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 .
Status of Claims
The present application is being examined under the claims filed 12/11/2025. The status of the claims are as follows:
Claims 1-7, 9-17, and 19-22 are pending.
Claims 1 and 11 are amended.
Claims 8 and 18 are cancelled.
Response to Amendment
The Office Action is in response to Applicant’s communication filled 12/11/2025 in response to office action mailed 06/13/2025. The Applicant’s remarks and any amendments to the claims or specification have been considered with the results that follow.
Response to Arguments
Regarding 35 U.S.C. § 102 (p. 11-12)
Applicant argues that Klaiman’s “digital images” are “not equivalent” to the amended claim requirement that “each of the datasets comprises data items arranged logically as an array of rows and columns …”, and therefore Klaiman allegedly does not teach “obtaining a first dataset and a plurality of candidate datasets”, nor other elements tied to those datasets. Applicant further argues Klaiman does not teach “appending the selected candidate dataset to the first dataset” and “training a machine learning model using the enriched dataset”.
Examiner response:
These arguments are not persuasive because the present Office action rejects amended claims 1 and 11 under 35 U.S.C. § 103 over Klaiman in view of Das and further in view of Maor (see below §103 rejection of claims in this Office action). Accordingly, Applicant’s arguments directed to Klaiman’s alleged deficiencies are not persuasive because the rejections of amended independent claims 1 and 11 are based on Klaiman in view of Das and further in view of Maor, and the contested limitations are taught by Das and Maor as set forth below.
Regarding U.S.C. § 103 (p. 12-13)
Applicant further traverses the prior §103 rejection of claims 6-9 and 16-19 over Klaiman in view of Le, asserting that Klaiman does not teach the limitations of claims 1 and 11 and that Le is silent on those limitations.
Examiner response:
These arguments are not persuasive for at least the reason that the present Office action addresses the amended independent-claim limitations (including the rows/columns dataset structure and the enrich/train limitations) through the applied combination as set forth in the §103 rejection of this Office action, including Maor for the rows/column’s dataset structure and Das for the append/enrich and training aspects.
Regarding new claims 21 and 22 (p. 13)
New claims 21 and 22 depend from amended claims 1 and 11. To the extent Applicant asserts allowability based on the allowability of the amended independent claims, such argument is not persuasive because amended claims 1 and 11 remain unpatentable for the reasons stated. Further, the additional “repeat until threshold” subject matter is taught by Das, which teaches determining whether the model accuracy exceeds a threshold and, if not, appending additional data to the training dataset and looping back to perform another training iteration (see §103 rejection in this Office action).
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1, 11, 21, are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claim 1 is indefinite because it recites, inter alia, “calculating a plurality of mathematical representations, each for one of the a first dataset and one of the a plurality of candidate datasets, wherein calculating the mathematical representation of a dataset of the first dataset and the plurality of candidate datasets comprises: calculating a set of features of the dataset;”. This language is grammatically unclear and fails to distinctly define what is being represented and for which dataset(s), at least because:
the phrase “each for one of the a first dataset and one of the a plurality …” does not clearly define what each mathematical representation corresponds to (e.g., one per dataset, one per dataset pair, etc.); and
“a dataset of the first dataset and the plurality of candidate datasets” does not distinctly identify which dataset is being represented (first dataset, each candidate dataset, a selected dataset, or a combination), thereby rendering the scope of the step indeterminate. Additionally, claim 1 introduces “a plurality of candidate datasets” but later recites “selecting the candidate dataset”, which does not distinctly specify whether a single candidate dataset is selected from the plurality or whether multiple candidate datasets may be selected.
Claim 11 is indefinite for the same reasons as claim 1 because it recites corresponding ambiguous language including “calculating the mathematical representation of a dataset of the first dataset and the plurality of candidate datasets”, which fails to distinctly identify which dataset is being represented. Claim 11 further similarly introduces “a plurality of candidate datasets” and later recites “selecting the candidate dataset”, which does not distinctly specify the selection scope.
Claims 21 and 22 are indefinite because each recites “repeating the enriching and training until the accuracy metric satisfies a threshold”, but the scope of “the enriching” is unclear, such that it is not distinctly clear what action(s) must be repeated (e.g., repeating the entire “generating the enriched dataset” sequence, only “appending”, or some other subset).
Claim 21 recites the limitation "the enriching" in the phrase " in the recited repetition clause. There is insufficient antecedent basis for this limitation in the claim. The lack of antecedent basis here renders the scope of what must be “repeated” indeterminate, and therefore constitutes an aggravated antecedent basis defect.
Claim 22 recites the limitation "the enriching" in the phrase "repeat the enriching and training ..." in the recited repetition clause. There is insufficient antecedent basis for this limitation in the claim. The lack of antecedent basis here renders the scope of what must be “repeated” indeterminate, and therefore constitutes an aggravated antecedent basis defect.
Claim Rejections - 35 USC § 101
Claims 1-7, 9-17, and 19-22 are rejected under 35 U.S.C. 101 as being directed to a judicial exception (i.e., an abstract idea) without significantly more.
Statutory Basis: 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 claim 1
Claim 1 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 1 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“calculating a plurality of mathematical representations, each for one of the a first dataset and one of the a plurality of candidate datasets, wherein calculating the mathematical representation of a dataset of the first dataset and the plurality of candidate datasets comprises: calculating a set of features of the dataset;” – this limitation recites mathematical/algorithmic operations of computing “mathematical representations” (i.e., numerical representations/embeddings) of datasets. Such computation of representations in a mathematical concept because it is a mathematical calculation/manipulation of information. See MPEP § 2106.04(a)(I). Additionally, the recited calculating “features” (derived values) from data is a mathematical/analytical computation over information. Feature calculation is a mathematical concept because it involves mathematical manipulation/derivation of numerical or categorical descriptors. See MPEP § 2106.04(a)(I).
“selecting a candidate dataset based on the similarity levels; and” – selecting a candidate dataset “based on” similarity levels is a judgment/evaluation step that can be performed in the human mind (e.g., reviewing similarity results and choosing). This is a mental process because it is a decision based on evaluated information. See MPEP § 2106.04(a)(2)(III).
Claim 1 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application. The additional elements:
“the method comprising, using a processor:” – this is generic computer implementation (a processor) and does not impose a meaningful limit on how the abstract idea is performed beyond “do it on a computer”. This does not integrate the exception into a practical application. See MPEP § 2106.05(f) (mere instructions to apply on a computer).
“obtaining a first dataset and a plurality of candidate datasets;” – this is data gathering/collection prior to performing the abstract computations. Data gathering is insignificant extra-solution activity and does not integrate the exception into a practical application. See MPEP § 2106.05(g).
“feeding the set of features of the dataset to a first neural network trained to generate the mathematical representation;” – this limitation invokes a neural network at a high level of generality as a tool for generating the mathematical representation, but it does not recite a specific improvement to the neural network’s operation or to computer functionality (e.g., a particular architecture or training technique that improves performance of the computer/ML technology). As claimed, it is simply using a generic ML model to perform the abstract calculation. See MPEP § 2106.05(a).
“appending the selected candidate dataset to the first dataset to generate the enriched dataset;” - appending/combining datasets is routine data organization/manipulation performed as part of implementing the abstract idea. This is an extra-solution data handling step that does not meaningfully limit or practically apply the abstract idea. See MPEP § 2106.05(g).
“training a machine learning model using the enriched dataset;” – this limitation recites the result of training a model using the data prepared by the abstract steps, without specifying a particular technical improvement to training (e.g., a specific training algorithm improvement, model architecture improvement, or computing resource improvement). It applies the abstract idea in a generic ML context and does not integrate the exception into a practical application. See MPEP § 2106.05(f) (mere instructions to apply).
“wherein each of the datasets comprises data items arranged logically as an array of rows and columns, wherein each of the rows relates to a single entity and each of the columns comprises data items that pertain to a single data category.” – this limitation specifies a conventional data format/organization (tabular rows/columns). This is, at most, a field-of-use / data presentation constraint and does not integrate the abstract idea into a practical application. See MPEP § 2106.05(h).
Claim 1 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception. The additional elements are:
“the method comprising, using a processor:”
“obtaining a first dataset and a plurality of candidate datasets;”
“feeding the set of features of the dataset to a first neural network trained to generate the mathematical representation;”
“appending the selected candidate dataset to the first dataset to generate the enriched dataset;”
“training a machine learning model using the enriched dataset;”
“wherein each of the datasets comprises data items arranged logically as an array of rows and columns, wherein each of the rows relates to a single entity and each of the columns comprises data items that pertain to a single data category.”
These additional elements amount to generic, well-understood, routine, and conventional (WURC) computer implementation, routine data gathering/organization, and applying conventional ML training/model usage at a high level of generality. Individually and as an ordered combination, they do not add a specific unconventional technical feature that transforms the claim into significantly more than the abstract idea. See MPEP § 2106.05(d); § 2106.05(g); §2106.05(h).
Claims 2-7, 9-10 and 12-17, 19-20 – 101 Analysis Grouped Dependent Claims:
These claims depend from independent Claims 1 and 11, which have been determined to be directed to a judicial exception (a mental process) and do not include additional elements that amount to significantly more than the exception.
Claims 2-7, 9-10 and 12-17, 19-20 - Step 2A Prong One – Abstract Idea Identification
Claims 2-10 and 12-20 depend from independent claims 1 and 11, respectively, and recite additional limitations such as:
Calculating specific types of features (e.g., column interaction features, column statistics, ontologies) (Claims 2, 3, 12, 13);
Updating dataset representations with new data or enriching datasets (Claims 6-7, 16-17);
Computing similarity via known mathematical functions (e.g., cosine similarity or Euclidean distance) (Claims 10, 20).
These limitations fall within the judicial exceptions of mathematical concepts and mental processes:
Limitations directed to calculating representations, similarity scores, or derived statistical features are mathematical concepts under MPEP 2106.04(a), as they involve the manipulation of numerical data without reciting any specific mathematical formula or algorithm.
Steps such as selecting a dataset based on similarity, enriching another dataset using selected data, or inferring ontologies reflect high-level data analysis and reasoning steps, which can be performed by the human mind or with pen and paper. There are mental processes under MPEP 2106.04(a)(2)(III).
None of these limitations introduce any technological structure, mathematical formula, or concrete transformation of data. All recite steps that may be implemented mentally or represent additional processing of data that does not change the nature of the abstract idea. Therefore, Claims 2-7, 9-10 and 12-17, 19-20 are directed to an abstract idea (a mental process) under MPEP 2106.05(a).
Claims 2-7, 9-10 and 12-17, 19-20 - Step 2A Prong Two – Practical Application:
The additional limitations of claims 2-7, 9-10 and 12-17, 19-20 do not integrate the abstract idea into a practical application. For example:
Training a neural network using labeled dataset pairs (Claims 4, 5, 14, 15) constitutes a generic application of conventional machine learning, without any disclosed architecture, training optimization, or novel use case. This is a mere instruction to apply the abstract idea and is not sufficient under MPEP 2106.05(f).
Feeding features into a neural network is a field-of-use limitation because the neural network is invoked as a black box, without any disclosed architecture or structural improvement (MPEP 2106.05(g)).
Obtaining a plurality of candidate datasets is merely a data gathering step for use in the subsequent analysis and does not represent a technological advance (MPEP 2106.05(g)).
Applying the method to a time series (Claims 9 and 19) simply identifies a domain of application and imposes no technological constraint. This is a field-of-use limitation under MPEP 2106.05(f).
Other limitations, such as using exponential decay or updating embeddings, are well-known conventional data manipulation steps that do not reflect a specific improvement to data storage, transmission, or representation.
There is no indication that any dependent claim improves the functioning of computer or modifies how data is stored or transmitted. Therefore, claims 2-7, 9-10 and 12-17, 19-20 do not integrate the abstract idea into a practical application and fail Step 2A Prong Two.
Claims 2-7, 9-10 and 12-17, 19-20 - Step 2B – Inventive Concept:
These claims depend from independent claims 1 and 11, which have been determined to be directed to a judicial exception and to lack additional elements amounting to significantly more.
The remaining elements – such as the use of a neural network, processor, or memory, as well as training a neural network using labeled dataset pairs, feeding features into a neural network, obtaining a plurality of candidate datasets, and applying the method to a time series – are assessed here under Step 2B. These elements are recited at a high level of generality and perform generic computer functions (e.g., receiving features, computing a representation). There is no indication that the network uses a novel architecture, training method, or model behavior, nor do the claims reflect any unconventional data transformation.
These conclusions are consistent with court decisions cited in MPEP 2106.05(d)(II), which identify that storing, processing, or transmitting data using conventional computer models does not render a claim eligible and are well-understood, routine, and conventional when recited at a high level of generality. See MPEP § 2106.05(d)(II) “receiving or transmitting data over a network”, "electronic record keeping”, and "storing and retrieving information in memory”.
More instructions to apply a judicial exception (see MPEP 2106.05(f)) and using a generic computer as a tool (see MPEP 2106.05(f)(II), 2106.05(d)) cannot amount to significantly more than the judicial exception itself.
Merely limiting a judicial exception to a particular field of use (see MPEP 2106.05(h)) cannot amount to significantly more than the judicial exception.
Claims 2-7, 9-10 and 12-17, 19-20 - Conclusion for Dependent Claims:
Accordingly, when considered individually or in combination, Claims 2-7, 9-10 and 12-17, 19-20 do not include additional elements that integrate the abstract idea into a practical application or amount to significantly more. Claims 2-7, 9-10 and 12-17, 19-20 fail Step 2B and are rejected under 35 U.S.C. 101.
Regarding claim 11
Claim 11 is the system claim directly analogous to method claim 1, reciting the same substantive subject matter in “a memory” and “a processor configured to” form (e.g., obtaining datasets, calculating mathematical representations and similarity levels, selecting a candidate dataset based on the similarity levels, appending to generate an enriched dataset, and training using the enriched dataset). Therefore, for the same reasons set forth with respect to claim 1, claim 11 is rejected under 35 U.S.C. § 101 as being directed to a judicial exception and as failing to integrate the exception into a practical application and failing to recite significantly more.
Regarding claim 21
Claim 21 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 21 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“testing the machine learning model;” – this limitation recites evaluating performance (a form of analysis/evaluation of information). Such evaluation can be performed mentally (at least in concept) by reviewing outputs and determining performance and therefore constitutes a mental process. See MPEP § 2106.04(a)(2)(III).
“calculating an accuracy metric;” – this limitation recites calculating a quantitative “accuracy metric”, which is a mathematical calculation/relationship over data. As claimed, it is a mathematical concept under MPEP § 2106.04(a)(2)(I).
“repeating the enriching and training until the accuracy metric satisfies a threshold.” – this limitation recites an iterative control loop based on the evaluation/accuracy calculation (“until … satisfies a threshold”), which is a mental decision rule and/or mathematical comparison applied to the computed metric. See MPEP § 2106.04(a)(2)(III); § 2106.04(a)(2)(I).
Claim 21 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application.
Claim 21 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception.
Regarding claim 22
Claim 22 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a system.
Claim 22 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“test the machine learning model;” – this limitation recites evaluating performance (a form of analysis/evaluation of information). Such evaluation can be performed mentally (at least in concept) by reviewing outputs and determining performance and therefore constitutes a mental process. See MPEP § 2106.04(a)(2)(III).
“calculate an accuracy metric;” – this limitation recites calculating a quantitative “accuracy metric”, which is a mathematical calculation/relationship over data. As claimed, it is a mathematical concept under MPEP § 2106.04(a)(2)(I).
“repeat the enriching and training until the accuracy metric satisfies a threshold.” – this limitation recites an iterative control loop based on the evaluation/accuracy calculation (“until … satisfies a threshold”), which is a mental decision rule and/or mathematical comparison applied to the computed metric. See MPEP § 2106.04(a)(2)(III); § 2106.04(a)(2)(I).
Claim 22 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application.
Claim 22 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception.
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 1-5, 10-15, and 20-22 are rejected under 35 U.S.C. 103 as being unpatentable over Eldad Klaiman et al. (US20220139072A1, henceforth “Klaiman”) in view of Mainak Das (US20220309250A1), henceforth “Das”) and further in view of Meir Maor et al. (US20200372414A1), henceforth “Maor”).
Regarding claim 1, Klaiman in view of Das and further in view of Maor, teach a method for training a machine learning model using an enriched dataset, the method comprising, using a processor:
“generating the enriched dataset by: obtaining a first dataset and a plurality of candidate datasets;” – Klaiman teaches this limitation. Klaiman teaches providing a search image (corresponding to the first dataset) and performing a similarity search in an image database containing a plurality of images (corresponding to candidate datasets):
“providing a digital search image … performing, by a similarity search engine, a similarity search in an image database of digital tissue sample images …” (Klaiman, p. 6, ¶[0060])
“calculating a plurality of mathematical representations, each for one of the a first dataset and one of the a plurality of candidate datasets, wherein calculating the mathematical representation of a dataset of the first dataset and the plurality of candidate datasets comprises: calculating a set of features of the dataset; – Klaiman teaches this limitation. Klaiman teaches extracting a search feature vector from the search image and comparing against feature vectors for images in the database (i.e., representations for the first dataset and candidate datasets):
“extracting … a search feature vector from the search image” (Klaiman, p. 6, ¶[0060])
And discloses determining similarity with:
“feature vectors extracted … for each of the images in the image database” (Klaiman, p. 6, ¶[0060])
Klaiman further teaches that such database feature vectors may be precomputed and stored with the images:
“feature vectors … are pre-computed and stored in the database in association with the image” (Klaiman, p. 6, ¶[0061])
“and feeding the set of features of the dataset to a first neural network trained to generate the mathematical representation;” – Klaiman teaches this limitation. Klaiman teaches a trained Siamese network / vector-output MLM that extracts a feature vector from each input image/tile:
“trained sub-networks … adapted to extract a feature vector from each of the input images” (Klaiman, p. 12, ¶[0145])
“calculating a plurality of similarity levels, each indicative of the similarity between the mathematical representation of one of the plurality of candidate datasets and the mathematical representation of the first dataset;” – Klaiman teaches this limitation. Klaiman teaches comparing feature vectors and outputting a similarity label/value:
“output layer … compares the two feature vectors and … outputs a similarity label 508 as a function of the two feature vectors …” (Klaiman, p. 12, ¶[0145])
“selecting the candidate dataset based on the similarity levels;” – Klaiman teaches this limitation. Klaiman teaches returning the images whose feature vectors are most similar to the search feature vector:
“returning the ones of the images in the database whose feature vectors are the most similar to the search feature vector …” (Klaiman, p. 6, ¶[0060])
“a similarity search … identifying similar images … most similar database image” (Klaiman, p. 13, ¶[0154])
Klaiman does not teach these limitations:
“and appending the selected candidate dataset to the first dataset to generate the enriched dataset;”
“and training a machine learning model using the enriched dataset;”
“wherein each of the datasets comprises data items arranged logically as an array of rows and columns, wherein each of the rows relates to a single entity and each of the columns comprises data items that pertain to a single data category.”
Das, however, teaches these limitations:
“and appending the selected candidate dataset to the first dataset to generate the enriched dataset;” – Das teaches extracting an additional batch of data and appending it to an existing training dataset:
“Extract an additional batch of randomly selected, unlabeled data from the historical case records and append to the existing training dataset 470” (Das, FIG. 4, block 470)
“and training a machine learning model using the enriched dataset;” – Das teaches training a classification model based on the training dataset, and retraining iteratively as the training dataset is appended/updated:
“the intermediate classification model is trained based on the current state of the training data set.” (Das, p. 6, ¶[0050])
Das does not teach this limitation:
“wherein each of the datasets comprises data items arranged logically as an array of rows and columns, wherein each of the rows relates to a single entity and each of the columns comprises data items that pertain to a single data category.”
Maor, however, teaches this limitation:
“wherein each of the datasets comprises data items arranged logically as an array of rows and columns, wherein each of the rows relates to a single entity and each of the columns comprises data items that pertain to a single data category.” – Maor discloses that a dataset may be represented as a table/matrix/database where each row represents a respective data instance (entity) and each column represents a respective field (data category):
“… a table wherein each row represents a secondary data instance and each column represents a secondary field” (Maor, p. 1, ¶[0012])
And further teaches the same row/column semantics in claim form:
“the … dataset is represented as a table, each row … indicates a respective … data instance, each column … indicatives a respective … field …” (Maor, p. 15, claim 4)
It would have been obvious to one of skill in the art to represent the datasets used in Klaiman’s similarity-search and Das’s enrichment/training workflow in the conventional tabular/matrix form taught by Maor (rows corresponding to entities/instances; columns corresponding to data categories/fields), because such representations are a routine and well-understood way to organize datasets for machine learning processing, storage, and manipulation. And it would have been obvious to a POSITA to modify Klaiman’s similarity-based selection framework to append the dataset(s) selected as similar to the existing dataset to form an enriched dataset, and then to train/retrain the machine learning model using the enriched dataset, as taught by Das, in order to increase the amount of training data and improve model performance/accuracy – a predictable result explicitly contemplated by Das’s iterative “append more data to the training dataset and train based on the updated dataset” approach.
Regarding Claim 2, Klaiman in view of Das and further in view of Maor, teaches the method of Claim 1, wherein:
calculating column interaction features, wherein the column interaction features are features related to interactions between different columns of data from the plurality of columns of data; - Klaiman teaches:
“According to embodiments, the MLM comprises routines for extracting a plurality of features that are low level descriptors… In addition, or alternatively, the extracted
features can comprise domain information descriptors giving information about objects and events in the respective biomedical domain.” (Klaiman, p. 5, paragraph [0051])
“A concrete example would be objects having been automatically identified to be particular cell components… Preferably, these domain information descriptors are identified in the received digital images fully automatically… ” (Klaiman, p.5, paragraph [0051])
This discloses that these features relate to inferred domain-specific relationships between columns, such as associations between cell morphology and biomarker channels, qualifying as “column interaction features” inferred by learned patterns across feature dimensions (i.e., columns).
calculating features related to statistics of a column of data from the plurality of columns of data – Klaiman teaches this:
“According to embodiments, the MLM comprises routines for extracting a plurality of features that are low level descriptors which give a description about color, shape, brightness, contrast, textures” (Klaiman, p.5, paragraph [0051])
These features correspond directly to statistical metrics of individual feature columns derived from image tiles (e.g., color intensity, brightness, contrast), satisfying the claim requirement for statistical column-level features.
predicting an ontology of a column of data from the plurality of columns of data – Klaiman teaches:
“[The MLM can] comprise domain information descriptors giving information about objects and events in the respective biomedical domain. A concrete example would be objects… identified to be particular cell components, e.g., nuclei, cell membranes…” (Klaiman, p.5, paragraph [0051])
The cited passages describe inferring biomedical ontology labels (e.g., “tumor cells”, “stroma cells”, or cells expressing biomarkers) from features, aligning with the notion of “predicting an ontology” for a data column (where a column could represent features derived from image regions).
Claim 2 depends from claim 1; therefore, the same motivation to combine recited in claim 1 applies to claim 2.
Regarding Claim 3, Klaiman in view of Das and further in view of Maor, teaches the method of Claim 2, wherein generating column interaction features for a pair of columns of data comprises:
inferring pairs of data items from different columns using a second neural network to generate inferred values – Klaiman teaches:
“Each sub-network of the Siamese network is adapted to extract a multi-dimensional feature vector from a respective one of two image tiles provided as input. The network is trained on a plurality of tile pairs… with the objective that tile pairs depicting similar tissue patterns should have outputs (feature vectors) that are close… and tile pairs depicting dissimilar tissue patterns should have outputs that are far.” (Klaiman, p.4, paragraph [0045])
This teaches that the Siamese network architecture computes intermediate outputs (feature vectors) from paired input data tiles (analogous to “columns of data”).
and providing the inferred values into a pooling layer – Klaiman teaches:
“A one-dimensional feature vector… is extracted from one of the two input images by a respective one of the two sub networks. Thereby, the last hidden layer… is adapted to compute the feature vector and provide the feature vector to the output layer.” (Klaiman, p.11, paragraph [0132])
This teaches the output layer functions as a pooling/comparison unit where the two feature vectors (inferred values) are aggregated to determine similarity, which corresponds to the pooling/comparing functionality claimed.
Claim 3 depends from claim 2 which depends from claim 1; therefore, the same motivation to combine recited in claim 1 applies to claim 3.
Regarding Claim 4, Klaiman in view of Das and further in view of Maor, teaches the method of Claim 1, wherein training the first neural network comprises:
using labeled pairs of sets of features to train a Siamese neural network, wherein a label of a pair indicates whether the two sets of features in the pair pertain to a same dataset – Klaiman teaches:
“the Siamese neuronal network is trained on the pairs of tiles using a loss function such that the similarity of the feature vectors extracted by the two sub-networks for the two tiles of the pair respectively correlates with the similarity of the tissue patterns depicted in the two tiles of the pair.” (Klaiman, p.4, paragraph [0044])
“The network is trained on a plurality of tile pairs having been automatically annotated with proximity-based tissue pattern-similarity labels… tile pairs depicting similar tissue patterns should have outputs (feature vectors) that are close… and tile pairs depicting dissimilar tissue patterns should have outputs that are far… ” (Klaiman, p.4, paragraph [0045])
“… label for the tile pair: 0 if they are labeled ‘similar’… 1 if they are labeled ‘dissimilar’” (Klaiman, p.4, paragraph [0046])
These disclosures teach training a Siamese network using labeled pairs where the label indicates the degree of similarity, effectively allowing the model to learn whether the data samples belong to the same semantic class – an analogue for determining whether two sets of features pertain to the same dataset.
Claim 4 depends from claim 1; therefore, the same motivation to combine recited in claim 1 applies to claim 4.
Regarding Claim 5, Klaiman in view of Das and further in view of Maor, teaches the method of Claim 4, comprising generating the labeled pairs of sets of features by:
obtaining a first labeled pair of datasets – Klaiman teaches:
“The creation of the annotated training data set comprises selecting a plurality of pairs of tiles and automatically assigning a label to each pair.” (Klaiman, p.9, paragraph [0107])
“The label is an indicator of the degree of similarity of the two tissue patterns depicted by the two tiles of the pair.” (Klaiman, p.9, paragraph [0107])
These labeled tile pairs constitute the “first labeled pair of datasets”.
selecting a subset of each dataset of the pair of datasets to generate a second labeled pair of datasets –
“Alternatively, this step can comprise randomly selecting a subset of the available distant tiles and creating a tile pair for each of the selected distant tiles by adding the start tile to the selected distant tile.” (Klaiman, p.10, paragraph [0117])
This discloses selection of subsets from each dataset (e.g., start tile and distant tile) to form a new labeled pair.
calculating the set of features for each dataset of the second pair of datasets –
“A one-dimensional feature vector 410, 420 is extracted from one of the two input images by a respective one of the two sub networks. […] The last hidden layer 408, 418 of each network is adapted to compute the feature vector and provide the feature vector to the output layer 424.” (Klaiman, p.11, paragraph [0132])
This confirms that features (in the form of feature vectors) are calculated for each of the datasets (tiles) in the second pair.
Claim 5 depends from claim 4 which depends from claim 1; therefore, the same motivation to combine recited in claim 1 applies to claim 5.
Regarding Claim 10, Klaiman in view of Das and further in view of Maor, teaches the method of Claim 1, wherein calculating the level of similarity between the first dataset and a candidate dataset comprises one of:
calculating Euclidean distance between the mathematical representation of the first dataset and the mathematical representation of the candidate dataset – Klaiman teaches this:
“a loss function that computes the loss in the form of the actual normalized distance between the two feature vectors (e.g. zero distance is label zero, i.e. similar and high distance is label one, i.e. dissimilar).” (Klaiman, p.5, paragraph [0052])
This passage confirms that Euclidean distance (or a direct L2-based proxy) is used as the similarity metric between representations (feature vectors).
calculating cosine similarity between the mathematical representation of the first dataset and the mathematical representation of the candidate dataset – while Klaiman does not use the words “cosine similarity” verbatim, it does teach:
“Each sub-network… is adapted to extract a multi-dimensional feature vector… The network is trained… with the objective that tile pairs depicting similar tissue patterns should have outputs (feature vectors) that are close… and tile pairs depicting dissimilar tissue patterns should have outputs that are far.” (Klaiman, p.4, paragraph [0045])
Because the learned feature vectors are numerical embeddings, and similarity is computed between them using geometric metrics, a POSITA at the time of the claimed invention would recognize cosine similarity as an obvious alternative to Euclidean distance.
or training a second machine learning model to calculate the level of similarity using labeled pairs of mathematical representations and feeding the mathematical representations of the first dataset and the mathematical representation of the candidate dataset to the trained second machine learning model to calculate the level of similarity – Klaiman explicitly teaches this:
“The Siamese neuronal network… is trained on a plurality of tile pairs having been automatically annotated with proximity-based tissue pattern similarity labels” (Klaiman, p.4, paragraph [0045])
“The output layer of the trained Siamese neuronal network is adapted to compute a label for each tile pair provided as input as a function of the two feature vectors.” (Klaiman, p.4, paragraph [0043])
“A one-dimensional feature vector… is extracted from one of the two input images by a respective one of the two sub networks. Thereby, the last hidden layer… is adapted to compute the feature vector and provide the feature vector to the output layer.” (Klaiman, p.11, paragraph [0132])
This discloses a second trained model (the Siamese architecture) that receives a pair of mathematical representations (feature vectors) and returns a similarity classification label, satisfying this limitation.
Claim 10 depends from claim 1; therefore, the same motivation to combine recited in claim 1 applies to claim 10.
Regarding Claims 11-15 and 20-22
Claims 11-15 and 20, 22 are the system claims directly analogous in scope to method Claims 1-5, 10, and 21, respectively. Each system claim reciting the same limitations as its corresponding method claim, rewritten in system form (e.g., Claim 11 mirrors Claim 1 but as a system; Claim 12 mirrors Claim 2, claim 13 mirrors claim 3; claim 14 mirrors claim 4; claim 15 mirrors claim 5; claim 20 mirrors claim 10; claim 22 mirrors claim 21).
Accordingly, the 35 U.S.C. 103 rejections for Claims 1-5, 10, and 21 over Klaiman in view of Das and further in view of Maor are equally applicable to Claims 11-15, 20, and 22, and the references cited apply with equal force.
Regarding claim 21, Klaiman in view of Das and further in view of Maor, teach the method of claim 1, comprising:
“testing the machine learning model;” – Klaiman teaches this limitation. Klaiman teaches testing/validation of the trained Siamese network on a testing set:
“the accuracy of the trained Siamese network was validated on the Camelyonl 6 testing set” (Klaiman, p. 12, ¶[0138])
“calculating an accuracy metric;” – Klaiman teaches this limitation. Klaiman teaches calculating an accuracy metric for the test tile pairs, including computing the Global Average Descriptor Distance Ratio (ADDR):
“Then the Global Average Descriptor Distance Ratio (ADDR) is computed for the test tile pairs.” (Klaiman, p. 12, ¶[0139])
Klaiman does not teach this limitation:
“and repeating the enriching and training until the accuracy metric satisfies a threshold.”
Das, however, teaches this limitation:
“and repeating the enriching and training until the accuracy metric satisfies a threshold.” – Das teaches determining whether a model accuracy exceeds an accuracy threshold, and if the threshold is not met, extracting an additional batch of data and appending it to the existing training dataset (i.e., enriching), and then looping back to perform another training iteration:
“determin[ing] whether the accuracy of the intermediate classification model exceeds a predetermined or configurable accuracy threshold.” – Das, p. 6, ¶[0051])
Das then discloses that if the model is:
“not yet sufficient to achieve the desired accuracy threshold ("No" at block 450) ... an additional batch … is extracted and appended to the existing training dataset” (Das, p. 6, ¶[0056]; FIG. 4 block 470)
“processing may loop back to block 440 to perform another training iteration.” (Das, p. 6, ¶[0058])
Thus, Das teaches repeating enrichment (by appending additional data to the training dataset) and repeating the training iterations until a threshold accuracy condition is satisfied.
It would have been obvious to a POSITA to incorporate Das’s threshold-based iterative retraining control into Klaiman’s trained-model evaluation framework (as applied to claim 1’s enrich/train process), in order to continue enrichment and retraining until a desired accuracy threshold is met, which is a predictable and routinely desired objective in machine learning development and model deployment workflows.
Regarding claim 22, , Klaiman in view of Das and further in view of Maor, teach the system of claim 11, wherein the processor is further configured to:
“test the machine learning model;” – Klaiman teaches this limitation. Klaiman teaches testing/validation of the trained Siamese network on a testing set:
“the accuracy of the trained Siamese network was validated on the Camelyonl 6 testing set” (Klaiman, p. 12, ¶[0138])
“calculate an accuracy metric;” – Klaiman teaches this limitation. Klaiman teaches calculating an accuracy metric for the test tile pairs, including computing the Global Average Descriptor Distance Ratio (ADDR):
“Then the Global Average Descriptor Distance Ratio (ADDR) is computed for the test tile pairs.” (Klaiman, p. 12, ¶[0139])
Klaiman does not teach this limitation:
“and repeat the enriching and training until the accuracy metric satisfies a threshold.”
Das, however, teaches this limitation:
“and repeat the enriching and training until the accuracy metric satisfies a threshold.” – Das teaches this limitation in system/processor form. Das discloses processor-executed instructions to determine whether accuracy exceeds an accuracy threshold and, if not, to enrich the training dataset by appending additional data and to repeat training iterations. Das teaches instructions to:
“determine whether the accuracy … exceeds a predetermined or configurable accuracy threshold.” – Das, p. 6, ¶[0051])
If the model is not sufficient to achieve the desired accuracy threshold, Das teaches extracting an additional batch and:
“append[ing] to the existing training dataset” (Das, p. 6, ¶[0056]; FIG. 4 block 470)
And Das teaches that:
“processing may loop back to block 440 to perform another training iteration.” (Das, p. 6, ¶[0058])
It would have been obvious to a POSITA to configure the processor of Klaiman’s system (as applied to claim 11) to implement Das’s threshold-based iterative retraining/enrichment control, in order to repeat enrichment and retraining until the accuracy metric satisfies a desired threshold, which is a predictable and routinely desirable objective in ML systems for ensuring acceptable performance prior to deployment.
Claims 6-7, 9, 16-17, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Klaiman in view of Das and further in view of Moar, and further in view of Meinh Le (US20220094649A1, henceforth "Le").
Regarding Claim 6, Klaiman in view of Le teaches the method of Claim 1, comprising updating a current mathematical representation of a first candidate dataset of at least one of the plurality of candidate datasets by:
obtaining new data pertaining to the first candidate dataset generating a new mathematical representation for the new data and combining the new mathematical representation with the current mathematical representation – Klaiman does not explicitly teach this, however LE teaches this:
“the system may determine a third output based on a weighted average of the first output and the second output.” (Le, p.10, paragraph [0086])
“… [wherein] the system may determine the third output based on the weighted average of the first output and the second output comprises determining a first weight for the first output and a second weight for the second output, wherein the first weight is greater than the second weight.” (Le, p.10, paragraph [0086])
This teaches the concept of incrementally updating or refining a model or representation using a combination (e.g., weighted average) of the old and the new representations, consistent with the limitation of combining current and new mathematical representations as claimed.
It would have been obvious to one of ordinary skill in the art at the time of the claimed invention to modify the system of Klaiman by incorporating the update mechanism taught by Le, wherein new data representations are combined with existing ones, in order to dynamically refine or enhance dataset similarity calculations over time. Such a combination would improve the adaptability and responsiveness of the enrichment system described by Klaiman, consistent with known benefits of incremental model updates in machine learning systems.
Regarding Claim 7, Klaiman in view of LE teaches the method of Claim 6, wherein:
combining the new mathematical representation with the current mathematical representation is performed using weighted average with an exponential decay factor – Klaiman does not expressly teach this, however LE teaches this:
“the system may determine a third output based on a weighted average of the first output and the second output.” (Le, p.10, paragraph [0086])
“… wherein the first weight is greater than the second weight” (Le, p.10, paragraph [0086])
“… the system may… produce a real-valued confidence score… ” (Le, p.7, paragraph [0061])
Although Le does not expressly use the phrase “exponential decay factor”, its disclosure of a recency-weighted averaging process, where older values are given progressively smaller weights, is recognized in the art as an exponential decay technique. The exponential decay function is a well-known implementation for assigning diminishing weights to older data in time-series or iterative machine learning updates.
It would have been obvious to one of ordinary skill in the art at the time of the claimed invention to modify the dataset representation update mechanism in Klaiman, as supplemented by Le, to apply a weighted average with exponential decay. This would allow the system to prioritize recent data while still retaining the influence of historical data in a controlled, mathematically stable way; an established benefit in machine learning systems that handle streaming or dynamic input.
Regarding Claim 9, Klaiman in view of LE teaches the method of Claim 1, wherein:
each of the first dataset and the candidate datasets comprises a time series – Klaiman doesn’t explicitly teach this, however Le teaches:
“In some embodiments, a second type of data (e.g., a time-dependent information) may include user account information, such as types of accounts the user has, other accounts on file, such as bank accounts for payment, information associated with accounts, such as credit limit, current
balance, due date, recent payments, recent transactions. The system may obtain this data in real-time for model prediction… ” (Le, p.3, paragraph [0028])
This paragraph confirms that LE’s system processes temporally ordered financial events, a textbook example of time series data, in both the first and candidate datasets used for model prediction.
It would have been obvious to a person of ordinary skill in the art at the time of the claimed invention to incorporate the time-dependent dataset structures of Le into the similarity-learning framework of Klaiman, in order to apply Klaiman’s embedding and similarity scoring approach to time-series data. A POSITA would have been motivated to do so in order to:
Enable Klaiman’s Siamese-based similarity model to operate on domains such as financial behavior, usage prediction, or temporal user modeling, as taught in Le;
Leverage the ability to compute similarity between chronologically structured datasets, which are common in practical applications like user analytics, transactions, or event logs;
Expand Klaiman’s architecture to new domains without altering the underlying system design.
Regarding Claims 16-17 and 19
Claims 16-17 and 19 are the system claims directly analogous in scope to method Claims 6-7 and 9 respectively, with each system claim reciting the same limitations as its corresponding method claims rewritten in system form (e.g., claim 16 mirrors claim 6 as a system; claim 17 mirrors claim 7; claim 19 mirrors claim 9). Accordingly, the 35 U.S.C. § 103 rejections for Claims 6-7 and 9 over Klaiman in view of Das and further in view of Maor, and further in view of Le are equally applicable to Claims 16-17 and 19, and the references cited apply with equal force.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to PAUL COLEMAN whose telephone number is (571)272-4687. The examiner can normally be reached Mon-Fri.
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, David Yi can be reached at 571-270-7519. 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.
/PAUL COLEMAN/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126