DETAILED ACTION
Notice of 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 statements submitted on 4/1/2024 and 5/20/2025 have been considered.
Claim Objections
Claim 14 and 20 are objected to because of the following informalities:
In Claim 14, line 2, the examiner suggests amending “processor set configured to” to recite “a processor set configured to”
In Claim 20, line 5, the examiner suggests amending “processor set configured to” to recite “a processor set configured to”
In claim 20, lines 18-19, “the ML model” should read “the pre-trained ML model”
Appropriate correction is required.
35 USC § 101 Analysis (Not a Rejection)
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Regarding Step 1 of the Alice/Mayo framework, Claims 1-13 are directed to a method (a process), Claims 14-19 are directed to a system (a machine), which each fall within one of the four statutory categories of inventions.
Claim 20 is directed to a computer program product comprising a computer-readable storage medium.” Para. 0036 explains that “A computer-readable storage medium, as that term is used in the present disclosure, is not to be construed as storage in the form of transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide, light pulses passing through a fiber optic cable, electrical signals communicated through a wire, and/or other transmission media.” Therefore, in view of this disclosure, the broadest reasonable interpretation of the recited “computer-readable storage medium” does not include transitory signals, and therefore claim 20 is considered to be eligible subject matter (an article of manufacture).
Regarding Independent Claims 1, 14, and 20
Step 2A, prong 1 (Is the claim directed to a law of nature, a natural phenomenon or an abstract idea).
The following calculations or determinations with respect to loss functions are each considered to be mental processes that can be performed in the human mind, or using physical aids such as pencil and paper:
calculating, ... a first surprise score for each target sample of the set of target samples;
calculating, ... a second surprise score for each of the retrieved set of supplemental samples associated with each target sample of the set of target samples;
determining, ... a first loss function based on the first surprise score for each target sample of the set of target samples and the second surprise score for each of the set of supplemental samples associated with each target sample of the set of target samples;
determining, ... a second loss function based on the second surprise score for each of the set of supplemental samples; and
Step 2A, prong 2 (Does the claim recite additional elements that integrate the judicial exception into a practical application?).
The examiner finds that the “updating, by the computer, the pre-trained ML model based on the first loss function and the second loss function” limitation of claim 1 integrates the judicial exception into a practical application, and that substantially similar limitations in independent claims 14 and 20 also integrate the judicial exceptions into a practical application.
MPEP 2106.04(d)(1) provides the following guidance:
In short, first the specification should be evaluated to determine if the disclosure provides sufficient details such that one of ordinary skill in the art would recognize the claimed invention as providing an improvement. The specification need not explicitly set forth the improvement, but it must describe the invention such that the improvement would be apparent to one of ordinary skill in the art. Conversely, if the specification explicitly sets forth an improvement but in a conclusory manner (i.e., a bare assertion of an improvement without the detail necessary to be apparent to a person of ordinary skill in the art), the examiner should not determine the claim improves technology. Second, if the specification sets forth an improvement in technology, the claim must be evaluated to ensure that the claim itself reflects the disclosed improvement. That is, the claim includes the components or steps of the invention that provide the improvement described in the specification.
First, the specification discloses a problem where in some instances, machine learning models may need to unlearn samples in certain situations, such as to comply with copyright and data privacy laws. (paras. 0001-0004). The specification recognizes the technical problem that re-training a model from scratch “may be expensive, cumbersome, and time-consuming.” (para. 0004). Figs. 4-5 (explained at paras. 0082-0109) describe, in detail, the technical details explaining how to update a machine learning model using unlearning techniques without having to re-train the entire model from scratch, and one of ordinary skill in the art would recognize this particular disclosure as providing an improvement to machine learning (or unlearning) technologies.
Second, the claim itself reflects the improvement. The independent claims recite “updating, ... the pre-trained ML model based on the first loss function and the second loss function” or something similar. The recited “first loss function” and “second loss function” are based on “surprise scores” based on “a set of target samples to be unlearned.” Therefore, this “updating” limitation reflects the improvement because now the pre-trained ML model is updated to unlearn the information from the “set of target samples to be unlearned” via the loss functions. Therefore, the examiner finds the independent claims to be subject matter-eligible.
The examiner further finds the present claims to be analogous to those in Ex Parte Desjardins. MPEP 2106.04(d), subsection III, which explains:
Specifically, the ARP upheld the Step 2A Prong One finding that the claims recited an abstract idea (i.e., mathematical concept). In Step 2A Prong Two, the ARP then determined that the specification identified improvements as to how the machine learning model itself operates, including training a machine learning model to learn new tasks while protecting knowledge about previous tasks to overcome the problem of “catastrophic forgetting” encountered in continual learning systems. Importantly, the ARP evaluated the claims as a whole in discerning at least the limitation “adjust the first values of the plurality of parameters to optimize performance of the machine learning model on the second machine learning task while protecting performance of the machine learning model on the first machine learning task” reflected the improvement disclosed in the specification. Accordingly, the claims as a whole integrated what would otherwise be a judicial exception instead into a practical application at Step 2A Prong Two, and therefore the claims were deemed to be outside any specific, enumerated judicial exception (Step 2A: NO). (emphasis added).
Here, the claimed invention also relates to preventing “catastrophic forgetting of supplemental samples.” (paras. 0005, 0010) And the “updating” step similarly reflects the improvement disclosed in the specification.
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.
Claims 1-3, 5-6, 12, 14-16, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over US 20250103878 A1, hereinafter referenced as LI, in view of Tarun, Ayush K., et al. "Deep regression unlearning." arXiv preprint arXiv:2210.08196 (2023), hereinafter referenced as TARUN.
Regarding Claim 1
LI teaches:
A computer-implemented method for unlearning samples by a pre-trained machine learning (ML) model, the computer-implemented method comprising: (LI, para. 0001: “Aspects generally relate to systems and methods for machine unlearning in generative models.”;
LI, para. 0064: “FIG. 3 further depicts exemplary computing device 302. Computing device 302 depicts exemplary hardware that executes the logic that drives the various system components described herein.”)
retrieving, by a computer, a set of target samples to be unlearned by the pre-trained ML model, wherein the set of target samples is retrieved from a data source; (LI, para. 0047: “In accordance with aspects, datum 112 from dataset 102 may be provided to encoder 123 of target model 122 as input to target model 122. Moreover, Gaussian noise sample 113, which is a sample drawn from Gaussian noise, may be provided as input to encoder 125 of original model 124. Output 130 is output from encoder 123 of target model 122 based on input datum 112. Output 134 is output from encoder 125 of target model 124 based on Gaussian noise sample 113. Output 130 along with output 134 may be used to compute a distance to normal noise”;
LI, para. 0049: “In accordance with aspects, the steps described with respect to the components of FIG. 1 may be iterated a number of times and a total loss may be accumulated based on the number of iterations. A machine unlearning process may then update the target model 122 with the generated accumulated loss.”
LI, para. 0051: “Step 210 includes providing a first datum to a target model as input to the target model, wherein the first datum is retrieved from a forget dataset.”;
Examiner’s Note: datum 112 corresponds to recited “set of target samples to be unlearned” and is retrieved from “forget database 112”, where datum 112 is part of the set of samples to be unlearned during the iterations of Fig. 1)
retrieving, by the computer, a set of supplemental samples associated with each target sample of the retrieved set of target samples; (LI, para. 0048: “In accordance with aspects, datum 114 from dataset 104 may be provided to encoder 123 of target model 122 as input to target model 122. Datum 114 may further be provided to encoder 125 of original model 124 as input to original model 124. Output 132 is output from encoder 123 of target model 122 based on input datum 114. Output 136 is output from encoder 125 of original model 124 based on input datum 114. Output 132 and output 136 may be used to compute a distance to original model 124 (i.e., loss L.sub.T, as discussed above).”;
LI, para. 0049: “In accordance with aspects, the steps described with respect to the components of FIG. 1 may be iterated a number of times and a total loss may be accumulated based on the number of iterations. A machine unlearning process may then update the target model 122 with the generated accumulated loss.”
LI, para. 0052: “Step 220 includes providing a sample drawn from Gaussian noise to an original model.”
Examiner’s Note: datum 114 corresponds to recited “set of supplemental samples associated with each target sample of the retrieved set of target samples”, where datum 114 is part of the set of samples during the iterations of Fig. 1)
calculating, by the computer, a first surprise score for each target sample of the set of target samples; (LI, para. 0047: “In accordance with aspects, datum 112 from dataset 102 may be provided to encoder 123 of target model 122 as input to target model 122. Moreover, Gaussian noise sample 113, which is a sample drawn from Gaussian noise, may be provided as input to encoder 125 of original model 124. Output 130 is output from encoder 123 of target model 122 based on input datum 112. Output 134 is output from encoder 125 of target model 124 based on Gaussian noise sample 113. Output 130 along with output 134 may be used to compute a distance to normal noise”;
Examiner’s Note: As depicted in Fig. 1, the distance to normal noise calculated from outputs 130 and 134 correspond to the recited “first surprise score”, which is the distance between the output 130 of the datum 112 with respect to the output 134 of the datum 113 with respect to Gaussian noise)
calculating, by the computer, a second surprise score for each of the retrieved set of supplemental samples associated with each target sample of the set of target samples; (LI, para. 0048: “In accordance with aspects, datum 114 from dataset 104 may be provided to encoder 123 of target model 122 as input to target model 122. Datum 114 may further be provided to encoder 125 of original model 124 as input to original model 124. Output 132 is output from encoder 123 of target model 122 based on input datum 114. Output 136 is output from encoder 125 of original model 124 based on input datum 114. Output 132 and output 136 may be used to compute a distance to original model 124 (i.e., loss LT, as discussed above).”;
Examiner’s Note: As depicted in Fig. 1, the distance to normal noise calculated from outputs 132 and 136 correspond to the recited “second surprise score”, which is the distance between the output 132 of the datum 114 from the target model, and output 136 of the datum 114 from the original model)
determining, by the computer, a first loss function based on the first surprise score for each target sample of the set of target samples ...; (LI, paras. 0036-0038:
PNG
media_image1.png
442
480
media_image1.png
Greyscale
LI, para. 0053: “Step 230 includes computing a first loss, wherein the first loss is based on target model output from processing the first datum and original model output from processing the sample drawn from Gaussian noise.”
Examiner’s Note: LN corresponds to the recited “first loss function” and relies on the difference between the forget sample and the Gaussian noise sample)
determining, by the computer, a second loss function based on the second surprise score for each of the set of supplemental samples; and (LI, paras. 0039-0040)
PNG
media_image2.png
376
484
media_image2.png
Greyscale
LI, para. 0056: “Step 260 includes computing a second loss, wherein the second loss is based on target model output from processing the second datum and original model output from processing the second datum.”;
Examiner’s Note: LT corresponds to the recited second loss function, which takes into account the difference between the use of the datum in the original and target models)
updating, by the computer, the pre-trained ML model based on the first loss function and the second loss function. (LI, para. 0041: “The overall loss function may be the weighted combination of the above two losses with the weight alpha (α)”;
LI, para. 0049: “In accordance with aspects, the steps described with respect to the components of FIG. 1 may be iterated a number of times and a total loss may be accumulated based on the number of iterations. A machine unlearning process may then update the target model 122 with the generated accumulated loss. Moreover, once an accumulated loss is secured and target model 122 is updated with the accumulated loss, the process may begin again using a new datum 112 from dataset 102 and a new datum 114 from dataset 104. The process may be iterated through several new datums from each dataset, and the process may minimize an expectation value. Because a minimal expectation value may not be known, the process may iterate a fixed number of times.”;
LI, claim 4: “updating the target model based on the total loss, wherein executing the total loss accumulation process and updating the target model based on the total loss are secondary steps of the machine unlearning process, and wherein a complete iteration of the machine unlearning process includes executing the initial steps of the machine unlearning process and the secondary steps of the machine unlearning process.”)
However, LI fails to explicitly teach:
... and the second surprise score for each of the set of supplemental samples associated with each target sample of the set of target samples
However, in a related field of endeavor (machine unlearning, see p. 2, section 1), TARUN teaches and makes obvious:
determining, by the computer, a first loss function based on the first surprise score for each target sample of the set of target samples and the second surprise score for each of the set of supplemental samples associated with each target sample of the set of target samples (TARUN, pp. 4-5, section 4.1:
PNG
media_image3.png
350
468
media_image3.png
Greyscale
PNG
media_image4.png
482
466
media_image4.png
Greyscale
Examiner’s Note: the LI-TARUN combination now modifies the first loss equation of LI (see para. 0053) with equation (6) of TARUN so that information relating to retained samples are considered as additional data when calculating the first loss)
Before the effective date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of LI with TARUN as explained above. As explained by TARUN, one of ordinary skill would have been motivated to do so use the blindspot unlearning techniques of TARUN to “give[] the model a vague idea about the output distribution in the absence of the forget set from the training data.” (p. 4, section 4.1). One of ordinary skill in the art would understand that including additional data into the loss metrics will improve the accuracy of the loss metrics, and therefore improve the unlearning process as a whole.
Regarding Claim 2
LI and TARUN teach the method of claim 1 as explained above. LI further teaches:
wherein the first surprise score is indicative of a surprise in the behavior of the pre-trained ML model when a target sample is provided as an input to the pre-trained ML model as compared to a training dataset used to train the pre-trained ML model. (LI, para. 0047: “In accordance with aspects, datum 112 from dataset 102 may be provided to encoder 123 of target model 122 as input to target model 122. Moreover, Gaussian noise sample 113, which is a sample drawn from Gaussian noise, may be provided as input to encoder 125 of original model 124. Output 130 is output from encoder 123 of target model 122 based on input datum 112. Output 134 is output from encoder 125 of target model 124 based on Gaussian noise sample 113. Output 130 along with output 134 may be used to compute a distance to normal noise”;
Examiner’s Note: As depicted in Fig. 1, the distance to normal noise calculated from outputs 130 and 134 correspond to the recited “first surprise score”, which is the distance between the output 130 of the datum 112 with respect to the output 134 of the datum 113 with respect to Gaussian noise, and this is a surprise because this takes into account how the model will react to random Gaussian noise being injected, as compared to a training dataset that lacks such Gaussian noise injection)
Regarding Claim 3
LI and TARUN teach the method of claim 1 as explained above. LI further teaches:
wherein the unlearning of the set of target samples from the pre-trained ML model corresponds to a removal of each target sample of the set of target samples from a knowledge base of the pre-trained ML model. (LI, para. 0033: “A machine unlearning process may be configured to partially or fully reverse any influence that the data included in the forget dataset previously had on the target model. That is, the machine unlearning process is configured to manipulate the model so as to reverse any training of the model with respect to the forget dataset that resulted in the model learning (i.e., being configured through the training process) to generate output based on the data in the forget dataset. This may be referred to as the model “forgetting” or “unlearning” the data in the forget dataset.”;
Examiner’s Note: the broadest reasonable interpretation of the “knowledge base of the pre-trained model” includes the “influence” of certain samples on the model during training)
Regarding Claim 5
LI and TARUN teach the method of claim 1 as explained above. However, LI fails to explicitly teach:
wherein the set of supplemental samples is retrieved from the data source, and wherein the data source comprises a training dataset used to train the pre-trained ML model.
However, in a related field of endeavor (machine unlearning, see p. 2, section 1), TARUN teaches and makes obvious:
wherein the set of supplemental samples is retrieved from the data source, and wherein the data source comprises a training dataset used to train the pre-trained ML model. (TARUN, p. 3, section 3.1:
PNG
media_image5.png
262
474
media_image5.png
Greyscale
Examiner’s Note: TARUN disclose a common dataset D, from which forgetting and retaining data come from; the LI-TARUN combination now retrieves the supplemental samples from the same initial dataset D as in TARUN)
Before the effective date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of LI with TARUN as explained above. As explained by TARUN, one of ordinary skill would have been motivated to do so use the blindspot unlearning techniques of TARUN to “give[] the model a vague idea about the output distribution in the absence of the forget set from the training data.” (p. 4, section 4.1). One of ordinary skill in the art would understand that including additional data into the loss metrics will improve the accuracy of the loss metrics, and therefore improve the unlearning process as a whole.
Regarding Claim 6
LI and TARUN teach the method of claim 1 as explained above. LI further teaches:
wherein the first surprise score for each target sample of the set of target samples is calculated based on a modality of at least one target sample of the set of target samples. (LI, para. 0029: “s used herein, a datum may include a picture or an image. A datum, such as a picture or an image, may be included in a dataset of that includes additional datums.”;
LI, para. 0046: “Datum 112 may be a datum (i.e., an image file) from dataset 102.”
Examiner’s Note: the distance to normal noise between outputs 130 and 134 (the first surprise score) is based on the type of datum being an image (corresponding to recited “modality”, or type of data)
Regarding Claim 12
LI and TARUN teach the method of claim 1 as explained above. However, LI fails to explicitly teach:
wherein the pre-trained ML model is trained for at least one epoch of a set of epochs, and wherein a count of the set of epochs corresponds to a second hyper-parameter associated with the training of the pre-trained ML model.
However, in a related field of endeavor (machine unlearning, see p. 2, section 1), TARUN teaches and makes obvious:
wherein the pre-trained ML model is trained for at least one epoch of a set of epochs, and wherein a count of the set of epochs corresponds to a second hyper-parameter associated with the training of the pre-trained ML model. (TARUN, p. 7, section 6.3: “We train the model for 100 epochs with initial learning rate of 0.01 and reduce it on plateau by a factor of 0.1. The models are optimized on L1-loss with Adam optimizer. In FineTune, 5 epochs of training is done with a learning rate of 0.001. We run gradient ascent for 1 epoch with a learning rate of 0.001 on the AgeDB dataset. In Gaussian Amnesiac, 1 epoch of amnesiac learning is done with a learning rate of 0.001. In Blindspot, the blindspot model is trained for 2 epochs with a learning rate of 0.01. Subsequently, 1 epoch of unlearning is performed on the original model with a learning rate of 0.001.”;
Examiner’s Note: TARUN teaches the concept of setting a certain number of epochs for training as a hyperparameter: the LI-TARUN combination now trains the underlying pre-trained model of LI using a set number of training epochs as in TARUN)
Before the effective date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of LI with TARUN as explained above. As explained by TARUN, one of ordinary skill would have been motivated to do in order to tune the number of training epochs and other hyper-parameters corresponding to such epochs, such as learning rates. (p. 7, section 6.3). One of ordinary skill would further understand that the number of training epochs relates to computing time and processing power, and it would therefore be beneficial for a designer to control the number of epochs to meet the computing time and processing power constraints, if any.
Regarding Claim 14
LI teaches:
A system comprising: processor set configured to: (LI, para. 0010: “In some aspects, the techniques described herein relate to a system including at least one computer including a processor and a memory, wherein the at least one computer is configured to:”)
The remaining limitations correspond to the method of claim 1, and therefore this claim 14 is rejected for the same reasons explained above with respect to claim 1.
Claim 15 depends from claim 14 and claims a system that corresponds to the method of claim 2, and is therefore rejected for the same reasons explained above with respect to claims 2 and 14.
Claim 16 depends from claim 14 and claims a system that corresponds to the method of claim 3, and is therefore rejected for the same reasons explained above with respect to claims 3 and 14.
Regarding Claim 20
LI teaches:
A computer program product for unlearning a first target sample of a set of target samples by a pre-trained machine learning (ML) model, the computer program product comprising a computer-readable storage medium having program instructions embodied therewith, the program instructions executable by a system to cause the system to, comprising: (LI, para. 0001: “Aspects generally relate to systems and methods for machine unlearning in generative models.”;
LI, para. 0017: “In some aspects, the techniques described herein relate to a non-transitory computer readable storage medium, including instructions stored thereon, which instructions, when read and executed by one or more computer processors, cause the one or more computer processors to perform steps including:”
processor set configured to: (LI, para. 0010: “In some aspects, the techniques described herein relate to a system including at least one computer including a processor and a memory, wherein the at least one computer is configured to:”)
retrieve the first target sample of a set of target samples to be unlearned by the pre- trained ML model, wherein the first target sample is retrieved from a data source; (LI, para. 0047: “In accordance with aspects, datum 112 from dataset 102 may be provided to encoder 123 of target model 122 as input to target model 122. Moreover, Gaussian noise sample 113, which is a sample drawn from Gaussian noise, may be provided as input to encoder 125 of original model 124. Output 130 is output from encoder 123 of target model 122 based on input datum 112. Output 134 is output from encoder 125 of target model 124 based on Gaussian noise sample 113. Output 130 along with output 134 may be used to compute a distance to normal noise”;
LI, para. 0049: “In accordance with aspects, the steps described with respect to the components of FIG. 1 may be iterated a number of times and a total loss may be accumulated based on the number of iterations. A machine unlearning process may then update the target model 122 with the generated accumulated loss.”
LI, para. 0051: “Step 210 includes providing a first datum to a target model as input to the target model, wherein the first datum is retrieved from a forget dataset.”;
Examiner’s Note: datum 112 corresponds to recited “set of target samples to be unlearned” and is retrieved from “forget database 112”, where datum 112 is part of the set of samples to be unlearned during the iterations of Fig. 1)
retrieve a first set of supplemental samples associated with the first target sample; (LI, para. 0048: “In accordance with aspects, datum 114 from dataset 104 may be provided to encoder 123 of target model 122 as input to target model 122. Datum 114 may further be provided to encoder 125 of original model 124 as input to original model 124. Output 132 is output from encoder 123 of target model 122 based on input datum 114. Output 136 is output from encoder 125 of original model 124 based on input datum 114. Output 132 and output 136 may be used to compute a distance to original model 124 (i.e., loss L.sub.T, as discussed above).”;
LI, para. 0049: “In accordance with aspects, the steps described with respect to the components of FIG. 1 may be iterated a number of times and a total loss may be accumulated based on the number of iterations. A machine unlearning process may then update the target model 122 with the generated accumulated loss.”
LI, para. 0052: “Step 220 includes providing a sample drawn from Gaussian noise to an original model.”
Examiner’s Note: datum 114 corresponds to recited “set of supplemental samples associated with each target sample of the retrieved set of target samples”, where datum 114 is part of the set of samples during the iterations of Fig. 1)
calculate a target surprise score for the first target sample; (LI, para. 0047: “In accordance with aspects, datum 112 from dataset 102 may be provided to encoder 123 of target model 122 as input to target model 122. Moreover, Gaussian noise sample 113, which is a sample drawn from Gaussian noise, may be provided as input to encoder 125 of original model 124. Output 130 is output from encoder 123 of target model 122 based on input datum 112. Output 134 is output from encoder 125 of target model 124 based on Gaussian noise sample 113. Output 130 along with output 134 may be used to compute a distance to normal noise”;
Examiner’s Note: As depicted in Fig. 1, the distance to normal noise calculated from outputs 130 and 134 correspond to the recited “first surprise score”, which is the distance between the output 130 of the datum 112 with respect to the output 134 of the datum 113 with respect to Gaussian noise)
calculate a supplemental surprise score associated with each of the retrieved first set of supplemental samples associated with the first target sample; (LI, para. 0048: “In accordance with aspects, datum 114 from dataset 104 may be provided to encoder 123 of target model 122 as input to target model 122. Datum 114 may further be provided to encoder 125 of original model 124 as input to original model 124. Output 132 is output from encoder 123 of target model 122 based on input datum 114. Output 136 is output from encoder 125 of original model 124 based on input datum 114. Output 132 and output 136 may be used to compute a distance to original model 124 (i.e., loss LT, as discussed above).”;
Examiner’s Note: As depicted in Fig. 1, the distance to normal noise calculated from outputs 132 and 136 correspond to the recited “second surprise score”, which is the distance between the output 132 of the datum 114 from the target model, and output 136 of the datum 114 from the original model)
determine a first loss function based on the target surprise score for the first target sample ...; (LI, paras. 0036-0038:
PNG
media_image1.png
442
480
media_image1.png
Greyscale
LI, para. 0053: “Step 230 includes computing a first loss, wherein the first loss is based on target model output from processing the first datum and original model output from processing the sample drawn from Gaussian noise.”
Examiner’s Note: LN corresponds to the recited “first loss function” and relies on the difference between the forget sample and the Gaussian noise sample)
determine a second loss function based on the supplemental surprise score for each of the set of supplemental samples; (LI, paras. 0039-0040)
PNG
media_image2.png
376
484
media_image2.png
Greyscale
LI, para. 0056: “Step 260 includes computing a second loss, wherein the second loss is based on target model output from processing the second datum and original model output from processing the second datum.”;
Examiner’s Note: LT corresponds to the recited second loss function, which takes into account the difference between the use of the datum in the original and target models)
determine a unified loss function based on the first loss function, the second loss function, and a first hyper-parameter; and (LI, para. 0041: “The overall loss function may be the weighted combination of the above two losses with the weight alpha (α)”;
LI, para. 0049: “In accordance with aspects, the steps described with respect to the components of FIG. 1 may be iterated a number of times and a total loss may be accumulated based on the number of iterations. A machine unlearning process may then update the target model 122 with the generated accumulated loss. Moreover, once an accumulated loss is secured and target model 122 is updated with the accumulated loss, the process may begin again using a new datum 112 from dataset 102 and a new datum 114 from dataset 104. The process may be iterated through several new datums from each dataset, and the process may minimize an expectation value. Because a minimal expectation value may not be known, the process may iterate a fixed number of times.”)
update the pre-trained ML model based on the unified loss function, (LI, para. 0041: “The overall loss function may be the weighted combination of the above two losses with the weight alpha (α)”;
LI, para. 0049: “In accordance with aspects, the steps described with respect to the components of FIG. 1 may be iterated a number of times and a total loss may be accumulated based on the number of iterations. A machine unlearning process may then update the target model 122 with the generated accumulated loss. Moreover, once an accumulated loss is secured and target model 122 is updated with the accumulated loss, the process may begin again using a new datum 112 from dataset 102 and a new datum 114 from dataset 104. The process may be iterated through several new datums from each dataset, and the process may minimize an expectation value. Because a minimal expectation value may not be known, the process may iterate a fixed number of times.”;
LI, claim 4: “updating the target model based on the total loss, wherein executing the total loss accumulation process and updating the target model based on the total loss are secondary steps of the machine unlearning process, and wherein a complete iteration of the machine unlearning process includes executing the initial steps of the machine unlearning process and the secondary steps of the machine unlearning process.”)
However, LI fails to explicitly teach:
and the supplemental surprise score for each of the set of supplemental samples
wherein the ML model is updated for at least one epoch of a set of epochs, and wherein a count of the set of epochs corresponds to a second hyper-parameter associated with the training of the pre-trained ML model (TARUN, p. 7, section 6.3: “We train the model for 100 epochs with initial learning rate of 0.01 and reduce it on plateau by a factor of 0.1. The models are optimized on L1-loss with Adam optimizer. In FineTune, 5 epochs of training is done with a learning rate of 0.001. We run gradient ascent for 1 epoch with a learning rate of 0.001 on the AgeDB dataset. In Gaussian Amnesiac, 1 epoch of amnesiac learning is done with a learning rate of 0.001. In Blindspot, the blindspot model is trained for 2 epochs with a learning rate of 0.01. Subsequently, 1 epoch of unlearning is performed on the original model with a learning rate of 0.001.”;
Examiner’s Note: TARUN teaches the concept of setting a certain number of epochs for training as a hyperparameter: the LI-TARUN combination now trains the underlying pre-trained model of LI using a set number of training epochs as in TARUN)
However, in a related field of endeavor (machine unlearning, see p. 2, section 1), TARUN teaches and makes obvious:
and the supplemental surprise score for each of the set of supplemental samples (TARUN, pp. 4-5, section 4.1:
PNG
media_image3.png
350
468
media_image3.png
Greyscale
PNG
media_image4.png
482
466
media_image4.png
Greyscale
Examiner’s Note: the LI-TARUN combination now modifies the first loss equation of LI (see para. 0053) with equation (6) of TARUN so that information relating to retained samples are considered as additional data when calculating the first loss)
wherein the ML model is updated for at least one epoch of a set of epochs, and wherein a count of the set of epochs corresponds to a second hyper-parameter associated with the training of the pre-trained ML model
Before the effective date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of LI with TARUN as explained above. As explained by TARUN, one of ordinary skill would have been motivated to do so use the blindspot unlearning techniques of TARUN to “give[] the model a vague idea about the output distribution in the absence of the forget set from the training data.” (p. 4, section 4.1). One of ordinary skill in the art would understand that including additional data into the loss metrics will improve the accuracy of the loss metrics, and therefore improve the unlearning process as a whole.
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over LI in view of TARUN and further in view of US 20240386188 A1, hereinafter referenced as BANDEL.
Regarding Claim 4
LI and TARUN teach the method of claim 1 as explained above. However, LI and TARUN fail to explicitly teach:
receiving, by the computer, a first input associated with a selection of at least one sampling policy of a set of sampling policies, wherein the set of sampling policies comprises at least one of a random sampling policy, a syntax-based sampling policy, or a semantic-based sampling policy;
selecting, by the computer, the at least one sampling policy of the set of sampling policies based on the first input; and
retrieving, by the computer, the set of supplemental samples based on the selected at least one sampling policy.
However, in a related field of endeavor (training machine learning models, see para. 0008), BANDEL teaches and makes obvious:
receiving, by the computer, a first input associated with a selection of at least one sampling policy of a set of sampling policies, wherein the set of sampling policies comprises at least one of a random sampling policy, a syntax-based sampling policy, or a semantic-based sampling policy; (BANDEL, para. 0004: “The quality of paraphrases is often evaluated based on both the semantic similarity and the lexical and/or syntactic diversity of the paraphrase, when compared to the original sentence.”;
BANDEL, para. 0011: “In some embodiments, the input quality control vector is a three-dimensional vector representing (i) desired semantic similarity, (ii) desired syntactic distance, and (iii) desired lexical distance, of the output paraphrase relative to the input sentence.”
BANDEL, para. 0029: “In some embodiments, the present paraphrase generator model further allows for selection of input control quality values. Generally, given an input sentence, not all paraphrase qualities are achievable, because some sentences are more amenable to paraphrasing than others. For example, sentences containing named entities and numbers are much more difficult to paraphrase while keeping sentence meaning, and thus the potential lexical diversity of paraphrases involving such terms is relatively limited. ... Accordingly, in some embodiments, the present technique provides for control values that can be adjusted based on the predicted expected quality of an input sentence. In some embodiments, the present technique provides for a method which optimizes the expected quality of the paraphrases generated by the present paraphrase generator. Accordingly, given a paraphrasing task requirements, the present method optimizes the input quality controls to yield the desired quality of paraphrases.”;
BANDEL, para. 0074: “In some embodiments, paraphrase generation module 306 receives as input sentence s and control vector c=(c.sub.sem, c.sub.syn, c.sub.tex), and outputs a paraphrase s′.”
Examiner’s Note: BANDEL discloses that a paragraph generator receives a control vector having user-selected parameters for paraphrasing content, where such parameters include whether such paraphrased contact should emphasize syntactic vs. semantic similarity; the LI-TARUN-BANDEL combination now utilizes the paraphrase generator to generate the supplemental samples of LI, for example, to replace inputs having PII with a paraphrase that lacks such PII.)
selecting, by the computer, the at least one sampling policy of the set of sampling policies based on the first input; and (BANDEL, para. 0011: “In some embodiments, the input quality control vector is a three-dimensional vector representing (i) desired semantic similarity, (ii) desired syntactic distance, and (iii) desired lexical distance, of the output paraphrase relative to the input sentence.”
Examiner’s Note: the LI-TARUN-BANDEL combination now determines to use semantic similarity, for example, if the vector is (1,0,0) or syntactic similarity if the vector is (0,1,0))
retrieving, by the computer, the set of supplemental samples based on the selected at least one sampling policy. (BANDEL, para. 0004: “The quality of paraphrases is often evaluated based on both the semantic similarity and the lexical and/or syntactic diversity of the paraphrase, when compared to the original sentence.”;
BANDEL, para. 0011: “In some embodiments, the input quality control vector is a three-dimensional vector representing (i) desired semantic similarity, (ii) desired syntactic distance, and (iii) desired lexical distance, of the output paraphrase relative to the input sentence.”
BANDEL, para. 0029: “In some embodiments, the present paraphrase generator model further allows for selection of input control quality values. Generally, given an input sentence, not all paraphrase qualities are achievable, because some sentences are more amenable to paraphrasing than others. For example, sentences containing named entities and numbers are much more difficult to paraphrase while keeping sentence meaning, and thus the potential lexical diversity of paraphrases involving such terms is relatively limited. ... Accordingly, in some embodiments, the present technique provides for control values that can be adjusted based on the predicted expected quality of an input sentence. In some embodiments, the present technique provides for a method which optimizes the expected quality of the paraphrases generated by the present paraphrase generator. Accordingly, given a paraphrasing task requirements, the present method optimizes the input quality controls to yield the desired quality of paraphrases.”;
BANDEL, para. 0074: “In some embodiments, paraphrase generation module 306 receives as input sentence s and control vector c=(c.sub.sem, c.sub.syn, c.sub.tex), and outputs a paraphrase s′.”
Examiner’s Note: BANDEL discloses that a paragraph generator receives a control vector having user-selected parameters for paraphrasing content, where such parameters include whether such paraphrased contact should emphasize syntactic vs. semantic similarity; the LI-TARUN-BANDEL combination now utilizes the paraphrase generator to generate the supplemental samples of LI, for example, to replace inputs having PII with a paraphrase that lacks such PII.)
Before the effective date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of LI with TARUN and BANDEL as explained above. As explained by BANDEL, one of ordinary skill would have been motivated to do so in order to provide high quality paraphrases to be used in “various downstream tasks” and one of ordinary skill would understand such downstream tasks to include machine unlearning. (para. 0003).
Claims 11 and 19 rejected under 35 U.S.C. 103 as being unpatentable over LI in view of TARUN and further in view of US 20230153607 A1, hereinafter referenced as BHUTANI.
Regarding Claim 11
LI and TARUN teach the method of claim 1 as explained above. LI further teaches:
determining, by the computer, a unified loss function based on the first loss function, the second loss function,...; and (LI, para. 0041: “The overall loss function may be the weighted combination of the above two losses with the weight alpha (α)”;
LI, para. 0057: “Step 270 includes combining the first loss and the second loss with an alpha weighting to generate a weighted combination of the first loss and the second loss.”)
updating, by the computer, the pre-trained ML model based on the unified loss function. ((LI, para. 0041: “The overall loss function may be the weighted combination of the above two losses with the weight alpha (α)”;
LI, para. 0049: “In accordance with aspects, the steps described with respect to the components of FIG. 1 may be iterated a number of times and a total loss may be accumulated based on the number of iterations. A machine unlearning process may then update the target model 122 with the generated accumulated loss. Moreover, once an accumulated loss is secured and target model 122 is updated with the accumulated loss, the process may begin again using a new datum 112 from dataset 102 and a new datum 114 from dataset 104. The process may be iterated through several new datums from each dataset, and the process may minimize an expectation value. Because a minimal expectation value may not be known, the process may iterate a fixed number of times.”;
LI, claim 4: “updating the target model based on the total loss, wherein executing the total loss accumulation process and updating the target model based on the total loss are secondary steps of the machine unlearning process, and wherein a complete iteration of the machine unlearning process includes executing the initial steps of the machine unlearning process and the secondary steps of the machine unlearning process.”)
However, LI and TARUN fail to explicitly teach:
... and a first hyper-parameter associated with the training of the pre-trained ML model
However, in a related field of endeavor (machine learning models, see para. 0004), BHUTANI teaches and makes obvious:
... and a first hyper-parameter associated with the training of the pre-trained ML model (BHUTANI, para. 0045: “Other hyper-parameters such as learning rate, batch size, lambda weights λ.sub.(emb,ctr) in the total loss function all are appropriately tuned while training the model.”;
Examiner’s Note: the LI-TARUN-BHUTANI combination now tunes the weight alpha (α) of LIN using the tuning techniques of CHUTANI when the machine learning model is trained)
Before the effective date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of LI with TARUN and BHUTANI as explained above. As explained by BHUTANI, one of ordinary skill would have been motivated to do so in order to use tuning techniques to determine the optimal weighted alpha in connection with other hyperparameters as in BHUTANI. (para. 0045).
Claim 19 depends from claim 14 and claims a system that corresponds to the method of claim 11, and is therefore rejected for the same reasons explained above with respect to claims 11 and 14.
Claim 13 is rejected under 35 U.S.C. 103 as being unpatentable over LI in view of TARUN and further in view of US 20230419075 A1, hereinafter referenced as KOIKE.
Regarding Claim 13
LI and TARUN teach the method of claim 1 as explained above. However, LI and TARUN fail to explicitly teach:
wherein the first loss function corresponds to one of a margin ranking loss function or a SoftMax loss function, and wherein the second loss function corresponds to a regularization loss function.
However, in a related field of endeavor (artificial networks), KOIKE teaches and makes obvious:
wherein the first loss function corresponds to one of a margin ranking loss function or a SoftMax loss function, and (KOIKE, para. 0055: “The objective function is a combination of various functions including but not limited to: L1 loss; Lp norm; mean-square error; cross entropy; connectionist temporal classification loss; negative log likelihood; Kullback-Leibler divergence (KLD); cross covariance; structural similarity; cosine similarity; clustering loss; margin ranking loss; hinge loss; Huber loss; negative sampling; Wasserstein distance; triplet loss.”)
wherein the second loss function corresponds to a regularization loss function. (KOIKE, para. 0081: “ the regularization DNN block is trained to minimize a loss function to estimate S from Z”;
Examiner’s Note: KOIKE teaches that there are a number of alternative types of loss metrics, including margin ranking loss, and further teaches using a regularization block to apply to a loss function; the LI-TARUN-KOIKE combination now modifies the first loss function of LI to be a margin ranking loss function as in KOIKE, and further updates the second loss function of LI to use the regularization block of KOIKE)
Before the effective date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of LI with TARUN and KOIKE as explained above. One of ordinary skill would have been motivated to do so in order to utilize well-known and tested optimization algorithms for loss functions, including at least a margin loss function, and further in order to normalize, or regularize, the gradients being used for backpropagation.
Allowable Subject Matter
Claims 7-10 and 17-18 are objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
The following is a statement of reasons for the indication of allowable subject matter:
Claim 7 would be considered allowable since none of the references of record either alone or in combination fairly disclose or suggest the combination of limitations specified in claim 7, including at least:
wherein the modality of each target sample of the set of target samples is unimodal, and wherein the calculation for the first surprise score for a target sample corresponds to at least one of a calculation of a loss of the pre-trained ML model on the corresponding target sample, or a calculation of a perplexity of the pre-trained ML model on the corresponding target sample.
The closest prior art of record discloses:
US 20250103878 A1, hereinafter referenced as LI discloses a particular technique for machine unlearning, using a combination of outputs from a first datum, second datum, which are used in first and second loss functions to determine how to update a machine learning model to unlearn information from a forgetting dataset. (See Figs. 1-2, paras. 0045-0057).
Tarun, Ayush K., et al. "Deep regression unlearning." arXiv preprint arXiv:2210.08196 (2023), hereinafter referenced as TARUN discloses, with respect to machine unlearning, using various loss metrics). (pp. 4-5, section 4.1, equations (6)-(9)).
However, the examiner has found that the distinct feature of the Applicant's claimed invention over the prior art is the explicit claiming of the aforementioned limitations in combination with all the other limitations as specified in claim 7. The examiner further finds that one of ordinary skill would not have been motivated to combine the teachings of the prior art in the manner specified in claim 7 without the hindsight aid of Applicant’s disclosure. Therefore, because the prior art of record does not anticipate or make obvious the limitations of claim 7, such claim would be allowed over the prior art if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Claim 8 would be considered allowable since none of the references of record either alone or in combination fairly disclose or suggest the combination of limitations specified in claim 8, including at least:
wherein the modality of each target sample of the set of target samples is multimodal, and wherein the calculation of the first surprise score for a target sample corresponds to a calculation of a dot product of at least a first portion of the corresponding target sample in a first modality and a second portion of the corresponding target sample in a second modality.
The closest prior art of record discloses:
US 20250103878 A1, hereinafter referenced as LI discloses a particular technique for machine unlearning, using a combination of outputs from a first datum, second datum, which are used in first and second loss functions to determine how to update a machine learning model to unlearn information from a forgetting dataset. (See Figs. 1-2, paras. 0045-0057).
Tarun, Ayush K., et al. "Deep regression unlearning." arXiv preprint arXiv:2210.08196 (2023), hereinafter referenced as TARUN discloses, with respect to machine unlearning, using various loss metrics). (pp. 4-5, section 4.1, equations (6)-(9)).
However, the examiner has found that the distinct feature of the Applicant's claimed invention over the prior art is the explicit claiming of the aforementioned limitations in combination with all the other limitations as specified in claim 8. The examiner further finds that one of ordinary skill would not have been motivated to combine the teachings of the prior art in the manner specified in claim 8 without the hindsight aid of Applicant’s disclosure. Therefore, because the prior art of record does not anticipate or make obvious the limitations of claim 8, such claim would be allowed over the prior art if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Claim 9 would be considered allowable since none of the references of record either alone or in combination fairly disclose or suggest the combination of limitations specified in claim 9, including at least:
calculating, by the computer, a first set of mean values based on an aggregation of the second surprise score for each supplemental sample of the set of supplemental samples associated with a corresponding target sample; and
determining, by the computer, the first loss function based on the first set of mean values and the first surprise score for the corresponding target sample.
The closest prior art of record discloses:
US 20250103878 A1, hereinafter referenced as LI discloses a particular technique for machine unlearning, using a combination of outputs from a first datum, second datum, which are used in first and second loss functions to determine how to update a machine learning model to unlearn information from a forgetting dataset. (See Figs. 1-2, paras. 0045-0057).
Tarun, Ayush K., et al. "Deep regression unlearning." arXiv preprint arXiv:2210.08196 (2023), hereinafter referenced as TARUN discloses, with respect to machine unlearning, using various loss metrics). (pp. 4-5, section 4.1, equations (6)-(9)).
However, the examiner has found that the distinct feature of the Applicant's claimed invention over the prior art is the explicit claiming of the aforementioned limitations in combination with all the other limitations as specified in claim 9. The examiner further finds that one of ordinary skill would not have been motivated to combine the teachings of the prior art in the manner specified in claim 9 without the hindsight aid of Applicant’s disclosure. Therefore, because the prior art of record does not anticipate or make obvious the limitations of claim 9, such claim would be allowed over the prior art if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Claim 10 would be considered allowable since none of the references of record either alone or in combination fairly disclose or suggest the combination of limitations specified in claim 10, including at least:
calculating, by the computer, a second set of mean values based on an aggregation of the second surprise score for each supplemental sample of the set of supplemental samples; and
determining, by the computer, the second loss function based on the second set of mean values.
The closest prior art of record discloses:
US 20250103878 A1, hereinafter referenced as LI discloses a particular technique for machine unlearning, using a combination of outputs from a first datum, second datum, which are used in first and second loss functions to determine how to update a machine learning model to unlearn information from a forgetting dataset. (See Figs. 1-2, paras. 0045-0057).
Tarun, Ayush K., et al. "Deep regression unlearning." arXiv preprint arXiv:2210.08196 (2023), hereinafter referenced as TARUN discloses, with respect to machine unlearning, using various loss metrics). (pp. 4-5, section 4.1, equations (6)-(9)).
However, the examiner has found that the distinct feature of the Applicant's claimed invention over the prior art is the explicit claiming of the aforementioned limitations in combination with all the other limitations as specified in claim 10. The examiner further finds that one of ordinary skill would not have been motivated to combine the teachings of the prior art in the manner specified in claim 10 without the hindsight aid of Applicant’s disclosure. Therefore, because the prior art of record does not anticipate or make obvious the limitations of claim 10, such claim would be allowed over the prior art if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Claim 17 recites a system that corresponds to the method of claim 7, and would be allowable for the same reasons explained with respect to claim 7, if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Claim 18 recites a system that corresponds to the method of claim 8, and would be allowable for the same reasons explained with respect to claim 8, if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
US 20250200300 A1 (Yao). “According to one aspect of the present disclosure, a computing system is provided, including one or more processing devices configured to receive a forgetting dataset including a plurality of forgetting-target prompt-output pairs.” (para. 0002).
US 20250103877 A1 (Martin). “As detailed herein, unlearning causes a previously-trained neural network to forget a previously learned concept by modifying how the model operates during the processing of a prompt to generate an output response.” (para. 0016).
Zhou, Juexiao, et al. "Audit to Forget: A Unified Method to Revoke Patients' Private Data in Intelligent Healthcare." arXiv preprint arXiv:2302.09813 (2023). See Fig. 1 that shows patient data to forget and 3 types of loss functions.
Shaik, Thanveer, et al. "Exploring the Landscape of Machine Unlearning: A Comprehensive Survey and Taxonomy." arXiv preprint arXiv:2305.06360 (2023). See Table V on page 11 for different machine unlearning metrics including an anamnesis index.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MICHAEL C LEE whose telephone number is (571)272-4933. The examiner can normally be reached M-F 12:00 pm - 8:00 pm ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Omar Fernandez Rivas can be reached at 571-272-2589. 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.
/MICHAEL C. LEE/Examiner, Art Unit 2128
Prosecution Insights