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 .
Claims 1-4, 6-8, 10-11, 13, 17, and 21-24 are pending in this application. Claims 1, 6, 11, and 17 are amended, claims 21-24 are new, and claims 5, 9, 15-16, 18, and 20 are canceled by applicant’s amendment filed 2 September 2026
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claims 1-2, 4-11, 15-18, 20-22, and 24 are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Hendrycks, Dan, et al. (“Natural adversarial examples,” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021; hereinafter “Hendrycks.” Papers from CVPR Conference 2021 were publicly available on 11 June 2021, as shown on their website, https://cvpr2021.thecvf.com/ {accessed by the examiner 10 June 2026}).
Regarding Claim 1, Hendrycks teaches a computer implemented method to establish a data collector (Abstract and section 1—a method establishes a data collector to collect a second dataset to improve the training of a machine learning model), the method comprising:
receiving, by a processor from a user device operated by a user utilizing a user interface, a collector identifier and information identifying an input data source comprising a prediction stream of a machine learning model, the prediction stream producing a first set of input data (section 1—an adversarial filtering method is received as a collector identifier that identifies a collection method. Section 3.1, in the “ImageNet-A Data Aggregation” portion, describes a process in more detail {second paragraph} that uses a prediction stream of ResNet-50 models as a first set of input data);
creating, by the processor, a collector queue with the collector identifier, the collector queue storing a second set of input data from the input data source, the collector queue decoupling ingestion of the first set of input data from asynchronous processing of the second set of input data by a post-queue workflow (section 3.1, the “ImageNet-A Data Aggregation” portion —a pre-queue workflow removes any images from the first set of input data that are correctly classified by the ResNet50 models or that either ResNet-50 model assigns greater than 15% confidence. The selected images are stored in a collector queue for subsequent filtering in a post-queue workflow, thus decoupling ingestion of the first set of input data from asynchronous processing of the second set of input data by a post-queue workflow);
receiving, by the processor from the user device, information identifying at least one of a pre-queue workflow and the post-queue workflow, the pre-queue workflow operating on the first set of input data to produce the second set of input data, the post-queue workflow asynchronously operating on the second set of input data stored in the collector queue to produce a first set of output data, and information identifying a data destination for the first set of output data (section 3.1, the “ImageNet-A Data Aggregation” portion—a pre-queue workflow removes any images from the first set of input data that are correctly classified by the ResNet50 models or that either ResNet-50 model assigns greater than 15% confidence. Consequently, a second round of filtering is performed in a post-queue workflow to create a shortlist, which is a first set of output data); and
outputting, by the processor to the data destination, the first set of output data for use in training a second machine learning model (sections 3.1 and 4—the first set of output data is a dataset that is used to train machine learning models, as evidenced by the experiments).
Regarding Claim 11, Hendrycks teaches a computing system (section 4—experiments that train and test a machine learning system imply a computing system) comprising:
a network interface communication device configured to receive an input data object, the input data object received from a data source comprising a prediction stream of a first machine learning model, wherein the input data object includes an input, at least a first predicted concept, and at least a first confidence score associated with the at least first predicted concept (section 1—an adversarial filtering method is received as a collector identifier that identifies a collection method. Section 3.1, in the “ImageNet-A Data Aggregation” portion, describes a process in more detail {second paragraph} that uses a prediction stream of ResNet-50 models as a first set of input data. The prediction stream comprises a classification {a first predicted concept} and a confidence score); and
a processor operably connected to the communication device (sections 3.1 and 4) and configured to:
identify at least a first collector associated with the data source; determine, based on a pre-queue workflow applying a first threshold to the at least first confidence score of the input data object, to add the input data object to a queue of the at least first collector, the queue decoupling ingestion of input data objects from asynchronous processing of the input data objects by a post-queue workflow (section 3.1, the “ImageNet-A Data Aggregation” portion —a pre-queue workflow removes any images from the first set of input data that are correctly classified by the ResNet50 models or that either ResNet-50 model assigns greater than 15% confidence; both the correct classification and the confidence score are thresholds. The selected images are stored in a collector queue for subsequent filtering in a post-queue workflow, thus decoupling ingestion of the first set of input data from asynchronous processing of the second set of input data by a post-queue workflow);
add the input data object to the queue of the at least first collector (section 3.1, as above);
determine, based on the post-queue workflow asynchronously applying a second threshold to attributes of the input data object, to pass the input data object from the queue to an output data sink (section 3.1, the “ImageNet-A Data Aggregation” portion—consequently, a second round of filtering is performed in a post-queue workflow to create a shortlist {a first set of output data} by applying a threshold number of appearances); and
transmit the input data object to the output data sink for use in training a second machine learning model (sections 3.1 and 4—the first set of output data is a dataset that is used to train machine learning models, as evidenced by the experiments).
Regarding Claim 2, Hendrycks teaches wherein the first set of input data is the same as the second set of input data (section 3.1—the selection of images to include in the second set of data is made using classifications and confidence scores of a classifier, so if all images in the first set of input data are classified incorrectly and with low confidence, then the second set of input data will be the entire first set of input data).
Regarding Claim 4, Hendrycks teaches wherein the pre-queue workflow applies at least a first threshold to attributes of the first set of input data (section 3.1, ImageNet-A Data Aggregation portion—a low-confidence threshold, such as 15%, is a first threshold applied to the first set of input data).
Regarding Claim 6, Hendrycks teaches wherein each set of input data from the prediction stream includes an input, at least a first predicted concept associated with the input, and at least a first confidence score associated with the at least first predicted concept (section 3.1, ImageNet-A Data Aggregation portion—the prediction stream from the ResNet-50 ensemble includes an input image data, a predicted classification {concept}, and a confidence score).
Regarding Claim 7, Hendrycks teaches wherein the pre-queue workflow applies a threshold to each set of input data (section 3.1, ImageNet-A Data Aggregation portion—thresholds are applied to the classification {correct/incorrect} and to the confidence score).
Regarding Claims 8 and 17, Hendrycks teaches wherein the threshold is a threshold associated with the at least first confidence score associated with the at least first predicted concept (section 3.1, ImageNet-A Data Aggregation portion—a threshold of 15% is applied to the confidence score).
Regarding Claim 10, Hendrycks teaches wherein the post-queue workflow is a threshold model (section 3.1, ImageNet-A Data Aggregation portion—the maximum number of images to include in the output set is a threshold model).
Regarding Claim 21, Hendrycks teaches wherein the pre-queue workflow and the post-queue workflow operate independently via the collector queue (section 3.1, ImageNet-A Data Aggregation portion—the post-flow workflow operates subsequent to the pre-queue workflow, therefore operating independently).
Regarding Claim 22, Hendrycks teaches wherein the collector queue stores the second set of input data for deferred processing by the post-queue workflow (section 3.1, ImageNet-A Data Aggregation portion—the post-flow workflow operates subsequent to the pre-queue workflow, so it is deferred processing).
Regarding Claim 24, Hendrycks teaches wherein a plurality of collectors receive input data objects from a plurality of data sources and transmit output to a common output data sink (section 3.1, ImageNet-A Data Aggregation portion—input data objects are received from an ensemble comprising a plurality of ResNet-50 models).
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 3 and 13 are rejected under 35 U.S.C. 103 as being unpatentable over Hendrycks, as applied to claims 1 and 11, above, in view of Cai et al. (U.S. 2020/0342265, hereinafter “Cai”).
Regarding Claim 3, Hendrycks does not specifically teach wherein the second set of input data is a random sample of the first set of input data. However, Cai teaches a second set of input data that is a random sample of a first set of input data (¶ [0024]).
All of the claimed elements were known in Hendrycks and Cai and could have been combined by known methods with no change in their respective functions. It therefore would have been obvious to a person of ordinary skill in the art at the time of filing of the applicant’s invention to combine the random sampling of Cai with the sets of input data of Hendrycks to yield the predictable result of wherein the second set of input data is a random sample of the first set of input data. One would be motivated to make this combination for the purpose of optimizing training of a machine learning model when a dataset is imbalanced (Cai, ¶ [0005]).
Regarding Claim 13, Hendrycks/Cai teaches wherein the pre-queue workflow is configured to select a random sample of data from the data source of the input data to be added to the queue, wherein the input data object is selected to be added to the queue (Cai, ¶ [0024]).
Claim 23 is rejected under 35 U.S.C. 103 as being unpatentable over Hendrycks, as applied to claim 11, above, in view of Duesterwald et al. (U.S. 2022/0172109, hereinafter “Duesterwald”).
Regarding Claim 23, Hendrycks does not specifically teach adjusting filtering criteria of the pre-queue workflow to mitigate data drift in the first set of output data. However, Duesterwald teaches adjusting filtering criteria of a workflow to mitigate data drift in a first set of output data (¶ [0038]—inputs from jobs can be filtered based on drift scenarios, indicating that the filtering criteria are adjusted based on drift).
All of the claimed elements were known in Hendrycks and Duesterwald and could have been combined by known methods with no change in their respective functions. It therefore would have been obvious to a person of ordinary skill in the art at the time of filing of the applicant’s invention to combine the filtering based on drift of Duesterwald with the filtering and first set of output data of Hendrycks to yield the predictable result of adjusting filtering criteria of the pre-queue workflow to mitigate data drift in the first set of output data. One would be motivated to make this combination for the purpose of improving performance by providing insights into drift scenarios (Duesterwald, ¶ [0004]).
Response to Arguments
Applicant’s arguments with respect to the rejections under 35 U.S.C. 101 have been fully considered and are persuasive. Specifying that the input data source comprises a prediction stream of a machine learning model introduces elements to the claims that cannot practically be performed in the human mind, so they are significantly more than a mental process The rejections under 35 U.S.C. 101 as an abstract idea have been withdrawn.
Applicant’s arguments filed 2 September 2026 have been fully considered but they are not persuasive. Upon further consideration, the examiner finds that Hendrycks teaches all of the amended limitations of claims 1 and 11. As detailed above, the rejection now relies upon different embodiments/operations of Hendrycks to map to the claims. The rejection above specifically references the “ImageNet-A Data Aggregation” portion of section 3.1. The pre-queue workflow takes a prediction stream from an ensemble of ResNet-50 machine learning models as input and filters this input to remove images that were correctly classified or were assigned a confidence score of over a 15% threshold. This produces the second set of input data, and are stored in a queue for subsequent processing. The subsequent processing is performed asynchronously by a post-queue workflow that further filters the second set of input data to produce a shortlist (the first set of output data) by applying a threshold to the number of images of each type.
Overall, the examiner finds that the claimed invention is essentially a two-stage filter that removes data from a dataset to produce a smaller dataset. Hendrycks teaches several methods of filtering datasets, including those recited by independent claims 1 and 11. Hendrycks also teaches most of the dependent claims, as also detailed above. Cai is still relied on to teach claims 3 and 13, and new prior art reference Duesterwald teaches new claim 23.
The rejections above no longer rely upon the citation to MPEP 2144.04 III, obviating the applicant’s arguments about this citation. The examiner notes, however, that the statement therein that “providing an automatic or mechanical means to replace a manual activity which accomplished the same result is not sufficient to distinguish over the prior art” is not a mere matter of the examiner’s opinion; it is a ruling by the courts about what is and is not patentable in view of the prior art.
Conclusion
Applicant’s amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HAL W SCHNEE whose telephone number is (571) 270-1918. The examiner can normally be reached M-F 7:30 a.m. - 6:00 p.m.
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, Michael Huntley can be reached at 303-297-4307. 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.
/HAL SCHNEE/ Primary Examiner, Art Unit 2129
Prosecution Insights