training data set and a known ground truth (Pg. 4, where Lt denotes individual prediction loss, Equation (3) and (4) demonstrate the use of some loss function that takes in a prediction x and ground truth y); determining an impact factor that measures a difference between the first prediction of the optimal model and a second prediction of the optimal model when trained with a sanitized training data set against the known ground truth, wherein the sanitized training data set comprises remaining data after removing the specific subset of data from the original training data set (Pg. 4, Eq (2-12), the influence of a specific subset of data is calculated and approximated, Eq (6) shows that this influence/impact factor, depicted in Eq (12) for an individual data point, depends on the difference between the optimal model without the subset θmod, which is the model with updated parameters aimed to remove the characteristics of the subset optimized against the ground truth as shown in Eq (4), and the optimal model with the subset); and fine-tuning a machine unlearning model on the impact factor to quantify an impact of removing the specific subset of data from the original training data set (Pg. 4, Performance Preserved Data Removal through Gradient Re-weighting, the model parameters are updated by the impact factor to account for data removal as shown in Eq (11) and Eq (12)).
Wu fails to teach receiving a data removal request identifying, for removal, a specific subset of data from an original training data set and wherein the machine unlearning model comprises a sharded, isolated, sliced, aggregated (SISA) unlearning model; and wherein the SISA unlearning model divides the original training data set into s data splits, wherein each of the s data splits are further split into r slices, and wherein the SISA unlearning model comprises a plurality of constituent models Ms, each trained on a unique slice; and wherein, responsive to receiving the data removal request: a constituent model is identified whose unique slice contains the specific subset of data to be unlearned; and retraining is initialized from last parameter values of the respective constituent model saved before the unique slice containing the specific subset of data to be unlearned was included.
Bourtoule teaches receiving a data removal request identifying, for removal, a specific subset of data from an original training data set (Pg. 4, Any user u ∈ U can revoke access to their individual data du ⊂ D…. Within a finite period of time, the service provider has to erase the revoker’s data and modify any trained models M) and wherein the machine unlearning model comprises a sharded, isolated, sliced, aggregated (SISA) unlearning model (Bourtoule Abs, we introduce SISA training, a framework that expedites the unlearning process); and wherein the SISA unlearning model divides the original training data set into s data splits, wherein each of the s data splits are further split into r slices, and wherein the SISA unlearning model comprises a plurality of constituent models Ms, each trained on a unique slice (Fig 2, the figure displays the SISA model). Wu and Bourtoule are analogous to the claimed invention because they are in the field of machine unlearning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have used the SISA model (including the slicing and splitting structure) to reduce computational overhead (Bourtoule Abs, SISA training reduces the computational overhead associated with unlearning); and wherein, responsive to receiving the data removal request: a constituent model is identified whose unique slice contains the specific subset of data to be unlearned; and retraining is initialized from last parameter values of the respective constituent model saved before the unique slice containing the specific subset of data to be unlearned was included (Fig 2, When data needs to be unlearned, only one of the constituent models whose shards contains the point to be unlearned needs to be retrained — retraining can start from the last parameter values saved before including the slice containing the data point to be unlearned., Section IV B, Thus, if the user desires for du to be unlearned, then the service provider has to (a) first locate the slice in which du is located, referred to as Dk,u, and (b) perform the training procedure as specified above from step u onwards using Dk,u \du; this will result in a new model Mk,u. For a single unlearning request, this provides a best-case speed-up up to R+1 2 × compared to using the strategy without slicing (we discus this in more detail in § V-C)).
Wu and Bourtoule are analogous to the claimed invention because they are in the field of machine unlearning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have included the process of receiving a request for data removal from Bourtoule into Wu to allow users to practice their right to be forgotten, privacy rights companies must follow in many nations (Bourtoule Pg. 1, include provisions that require the so-called right to be forgotten….mandates that companies take reasonable steps to achieve the erasure of personal data concerning [the individual]) and to have used SISA training as done in Bourtoule to reduce the computational overhead associated with unlearning (Bourtoule Abs).
Regarding claim 2, Wu in view of Bourtoule teaches the method of claim 1, further Wu teaches wherein minimizing the prediction error over the original data comprises minimizing a regularized empirical risk (Pg. 3 Eq (4), the optimal model minimizes the loss averaged over the dataset divided by the size of the dataset which is the empirical risk).
Regarding claim 6, Wu in view of Bourtoule teaches the method of claim 1 and Wu teaches wherein fine-tuning the machine unlearning model comprises determining a data impact factor Ri (Pg. 4, Eq (2-12), the influence of a specific subset of data is calculated and approximated, Eq (6) shows that this influence/impact factor, depicted in Eq (12) for an individual data point, depends on the difference between the optimal model without the subset θmod, which is the model with updated parameters aimed to remove the characteristics of the subset optimized against the ground truth as shown in Eq (4), and the optimal model with the subset) and Bourtoule teaches, which Wu is silent on, for each slice comprising data of the specific subset of data (Fig 2). Wu and Bourtoule are analogous to the claimed invention because they are in the field of machine unlearning. Furthermore, Wu calculates the impact factor to determine the influence of removed data and given the SISA model divides data into slices used on individual models, it would have been obvious to one of ordinary skill in the art before the effective filling data of the claimed invention to have calculated an impact factor for each individual slice containing removed data so that model parameters can be updated only for models containing removed points maintaining the SISA model and reducing computational overhead (Bourtoule Fig 2, When data needs to be unlearned, only one of the constituent models whose shards contains the point to be unlearned needs to be retrained, data unlearning is done for only some of the models Mi therefore the method in Wu must be altered to be done for only specific models meaning model parameters must be updated differently and in turn different impact factors must be used to fine tune the models derived from different slices).
Regarding claim 7, Wu in view of Bourtoule teaches the method of claim 6, wherein each of the data impact factors Ri for each respective slice comprising data of the specific subset of data are normalized and applied as a weight to a respective constituent model Ms when fine-tuning the SISA unlearning model. As stated in the rejection of claim 6, Wu in view of Bourtoule teaches determining the data impact factors Ri for each of the respective slices comprising data of the specific subset of data, furthermore, Wu teaches applying the normalized impact factor as weight to a model when fine tuning (Pg. 4, Performance Preserved Data Removal through Gradient Re-weighting, the model parameters are updated by the impact factor to account for data removal as shown in Eq (11) and Eq (12), Pg 4, By combining Equation 2 and Equation 8, we note they form an implicit constraint on the data up-scaling factors λ, the optimized contribution factor λ∗ normalizes the impact factors). Given Wu and Bourtoule are analogous to the claimed invention, it would have been obvious to one of ordinary skill in the art to apply each of the data impact factors as a weight to a respective constituent model when fine-tuning the SISA unlearning model to address model degradation by “complementing the negative impact of removing marked data by reweighting” (Wu Abs) while maintaining the SISA model to “reduce the computational overhead associated with unlearning” (Bourtoule Abs).
Regarding claims 8-14, excluding canceled claims, the inventive concept is essentially the same as claims 1-7 except for the system having a memory, computer readable instructions, and one or more processors for executing the computer readable instructions, the computer readable instructions controlling the one or more processors to perform operations comprising. Wu teaches a system having a memory, computer readable instructions, and one or more processors for executing the computer readable instructions, the computer readable instructions controlling the one or more processors to perform operations comprising (Appendix A, The following pytorch code shows the actual architecture we used in the experiments, the use of pytorch to conduct the experiment implies the existences of computer and in turn a memory, processor, and computer readable instructions). Wu in view of Bourtoule teaches the remaining elements of the claims as shown in the rejection of claims 1-7.
Regarding claims 15-20, excluding canceled claims, the inventive concept is essentially the same as claims 1-6 except for a computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by one or more processors to cause the one or more processors to perform operations comprising. Wu teaches a computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by one or more processors to cause the one or more processors to perform operations comprising (Appendix A, The following pytorch code shows the actual architecture we used in the experiments, the use of pytorch to conduct the experiment implies the existences of a computer and in turn a processor, computer readable storage, and computer program product). Wu in view of Bourtoule teaches the remaining elements of the claims as shown in the rejection of claims 1-6.
Response to Arguments
Applicant’s arguments, see Pg.2, filed 08/25/2026, with respect to the 35 U.S.C 101 rejection have been fully considered and are persuasive. The 101 rejection of 05/26/2026 has been withdrawn.
Applicant's arguments regarding the 35 U.S.C 103 rejection filed 08/25/2026 have been fully considered but they are not persuasive. The applicant asserts “Bourtoule does not offer a mechanism whereby, responsive to receiving a data removal request: a constituent model is identified whose unique slice contains the specific subset of data to be unlearned; and retraining is initialized from last parameter values of the respective constituent model saved before the unique slice containing the specific subset of data to be unlearned was included” and as such Wu in view of Bourtoule fails to disclose each and every element of the independent claims, the examiner respectfully disagrees. Bourtoule not only discusses identifying a model and retraining using the last parameter saved (as shown in the 103 rejection for claim 1 above), it also mentions unlearning requests from users. Therefore, Wu in view of Bourtoule teaches all the newly added elements.
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.
/NATNAEL A ASEGDEW/ Examiner, Art Unit 2122 /MICHAEL H HOANG/PRIMARY EXAMINER, Art Unit 2122
Prosecution Insights