DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 2025/05/30. The submission is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
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 1–20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to a judicial exception (an abstract idea) without reciting significantly more.
Regarding independent claims 1 and 11
Step 1 -- whether the claim falls within any statutory category. See MPEP 2106.03.
Claim 1 is drawn to a computer-implemented method (a process). Claim 11 is drawn to a system comprising data processing hardware and memory hardware (a machine). Therefore, each of these claims falls under at least one of the four categories of statutory subject matter (process/method, machine/product/apparatus, manufacture, or composition of matter). (Step 1: YES.)
Step 2A Prong One -- whether the claim recites a judicial exception. See MPEP 2106.04, subsection II.
Regarding independent Claim 1, the claim is directed to a method of selectively predicting on unlabeled data using a deep ensemble model and self-training. The claim recites the following limitations:
The limitations of "determining, using a deep ensemble model pre-trained on a plurality of source training samples, a first average output for each unlabeled test data sample of the set of unlabeled test data samples," "determining, using the fine-tuned deep ensemble model, a second average output for each unlabeled test data sample of the set of unlabeled test data samples," "fine-tuning the deep ensemble model using the subset of labeled test data samples," and "training the deep ensemble model using the pseudo-labeled set of training data samples" are directed towards the abstract idea of mathematical concepts, specifically mathematical calculations and the manipulation of information through mathematical relationships. See MPEP § 2106.04(a)(2), subsection I. Computing an average of a model's outputs, and fine-tuning and training a model by minimizing a loss function, are performed through mathematical calculations. As recognized in the July 2024 Subject Matter Eligibility Examples (Example 47), the training of a neural network (e.g., by gradient descent) is a mathematical calculation.
The limitations of "selecting, from the set of unlabeled test data samples, a subset of unlabeled test data samples based on the determined first average outputs" and "generating, using the set of unlabeled test data samples and the determined second average outputs, a pseudo-labeled set of training data samples" are directed towards the abstract idea of a mental process, i.e., a concept that can be performed in the human mind, including observation, evaluation, judgment, and opinion. See MPEP § 2106.04(a)(2), subsection III. Selecting a subset of samples based on determined output values, and assigning pseudo-labels to samples based on determined output values, are evaluations and judgments that can be practically performed in the human mind or with pen and paper.
Accordingly, independent Claim 1 recites an abstract idea under Step 2A Prong 1 because the claim recites limitations falling within the mathematical concepts and mental processes groupings of abstract ideas.
Independent Claim 11 is a system claim reciting operations commensurate in scope with the method of Claim 1 and recites the same abstract idea for the same reasons.
Step 2A Prong Two -- whether the claim as a whole integrates the recited judicial exception into a practical application, or whether the claim is "directed to" the judicial exception. See MPEP 2106.04(d).
Regarding independent Claim 1, beyond the abstract idea identified above, the claim recites the additional elements of "data processing hardware," a "deep ensemble model," "obtaining a set of unlabeled test data samples," and "labeling each respective unlabeled test data sample in the subset of unlabeled test data samples."
The "data processing hardware" and the "deep ensemble model" are recited at a high level of generality and amount to no more than mere instructions to implement the abstract idea on a generic computer or to use a generic computer/model as a tool to perform the abstract idea. See MPEP § 2106.05(f). The limitation of "obtaining a set of unlabeled test data samples" and the "labeling" of samples amount to mere data gathering, which is insignificant extra-solution activity. See MPEP § 2106.05(g). Reciting that the operations are performed using a deep ensemble model further amounts to no more than generally linking the use of the judicial exception to a particular technological environment or field of use. See MPEP § 2106.05(h). These additional elements, considered individually and in combination, do not impose any meaningful limit on the abstract idea, do not reflect an improvement to the functioning of a computer or to any other technology or technical field, and therefore fail to integrate the exception into a practical application. The claim is therefore directed to the abstract idea.
Independent Claim 11 is drawn to a system reciting limitations commensurate in scope with Claim 1 and is rejected under the same rationale. Claim 11 additionally recites "data processing hardware; and memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations." These additional elements amount to no more than a generic computer and generic computer components that merely act as a tool on which the abstract idea is performed, and thus fail to integrate the exception into a practical application. See MPEP § 2106.05(f).
Step 2B -- whether the claim amounts to significantly more than the judicial exception. See MPEP § 2106.05.
Regarding independent Claims 1 and 11, the additional elements identified above, considered individually and as an ordered combination, do not provide an inventive concept. The "data processing hardware," "memory hardware," and "deep ensemble model" are generic computer components performing generic computer functions (storing instructions, executing operations, computing model outputs), which the courts have recognized as well-understood, routine, and conventional. See MPEP § 2106.05(d). The "obtaining" and "labeling" of data are insignificant extra-solution activities of the type the courts have recognized as well-understood, routine, and conventional data gathering. See MPEP § 2106.05(d)(II) and § 2106.05(g). Considered as an ordered combination, the additional elements add nothing that is not already present when the elements are considered separately, and merely apply the abstract idea using generic components. Accordingly, Claims 1 and 11 do not amount to significantly more than the abstract idea and are not patent eligible.
Regarding dependent Claims 2–10 and 12–20
Claims 2–10 and 12–20 depend from Claims 1 and 11, respectively, and merely narrow the previously identified abstract idea or add additional elements that, individually and in combination, fail to integrate the exception into a practical application and fail to amount to significantly more than the abstract idea.
Step 1 - whether the claim falls within any statutory category. See MPEP 2106.03.
Claims 2–10 are drawn to a method (a process). Claims 12–20 are drawn to a system (a machine). Therefore, each of these claims falls under at least one of the four categories of statutory subject matter. (Step 1: YES.)
Step 2A Prong 1 - whether the claim recites a judicial exception. See MPEP 2106.04, subsection II.
Regarding Claim 2, this claim recites the limitation of "wherein labeling each respective unlabeled test data sample in the set of unlabeled test data samples comprises obtaining, from an oracle, for each respective unlabeled test data sample in the set of unlabeled test data samples, a corresponding label for the respective unlabeled test data sample." This limitation is directed towards the abstract idea of a mental process, i.e., a concept that can be performed in the human mind, including observation, evaluation, judgment, and opinion. See MPEP § 2106.04(a)(2), subsection III. Assigning a corresponding label to a data sample is an evaluation or judgment that can be practically performed by a human being.
Regarding Claim 3, this claim recites the limitation of "wherein the oracle comprises a human annotator." This limitation is directed towards the abstract idea of a mental process, as it expressly recites that the labeling (assigning of labels) is performed by a human annotator. See MPEP § 2106.04(a)(2), subsection III.
Regarding Claim 4, this claim recites the limitation of "wherein training the deep ensemble model using the pseudo-labeled set of training data samples comprises using a stochastic gradient descent technique." This limitation is directed towards the abstract idea of a mathematical concept, specifically a mathematical calculation, because stochastic gradient descent is a mathematical optimization calculation used to train the model. See MPEP § 2106.04(a)(2), subsection I; see also July 2024 Subject Matter Eligibility Examples, Example 47 (gradient descent is a mathematical calculation).
Regarding Claim 5, this claim recites the limitations of "the deep ensemble model comprises an ensemble of one or more machine learning models" and "training the deep ensemble model using the pseudo-labeled set of training data samples comprises training each machine learning model with a different randomly selected subset of the pseudo-labeled set of training data samples." The latter limitation is directed towards the abstract idea of a mathematical concept, specifically the manipulation of information through mathematical calculations used to train each machine learning model. See MPEP § 2106.04(a)(2), subsection I. The recited "ensemble of one or more machine learning models" is an additional element addressed under Step 2A Prong 2 below.
Regarding Claim 6, this claim recites the limitations of "for each machine learning model of the one or more machine learning models, determining a prediction and a confidence value indicating a likelihood that the prediction is correct" and "averaging the confidence values determined by each machine learning model for the respective unlabeled test data sample." These limitations are directed towards the abstract idea of a mathematical concept, specifically mathematical calculations and the manipulation of information through mathematical relationships, because determining a confidence value and averaging the confidence values are mathematical calculations. See MPEP § 2106.04(a)(2), subsection I.
Regarding Claim 7, this claim recites the limitation of "wherein selecting, from the set of unlabeled test data samples, the subset of unlabeled test data samples based on the determined first average outputs comprises selecting the unlabeled test data samples comprising the lowest determined first average outputs." This limitation is directed towards the abstract idea of a mental process, i.e., a concept that can be performed in the human mind, including observation, evaluation, and judgment, because comparing the determined output values and selecting those that are lowest is an evaluation and judgment that can be practically performed by a human being. See MPEP § 2106.04(a)(2), subsection III.
Regarding Claim 8, this claim recites the limitation of "wherein fine-tuning the deep ensemble model using the subset of labeled test data samples comprises jointly fine-tuning the deep ensemble model using the subset of labeled test data samples and the plurality of source training samples." This limitation is directed towards the abstract idea of a mathematical concept, specifically a mathematical calculation, because jointly fine-tuning (optimizing) the model over the labeled subset and the source training samples is performed through mathematical calculations. See MPEP § 2106.04(a)(2), subsection I.
Regarding Claim 9, this claim recites the limitation of "wherein fine-tuning the deep ensemble model using the subset of labeled test data samples comprises determining a cross-entropy loss." This limitation is directed towards the abstract idea of a mathematical concept, specifically a mathematical calculation, because a cross-entropy loss is a mathematical calculation. See MPEP § 2106.04(a)(2), subsection I.
Regarding Claim 10, this claim recites the limitation of "wherein training the deep ensemble model using the pseudo-labeled set of training data samples comprises determining a KL-Divergence loss." This limitation is directed towards the abstract idea of a mathematical concept, specifically a mathematical calculation, because a KL-Divergence loss is a mathematical calculation. See MPEP § 2106.04(a)(2), subsection I.
Claims 12–20 recite limitations commensurate in scope with Claims 2–10, respectively, and recite the same abstract ideas for the same reasons set forth above for their respective counterparts.
Step 2A Prong 2 - whether the claim as a whole integrates the recited judicial exception into a practical application. See MPEP 2106.04(d).
Regarding Claims 2–10 and 12–20, beyond the abstract ideas identified above, these claims recite the additional elements of a "deep ensemble model" and "machine learning models" (Claims 5–6, 15–16), and "obtaining, from an oracle, … a corresponding label" / a "human annotator" (Claims 2–3, 12–13). The recited deep ensemble model and machine learning models are recited at a high level of generality and amount to no more than generic models used as a tool to perform the abstract idea. See MPEP § 2106.05(f). The obtaining of a corresponding label from an oracle/human annotator amounts to mere data gathering, which is insignificant extra-solution activity. See MPEP § 2106.05(g). Reciting that the operations are performed using the deep ensemble model further amounts to no more than generally linking the use of the judicial exception to a particular technological environment or field of use. See MPEP § 2106.05(h). Considered individually and in combination, these additional elements do not impose any meaningful limit on the abstract idea, do not reflect an improvement to the functioning of a computer or to any other technology or technical field, and therefore fail to integrate the exception into a practical application. The claims are therefore directed to the abstract idea.
Step 2B - whether the claim provides an inventive concept. See MPEP 2106.05.
Regarding Claims 2–10 and 12–20, the additional elements identified above, considered individually and as an ordered combination, do not provide an inventive concept. The deep ensemble model and machine learning models are generic components performing generic functions, and the obtaining of labels from an oracle/human annotator is insignificant extra-solution data gathering, each of which the courts have recognized as well-understood, routine, and conventional. See MPEP § 2106.05(d). Considered as an ordered combination, these additional elements add nothing that is not already present when the elements are considered separately and merely apply the abstract idea using generic components and conventional data gathering. Accordingly, Claims 2–10 and 12–20 do not amount to significantly more than the abstract idea and are not patent eligible.
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 text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1, 2, 3, 7, 8, 11, 12, 13, 17, and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Chen et al. (Chen), Non-Patent Literature, "Detecting Errors and Estimating Accuracy on Unlabeled Data with Self-Training Ensembles," arXiv:2106.15728v3 [cs.LG], published February 7, 2022, in view of Beluch et al. (Beluch), Non-Patent Literature, "The Power of Ensembles for Active Learning in Image Classification," CVPR 2018, published June 2018 and cited in the IDS filed on 5/30/25, and relied upon at pages 9368, 9370, and further in view of Snow et al. (Snow), U.S. Patent Application Publication No. 2022/0391765 A1.
Regarding independent Claim 1, Chen teaches a computer-implemented method executed by data processing hardware that causes the data processing hardware to perform operations comprising:
obtaining a set of unlabeled test data samples (Chen, p. 3, §3, and p. 4, Framework 1, Input: "an unlabeled test dataset U_X"). Chen obtains a set of unlabeled test inputs U_X drawn from the test distribution.
for each respective initial training step of a plurality of initial training steps, determining, using a deep ensemble model pre-trained on a plurality of source training samples, a first average output for each unlabeled test data sample of the set of unlabeled test data samples (Chen, p. 7, Algorithm 2, line 1 ["Pre-train {h′_i} … from different random initialization"] and Algorithm 3, line 2 [checkpoint models "at the end of each training epoch" used as the ensemble]; p. 6, Algorithm 1, line 4; p. 8, §7.1 ["Ensemble Average Confidence"]). Chen pre-trains a deep ensemble of models on the training dataset D (the source training samples) and, at each training epoch/checkpoint, evaluates the ensemble on each unlabeled test sample to produce an aggregated (average) ensemble output for that sample.
for each round of a plurality of rounds: selecting, from the set of unlabeled test data samples, a subset of unlabeled test data samples based on the determined first average outputs (Chen, p. 6, Algorithm 1, lines 2–4 ["for t = 1, … , T"]; p. 4, Framework 1, line 4). Chen iterates over a plurality of rounds t = 1 … T and, in each round, selects a subset of the unlabeled test samples on which the ensemble output diverges from the pre-trained model f (i.e., based on the determined ensemble outputs).
determining, using the fine-tuned deep ensemble model, a second average output for each unlabeled test data sample of the set of unlabeled test data samples (Chen, p. 5, ¶ "executes in T iterations"; p. 6, Algorithm 1). In each subsequent round, Chen regenerates the ensemble and recomputes the aggregated (average) ensemble output for each unlabeled test sample using the updated ensemble.
generating, using the set of unlabeled test data samples and the determined second average outputs, a pseudo-labeled set of training data samples (Chen, p. 6, Algorithm 1, lines 4–5 [ỹ_x = majority vote of the ensemble; R = {(x, ỹ_x) : x ∈ U_X}]; p. 2, §1). Chen assigns a pseudo-label to each unlabeled test sample based on the ensemble's aggregated output and forms a pseudo-labeled set R.
training the deep ensemble model using the pseudo-labeled set of training data samples (Chen, p. 4, ¶ "train … a new ensemble"; p. 5; p. 7, Algorithm 2, line 3 and Eq. (3)). Chen performs self-training by training the deep ensemble on the pseudo-labeled set R.
Chen further teaches fine-tuning the deep ensemble model (Chen, p. 7, Algorithm 2, line 3 and Eq. (3); “fine-tunes h’I on D and R for one epoch”) and selecting uncertain unlabeled test samples and assigning them labels by majority vote of the ensemble for self-training (Chen, p. 4, Framework 1, line 5; p. 6, Algorithm 1, lines 4–5). However, Chen does not teach labeling each respective unlabeled test data sample in the subset of unlabeled test data samples and fine-tuning the deep ensemble model using the subset of labeled test data samples, because Chen fine-tunes the ensemble on the source training data and the pseudo-labeled set rather than on an oracle-labeled subset of the test data, and assigns pseudo-labels by majority vote rather than obtaining labels for the selected subset.
In the same field of endeavor, Beluch teaches labeling each respective unlabeled test data sample in the subset of unlabeled test data samples and fine-tuning the deep ensemble model using the subset of labeled test data samples (Beluch, p. 9370, §3.1; p. 9368, col. 2). Beluch discloses pool-based active learning in which an acquisition function selects a subset of the most informative unlabeled samples (p. 9370, §3.3, entropy and variation-ratio measures), the selected subset is "labeled by an external oracle" - e.g., "by a human expert" - and the newly labeled subset is added to the labeled set and used to (re)train the ensemble model, "This process is repeated" over rounds (p. 9370, §3.1). Beluch further uses an ensemble of classifiers trained from different random weight initializations and takes the averaged softmax vectors of the ensemble members as the output (p. 9370, §3.2, Eq. (1)).
Chen and Beluch are analogous art, as both are from the same field of endeavor of training deep neural-network ensembles for prediction under limited labeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-training-ensemble framework of Chen with the ensemble-based active-learning acquisition and oracle labeling of Beluch, so as to select the most informative uncertain test samples, obtain their labels from an oracle, and fine-tune the deep ensemble on the labeled subset. The motivation to combine Chen and Beluch is as recited by Beluch (p. 9368, Abstract; p. 9370, §3.1) - ensemble-based active learning yields better-calibrated uncertainty estimates and reaches a target accuracy with the minimally required set of labeled images, thereby reducing labeling effort - such that incorporating Beluch's acquisition-and-oracle-labeling step into Chen's ensemble would predictably improve the ensemble's accuracy on the shifted test distribution before fine-tuning and self-training.
The combination of Chen and Beluch, however, does not expressly teach performing, on data processing hardware, a single iterative process that both labels each respective unlabeled test data sample in the subset of unlabeled test data samples (by oracle) and generates … a pseudo-labeled set of training data samples, using a deep ensemble whose per-sample outputs are averaged.
In the same field of endeavor, Snow teaches this integrated scheme on data processing hardware (Snow, ¶ [0023], processor(s) and memory storing instructions). Snow discloses a machine-learning model comprising an ensemble of sub-models that are neural networks (¶¶ [0020]–[0022]) for which a confidence measure is determined as an average of the sub-models' predictions (¶¶ [0016], [0042]); selecting unlabeled samples based on uncertainty (¶ [0038]); labeling each respective unlabeled test data sample in the subset of unlabeled test data samples by "quer[ying] an oracle," e.g., "a human domain expert," to acquire a ground-truth label (¶ [0045]); generating … a pseudo-labeled set of training data samples by assigning a pseudo-label to a sample based on the prediction (¶ [0044]); and fine-tuning / training the deep ensemble model over a plurality of rounds (¶¶ [0046], [0047]).
Chen, Beluch, and Snow are analogous art, as all are from the same field of endeavor of training deep neural-network models/ensembles over labeled and unlabeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-training ensemble of Chen and the ensemble-based active learning of Beluch with the integrated scheme of Snow. The motivation to combine Chen, Beluch, and Snow is as recited by Snow (¶ [0031]) - assigning labels to unlabeled data by selecting between active learning (oracle) and semi-supervised pseudo-labeling based on prediction uncertainty "reduces the load on the oracle" while improving the model's calibration as training progresses - such that one of ordinary skill would integrate the oracle-labeling and pseudo-labeling steps within a single iterative ensemble training process executed on data processing hardware to obtain the predictable benefit of reduced labeling cost and improved accuracy under distribution shift.
Regarding Claim 2, the claim depends from Claim 1 and incorporates all of the limitations of Claim 1. Those limitations are rejected under the same rationale set forth above for Claim 1.
Claim 2 further recites:
The method of claim 1, wherein labeling each respective unlabeled test data sample in the set of unlabeled test data samples comprises obtaining, from an oracle, for each respective unlabeled test data sample in the set of unlabeled test data samples, a corresponding label for the respective unlabeled test data sample.
As established above with respect to Claim 1, Chen assigns pseudo-labels by majority vote of the ensemble and does not teach obtaining, from an oracle, for each respective unlabeled test data sample in the set of unlabeled test data samples, a corresponding label for the respective unlabeled test data sample.
In the same field of endeavor, Beluch teaches labeling each respective unlabeled test data sample in the set of unlabeled test data samples comprises obtaining, from an oracle, for each respective unlabeled test data sample in the set of unlabeled test data samples, a corresponding label for the respective unlabeled test data sample (Beluch, p. 9370, §3.1; p. 9368, col. 2). Beluch discloses that, in each round, an acquisition function selects a subset of unlabeled samples that are "labeled by an external oracle" - e.g., "by a human expert" - to obtain, for each selected unlabeled sample, a corresponding label that is then added to the labeled set.
Chen and Beluch are analogous art, as both are from the same field of endeavor of training deep neural-network ensembles for prediction under limited labeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-training-ensemble framework of Chen with the oracle labeling of Beluch, so as to obtain, from an oracle, a corresponding label for each unlabeled test data sample in the selected subset. The motivation to combine Chen and Beluch is as recited by Beluch (p. 9368, Abstract; p. 9370, §3.1) - obtaining labels from an oracle for the most informative samples allows a target accuracy to be reached with the minimally required set of labeled data, thereby reducing labeling effort - such that obtaining oracle labels for the selected uncertain test samples would predictably improve the accuracy of Chen's ensemble before fine-tuning and self-training.
The combination of Chen and Beluch further teaches the above limitation as evidenced by Snow, which likewise teaches obtaining, from an oracle, … a corresponding label for the respective unlabeled test data sample (Snow, ¶ [0045]). Snow discloses that the processor "queries an oracle," e.g., "a human domain expert," to "acquire a ground-truth label" for a selected unlabeled sample. Chen, Beluch, and Snow are analogous art, as all are from the same field of endeavor of training deep neural-network models/ensembles over labeled and unlabeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine Chen, Beluch, and Snow, the motivation being as recited by Snow (¶ [0031]) - selectively querying an oracle for a corresponding label reduces the load on the oracle while improving the model's calibration as training progresses.
Regarding Claim 3, the claim depends from Claim 2 and incorporates all of the limitations of Claim 2. Those limitations are rejected under the same rationale set forth above for Claims 1 and 2.
Claim 3 further recites:
The method of claim 2, wherein the oracle comprises a human annotator.
As established above with respect to Claims 1 and 2, Chen assigns pseudo-labels by majority vote of the ensemble and does not teach an oracle, and therefore does not teach wherein the oracle comprises a human annotator.
In the same field of endeavor, Beluch teaches wherein the oracle comprises a human annotator (Beluch, p. 9368, col. 2). Beluch discloses that the unlabeled data-points selected by the acquisition function are labeled "by a human expert," i.e., the oracle that supplies the corresponding labels is a human annotator.
Chen and Beluch are analogous art, as both are from the same field of endeavor of training deep neural-network ensembles for prediction under limited labeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-training-ensemble framework of Chen with the human-annotator oracle of Beluch, such that the corresponding label obtained for each selected unlabeled test data sample is provided by a human annotator. The motivation to combine Chen and Beluch is as recited by Beluch (p. 9368, Abstract; p. 9370, §3.1) - obtaining labels from a human expert for the most informative samples allows a target accuracy to be reached with the minimally required set of labeled data, thereby reducing labeling effort - such that using a human annotator as the oracle would predictably provide reliable ground-truth labels and improve the accuracy of Chen's ensemble before fine-tuning and self-training.
The combination of Chen and Beluch further teaches the above limitation as evidenced by Snow, which likewise teaches wherein the oracle comprises a human annotator (Snow, ¶ [0045]). Snow discloses that "the oracle may comprise a human domain expert" that supplies high-accuracy ground-truth labels for the selected unlabeled samples. Chen, Beluch, and Snow are analogous art, as all are from the same field of endeavor of training deep neural-network models/ensembles over labeled and unlabeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine Chen, Beluch, and Snow, the motivation being as recited by Snow (¶ [0031]) - querying a human annotator as the oracle for selected uncertain samples reduces the load on the oracle while improving the model's calibration as training progresses.
Regarding Claim 7, the claim depends from Claim 1 and incorporates all of the limitations of Claim 1. Those limitations are rejected under the same rationale set forth above for Claim 1.
Claim 7 further recites:
wherein selecting, from the set of unlabeled test data samples, the subset of unlabeled test data samples based on the determined first average outputs comprises selecting the unlabeled test data samples comprising the lowest determined first average outputs.
As established above with respect to Claim 1, Chen teaches selecting, from the set of unlabeled test data samples, a subset of unlabeled test data samples based on the determined first average outputs (Chen, p. 4, Framework 1, line 4; p. 6, Algorithm 1). However, Chen does not teach selecting the unlabeled test data samples comprising the lowest determined first average outputs, because Chen selects samples based on disagreement between the ensemble and the pre-trained model f rather than by selecting the samples having the lowest determined average outputs.
In the same field of endeavor, Beluch teaches this limitation (Beluch, p. 9370, §3.3). Beluch discloses uncertainty-based acquisition functions in which the samples selected for labeling are those whose averaged ensemble predictions are least confident - e.g., choosing the points "whose predicted classification probability distributions have the highest entropy" (and, by the variation-ratio and predictive-variance measures of Eq. (4)–(5)) - i.e., selecting the unlabeled test data samples comprising the lowest determined average outputs.
Chen and Beluch are analogous art, as both are from the same field of endeavor of training deep neural-network ensembles for prediction under limited labeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the ensemble selection of Chen with the uncertainty-based acquisition of Beluch, such that selecting the subset based on the determined first average outputs comprises selecting the unlabeled test data samples comprising the lowest determined first average outputs. The motivation to combine Chen and Beluch is as recited by Beluch (p. 9368, Abstract; p. 9370, §3.3) - selecting the least-confident (most uncertain) samples for labeling allows a target accuracy to be reached with the minimally required set of labeled data - such that selecting the lowest-average-output samples in Chen would predictably improve label efficiency and ensemble accuracy.
The combination of Chen and Beluch further teaches the above limitation as evidenced by Snow, which likewise teaches selecting the lowest-average-output samples (Snow, ¶ [0038]). Snow discloses sampling from the unlabeled dataset "based on a measure of uncertainty" by "selecting elements having high uncertainty as samples," i.e., selecting those samples having the lowest confidence (lowest average output). Chen, Beluch, and Snow are analogous art, as all are from the same field of endeavor of training deep neural-network models/ensembles over labeled and unlabeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine Chen, Beluch, and Snow, the motivation being as recited by Snow (¶ [0031]) - selecting the most uncertain samples for oracle labeling, while pseudo-labeling the more confident samples, reduces the load on the oracle while improving the model's calibration as training progresses.
Regarding Claim 8, the claim depends from Claim 1 and incorporates all of the limitations of Claim 1. Those limitations are rejected under the same rationale set forth above for Claim 1.
Claim 8 further recites:
wherein fine-tuning the deep ensemble model using the subset of labeled test data samples comprises jointly fine-tuning the deep ensemble model using the subset of labeled test data samples and the plurality of source training samples.
Chen teaches jointly fine-tuning the deep ensemble model using [a labeled set] and the plurality of source training samples (Chen, p. 7, Algorithm 2, line 3 and Eq. (3)). Chen discloses fine-tuning each ensemble member h_i for one epoch by minimizing a combined objective E₍ₓ,ᵧ₎∈D[ℓ(h(x), y)] + γ · E₍ₓ,ᵧ₎∈R[ℓ(h(x), y)] - i.e., fine-tuning the ensemble jointly over the source training dataset D and an additional labeled set R simultaneously (p. 7, "fine-tunes h′_i on D and R for one epoch"). As established above with respect to Claims 1 and 2, the subset of labeled test data samples is obtained from an oracle per Beluch.
Accordingly, the combination of Chen and Beluch teaches fine-tuning the deep ensemble model using the subset of labeled test data samples comprises jointly fine-tuning the deep ensemble model using the subset of labeled test data samples and the plurality of source training samples, wherein Chen supplies the joint objective over the source training samples and an additional labeled set, and Beluch supplies the subset of labeled (oracle-labeled) test data samples used in that joint objective.
Chen and Beluch are analogous art, as both are from the same field of endeavor of training deep neural-network ensembles for prediction under limited labeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to jointly fine-tune Chen's deep ensemble using the source training samples together with the subset of labeled test data samples obtained from the oracle of Beluch. The motivation to combine Chen and Beluch is as recited by Chen (p. 7, Eq. (3); p. 5, condition (A)) - retaining the source training dataset D in the fine-tuning objective keeps the ensemble accurate on the inputs where the pre-trained model is correct - such that jointly fine-tuning the oracle-labeled subset together with the source training samples would predictably preserve the ensemble's accuracy on the source distribution while adapting it to the newly labeled test samples.
Regarding Claim 11, the claim is a system claim reciting operations commensurate in scope with the method of Claim 1, and is rejected under the same rationale set forth above for Claim 1. Claim 11 additionally recites A system comprising: data processing hardware; and memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising. Snow teaches this limitation (Snow, ¶ [0023]), disclosing a system comprising one or more processors and a memory storing instructions which cause the one or more processors to perform the operations. It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to implement the method of Chen (as combined with Beluch and Snow) on the data processing hardware and memory hardware of Snow, the motivation being to provide a system that executes the recited operations.
Regarding Claim 12, the claim recites limitations commensurate in scope with Claim 2 and is rejected under the same rationale set forth above for Claim 2.
Regarding Claim 13, the claim recites limitations commensurate in scope with Claim 3 and is rejected under the same rationale set forth above for Claim 3.
Regarding Claim 17, the claim recites limitations commensurate in scope with Claim 7 and is rejected under the same rationale set forth above for Claim 7.
Regarding Claim 18, the claim recites limitations commensurate in scope with Claim 8 and is rejected under the same rationale set forth above for Claim 8.
Claims 4 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Beluch, in view of Snow, and further in view of Luong et al. (Luong), U.S. Patent Application Publication No. 2022/0083840 A1.
Regarding Claim 4, the claim depends from Claim 1 and incorporates all of the limitations of Claim 1. Those limitations are rejected under the same rationale set forth above for Claim 1.
Claim 4 further recites:
The method of claim 1, wherein training the deep ensemble model using the pseudo-labeled set of training data samples comprises using a stochastic gradient descent technique.
As established above with respect to Claim 1, Chen teaches training the deep ensemble model using the pseudo-labeled set of training data samples (Chen, p. 4, ¶ "train … a new ensemble"; p. 5). However, Chen does not explicitly teach wherein training the deep ensemble model using the pseudo-labeled set of training data samples comprises using a stochastic gradient descent technique.
In the same field of endeavor, (Luong) teaches wherein training the deep ensemble model using the pseudo-labeled set of training data samples comprises using a stochastic gradient descent technique (Luong, Abstract and ¶ [0096]). (Luong) discloses a self-training technique in which a neural network model is trained on a combined dataset that includes labeled data and a corresponding pseudo-label for each item of unlabeled data (Abstract), and that this training process uses "stochastic gradient descent (SGD)" (¶ [0096]).
Chen and (Luong) are analogous art, as both are from the same field of endeavor of self-training deep neural-network models on labeled and pseudo-labeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-training-ensemble framework of Chen with the stochastic gradient descent training technique of (Luong), such that training the deep ensemble model on the pseudo-labeled set is performed using a stochastic gradient descent technique. The motivation to combine Chen and (Luong) is as recited by (Luong) (¶ [0096]) - stochastic gradient descent "introduces stochasticity into the training process," which (Luong) attributes to improved model performance - such that using a stochastic gradient descent technique to train Chen's deep ensemble on the pseudo-labeled set would predictably provide effective optimization and improved generalization of the trained ensemble.
Regarding Claim 14, the claim recites limitations commensurate in scope with Claim 4 and is rejected under the same rationale set forth above for Claim 4.
Claims 5, 6, 15, and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Beluch, in view of Snow, and further in view of Ghosh et al. (Ghosh), U.S. Patent Application Publication No. 2021/0374500 A1.
Regarding Claim 5, the claim depends from Claim 1 and incorporates all of the limitations of Claim 1. Those limitations are rejected under the same rationale set forth above for Claim 1.
Claim 5 further recites:
the deep ensemble model comprises an ensemble of one or more machine learning models; and
training the deep ensemble model using the pseudo-labeled set of training data samples comprises training each machine learning model with a different randomly selected subset of the pseudo-labeled set of training data samples.
Chen teaches the deep ensemble model comprises an ensemble of one or more machine learning models (Chen, p. 6, Algorithm 1 [the ensemble {h_i} for i = 1 … N]; p. 7, Algorithm 2). Chen discloses that the deep ensemble is composed of a plurality of machine learning models (deep networks) h_1 … h_N.
Chen does not teach training the deep ensemble model using the pseudo-labeled set of training data samples comprises training each machine learning model with a different randomly selected subset of the pseudo-labeled set of training data samples, because Chen trains each ensemble member from a different random initialization on the same data rather than on a different randomly selected subset of the training data.
In the same field of endeavor, (Ghosh) teaches this limitation (Ghosh, ¶¶ [0104], [0106]–[0108]). (Ghosh) discloses creating an ensemble of trained learners by "random splitting of training data," wherein the training data is randomly split multiple times - once for each component to be obtained - so that "for every trained learner component to be obtained we have a different training dataset" (¶ [0107]); for example, the training data is randomly split five times to obtain five trained learner components, each trained on a different randomly selected subset of the data (¶ [0108]). That is, each machine learning model of the ensemble is trained with a different randomly selected subset of the training data samples.
Chen and (Ghosh) are analogous art, as both are from the same field of endeavor of constructing and training deep neural-network ensembles. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-training-ensemble framework of Chen with the random-data-splitting ensemble-construction technique of (Ghosh), such that each machine learning model of Chen's deep ensemble is trained with a different randomly selected subset of the pseudo-labeled set of training data samples. The motivation to combine Chen and (Ghosh) is as recited by (Ghosh) (Abstract; ¶ [0104]) - generating ensemble components from random splits of the training data improves the reproducibility and accuracy of the ensemble - which is consistent with Chen's own recognition that ensemble diversity improves performance (Chen, p. 4), such that training each member of Chen's ensemble on a different randomly selected subset would predictably increase ensemble diversity and improve accuracy.
Regarding Claim 6, the claim depends from Claim 5 and incorporates all of the limitations of Claims 1 and 5. Those limitations are rejected under the same rationale set forth above for Claims 1 and 5.
Claim 6 further recites:
wherein determining the first average output for each unlabeled test data sample comprises, for each respective unlabeled test data sample: for each machine learning model of the one or more machine learning models, determining a prediction and a confidence value indicating a likelihood that the prediction is correct; and
averaging the confidence values determined by each machine learning model for the respective unlabeled test data sample.
Chen teaches wherein determining the first average output for each unlabeled test data sample comprises, for each respective unlabeled test data sample: for each machine learning model of the one or more machine learning models, determining a prediction and a confidence value indicating a likelihood that the prediction is correct (Chen, p. 4, footnote 2). Chen discloses that each model h of the ensemble outputs h(x) = [h¹(x), … , hᴷ(x)], where hⁱ(x) is the predicted probability that input x belongs to class i - i.e., for each machine learning model, a prediction (the class) and an associated confidence value (the predicted probability indicating the likelihood that the prediction is correct) are determined.
Chen does not teach averaging the confidence values determined by each machine learning model for the respective unlabeled test data sample, because Chen aggregates the ensemble's per-model outputs by majority vote rather than by averaging the per-model confidence values.
In the same field of endeavor, (Ghosh) teaches this limitation (Ghosh, ¶ [0147]). (Ghosh) discloses that "all the probability vectors obtained from each component are combined by calculating the mean/average," and "the resultant probability vector is then used to obtain the final predictions" - i.e., the confidence values (probability vectors) determined by each machine learning model are averaged for the respective sample to produce the aggregate (average) output.
Chen and (Ghosh) are analogous art, as both are from the same field of endeavor of constructing and aggregating the outputs of deep neural-network ensembles. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the deep ensemble of Chen with the confidence-averaging aggregation of (Ghosh), such that determining the first average output comprises averaging the confidence values determined by each machine learning model for the respective unlabeled test data sample. The motivation to combine Chen and (Ghosh) is as recited by (Ghosh) (Abstract; ¶ [0147]) - averaging the per-component probability vectors yields a single resultant prediction that improves the reproducibility and accuracy of the ensemble - which is consistent with Chen's own use of an ensemble-average-confidence measure (Chen, p. 8), such that averaging the per-model confidence values to determine Chen's ensemble output would predictably provide a more reliable aggregate output.
Regarding Claim 15, the claim recites limitations commensurate in scope with Claim 5 and is rejected under the same rationale set forth above for Claim 5.
Regarding Claim 16, the claim recites limitations commensurate in scope with Claim 6 and is rejected under the same rationale set forth above for Claim 6.
Claims 9 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Beluch, in view of Snow, and further in view of Lee (Lee), Non-Patent Literature, "Pseudo-Label: The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks," ICML Workshop on Challenges in Representation Learning, published June 20, 2013 and cited in the IDS filed on 5/30/25, and relied upon at pages 2-3.
Regarding Claim 9, the claim depends from Claim 1 and incorporates all of the limitations of Claim 1. Those limitations are rejected under the same rationale set forth above for Claim 1.
Claim 9 further recites:
wherein fine-tuning the deep ensemble model using the subset of labeled test data samples comprises determining a cross-entropy loss.
As established above with respect to Claims 1 and 8, Chen teaches fine-tuning the deep ensemble model using the subset of labeled test data samples by minimizing a supervised loss ℓ (Chen, p. 7, Algorithm 2, line 3 and Eq. (3)). However, Chen does not explicitly teach determining a cross-entropy loss, because Chen describes the fine-tuning objective in terms of a generic supervised loss ℓ without specifying that the loss is a cross-entropy loss.
In the same field of endeavor, (Lee) teaches this limitation (Lee, p. 2; p. 3, Eq. (15)). (Lee) discloses choosing "Cross Entropy as a loss function" for training the deep neural network, wherein the supervised training (fine-tuning) objective L for the labeled data is a cross-entropy loss (Eq. (15)).
Chen and (Lee) are analogous art, as both are from the same field of endeavor of training deep neural-network models on labeled and unlabeled data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the fine-tuning objective of Chen with the cross-entropy loss of (Lee), such that fine-tuning the deep ensemble model using the subset of labeled test data samples comprises determining a cross-entropy loss. The motivation to combine Chen and (Lee) is as recited by (Lee) (p. 2; p. 3) - cross-entropy is a standard and effective loss function for training deep neural-network classifiers on labeled data - such that implementing Chen's generic supervised fine-tuning loss ℓ as a cross-entropy loss would predictably provide effective optimization of the ensemble on the labeled test data samples.
Regarding Claim 19, the claim recites limitations commensurate in scope with Claim 9 and is rejected under the same rationale set forth above for Claim 9.
Claims 10 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Beluch, in view of Snow, and further in view of Lan et al. (Lan), Non-Patent Literature, "Knowledge Distillation by On-the-Fly Native Ensemble," arXiv:1806.04606v2 [cs.CV], published September 8, 2018 and cited in the IDS filed on 5/30/25, and relied upon at page 4.
Regarding Claim 10, the claim depends from Claim 1 and incorporates all of the limitations of Claim 1. Those limitations are rejected under the same rationale set forth above for Claim 1.
Claim 10 further recites:
wherein training the deep ensemble model using the pseudo-labeled set of training data samples comprises determining a KL-Divergence loss.
As established above with respect to Claim 1, Chen teaches training the deep ensemble model using the pseudo-labeled set of training data samples (Chen, p. 4, ¶ "train … a new ensemble"; p. 5; p. 7, Eq. (3)). However, Chen does not explicitly teach determining a KL-Divergence loss, because Chen describes training the ensemble using a generic supervised loss ℓ without specifying a KL-Divergence loss.
In the same field of endeavor, (Lan) teaches this limitation (Lan, p. 4, Eq. (6); Eq. (7)). (Lan) discloses training a deep neural-network ensemble - an on-the-fly native ensemble of branch models - by determining a Kullback-Leibler (KL) divergence loss L_kl that aligns each branch model's prediction distribution with the ensemble teacher's prediction distribution (Eq. (6)), wherein the overall training objective combines a softmax cross-entropy loss and the KL-Divergence loss (Eq. (7)).
Chen and (Lan) are analogous art, as both are from the same field of endeavor of training deep neural-network ensembles. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the ensemble training of Chen with the KL-Divergence loss of (Lan), such that training the deep ensemble model using the pseudo-labeled set of training data samples comprises determining a KL-Divergence loss. The motivation to combine Chen and (Lan) is as recited by (Lan) (Abstract; p. 4) - determining a KL-Divergence loss that aligns the individual ensemble members with the ensemble's aggregate (teacher) prediction trains more generalizable models - such that determining a KL-Divergence loss when training Chen's deep ensemble on the pseudo-labeled set (whose pseudo-labels are themselves derived from the ensemble's aggregate prediction) would predictably improve the generalization and accuracy of the trained ensemble.
Regarding Claim 20, the claim recites limitations commensurate in scope with Claim 10 and is rejected under the same rationale set forth above for Claim 10.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HUNG VAN LE whose telephone number is (571)270-0164. The examiner can normally be reached 8 a.m. - 5 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, Cesar Paula can be reached at (571) 272-4128. 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.
/HUNG VAN LE/Examiner, Art Unit 2145
/CHAU T NGUYEN/Primary Examiner, Art Unit 2145
Prosecution Insights