DETAILED ACTION
This Final Office action is in response to Applicant’s Amendment filed on 07/28/2026. Claims 1, 4, 10-17, and 20-27 are pending. The effective filing date of the claimed invention is 10/13/2022.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 101
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.
Claims 1, 4, 10-17, 20-27 are rejected under 35 U.S.C. 101 because the claims are found to be directed to abstract idea.
Step 1 – Claims 1, 4, 10-17, 22-27 are process claims; claim 20 is machine claim; claim 21 is manufacture. Step 1 is satisfied.
Step 2A Prong 1 – Exemplary claim 1 (and similarly claims 20-21) recites the following abstract idea:
A method comprising:
accessing a preconfigured dataset comprised of synthetic samples and real world samples, wherein the synthetic samples and the real world samples are labeled in accordance with a downstream task (e.g. MPEP 2106.04(a)(2)(III));
using a reweighting algorithm and the real world samples to learn weights for the synthetic samples included in the preconfigured dataset, wherein each of the weights is indicative of an importance of a corresponding one of the synthetic samples with regard to using the synthetic sample to train a model for the downstream task (e.g. MPEP 2106.04(a)(2)(I); Recentive v. Fox);
generating a training synthetic dataset comprised of training synthetic samples, based on the weights learned for the synthetic samples included in the preconfigured dataset (e.g. MPEP 2106.04(a)(2)(I); Recentive v. Fox); and
training a machine learning model to perform the downstream task, using the training synthetic dataset (e.g. MPEP 2106.04(a)(2)(I); Recentive v. Fox; see also MPEP 2106.04(a)(2)(II)(A-C)).
When viewed alone and in ordered combination, these limitations are found to recite abstract idea.
Step 2A Prong 2 – Exemplary claim 1 is not found to integrate the abstract idea with practical application. Claim 1 recites additional limitations such as where the abstract idea is performed “at a device.” See MPEP 2106.05(f) apply it rationale. See also Recentive for a discussion of the machine learning, training aspects and how they relate to abstract idea, and were not found to integrate the abstract idea into practical application. When viewed alone and in ordered combination, these additional limitations are found to be directed to abstract idea.
Step 2B – Exemplary claim 1 is not found to recite significantly more. The additional limitation analysis of Step 2A Prong 2 is equally applied to Step 2B. MPEP 2106.05(d) indicates that “Another consideration when determining whether a claim recites significantly more than a judicial exception is whether the additional element(s) are well-understood, routine, conventional activities previously known to the industry. This consideration is only evaluated in Step 2B of the eligibility analysis.” In this regard, The courts have recognized the following computer functions as well‐understood, routine, and conventional functions when they are claimed in a merely generic manner (e.g., at a high level of generality) or as insignificant extra-solution activity.
i. Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016) (using a telephone for image transmission); OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363, 115 USPQ2d 1090, 1093 (Fed. Cir. 2015) (sending messages over a network); buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network); but see DDR Holdings, LLC v. Hotels.com, L.P., 773 F.3d 1245, 1258, 113 USPQ2d 1097, 1106 (Fed. Cir. 2014) (“Unlike the claims in Ultramercial, the claims at issue here specify how interactions with the Internet are manipulated to yield a desired result‐‐a result that overrides the routine and conventional sequence of events ordinarily triggered by the click of a hyperlink.” (emphasis added));
ii. Performing repetitive calculations, Flook, 437 U.S. at 594, 198 USPQ2d at 199 (recomputing or readjusting alarm limit values); Bancorp Services v. Sun Life, 687 F.3d 1266, 1278, 103 USPQ2d 1425, 1433 (Fed. Cir. 2012) (“The computer required by some of Bancorp’s claims is employed only for its most basic function, the performance of repetitive calculations, and as such does not impose meaningful limits on the scope of those claims.”).
The examiner finds that the claims touch upon these limitations that have been found to be WURC activity. Accordingly, the examiner finds that exemplary claim 1 is directed to abstract idea.
Dependent Claims – Claim 4 includes more abstract idea along with insignificant extra solution activity, selecting a particular data source or type of data to be manipulated. See MPEP 2106.05(g). Claim 10-17 is more insignificant extra solution activity, selecting a particular data source or type of data to be manipulated. See MPEP 2106.05(g). Claim 22 includes more abstract idea (e.g. MPEP 2106.04(a)(2)(I) and (III)), with iterative calculations under WURC activity Step 2B. Claims 23-27 include more insignificant extra solution activity. See MPEP 2106.05(g).
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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, 4, 10, 11, 12, 13, 14, 16, 20-24 are rejected under 35 U.S.C. 103 as being unpatentable over Jiarong Ye et al. (2020). Synthetic sample selection via reinforcement learning. Retrieved from arXiv:2008.11331v1 [cs.CV] 26 Aug 2020 (referred to as “Ye”) in view of Ren et al. (May 5, 2019). Learning to reweight examples for robust deep learning. Retrieved at arXiv:1803.09050v3 [cs.LG] 5 May 2019 (referred to as “Ren”)
With regard to claims 1, 20-21, Ye discloses the claimed method comprising: at a device (Ye, e.g. Fig. 1): accessing a preconfigured dataset comprised of synthetic samples and real world samples, wherein the synthetic samples and the real world samples are labeled in accordance with a downstream task (Ye e.g. page 1 The selected images are mixed with the original training data for improved training of image recognition systems, and describes synthetic data generated by GANs for augmenting real medical training datasets. Ye page 3 After sample selection, we train the classifier on the expanded dataset; for the downstream task at Ye, e.g., page 2 conditional GAN generates fake images from class label and noise vector, Ye page 6, As illustrated in Fig. 1, the candidate pool contains 1,024 HistoGAN-generated images for each class (CIN1, CIN2, CIN3, NORMAL) of the cervical histopathology dataset, and 2,048 images for each class (Negative, Positive) of the PCam dataset. Ye page 5 images annotated by same pathologist)
using a reweighting algorithm and the real world samples to learn weights for the synthetic samples included in the preconfigured dataset, wherein each of the weights is indicative of an importance of a corresponding one of the synthetic samples with regard to using the synthetic sample to train a model for the downstream task (See Ye page 2, describes that synthetic images have unequal quality and that the goal is “how to select high quality synthetic samples to improve medical image recognition systems”, Ye page 2, “a controller determines the actions applied to the candidate synthetic images and keeps updating based on the reward of downstream task performance”; Ye page 3, “outputs the binary action (select or discard) for each candidate synthetic image.” Ye does not disclose using a reweighting algorithm and the real world samples and optimize those example weights according to downstream validation performance. Ren teaches calculate a learned numerical importance weight for training sample, see Ren e.g. section 3.2, abstract. Therefore, it would have been obvious to one of ordinary skill in the reweighting art before the effective filing date of the claimed invention to modify Ye’s decision as to whether synthetic sample should be used, with Ren’s ability to calculate a learned numerical importance weight for training sample i, where this is beneficial in that Ren provides a known quantitative replacement/refinement for Ye’s binary sample-selection signal. See Ren, section 3.4, This helps both generalization and robustness to biases in the training set, which will be shown in our experiments);
generating a training synthetic dataset comprised of training synthetic samples, based on the weights learned for the synthetic samples included in the preconfigured dataset (see Ye, page 5, describing the action for every synthetic candidate as “If . . . candidate i is discarded, otherwise it is added to the original training set.”; Ye, page 6, the controller “outputs a binary action vector . . . for the further selection of the augmented training set” and the number of selected images is determined by the controller. This is essentially contrasting the training dataset according to the learned sample-level values/decisions. Then see Ren’s teaching that the usefulness of individual samples is represented as learned weights, and the combination satisfies the claimed invention, e.g. learn per-synthetic-sample importance weights according to real validation performance, then select the useful/high weight synthetic examples, then construct the resulting synthetic training dataset. See motivation to combine above.); and
training a machine learning model to perform the downstream task, using the training synthetic dataset (see Ye, page 3, “After sample selection, we train the classifier on the expanded dataset.).
With regard to claim 4, Ye further discloses the input synthetic preconfigured dataset includes a greater number of the synthetic samples than the input real world dataset real world samples (Ye e.g. page 5-6, . In total, there are 1,284Normal, 410CIN1, 481 CIN2,472CIN3 patches).
With regard to claim 10, Ye further discloses the training synthetic dataset is curated from the input synthetic samples included in the preconfigured dataset (Ye page 2-3, 6).
With regard to claim 11, Ye further discloses the training synthetic dataset is curated from the synthetic samples included in the preconfigured dataset by:determining a defined number of top-weighted synthetic samples included in the synthetic samples included in the preconfigured dataset, and selecting the top-weighted synthetic samples as the training synthetic dataset (Ye, e.g. page 2 selected or discarded samples, page 3 In each training iteration, it takes the feature vectors extracted from a ResNet34 [7] model trained on the original training images as the input, then outputs the binary action (select or discard) for each candidate synthetic image. Ren clearly teaches per-example learned weights, assigning weights to training examples based on how they affect performance on a clean validation set. See above. Ye expressly selects synthetic images and uses them in the augmented training set. Choosing the top-weighted data samples is an obvious modification of the combined references).
With regard to claim 12, Ye teaches the training synthetic dataset is actively synthesized from the input preconfigured dataset (Ye e.g. page 1 The selected images are mixed with the original training data for improved training of image recognition systems, and describes synthetic data generated by GANs for augmenting real medical training datasets.).
With regard to claim 13, Ye teaches the training synthetic dataset includes newly generated synthetic samples that augment the input preconfigured dataset (Ye e.g. page 1 The selected images are mixed with the original training data for improved training of image recognition systems, and describes synthetic data generated by GANs for augmenting real medical training datasets.).
With regard to claim 14, Ye teaches the newly generated synthetic samples include additional synthetic samples generated over a plurality of iterations (see Ye page 3, in each training iteration…).
With regard to claim 16, Ye teaches the target downstream task is a computer vision task (Ye Abstract, image classification performance is improved).
With regard to claim 22, Ye supplies thesynthetic/real context. Ren provides the rest at: wherein using the reweighting algorithm and the real world samples to learn the weights for the synthetic samples included in the preconfigured dataset includes, for at least iteration: selecting a subset of the synthetic samples from the preconfigured dataset, selecting a subset of the real world samples from the preconfigured dataset (Ren e.g. page 4, referring to SampleMiniBatch, Algorithm 1 steps 2-3), processing the subset of the synthetic samples, by a second machine learning model, to obtain predicted labels for the subset of the synthetic samples (Ren page 4-5, : ˆyg←Forward(Xg,yg,ˆθt)), computing a first loss between the labels of the synthetic samples in the subset of the synthetic samples and the predicted labels of the synthetic samples in the subset of the synthetic samples (Ren e.g. page 4-5, and then forms the training loss at 12: ˆ lf← n i=1wiC(yi,ˆyf,i)), computing a first gradient vector from the first loss, updating parameters of the second machine learning model based on the first gradient vector (Ren e.g. page 4-5, 13: ∇θt←BackwardAD(ˆ lf,θt)), processing the subset of the real world samples, by the second machine learning model with the updated parameters, to obtain predicted labels for the subset of the real world samples (Ren page 4-5, 8: ˆyg←Forward(Xg,yg,ˆθt)), computing a second loss between the labels of the real world samples in the subset of the real world samples and the predicted labels of the real world samples in the subset of the real world samples (Ren page 4, step 9: lg←1 m m i=1C(yg,i,ˆyg,i)), computing a second gradient vector from the second loss (Ren step 10 page 4), and determining the weights for the synthetic samples in the subset of the synthetic samples as a function of the second gradient vector (Ren page 4, step 11: ˜w←max(−∇,0);w← ˜w j ˜w+δ( j ˜w)). See combination above.
With regard to claim 23-24, Ye does not disclose claims 23-24. Ren does disclose negative validation loss gradient (section 3.3, 10) and batch normalization (section 3.1 normalizing the weights of all examples in a training batch so that they sum up to one. In other words,we….)
Claim 15 is rejected under 35 U.S.C. 103 as being unpatentable over Ye, Ren, in view of Amos et al. (2020). The differentiable cross-entropy method. PMLR (referred to as “Amos”).
With regard to claim 15, Ye teaches the newly generated synthetic samples include additional synthetic samples generated by: determining a defined number of top-weighted synthetic samples included in the synthetic samples included in the preconfigured dataset (Ye teaches the synthetic-sample context; Ren is used for the weighting aspect; Further Amos is referred to at page 3, “the number of elite candidates k to use to fit the new sampling distribution to, and the num ber of iterations T. The iterates of CEM are the parameters φ of the sampling distribution. CEM starts with an initial sampling distribution gφ1 (X) 2 Rn, and in each iteration t generates N samples from the domain [Xt,i]N i=1 ⇠ gφt (·), evaluates the function at those points vt,i := f✓(Xt,i), and re-fits the sampling distribution to the top-k samples by solving the maximum-likelihood problem1”), computing a generative parameter distribution of the top-weighted synthetic samples included in the synthetic samples included in the preconfigured dataset (see Amos, page 3, Proposition 1. For multivariate isotropic Gaussian sam pling distributions we have that φ = {µ,σ2} and eq. (2) has a closed-form solution given by the sample mean and variance of the top-k samples as µt+1 := 1/k P i2It Xt,i), selecting a plurality of synthesis parameters, based on the generative parameter distribution (see Ye, page 3, The iterates of CEM are the parameters φ of the sampling distribution. CEM starts with an initial sampling distribution gφ1 (X) 2 Rn, and in each iteration t generates N samples from the domain [Xt,i]N i=1 ⇠ gφt (·), evaluates the function at those points vt,i := f✓(Xt,i), and re-fits the sampling distribution to the top-k samples by solving the maximum-likelihood problem1), and generating the additional synthetic samples based on the plurality of synthesis parameters (see Ye, page 5). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to modify Ye’s synthetic sample context invention to include such sequential algorithm as shown in Amos, where this is beneficial in that “Optimization-based modeling is a way of integrating specialized operations and domain knowledge into end-to-end machine learning pipelines, typically in the form of a parameterized argmin operation.” Amos, page 3.
Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable over Ye, Ren, in view of Wang et al. (2021) Towards zero-label language learning, retrieved at arXiv:2109.09193v1 [cs.CL] 19 Sep 2021 (referred to as Wang).
With regard to claim 17, Ye and Ren do not teach where the target downstream task is a natural language processing task. See Wang, e.g. Abstract Section 3.2 etc.. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to modify Ye to include where the downstream task is NLP, as shown in Wang, as the benefit is such that allows task specific models to be trained from synthesized examples. Wang, 3.2.
Claim 25 is rejected under 35 U.S.C. 103 as being unpatentable over Ye, Ren, in view of Hinterstoisser et al. (2019). An annotation saved is an annotation earned: using fully synthetic training for object instance detection, retrieved from arXiv:1902.09967v1 [cs.CV] 26 Feb 2019 (referred to as “Hinterstoisser”).
With regard to claim 25, Ye does not teach claim 25. See Hinterstoisser throughout, and particularly at abstract, page 4-10. Therefore, it would have been obvious to one of ordinary skill in the retail object instance art before the effective filing date of the claimed invention to modify Ye to include such ability as this provides the improvement/benefit in that “This yields background images with realistic shapes and texture on top of which we render the objects of interest. During training, the data generation process follows a curriculum strategy guaranteeing that all foreground models are pre sented to the network equally under all possible poses and conditions with increasing complexity. As a result, we en tirely control the underlying statistics and we create optimal training samples at every stage of training. Using a set of 64 retail objects, we demonstrate that our simple approach enables the training of detectors that outperform models trained with real data on a challenging evaluation dataset.”
Claim 26 is rejected under 35 U.S.C. 103 as being unpatentable over Ye, Ren, in view of Erroll Wood et al. (2016). Learning an appearance-based gaze estimator from one million synthesized images. ETRA ’16. Referred to as “Wood”.
With regard to claim 26, Ye does not teach claim 26. See Wood at e.g. abstract, pages 1-2. Therefore, it would have been obvious to one of ordinary skill in the eye gaze detection art to modify Ye to include such eye gaze detections limitations, as shown in Wood, as the advantage/benefit of this combination is “for increased appearance variation.” Wood, page 2. See also , Wood, page 7, “Our results demonstrate that using our morphable model shape vari ation was beneficial for both the pixel errors on the MPIIGaze dataset (0.456 vs 0.477) and the angle estimation on the same dataset (9.95 vs 10.62). Angle error differences are statistically sig nificant (p < 0.001) according to a pair-wise t-test. See Figure 15 for more detailed comparisons.”
Claim 27 is rejected under 35 U.S.C. 103 as being unpatentable over Ye, Ren, in view of Riabi et al. (2021). Synthetic data augmentation for zero-shot cross-lingual question answering. ACL (referred to as “Riabi”).
With regard to claim 27, Ye does not teach claim 27. See Riabi at e.g. abstract, page 7018-7023. Therefore, it would have been obvious to one of ordinary skill in the cross-lingual training datasets art to include such cross-lingual ability as the benefits/improvements are shown in Riabi e.g. 7023, “In this work, we presented a method to gener ate synthetic QA dataset in a multilingual fashion, showing how QA models can benefit from it and reporting large improvements over the baselines. The proposed approach contributes to fill the gap between English and other languages, and is shown to generalize for languages not present in the syn thetic corpus (e.g. French, Italian, Korean).”
Response to Arguments
Applicant’s arguments with respect to claim(s) have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Peter Ludwig whose telephone number is (571)270-5599. The examiner can normally be reached Mon-Fri 9-5.
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, Fahd Obeid can be reached at 571-270-3324. 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.
/PETER LUDWIG/Primary Examiner, Art Unit 3627
Prosecution Insights