DETAILED ACTION
This communication is in response to the Application No. 18/163,162 filed on April 22, 2026
in which Claims 1-20 are presented for examination.
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 .
Response to Amendment
The amendments filed on April 22, 2026 have been considered. Claims 1, 5, 10, 13, 15, and
18 have been amended. Thus, Claims 1-20 are pending and presented for examination.
Applicant’s arguments filled April 22, 2026 with respect to the 35 U.S.C. 101 rejections have been fully considered and are persuasive. Thus, the 35 U.S.C. 101 rejections have been withdrawn.
Applicant’s arguments filled April 22, 2026 with respect to the 35 U.S.C. 103 rejections have
been fully considered but are moot because of the new ground of rejection.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
Claim 1-20 rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the enablement requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to enable one skilled in the art to which it pertains, or with which it is most nearly connected, to make and/or use the invention. The claimed subject matter for which the specification is not enabling is, “identifying…one or more pathways that decrease an accuracy of the machine learning model outputting the label relative to a ground-truth label, wherein the one or more pathways comprise a continuous plurality of nodes, from an input layer of the machine learning model to an output layer of the machine learning model, that are not used in a determination to output the label," and removing the datapoints correlated with such not-used pathways to generate a reduced training dataset that trains the model to be at least as accurate. The specification explains that a pathway “not used” in determining the output label may include nodes having weights below a threshold, such that the nodes “do not contribute” to the determination to output the label. See Par. [0073]. However, if the nodes/pathway do not contribute to the output label determination, the specification does not explain how such a pathway would decrease the accuracy of that output label relative to a ground-truth label. The specification also does not teach how datapoints correlated with a non-contributing pathway are identified and removed in a manner that would cause the reduced training dataset to train the model to be at least as accurate as the original training dataset. Although the specification describes examples of removing datapoints having features that affect the model, such as ink blots or scratches erroneously interpreted by the model, or irrelevant consumer data features that make a significant contribution to training, those examples concern features that contribute to model operation rather than pathways that are not used in the label determination. See Par. [0051], [0059].
Going through the Wands factors set out in MPEP 2164:
(A) The breadth of the claims; The claims are broad because they require identifying any “continuous plurality of nodes” from an input layer to an output layer that are not used in determining the output label, but that nevertheless decrease the accuracy of the model output relative to a ground-truth label. The claims also broadly require removing datapoints correlated with such pathways to generate a reduced training dataset that trains the model to be at least as accurate as the original training dataset. The specification does not provide sufficient guidance for the full scope of this broad functional result.
(B) The nature of the invention; The invention concerns training machine learning models, identifying internal model pathways, correlating those pathways with training datapoints, and reducing a training dataset while maintaining accuracy. This involves internal model behavior, node weights, pathway selection, explainability analysis, and model validation. Because the claimed invention depends on how datapoints affect internal model pathways and output accuracy, the disclosure must explain how a pathway that is not used in the output determination can still decrease output accuracy.
(C) The state of the prior art; The prior art may generally teach neural-network nodes, pathways, contiguous neurons, explainability techniques, and dataset reduction. However, the issue is not whether contiguous nodes or explainability techniques were generally known. The issue is whether the specification enables the particular claimed relationship: identifying pathways that are not used in determining the label, treating those pathways as decreasing label accuracy, and removing correlated datapoints to obtain a model that is at least as accurate. The specification does not provide sufficient teaching for that specific claimed result.
(D) The level of one of ordinary skill; A person of ordinary skill in the art would likely understand machine-learning models, node weights, training datasets, and explainability methods such as LIME or SHAP. However, even a skilled artisan would still need the specification to explain how to practice the claimed contradiction between a non-contributing pathway and a pathway that decreases output accuracy. General skill in machine learning does not cure the lack of guidance for identifying and removing datapoints tied to pathways that are not used in the output determination.
(E) The level of predictability in the art; Machine-learning training is not fully predictable because the effect of removing particular datapoints depends on the model architecture, weights, training process, dataset, and validation metric. It would not be predictable that removing datapoints correlated with pathways that do not contribute to the output label would improve or maintain model accuracy. This factor weighs against enablement because the claimed result would require more than routine application of known ML techniques.
(F) The amount of direction provided by the inventor; The specification provides some general direction for identifying pathways, correlating pathways with datapoints, and removing datapoints. However, the specification states that nodes below a threshold “do not contribute” to the determination to output the label. See Par. [0073]. The specification does not provide sufficient direction explaining how such non-contributing nodes or pathways can decrease accuracy, or how datapoints correlated with those non-contributing pathways are identified and removed to produce the claimed accuracy result.
(G) The existence of working examples; The specification provides conceptual examples involving checks, customer feedback, and consumer data. See Par. [0051], [0059]. Those examples involve features that affect the model, such as ink blots, scratches, or irrelevant consumer-data features that make a significant contribution to training. The examples do not show a working implementation where datapoints are removed because they correlate with pathways that are not used in the determination to output the label, yet the reduced dataset still trains the model to be at least as accurate.
(H) The quantity of experimentation needed to make or use the invention based on the content of the disclosure. One of ordinary skill in the art would be required to determine, without sufficient guidance from the specification, how to identify pathways that are not used in the output determination, how such non-used pathways could nevertheless decrease output accuracy, how to associate those pathways with particular datapoints, and how to remove those datapoints while still maintaining or improving model accuracy. Because the specification does not explain how a non-contributing pathway can reduce the accuracy of the model’s output, practicing the full scope of the claimed invention would require undue experimentation rather than routine implementation
After reviewing all the Wands factors, the examiner finds that the specification lacks
enablement for “identifying one or more pathways that decrease an accuracy of the machine learning model outputting the label relative to a ground-truth label, wherein the one or more pathways comprise a continuous plurality of nodes, from an input layer of the machine learning model to an output layer of the machine learning model, that are not used in a determination to output the label”. As such, Claims 1-20 are rejected under 35 U.S.C. 112(a).
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.
Claim 1-20 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, 10, and 15 recite the limitation "identifying, using one or more model explainablity techniques and based on the one or more changes to the weight associated with each node of the plurality of nodes, one or more pathways that decreases an accuracy of the machine learning model outputting the label relative to the ground-truth label, wherein the one or more pathways comprise a continuous plurality of nodes, from an input layer of the machine learning model to an outout layer of the machine learning model, that are not used in a determination to output the label based on the training dataset inputted into the machine learning model". The metes and bounds of “not used in a determination to output the label” cannot be determined with reasonable clarity when the limitation is read considering the specification.
The specification explains that a node may be “not used” when its weight is below a threshold and therefore does not contribute to the determination to output the label. See Par. [0073]. Under that description, the claimed pathway would be inactive or non-contributing. However, the same independent claims also require the pathway to “decrease an accuracy” of the machine learning model outputting the label relative to a ground-truth label. The specification’s examples of accuracy-decreasing datapoints involve features that affect the model, such as ink blots or scratches erroneously interpreted by the model, or irrelevant consumer-data features making a significant contribution to training. See Par. [0050]-[0051], [0059].
These requirements create uncertainty as to the scope of the claimed pathway. A pathway that is below the contribution threshold and does not contribute to outputting the label would not be understood as the same type of pathway that decreases output accuracy by actively affecting the model. Thus, it is unclear whether the claimed “not used” pathway refers to nodes that do not contribute to the label determination, as described in Par. [0073], or to nodes/features that actively contribute to an inaccurate output, as described in Par. [0050]-[0051] and [0059]. Because the claim requires both that the pathway decreases the accuracy of the output and that the pathway not be used in determining that output, a person of ordinary skill in the art would not be able to determine the scope of the claim with reasonable clarity.
Claims 2–9, 11–14, and 16–20 are rejected for the same reasons by virtue of their dependence from the rejected independent claims.
Claim 5 recites the limitation "the mismatched label" in line 2. There is insufficient antecedent basis for this limitation in the claim. The phrase “mismatched label” was not previously introduced.
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 - 20 are rejected under 35 U.S.C. 103 as being unpatentable over Prendki et al.
(hereafter Prendki) (US 20220138561), in view of Shachar et al. (hereinafter Shachar) (US 12124933) and further in view of Wang et al. (hereinafter Wang), a non-patent literature reference titled “Training with More Confidence: Mitigating Injected and Natural Backdoors During Training”).
As best understood:
Regarding Claim 1, Prendki teaches a method (Prendki, Par. [0018], “A computer-implemented process or method”, thus a method is disclosed) comprising:
inputting, by a computing device, a training dataset into a machine learning model to train the machine learning model to output a label […] (Prendki, Par. [0053], “At block 150, using a hardware processor for example, the method is programmed for executing computer instructions that are programmed to receive an input dataset of training data, the input dataset comprising a plurality of records, the input dataset having been previously used to train the second machine learning model”, thus inputting, by a computing device/a hardware processor, a training dataset into a machine learning model to train the machine learning model to output a label is disclosed)
determining, based on one or more datapoints of the training dataset […] (Prendki, Par. [0022], “In an embodiment, the disclosure provides a computer-implemented process of building a predictive (ML) model to predict the usefulness of a record (data point) in the context of the training process of a machine learning model”, thus Prendki discloses evaluating individual training datapoints by predicting their usefulness with respect to a machine learning model’s training and performance, where determinations about model behavior are made by analyzing how specific datapoints influence training outcomes)
identifying […] one or more pathways that decrease an accuracy of the machine learning model outputting the label relative to a ground-truth label, wherein the one or more pathways comprise a continuous plurality of nodes […] (Prendki, Par. [0021], “Training Set Optimization refers to the process of modifying a training set by removing redundant, useless, or harmful data rows; it differs from conventional compression in which each row is compressed by reducing its individual size and is more accurately described as denoising. Filter refers to a classifier (in most cases, binary) that separates a first subset of data having high information value from a second subset of data having less or no information value.”, & Par. [0048], “The process described thus far offers many benefits and improvements over prior approaches. First, the process is agnostic concerning models. Using several models built for the same task, an implementation can build a more robust filter that will work for any model within the same family of tasks. By using models for different tasks on the same dataset, it is possible to build a map of the data in terms of its absolute value; data that is useless across all tasks is useless in the absolute”, thus identifying one or more pathways that decrease an accuracy of the machine learning model outputting the label is disclosed, because, as stated in the applicant’s specification – Par [0071], the one or more pathways may comprise a contiguous plurality of nodes forming a path from a node that receives input based on a datapoint from the training dataset to a node that outputs the label. Prendki identifies datapoints that are harmful to model accuracy by evaluating their effect on model performance, with datapoints classified as harmful corresponding to contiguous computational routes within the machine learning model that lead to inaccurate label outputs)
determining a first set of the one or more datapoints that correlate with the one or more pathways (Prendki, Par. [0021], “Training Set Optimization refers to the process of modifying a training set by removing redundant, useless, or harmful data rows; it differs from conventional compression in which each row is compressed by reducing its individual size and is more accurately described as denoising. Filter refers to a classifier (in most cases, binary) that separates a first subset of data having high information value from a second subset of data having less or no information value”, & Par. [0077], “The goal of the disclosed system is to identify which data records (rows) from the training set are creating such confusion and classify them as “harmful” to the model, in order to eliminate them in future retraining of the model.”, thus Prendki discloses determining a first set of datapoints that correlate with pathways that decrease model accuracy by evaluating individual training records based on their effect on a trained machine learning model’s performance and classifying those records as harmful when they introduce confusion or reduce predictive accuracy. Because, as stated in the applicant’s specification, a pathway may be a contiguous set of nodes from an input node receiving a datapoint to an output node generating a label, Prendki’s identification of harmful datapoints necessarily corresponds to identifying the specific input-to-output computational routes within the model through which those datapoints lead to incorrect label outputs.)
removing the first set of the one or more datapoints from the training dataset to generate a reduced training dataset (Prendki, Par. [0021], “Training Set Optimization refers to the process of modifying a training set by removing redundant, useless, or harmful data rows; it differs from conventional compression in which each row is compressed by reducing its individual size and is more accurately described as denoising.”, thus removing the first set of the one or more datapoints from the training dataset to generate a reduced training dataset is disclosed)
Prendki does not explicitly teach the machine learning model comprising a plurality of nodes and each node, of the plurality of nodes, being associated with a weight, one or more changes to the weight associated with each node of the plurality of nodes, and one or more changes to the weight associated with each node of the plurality of nodes, and using one or more model explainability techniques, relative to a ground-truth label, […] comprise a continuous plurality of nodes from an input layer of [….a machine learning model….] to an output layer of [….a machine learning model….], that are not used in a determination to output […a label…] based on [….a training dataset…] inputted into […a machine learning model…]
However, Shachar teaches wherein the machine learning model comprises a plurality of nodes and each node, of the plurality of nodes, is associated with a weight (Shachar, Par. [0045], “As shown in FIGS. 3 and 4, models 300 and 400 may include different layers, such as an input layer, a hidden layer, and an output layer, each having one or more nodes, however, different layers may also be utilized. For model 300, layers 1104 are shown, while for model 400, layers 1204 are shown”, & Par. [0046], “By continuously providing different sets of training data and penalizing models 300 and 400 when the output is incorrect, models 300 and 400 (and specifically, the representations of the nodes in the hidden layer) may be trained (adjusted) to improve performance of models 300, 400 in data classification. Adjusting models 300 and 400 may include separately adjusting the weights associated with each node in the hidden layer.”, thus Sharchar discloses a machine learning model having multiple layers of nodes in which each node participates in weighted connections that are adjusted during training, such that each node is associated with at least one weight used to compute the model’s output)
Shachar also teaches one or more changes to the weight associated with each node of the plurality of nodes (Shachar, Par. [0046], “By continuously providing different sets of training data and penalizing models 300 and 400 when the output is incorrect, models 300 and 400 (and specifically, the representations of the nodes in the hidden layer) may be trained (adjusted) to improve performance of models 300, 400 in data classification. Adjusting models 300 and 400 may include separately adjusting the weights associated with each node in the hidden layer.”, thus changes/adjustments to the weight associated with each node of plurality of nodes is disclosed)
Additionally, Shachar teaches using one or mode model explainability techniques (Shachar, Par. [0039], “After creation of the models, at block 9, model explanation is performed to understand the importance of micromodels and, inside each micromodel, the importance of the features to the micromodels. Thus, after building the models, an ML model explainer, such as an explanation algorithm, may be used to verify the added value of each separate feature. This may include utilizing SHAP or LIME to obtain a measure of importance of each feature in each classification task. Thereafter, an average of those contributions is determined to obtain a total significance level of each feature”, thus using one or mode model explainability techniques is disclosed)
It would have been obvious for one of ordinary skill in the art before the effective filing date
of the claimed invention to combine Prendki’s approach of identifying training datapoints whose processing through the machine learning model results in decreased output accuracy, which reads on identifying one or more pathways that decrease an accuracy of the machine learning model outputting the label, with Shachar’s use of ML model explainers to measure the importance of each feature, which corresponds to using one or more model explainability techniques, thereby improving dataset quality and resulting in a more accurate machine learning model (Shachar, Par. [0051], “An ML model explainer may be used to determine an added value of each feature to the ML models' classifications, such as a measure of importance of each feature in the classification tasks. This may be done using SHAP, LIME, or a lift ratio per each feature separately. Thereafter, an administrator may determine whether the risk scores used to enrich the data set provide a more accurate ML model for anomaly detection based on the comparison and feature importance.”) Thus, the combined teachings disclose using explainability feature importance to associate reductions in model accuracy with specific computational pathways activated by particular training datapoints, allowing identification of datapoints that negatively affect model performance and their removal to improve model accuracy.
Sharcar does not explicitly teach relative to a ground-truth label, […] comprise a continuous plurality of nodes from an input layer of [….a machine learning model….] to an output layer of [….a machine learning model….], that are not used in a determination to output […a label…] based on [….a training dataset…] inputted into […a machine learning model…].
However, Wang teaches:
relative to a ground-truth label, […] comprise a continuous plurality of nodes from an input layer of [….a machine learning model….] to an output layer of [….a machine learning model….], that are not used in a determination to output […a label…] based on [….a training dataset…] inputted into […a machine learning model…] (Wang, Page 5 – Section 4, “The first step is to identify comprised neurons, namely the neurons that carry Trojans. Based on our discussion on § 3.1, we do this by checking the activation values of neuron n, denoted as An, to see if its function is highly linear using the condition P(An ≥ 0) ≥ θ. If so, we make the neuron n as potentially compromised and add it to the candidate set C. The second step is to identify highly biased samples or poisoning samples”, & Page 5 – Section 4, “If so, we make the neuron n as potentially compromised and add it to the candidate set C. The second step is to identify highly biased samples or poisoning samples. The overall design is a statistical testing process: we first find a reference distribution of a particular neuron and then mark all inputs whose activation values do not follow such distributions as potentially biased or poisoning samples”, & Page 8 – Section 5.2, “compromised neurons are highly relevant to the Trojan behavior but less related to the model’s original task. Therefore, most benign knowledge is preserved when applying NONE, and the model can still perform well on its original tasks. Meanwhile, NONE fine-tunes the model on purified data after the reset process, further strengthening the model’s capabilities and reducing defense costs”, thus Wang teaches identifying neurons associated with undesirable model behavior by checking neuron activation values and marking neurons as potentially compromised when their activations satisfy a statistical condition. Wang further teaches identifying biased or poisoning samples by finding a reference distribution for a neuron and marking inputs whose activation values deviate from that distribution as potentially biased or poisoning samples. Because Wang explains that compromised neurons are highly relevant to Trojan behavior but less related to the model’s original task, Wang teaches or suggests identifying neurons that are associated with inaccurate or undesired model behavior relative to the original task while preserving the model’s benign knowledge. Wang also teaches using purified data after identifying such compromised neurons and poisoning samples, thereby teaching or suggesting removing datapoints associated with the accuracy-decreasing behavior so that the model can continue to perform well on its original task)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to combine Wang’s teaching of identifying compromised neurons and poisoning samples with Prendki and Shachar’s method for reducing a training dataset. Wang teaches identifying internal model structures associated with undesirable or inaccurate model behavior by analyzing neuron activation values, marking neurons as potentially compromised, and identifying biased or poisoning samples based on activation distributions. Wang further teaches that compromised neurons are highly relevant to Trojan behavior but less related to the model’s original task, and that using purified data after identifying such compromised neurons and poisoning samples preserves benign knowledge and allows the model to continue performing well on its original tasks. Therefore, a POSITA would have been motivated to incorporate Wang’s neuron/pathway-based identification of compromised model behavior and correlated poisoning samples into the Prendki/Shachar dataset-reduction method in order to more accurately identify the specific training datapoints responsible for degraded accuracy, instead of relying only on record-level usefulness scoring (Wang, Page 19 – Section 8.3, “DP-SGD [24] improves existing SGD methods by removing the noise added to poisoned training samples and shows promising results in defending against Trojans”)
Regarding Claim 2, Prendki and Shachar combined with Wang teaches all of the limitations of claim 1 as cited above and Prendki further teaches:
inputting the reduced training dataset into the machine learning model to determine whether the machine learning model outputs the label (Prendki, Par. [0046], “7: m.sub.select ← train(m, selected) (train a new model on selected) accuracy.sub.selected = test(m.sub.select, S.sub.test) “, & Par. [0058], “At block 160, the process executes computer instructions that are programmed to filter the second dataset of prospective training data using the data filter, and to output a refined training dataset comprising fewer records than the second dataset, the refined training dataset comprising only records of the second dataset having the usefulness value greater than a specified threshold.”, thus inputting the reduced training dataset into the machine learning model to determine whether the machine learning model outputs the label is disclosed, because Prendki retrains a machine learning model using a reduced (refined) training dataset and evaluates the model’s predictive output and accuracy on a test set, which corresponds to determining whether the machine learning model outputs the label when trained on the reduced training dataset)
comparing a first label outputted by the machine learning model trained on the training dataset to a second label outputted by the machine learning model trained on the reduced training dataset to determine whether the reduced training dataset causes the machine learning model to render a determination at least as accurate as the training dataset (Prendki, Par. [102], “The next step of filter validation 204 (FIG. 2) includes testing if the data filter does not generate biases, and that the accuracy obtained is as expected (function of how the threshold was set). For a more thorough estimation of the filter's efficacy, there can be a held-out training dataset which is filtered down. Two versions of the model are trained, one on the entire dataset and the other on the filtered down version. If the filtered down version achieves a similar accuracy level to the full version, then the data filter is useful.”, thus comparing a first label outputted by a model trained on the full training dataset with a second label outputted by a model trained on a reduced training dataset to evaluate whether the reduced dataset maintains at least comparable accuracy is disclosed, because Prendki trains two separate models on the full and filtered datasets and compares their resulting predictive accuracy to determine whether the reduced training dataset yields determinations that are at least as accurate as those produced using the full training dataset)
in response to a determination that the reduced training dataset causes the machine learning model to render a determination at least as accurate as the training dataset, determining that the reduced training dataset is valid (Prendki, Par. [102], “Two versions of the model are trained, one on the entire dataset and the other on the filtered down version. If the filtered down version achieves a similar accuracy level to the full version, then the data filter is useful”, &Par. [103], “For instance, if the filter predicts 10% of the data as “useful”, the future versions of the model will be able to be trained with only 10% of the data (note that the amount of data used when training a model is not necessarily linear with the amount of time it takes to train; the disclosed system also provides customers with the capability to review this relationship)”, thus determining that the reduced training dataset is valid is disclosed, because Prendki evaluates the performance of a machine learning model trained on a reduced training dataset relative to a model trained on the full dataset and, upon determining that comparable accuracy is achieved, deems the reduced dataset useful and acceptable for future training, which corresponds to determining the validity of the reduced training dataset in response to achieving at least equivalent model accuracy)
Regarding Claim 3, Prendki and Shachar combined with Wang teaches all of the limitations of claim 1 as cited above and Prendki further teaches:
determining that the first set of the one or more datapoints causes the weight associated with each node associated with the one or more pathways to change by more than a threshold amount (Prendki, Par. [0022 - 0026], “In an embodiment, the disclosure provides a computer-implemented process of building a predictive (ML) model to predict the usefulness of a record (data point) in the context of the training process of a machine learning model. According to one embodiment, the following algorithmic flow is programmed [0023] 1. Collect/acquire (historical) training data. In the pseudocode algorithm examples set forth below, training data is denoted Strain. [0024] 2. Run process to measure usefulness of records within this training dataset (*); measurement of usefulness can be categorical or a score (number) [0025] 3. Categorize training data into groups of usefulness (*) [0026] This can be binary (useful/not useful), and can use a process to establish a threshold above which data is useful”, & Par. [0073], “All the details computed during the metadata generation phase are referred to as “metadata”—they are not data per se, but by-products of the training of the customer's model using a fraction of the customer's data that the disclosed system will use in the next stages of the process. Examples of metadata include, but are not limited to: Inference, Binary “correctness” (correctly/incorrectly predicted), Unlikelihood of prediction (if a record is predicted to be of a class that is rarely confused with ‘true’ class confusion matrix), Confidence level, First margin (difference between confidence of predicted class and next best class), Subsequent margins, Consensus between multiple models (can be perturbed versions of the same model) “Bayesian” confidence, List of activated neurons (if neural net), Activation functions, Weights and biases in model, and/or their derivatives, “Path length” (if decision tree)”, thus Prendki discloses determining, based on training-generated metadata, that particular datapoints produce changes in model weights and activations along specific computational pathways, and further discloses applying threshold-based criteria to classify those datapoints, which corresponds to determining that the first set of datapoints causes the weight associated with each node associated with the one or more pathways to change by more than a threshold amount.)
Regarding Claim 4, Prendki and Shachar combined with Wang teaches all of the limitations of claim 1 as cited above and Sharchar further teaches:
wherein the one or more changes comprise at least one of: a magnitude by which the weight associated with each node of the plurality of nodes changes; or a direction in which the weight associated with each node of the plurality of nodes changes (Sharchar, Par. [0033], “Normalizing may also occur where data sets are normalized to reduce their means and then scaled by the standard deviation of each feature. Normalizing may be performed due to two main reasons. First, gradient descent-based algorithms introduce exploding gradients if the features are not normalized”, & Par. [0046], “Models 300 and 400 may be separately trained using training data, where the nodes in the hidden layer may be trained (adjusted) such that an optimal output (e.g., a classification) is produced in the output layer based on the training data”, thus the one or more changes comprise at least one of: a magnitude by which the weight associated with each node of the plurality of nodes changes; or a direction in which the weight associated with each node of the plurality of nodes changes is disclosed, because Shachar teaches gradient-descent-based training in which node weights in the hidden layers are iteratively adjusted in response to training error, and such adjustments involve changes in both the magnitude and the direction of the weights as part of improving model output accuracy)
Regarding Claim 5, Prendki and Shachar combined with Wang teaches all of the limitations of claim 1 as cited above and Wang further teaches:
wherein each of the one or more pathways comprises a plurality of nodes that are used to output the mismatched label relative to the ground-truth label based on the training dataset inputted into the machine learning model (Wang, Page 5 – Section 4, “The first step is to identify compromised neurons, namely the neurons that carry Trojans. Based on our discussion on § 3.1, we do this by checking the activation values of neuron n, denoted as An, to see if its function is highly linear using the condition P(An ≥ 0) ≥ θ. If so, we mark the neuron n as potentially compromised and add it to the candidate set C. The second step is to identify highly biased samples or poisoning samples”, & Page 5 – section 4, “The overall design is a statistical testing process: we first find a reference distribution of a particular neuron and then mark all inputs whose activation values do not follow such distributions as potentially biased or poisoning samples”, & Page 8 – Section 5.2, “compromised neurons are highly relevant to the Trojan behavior but less related to the model’s original task”, thus Wang teaches a plurality of nodes that are used to output a mismatched label relative to a ground-truth label because Wang identifies compromised neurons as neurons that carry Trojans and are highly relevant to Trojan behavior. Wang further teaches analyzing activation values of particular neurons and identifying inputs whose activation values deviate from a reference distribution as potentially biased or poisoning samples. Since the compromised neurons are associated with Trojan behavior and the poisoning samples cause the model to produce an undesired output rather than the correct output for the original task, Wang teaches nodes that are used to output a mismatched label relative to the ground-truth label based on the poisoned training data inputted into the machine learning mode)
Regarding Claim 6, Prendki and Shachar combined with Wang teaches all of the limitations of claim 1 as cited above and Shachar further teaches:
determining the first set of the one or more datapoints that correlate with the one or more pathways that cause the machine learning model to have a net decrease in outputting the label over the plurality of epochs (Shachar, Par. [0077], “Other tests with Active Learning (e.g., a process where the model is trained iteratively after gradually incrementing the size of the training set) have shown that, at times, the model oscillates from a state where it seems to have understood a class, back to a state where it is clearly confused”, &Par. [0088], “In some embodiments, this approach is simplistic because whenever a training record ends up helping for one class (typically, the one it belongs to) and hurting another, the formula would annihilate those different effects on different test records; which is why in practice, the system may use other approaches to correlate the absence/presence of a record from the training set to its effect on the training (inferred on the test set). Assuming that the ground truth is available for the training set also, it is possible to correlate those effects with more precision”, &Par. [0098], “The “threshold” can either reflect the maximum amount of the data that is desired to be used when training future versions of the model, or the limit (value) under which data seems to become useless (flat learning curve) or harmful (decreasing learning curve)”, thus determining the first set of the one or more datapoints that correlate with the one or more pathways that cause the machine learning model to have a net decrease in outputting the label over the plurality of epochs is disclosed, because Shachar teaches iterative, epoch-based training in which the presence or absence of individual training records is correlated with model behavior across training iterations, including oscillations, confusion, and decreasing learning curves, and further teaches identifying datapoints whose inclusion results in harmful or degrading effects on model performance over time based on thresholded learning trends, which corresponds to correlating specific datapoints with model pathways that negatively affect label output across epochs, thereby enabling removal of datapoints that degrade training performance to improve model accuracy)
Regarding Claim 7, Prendki and Shachar combined with Wang teaches all of the limitations of claim 1 as cited above and Prendki further teaches:
wherein the accuracy of the machine learning model is based on at least one of: a classification accuracy of the machine learning model; or a logarithmic loss of the machine learning model (Prendki, Claim 6, “determining a first accuracy value representing a first classification accuracy of the another machine learning model that has been trained using the refined training dataset”, thus the accuracy of the machine learning model being based on at least one of: a classification accuracy of the machine learning model; or a logarithmic loss of the machine learning model is disclosed)
Regarding Claim 8, Prendki and Shachar combined with Wang teaches all of the limitations of claim 1 as cited above and Prendki further teaches:
identifying, using the one or more model explainability techniques and the one or more changes to the weight associated with each node of the plurality of nodes, one or more second pathways, wherein the one or more second pathways increase the accuracy of the machine learning model outputting the label (Pendki, Par. [0058], “At block 160, the process executes computer instructions that are programmed to filter the second dataset of prospective training data using the data filter, and to output a refined training dataset comprising fewer records than the second dataset, the refined training dataset comprising only records of the second dataset having the usefulness value greater than a specified threshold”, &Par. [0079], “One way to do so is to simply average, for each data record from training set, the confidence level achieved for each data record from within the test set and each sample (run) with a weight of +1 if the prediction for that record is correct, and −1 if it's incorrect, whenever this data record has been used to train the model. The metadata can be used to improve the confidence level. By doing so, the disclosed system will have high scores for each training record if they consistently help the model learn correctly”, thus identifying one or more second pathways that increase model accuracy, because Prendki evaluates the contribution of training datapoints to correct predictions using confidence and correctness metadata generated during model training, and selects datapoints that positively influence learned weights and model behavior, which correspond to contiguous computational pathways through nodes whose weight updates improve label output accuracy)
determining a second set of the one or more datapoints that correlate with the one or more second pathways (Prendki, Par. [0058], “At block 160, the process executes computer instructions that are programmed to filter the second dataset of prospective training data using the data filter, and to output a refined training dataset comprising fewer records than the second dataset, the refined training dataset comprising only records of the second dataset having the usefulness value greater than a specified threshold”, thus determining a second set of the one or more datapoints that correlate with the one or more second pathways is disclosed, because Prendki selects and retains training records based on their positive usefulness values derived from training behavior, which correspond to datapoints associated with computational pathways)
generating a second reduced training dataset comprising the second set of the one or more datapoints (Prendki, Par. [0058], “At block 160, the process executes computer instructions that are programmed to filter the second dataset of prospective training data using the data filter, and to output a refined training dataset comprising fewer records than the second dataset, the refined training dataset comprising only records of the second dataset having the usefulness value greater than a specified threshold”, thus generating a second reduced training dataset comprising the second set of the one or more datapoints is disclosed)
Regarding Claim 9, Prendki and Shachar combined with Wang teaches all of the limitations of claim 1 as cited above and Shachar further teaches:
wherein the one or more model explainability techniques comprise at least one of: a local interpretable model-agnostic explanations technique; or a Shapley additive explanations technique (Sharchar, Par. [0039], “After creation of the models, at block 9, model explanation is performed to understand the importance of micromodels and, inside each micromodel, the importance of the features to the micromodels. Thus, after building the models, an ML model explainer, such as an explanation algorithm, may be used to verify the added value of each separate feature. This may include utilizing SHAP or LIME to obtain a measure of importance of each feature in each classification task. Thereafter, an average of those contributions is determined to obtain a total significance level of each feature”, thus the one or more model explainability techniques comprising at least one of: a local interpretable model-agnostic explanations technique; or a Shapley additive explanations technique is disclosed)
Regarding Claim 10, Prendki teaches a non-transitory machine-readable medium (Prendki, Claim 18, “One or more non-transitory computer-readable media”, thus a non-transitory machine-readable medium is disclosed) storing instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:
inputting, by a computing device, a training dataset into a machine learning model to train the machine learning model to output a label […] (Prendki, Par. [0053], “At block 150, using a hardware processor for example, the method is programmed for executing computer instructions that are programmed to receive an input dataset of training data, the input dataset comprising a plurality of records, the input dataset having been previously used to train the second machine learning model”, thus inputting, by a computing device/a hardware processor, a training dataset into a machine learning model to train the machine learning model to output a label is disclosed)
determining, based on one or more datapoints of the training dataset […] (Prendki, Par. [0022], “In an embodiment, the disclosure provides a computer-implemented process of building a predictive (ML) model to predict the usefulness of a record (data point) in the context of the training process of a machine learning model”, thus Prendki discloses evaluating individual training datapoints by predicting their usefulness with respect to a machine learning model’s training and performance, where determinations about model behavior are made by analyzing how specific datapoints influence training outcomes)
identifying […] one or more pathways that decrease an accuracy of the machine learning model outputting the label relative to a ground-truth label, wherein the one or more pathways comprise a continuous plurality of nodes […] (Prendki, Par. [0021], “Training Set Optimization refers to the process of modifying a training set by removing redundant, useless, or harmful data rows; it differs from conventional compression in which each row is compressed by reducing its individual size and is more accurately described as denoising. Filter refers to a classifier (in most cases, binary) that separates a first subset of data having high information value from a second subset of data having less or no information value.”, & Par. [0048], “The process described thus far offers many benefits and improvements over prior approaches. First, the process is agnostic concerning models. Using several models built for the same task, an implementation can build a more robust filter that will work for any model within the same family of tasks. By using models for different tasks on the same dataset, it is possible to build a map of the data in terms of its absolute value; data that is useless across all tasks is useless in the absolute”, thus identifying one or more pathways that decrease an accuracy of the machine learning model outputting the label is disclosed, because, as stated in the applicant’s specification – Par [0071], the one or more pathways may comprise a contiguous plurality of nodes forming a path from a node that receives input based on a datapoint from the training dataset to a node that outputs the label. Prendki identifies datapoints that are harmful to model accuracy by evaluating their effect on model performance, with datapoints classified as harmful corresponding to contiguous computational routes within the machine learning model that lead to inaccurate label outputs)
determining a first set of the one or more datapoints that correlate with the one or more pathways (Prendki, Par. [0021], “Training Set Optimization refers to the process of modifying a training set by removing redundant, useless, or harmful data rows; it differs from conventional compression in which each row is compressed by reducing its individual size and is more accurately described as denoising. Filter refers to a classifier (in most cases, binary) that separates a first subset of data having high information value from a second subset of data having less or no information value”, & Par. [0077], “The goal of the disclosed system is to identify which data records (rows) from the training set are creating such confusion and classify them as “harmful” to the model, in order to eliminate them in future retraining of the model.”, thus Prendki discloses determining a first set of datapoints that correlate with pathways that decrease model accuracy by evaluating individual training records based on their effect on a trained machine learning model’s performance and classifying those records as harmful when they introduce confusion or reduce predictive accuracy. Because, as stated in the applicant’s specification, a pathway may be a contiguous set of nodes from an input node receiving a datapoint to an output node generating a label, Prendki’s identification of harmful datapoints necessarily corresponds to identifying the specific input-to-output computational routes within the model through which those datapoints lead to incorrect label outputs.)
removing the first set of the one or more datapoints from the training dataset to generate a reduced training dataset (Prendki, Par. [0021], “Training Set Optimization refers to the process of modifying a training set by removing redundant, useless, or harmful data rows; it differs from conventional compression in which each row is compressed by reducing its individual size and is more accurately described as denoising.”, thus removing the first set of the one or more datapoints from the training dataset to generate a reduced training dataset is disclosed)
Prendki does not explicitly teach the machine learning model comprising a plurality of nodes and each node, of the plurality of nodes, being associated with a weight, one or more changes to the weight associated with each node of the plurality of nodes, and one or more changes to the weight associated with each node of the plurality of nodes, and using one or more model explainability techniques, relative to a ground-truth label, […] comprise a continuous plurality of nodes from an input layer of [….a machine learning model….] to an output layer of [….a machine learning model….], that are not used in a determination to output […a label…] based on [….a training dataset…] inputted into […a machine learning model…]
However, Shachar teaches wherein the machine learning model comprises a plurality of nodes and each node, of the plurality of nodes, is associated with a weight (Shachar, Par. [0045], “As shown in FIGS. 3 and 4, models 300 and 400 may include different layers, such as an input layer, a hidden layer, and an output layer, each having one or more nodes, however, different layers may also be utilized. For model 300, layers 1104 are shown, while for model 400, layers 1204 are shown”, & Par. [0046], “By continuously providing different sets of training data and penalizing models 300 and 400 when the output is incorrect, models 300 and 400 (and specifically, the representations of the nodes in the hidden layer) may be trained (adjusted) to improve performance of models 300, 400 in data classification. Adjusting models 300 and 400 may include separately adjusting the weights associated with each node in the hidden layer.”, thus Sharchar discloses a machine learning model having multiple layers of nodes in which each node participates in weighted connections that are adjusted during training, such that each node is associated with at least one weight used to compute the model’s output)
Shachar also teaches one or more changes to the weight associated with each node of the plurality of nodes (Shachar, Par. [0046], “By continuously providing different sets of training data and penalizing models 300 and 400 when the output is incorrect, models 300 and 400 (and specifically, the representations of the nodes in the hidden layer) may be trained (adjusted) to improve performance of models 300, 400 in data classification. Adjusting models 300 and 400 may include separately adjusting the weights associated with each node in the hidden layer.”, thus changes/adjustments to the weight associated with each node of plurality of nodes is disclosed)
Additionally, Shachar teaches using one or mode model explainability techniques (Shachar, Par. [0039], “After creation of the models, at block 9, model explanation is performed to understand the importance of micromodels and, inside each micromodel, the importance of the features to the micromodels. Thus, after building the models, an ML model explainer, such as an explanation algorithm, may be used to verify the added value of each separate feature. This may include utilizing SHAP or LIME to obtain a measure of importance of each feature in each classification task. Thereafter, an average of those contributions is determined to obtain a total significance level of each feature”, thus using one or mode model explainability techniques is disclosed)
It would have been obvious for one of ordinary skill in the art before the effective filing date
of the claimed invention to combine Prendki’s approach of identifying training datapoints whose processing through the machine learning model results in decreased output accuracy, which reads on identifying one or more pathways that decrease an accuracy of the machine learning model outputting the label, with Shachar’s use of ML model explainers to measure the importance of each feature, which corresponds to using one or more model explainability techniques, thereby improving dataset quality and resulting in a more accurate machine learning model (Shachar, Par. [0051], “An ML model explainer may be used to determine an added value of each feature to the ML models' classifications, such as a measure of importance of each feature in the classification tasks. This may be done using SHAP, LIME, or a lift ratio per each feature separately. Thereafter, an administrator may determine whether the risk scores used to enrich the data set provide a more accurate ML model for anomaly detection based on the comparison and feature importance.”) Thus, the combined teachings disclose using explainability feature importance to associate reductions in model accuracy with specific computational pathways activated by particular training datapoints, allowing identification of datapoints that negatively affect model performance and their removal to improve model accuracy.
Sharcar does not explicitly teach relative to a ground-truth label, […] comprise a continuous plurality of nodes from an input layer of [….a machine learning model….] to an output layer of [….a machine learning model….], that are not used in a determination to output […a label…] based on [….a training dataset…] inputted into […a machine learning model…].
However, Wang teaches:
relative to a ground-truth label, […] comprise a continuous plurality of nodes from an input layer of [….a machine learning model….] to an output layer of [….a machine learning model….], that are not used in a determination to output […a label…] based on [….a training dataset…] inputted into […a machine learning model…] (Wang, Page 5 – Section 4, “The first step is to identify comprised neurons, namely the neurons that carry Trojans. Based on our discussion on § 3.1, we do this by checking the activation values of neuron n, denoted as An, to see if its function is highly linear using the condition P(An ≥ 0) ≥ θ. If so, we make the neuron n as potentially compromised and add it to the candidate set C. The second step is to identify highly biased samples or poisoning samples”, & Page 5 – Section 4, “If so, we make the neuron n as potentially compromised and add it to the candidate set C. The second step is to identify highly biased samples or poisoning samples. The overall design is a statistical testing process: we first find a reference distribution of a particular neuron and then mark all inputs whose activation values do not follow such distributions as potentially biased or poisoning samples”, & Page 8 – Section 5.2, “compromised neurons are highly relevant to the Trojan behavior but less related to the model’s original task. Therefore, most benign knowledge is preserved when applying NONE, and the model can still perform well on its original tasks. Meanwhile, NONE fine-tunes the model on purified data after the reset process, further strengthening the model’s capabilities and reducing defense costs”, thus Wang teaches identifying neurons associated with undesirable model behavior by checking neuron activation values and marking neurons as potentially compromised when their activations satisfy a statistical condition. Wang further teaches identifying biased or poisoning samples by finding a reference distribution for a neuron and marking inputs whose activation values deviate from that distribution as potentially biased or poisoning samples. Because Wang explains that compromised neurons are highly relevant to Trojan behavior but less related to the model’s original task, Wang teaches or suggests identifying neurons that are associated with inaccurate or undesired model behavior relative to the original task while preserving the model’s benign knowledge. Wang also teaches using purified data after identifying such compromised neurons and poisoning samples, thereby teaching or suggesting removing datapoints associated with the accuracy-decreasing behavior so that the model can continue to perform well on its original task)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to combine Wang’s teaching of identifying compromised neurons and poisoning samples with Prendki and Shachar’s method for reducing a training dataset. Wang teaches identifying internal model structures associated with undesirable or inaccurate model behavior by analyzing neuron activation values, marking neurons as potentially compromised, and identifying biased or poisoning samples based on activation distributions. Wang further teaches that compromised neurons are highly relevant to Trojan behavior but less related to the model’s original task, and that using purified data after identifying such compromised neurons and poisoning samples preserves benign knowledge and allows the model to continue performing well on its original tasks. Therefore, a POSITA would have been motivated to incorporate Wang’s neuron/pathway-based identification of compromised model behavior and correlated poisoning samples into the Prendki/Shachar dataset-reduction method in order to more accurately identify the specific training datapoints responsible for degraded accuracy, instead of relying only on record-level usefulness scoring (Wang, Page 19 – Section 8.3, “DP-SGD [24] improves existing SGD methods by removing the noise added to poisoned training samples and shows promising results in defending against Trojans”)
Regarding Claim 11, Prendki and Shachar combined with Wang teaches all of the limitations of claim 10 as cited above and Prendki further teaches:
inputting the reduced training dataset into the machine learning model to determine whether the machine learning model outputs the label (Prendki, Par. [0046], “7: m.sub.select ← train(m, selected) (train a new model on selected) accuracy.sub.selected = test(m.sub.select, S.sub.test) “, & Par. [0058], “At block 160, the process executes computer instructions that are programmed to filter the second dataset of prospective training data using the data filter, and to output a refined training dataset comprising fewer records than the second dataset, the refined training dataset comprising only records of the second dataset having the usefulness value greater than a specified threshold.”, thus inputting the reduced training dataset into the machine learning model to determine whether the machine learning model outputs the label is disclosed, because Prendki retrains a machine learning model using a reduced (refined) training dataset and evaluates the model’s predictive output and accuracy on a test set, which corresponds to determining whether the machine learning model outputs the label when trained on the reduced training dataset)
comparing a first label outputted by the machine learning model trained on the training dataset to a second label outputted by the machine learning model trained on the reduced training dataset to determine whether the reduced training dataset causes the machine learning model to render a determination at least as accurate as the training dataset (Prendki, Par. [102], “The next step of filter validation 204 (FIG. 2) includes testing if the data filter does not generate biases, and that the accuracy obtained is as expected (function of how the threshold was set). For a more thorough estimation of the filter's efficacy, there can be a held-out training dataset which is filtered down. Two versions of the model are trained, one on the entire dataset and the other on the filtered down version. If the filtered down version achieves a similar accuracy level to the full version, then the data filter is useful.”, thus comparing a first label outputted by a model trained on the full training dataset with a second label outputted by a model trained on a reduced training dataset to evaluate whether the reduced dataset maintains at least comparable accuracy is disclosed, because Prendki trains two separate models on the full and filtered datasets and compares their resulting predictive accuracy to determine whether the reduced training dataset yields determinations that are at least as accurate as those produced using the full training dataset)
and in response to a determination that the reduced training dataset causes the machine learning model to render a determination at least as accurate as the training dataset, validating the reduced training dataset (Prendki, Par. [102], “Two versions of the model are trained, one on the entire dataset and the other on the filtered down version. If the filtered down version achieves a similar accuracy level to the full version, then the data filter is useful”, &Par. [103], “For instance, if the filter predicts 10% of the data as “useful”, the future versions of the model will be able to be trained with only 10% of the data (note that the amount of data used when training a model is not necessarily linear with the amount of time it takes to train; the disclosed system also provides customers with the capability to review this relationship)”, thus validating the reduced training dataset is disclosed, because Prendki evaluates the performance of a machine learning model trained on a reduced training dataset relative to a model trained on the full dataset and, upon determining that comparable accuracy is achieved, deems the reduced dataset useful and acceptable for future training, which corresponds to determining the validity of the reduced training dataset in response to achieving at least equivalent model accuracy)
Regarding Claim 12, Prendki and Shachar combined with Wang teaches all of the limitations of claim 10 as cited above and Prendki further teaches:
determining the first set of the one or more datapoints that causes the weight associated with each node associated with the one or more pathways to change by more than a threshold amount (Prendki, Par. [0022 - 0026], “In an embodiment, the disclosure provides a computer-implemented process of building a predictive (ML) model to predict the usefulness of a record (data point) in the context of the training process of a machine learning model. According to one embodiment, the following algorithmic flow is programmed [0023] 1. Collect/acquire (historical) training data. In the pseudocode algorithm examples set forth below, training data is denoted Strain. [0024] 2. Run process to measure usefulness of records within this training dataset (*); measurement of usefulness can be categorical or a score (number) [0025] 3. Categorize training data into groups of usefulness (*) [0026] This can be binary (useful/not useful), and can use a process to establish a threshold above which data is useful”, & Par. [0073], “All the details computed during the metadata generation phase are referred to as “metadata”—they are not data per se, but by-products of the training of the customer's model using a fraction of the customer's data that the disclosed system will use in the next stages of the process. Examples of metadata include, but are not limited to: Inference, Binary “correctness” (correctly/incorrectly predicted), Unlikelihood of prediction (if a record is predicted to be of a class that is rarely confused with ‘true’ class confusion matrix), Confidence level, First margin (difference between confidence of predicted class and next best class), Subsequent margins, Consensus between multiple models (can be perturbed versions of the same model) “Bayesian” confidence, List of activated neurons (if neural net), Activation functions, Weights and biases in model, and/or their derivatives, “Path length” (if decision tree)”, thus Prendki discloses determining, based on training-generated metadata, that particular datapoints produce changes in model weights and activations along specific computational pathways, and further discloses applying threshold-based criteria to classify those datapoints, which corresponds to determining that the first set of datapoints causes the weight associated with each node associated with the one or more pathways to change by more than a threshold amount.)
Regarding Claim 13, Prendki and Shachar combined with Wang teaches all of the limitations of claim 10 as cited above and Wang further teaches:
wherein each of the one or more pathways comprises a plurality of nodes that are used to output a mismatched label relative to the ground-truth label based on the training dataset inputted into the machine learning model (Wang, Page 5 – Section 4, “The first step is to identify compromised neurons, namely the neurons that carry Trojans. Based on our discussion on § 3.1, we do this by checking the activation values of neuron n, denoted as An, to see if its function is highly linear using the condition P(An ≥ 0) ≥ θ. If so, we mark the neuron n as potentially compromised and add it to the candidate set C. The second step is to identify highly biased samples or poisoning samples”, & Page 5 – section 4, “The overall design is a statistical testing process: we first find a reference distribution of a particular neuron and then mark all inputs whose activation values do not follow such distributions as potentially biased or poisoning samples”, & Page 8 – Section 5.2, “compromised neurons are highly relevant to the Trojan behavior but less related to the model’s original task”, thus Wang teaches a plurality of nodes that are used to output a mismatched label relative to a ground-truth label because Wang identifies compromised neurons as neurons that carry Trojans and are highly relevant to Trojan behavior. Wang further teaches analyzing activation values of particular neurons and identifying inputs whose activation values deviate from a reference distribution as potentially biased or poisoning samples. Since the compromised neurons are associated with Trojan behavior and the poisoning samples cause the model to produce an undesired output rather than the correct output for the original task, Wang teaches nodes that are used to output a mismatched label relative to the ground-truth label based on the poisoned training data inputted into the machine learning mode)
Regarding Claim 14, Prendki combined with Shachar teaches all of the limitations of claim 10 as cited above and Shachar further teaches:
determining that the first set of the one or more datapoints causes the one or more pathways to have a net decrease in outputting the label over the plurality of epochs (Shachar, Par. [0077], “Other tests with Active Learning (e.g., a process where the model is trained iteratively after gradually incrementing the size of the training set) have shown that, at times, the model oscillates from a state where it seems to have understood a class, back to a state where it is clearly confused”, &Par. [0088], “In some embodiments, this approach is simplistic because whenever a training record ends up helping for one class (typically, the one it belongs to) and hurting another, the formula would annihilate those different effects on different test records; which is why in practice, the system may use other approaches to correlate the absence/presence of a record from the training set to its effect on the training (inferred on the test set). Assuming that the ground truth is available for the training set also, it is possible to correlate those effects with more precision”, &Par. [0098], “The “threshold” can either reflect the maximum amount of the data that is desired to be used when training future versions of the model, or the limit (value) under which data seems to become useless (flat learning curve) or harmful (decreasing learning curve)”, thus determining that the first set of the one or more datapoints causes the one or more pathways to have a net decrease in outputting the label over the plurality of epochs is disclosed, because Shachar teaches iterative, epoch-based training in which the presence or absence of individual training records is correlated with model behavior across training iterations, including oscillations, confusion, and decreasing learning curves, and further teaches identifying datapoints whose inclusion results in harmful or degrading effects on model performance over time based on thresholded learning trends, which corresponds to correlating specific datapoints with model pathways that negatively affect label output across epochs, thereby enabling removal of datapoints that degrade training performance to improve model accuracy)
Regarding Claim 15, Prendki teaches a computing device (Prendki, Par. [0125], “a computing device”, thus a computing device is disclosed) comprising: one or more processors (Prendki, Par. [0127], “a processor”, thus one or more processors is disclosed); and memory (Prendki, Par. [0127], “memory”, memory is disclosed storing instructions that, when executed by the one or more processors, cause the computing device to:
input a training dataset into a machine learning model to train the machine learning model to output a label […] (Prendki, Par. [0053], “At block 150, using a hardware processor for example, the method is programmed for executing computer instructions that are programmed to receive an input dataset of training data, the input dataset comprising a plurality of records, the input dataset having been previously used to train the second machine learning model”, thus inputting a training dataset into a machine learning model to train the machine learning model to output a label is disclosed)
determine, based on one or more datapoints of the training dataset […] (Prendki, Par. [0022], “In an embodiment, the disclosure provides a computer-implemented process of building a predictive (ML) model to predict the usefulness of a record (data point) in the context of the training process of a machine learning model”, thus Prendki discloses evaluating individual training datapoints by predicting their usefulness with respect to a machine learning model’s training and performance, where determinations about model behavior are made by analyzing how specific datapoints influence training outcomes)
identify […] one or more pathways that decrease an accuracy of the machine learning model outputting the label relative to a ground-truth label, wherein the one or more pathways comprise a continuous plurality of nodes […] (Prendki, Par. [0021], “Training Set Optimization refers to the process of modifying a training set by removing redundant, useless, or harmful data rows; it differs from conventional compression in which each row is compressed by reducing its individual size and is more accurately described as denoising. Filter refers to a classifier (in most cases, binary) that separates a first subset of data having high information value from a second subset of data having less or no information value.”, & Par. [0048], “The process described thus far offers many benefits and improvements over prior approaches. First, the process is agnostic concerning models. Using several models built for the same task, an implementation can build a more robust filter that will work for any model within the same family of tasks. By using models for different tasks on the same dataset, it is possible to build a map of the data in terms of its absolute value; data that is useless across all tasks is useless in the absolute”, thus identifying one or more pathways that decrease an accuracy of the machine learning model outputting the label is disclosed, because, as stated in the applicant’s specification – Par [0071], the one or more pathways may comprise a contiguous plurality of nodes forming a path from a node that receives input based on a datapoint from the training dataset to a node that outputs the label. Prendki identifies datapoints that are harmful to model accuracy by evaluating their effect on model performance, with datapoints classified as harmful corresponding to contiguous computational routes within the machine learning model that lead to inaccurate label outputs)
determine a first set of the one or more datapoints that correlate with the one or more pathways (Prendki, Par. [0021], “Training Set Optimization refers to the process of modifying a training set by removing redundant, useless, or harmful data rows; it differs from conventional compression in which each row is compressed by reducing its individual size and is more accurately described as denoising. Filter refers to a classifier (in most cases, binary) that separates a first subset of data having high information value from a second subset of data having less or no information value”, & Par. [0077], “The goal of the disclosed system is to identify which data records (rows) from the training set are creating such confusion and classify them as “harmful” to the model, in order to eliminate them in future retraining of the model.”, thus Prendki discloses determining a first set of datapoints that correlate with pathways that decrease model accuracy by evaluating individual training records based on their effect on a trained machine learning model’s performance and classifying those records as harmful when they introduce confusion or reduce predictive accuracy. Because, as stated in the applicant’s specification, a pathway may be a contiguous set of nodes from an input node receiving a datapoint to an output node generating a label, Prendki’s identification of harmful datapoints necessarily corresponds to identifying the specific input-to-output computational routes within the model through which those datapoints lead to incorrect label outputs.)
remove the first set of the one or more datapoints from the training dataset to generate a reduced training dataset (Prendki, Par. [0021], “Training Set Optimization refers to the process of modifying a training set by removing redundant, useless, or harmful data rows; it differs from conventional compression in which each row is compressed by reducing its individual size and is more accurately described as denoising.”, thus removing the first set of the one or more datapoints from the training dataset to generate a reduced training dataset is disclosed)
Prendki does not explicitly teach the machine learning model comprising a plurality of nodes and each node, of the plurality of nodes, being associated with a weight, one or more changes to the weight associated with each node of the plurality of nodes, and one or more changes to the weight associated with each node of the plurality of nodes, and using one or more model explainability techniques, relative to a ground-truth label, […] comprise a continuous plurality of nodes from an input layer of [….a machine learning model….] to an output layer of [….a machine learning model….], that are not used in a determination to output […a label…] based on [….a training dataset…] inputted into […a machine learning model…].
However, Shachar teaches wherein the machine learning model comprises a plurality of nodes and each node, of the plurality of nodes, is associated with a weight (Shachar, Par. [0045], “As shown in FIGS. 3 and 4, models 300 and 400 may include different layers, such as an input layer, a hidden layer, and an output layer, each having one or more nodes, however, different layers may also be utilized. For model 300, layers 1104 are shown, while for model 400, layers 1204 are shown”, & Par. [0046], “By continuously providing different sets of training data and penalizing models 300 and 400 when the output is incorrect, models 300 and 400 (and specifically, the representations of the nodes in the hidden layer) may be trained (adjusted) to improve performance of models 300, 400 in data classification. Adjusting models 300 and 400 may include separately adjusting the weights associated with each node in the hidden layer.”, thus Sharchar discloses a machine learning model having multiple layers of nodes in which each node participates in weighted connections that are adjusted during training, such that each node is associated with at least one weight used to compute the model’s output)
Shachar also teaches one or more changes to the weight associated with each node of the plurality of nodes (Shachar, Par. [0046], “By continuously providing different sets of training data and penalizing models 300 and 400 when the output is incorrect, models 300 and 400 (and specifically, the representations of the nodes in the hidden layer) may be trained (adjusted) to improve performance of models 300, 400 in data classification. Adjusting models 300 and 400 may include separately adjusting the weights associated with each node in the hidden layer.”, thus changes/adjustments to the weight associated with each node of plurality of nodes is disclosed)
Additionally, Shachar teaches using one or mode model explainability techniques (Shachar, Par. [0039], “After creation of the models, at block 9, model explanation is performed to understand the importance of micromodels and, inside each micromodel, the importance of the features to the micromodels. Thus, after building the models, an ML model explainer, such as an explanation algorithm, may be used to verify the added value of each separate feature. This may include utilizing SHAP or LIME to obtain a measure of importance of each feature in each classification task. Thereafter, an average of those contributions is determined to obtain a total significance level of each feature”, thus using one or mode model explainability techniques is disclosed)
It would have been obvious for one of ordinary skill in the art before the effective filing date
of the claimed invention to combine Prendki’s approach of identifying training datapoints whose processing through the machine learning model results in decreased output accuracy, which reads on identifying one or more pathways that decrease an accuracy of the machine learning model outputting the label, with Shachar’s use of ML model explainers to measure the importance of each feature, which corresponds to using one or more model explainability techniques, thereby improving dataset quality and resulting in a more accurate machine learning model (Shachar, Par. [0051], “An ML model explainer may be used to determine an added value of each feature to the ML models' classifications, such as a measure of importance of each feature in the classification tasks. This may be done using SHAP, LIME, or a lift ratio per each feature separately. Thereafter, an administrator may determine whether the risk scores used to enrich the data set provide a more accurate ML model for anomaly detection based on the comparison and feature importance.”) Thus, the combined teachings disclose using explainability feature importance to associate reductions in model accuracy with specific computational pathways activated by particular training datapoints, allowing identification of datapoints that negatively affect model performance and their removal to improve model accuracy.
Sharcar does not explicitly teach relative to a ground-truth label, […] comprise a continuous plurality of nodes from an input layer of [….a machine learning model….] to an output layer of [….a machine learning model….], that are not used in a determination to output […a label…] based on [….a training dataset…] inputted into […a machine learning model…].
However, Wang teaches:
relative to a ground-truth label, […] comprise a continuous plurality of nodes from an input layer of [….a machine learning model….] to an output layer of [….a machine learning model….], that are not used in a determination to output […a label…] based on [….a training dataset…] inputted into […a machine learning model…] (Wang, Page 5 – Section 4, “The first step is to identify comprised neurons, namely the neurons that carry Trojans. Based on our discussion on § 3.1, we do this by checking the activation values of neuron n, denoted as An, to see if its function is highly linear using the condition P(An ≥ 0) ≥ θ. If so, we make the neuron n as potentially compromised and add it to the candidate set C. The second step is to identify highly biased samples or poisoning samples”, & Page 5 – Section 4, “If so, we make the neuron n as potentially compromised and add it to the candidate set C. The second step is to identify highly biased samples or poisoning samples. The overall design is a statistical testing process: we first find a reference distribution of a particular neuron and then mark all inputs whose activation values do not follow such distributions as potentially biased or poisoning samples”, & Page 8 – Section 5.2, “compromised neurons are highly relevant to the Trojan behavior but less related to the model’s original task. Therefore, most benign knowledge is preserved when applying NONE, and the model can still perform well on its original tasks. Meanwhile, NONE fine-tunes the model on purified data after the reset process, further strengthening the model’s capabilities and reducing defense costs”, thus Wang teaches identifying neurons associated with undesirable model behavior by checking neuron activation values and marking neurons as potentially compromised when their activations satisfy a statistical condition. Wang further teaches identifying biased or poisoning samples by finding a reference distribution for a neuron and marking inputs whose activation values deviate from that distribution as potentially biased or poisoning samples. Because Wang explains that compromised neurons are highly relevant to Trojan behavior but less related to the model’s original task, Wang teaches or suggests identifying neurons that are associated with inaccurate or undesired model behavior relative to the original task while preserving the model’s benign knowledge. Wang also teaches using purified data after identifying such compromised neurons and poisoning samples, thereby teaching or suggesting removing datapoints associated with the accuracy-decreasing behavior so that the model can continue to perform well on its original task)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to combine Wang’s teaching of identifying compromised neurons and poisoning samples with Prendki and Shachar’s method for reducing a training dataset. Wang teaches identifying internal model structures associated with undesirable or inaccurate model behavior by analyzing neuron activation values, marking neurons as potentially compromised, and identifying biased or poisoning samples based on activation distributions. Wang further teaches that compromised neurons are highly relevant to Trojan behavior but less related to the model’s original task, and that using purified data after identifying such compromised neurons and poisoning samples preserves benign knowledge and allows the model to continue performing well on its original tasks. Therefore, a POSITA would have been motivated to incorporate Wang’s neuron/pathway-based identification of compromised model behavior and correlated poisoning samples into the Prendki/Shachar dataset-reduction method in order to more accurately identify the specific training datapoints responsible for degraded accuracy, instead of relying only on record-level usefulness scoring (Wang, Page 19 – Section 8.3, “DP-SGD [24] improves existing SGD methods by removing the noise added to poisoned training samples and shows promising results in defending against Trojans”)
Regarding Claim 16, Prendki combined with Shachar teaches all of the limitations of claim 15 as cited above and Prendki further teaches:
input the reduced training dataset into the machine learning model to determine whether the machine learning model outputs the label (Prendki, Par. [0046], “7: m.sub.select ← train(m, selected) (train a new model on selected) accuracy.sub.selected = test(m.sub.select, S.sub.test) “, & Par. [0058], “At block 160, the process executes computer instructions that are programmed to filter the second dataset of prospective training data using the data filter, and to output a refined training dataset comprising fewer records than the second dataset, the refined training dataset comprising only records of the second dataset having the usefulness value greater than a specified threshold.”, thus inputting the reduced training dataset into the machine learning model to determine whether the machine learning model outputs the label is disclosed, because Prendki retrains a machine learning model using a reduced (refined) training dataset and evaluates the model’s predictive output and accuracy on a test set, which corresponds to determining whether the machine learning model outputs the label when trained on the reduced training dataset)
compare a first label outputted by the machine learning model trained on the training dataset to a second label outputted by the machine learning model trained on the reduced training dataset to determine whether the reduced training dataset causes the machine learning model to render a determination at least as accurate as the training dataset (Prendki, Par. [102], “The next step of filter validation 204 (FIG. 2) includes testing if the data filter does not generate biases, and that the accuracy obtained is as expected (function of how the threshold was set). For a more thorough estimation of the filter's efficacy, there can be a held-out training dataset which is filtered down. Two versions of the model are trained, one on the entire dataset and the other on the filtered down version. If the filtered down version achieves a similar accuracy level to the full version, then the data filter is useful.”, thus comparing a first label outputted by a model trained on the full training dataset with a second label outputted by a model trained on a reduced training dataset to evaluate whether the reduced dataset maintains at least comparable accuracy is disclosed, because Prendki trains two separate models on the full and filtered datasets and compares their resulting predictive accuracy to determine whether the reduced training dataset yields determinations that are at least as accurate as those produced using the full training dataset)
and in response to a determination that the reduced training dataset causes the machine learning model to render a determination at least as accurate as the training dataset, validate the reduced training dataset (Prendki, Par. [102], “Two versions of the model are trained, one on the entire dataset and the other on the filtered down version. If the filtered down version achieves a similar accuracy level to the full version, then the data filter is useful”, &Par. [103], “For instance, if the filter predicts 10% of the data as “useful”, the future versions of the model will be able to be trained with only 10% of the data (note that the amount of data used when training a model is not necessarily linear with the amount of time it takes to train; the disclosed system also provides customers with the capability to review this relationship)”, thus validating the reduced training dataset is disclosed, because Prendki evaluates the performance of a machine learning model trained on a reduced training dataset relative to a model trained on the full dataset and, upon determining that comparable accuracy is achieved, deems the reduced dataset useful and acceptable for future training, which corresponds to determining the validity of the reduced training dataset in response to achieving at least equivalent model accuracy)
Regarding Claim 17, Prendki combined with Shachar teaches all of the limitations of claim 15 as cited above and Prendki further teaches:
determine the first set of the one or more datapoints causes the weight associated with each node associated with the one or more pathways to change by more than a threshold amount (Prendki, Par. [0022 - 0026], “In an embodiment, the disclosure provides a computer-implemented process of building a predictive (ML) model to predict the usefulness of a record (data point) in the context of the training process of a machine learning model. According to one embodiment, the following algorithmic flow is programmed [0023] 1. Collect/acquire (historical) training data. In the pseudocode algorithm examples set forth below, training data is denoted Strain. [0024] 2. Run process to measure usefulness of records within this training dataset (*); measurement of usefulness can be categorical or a score (number) [0025] 3. Categorize training data into groups of usefulness (*) [0026] This can be binary (useful/not useful), and can use a process to establish a threshold above which data is useful”, & Par. [0073], “All the details computed during the metadata generation phase are referred to as “metadata”—they are not data per se, but by-products of the training of the customer's model using a fraction of the customer's data that the disclosed system will use in the next stages of the process. Examples of metadata include, but are not limited to: Inference, Binary “correctness” (correctly/incorrectly predicted), Unlikelihood of prediction (if a record is predicted to be of a class that is rarely confused with ‘true’ class confusion matrix), Confidence level, First margin (difference between confidence of predicted class and next best class), Subsequent margins, Consensus between multiple models (can be perturbed versions of the same model) “Bayesian” confidence, List of activated neurons (if neural net), Activation functions, Weights and biases in model, and/or their derivatives, “Path length” (if decision tree)”, thus Prendki discloses determining, based on training-generated metadata, that particular datapoints produce changes in model weights and activations along specific computational pathways, and further discloses applying threshold-based criteria to classify those datapoints, which corresponds to determining that the first set of datapoints causes the weight associated with each node associated with the one or more pathways to change by more than a threshold amount.)
Regarding Claim 18, Prendki and Shachar combined with Wang teaches all of the limitations of claim 15 as cited above and Wang further teaches:
wherein each of the one or more pathways comprises a plurality of nodes that are used to output a mismatched label relative to the ground-truth label based on the training dataset inputted into the machine learning model (Wang, Page 5 – Section 4, “The first step is to identify compromised neurons, namely the neurons that carry Trojans. Based on our discussion on § 3.1, we do this by checking the activation values of neuron n, denoted as An, to see if its function is highly linear using the condition P(An ≥ 0) ≥ θ. If so, we mark the neuron n as potentially compromised and add it to the candidate set C. The second step is to identify highly biased samples or poisoning samples”, & Page 5 – section 4, “The overall design is a statistical testing process: we first find a reference distribution of a particular neuron and then mark all inputs whose activation values do not follow such distributions as potentially biased or poisoning samples”, & Page 8 – Section 5.2, “compromised neurons are highly relevant to the Trojan behavior but less related to the model’s original task”, thus Wang teaches a plurality of nodes that are used to output a mismatched label relative to a ground-truth label because Wang identifies compromised neurons as neurons that carry Trojans and are highly relevant to Trojan behavior. Wang further teaches analyzing activation values of particular neurons and identifying inputs whose activation values deviate from a reference distribution as potentially biased or poisoning samples. Since the compromised neurons are associated with Trojan behavior and the poisoning samples cause the model to produce an undesired output rather than the correct output for the original task, Wang teaches nodes that are used to output a mismatched label relative to the ground-truth label based on the poisoned training data inputted into the machine learning mode)
Regarding Claim 19, Prendki and Shachar combined with Wang teaches all of the limitations of claim 15 as cited above and Shachar further teaches:
determine that the first set of the one or more datapoints causes the one or more pathways to have a net decrease in outputting the label over the plurality of epochs (Shachar, Par. [0077], “Other tests with Active Learning (e.g., a process where the model is trained iteratively after gradually incrementing the size of the training set) have shown that, at times, the model oscillates from a state where it seems to have understood a class, back to a state where it is clearly confused”, &Par. [0088], “In some embodiments, this approach is simplistic because whenever a training record ends up helping for one class (typically, the one it belongs to) and hurting another, the formula would annihilate those different effects on different test records; which is why in practice, the system may use other approaches to correlate the absence/presence of a record from the training set to its effect on the training (inferred on the test set). Assuming that the ground truth is available for the training set also, it is possible to correlate those effects with more precision”, &Par. [0098], “The “threshold” can either reflect the maximum amount of the data that is desired to be used when training future versions of the model, or the limit (value) under which data seems to become useless (flat learning curve) or harmful (decreasing learning curve)”, thus determining the first set of the one or more datapoints that correlate with the one or more pathways that cause the machine learning model to have a net decrease in outputting the label over the plurality of epochs is disclosed, because Shachar teaches iterative, epoch-based training in which the presence or absence of individual training records is correlated with model behavior across training iterations, including oscillations, confusion, and decreasing learning curves, and further teaches identifying datapoints whose inclusion results in harmful or degrading effects on model performance over time based on thresholded learning trends, which corresponds to correlating specific datapoints with model pathways that negatively affect label output across epochs, thereby enabling removal of datapoints that degrade training performance to improve model accuracy)
Regarding Claim 20, Prendki and Shachar combined with Wang teaches all of the limitations of claim 15 as cited above and Prendki further teaches:
identify, using the one or more model explainability techniques and the one or more changes to the weight associated with each node of the plurality of nodes, one or more second pathways, wherein the one or more second pathways increase the accuracy of the machine learning model outputting the label (Pendki, Par. [0058], “At block 160, the process executes computer instructions that are programmed to filter the second dataset of prospective training data using the data filter, and to output a refined training dataset comprising fewer records than the second dataset, the refined training dataset comprising only records of the second dataset having the usefulness value greater than a specified threshold”, &Par. [0079], “One way to do so is to simply average, for each data record from training set, the confidence level achieved for each data record from within the test set and each sample (run) with a weight of +1 if the prediction for that record is correct, and −1 if it's incorrect, whenever this data record has been used to train the model. The metadata can be used to improve the confidence level. By doing so, the disclosed system will have high scores for each training record if they consistently help the model learn correctly”, thus identifying one or more second pathways that increase model accuracy, because Prendki evaluates the contribution of training datapoints to correct predictions using confidence and correctness metadata generated during model training, and selects datapoints that positively influence learned weights and model behavior, which correspond to contiguous computational pathways through nodes whose weight updates improve label output accuracy)
determine a second set of the one or more datapoints that correlate with the one or more second pathways (Prendki, Par. [0058], “At block 160, the process executes computer instructions that are programmed to filter the second dataset of prospective training data using the data filter, and to output a refined training dataset comprising fewer records than the second dataset, the refined training dataset comprising only records of the second dataset having the usefulness value greater than a specified threshold”, thus determining a second set of the one or more datapoints that correlate with the one or more second pathways is disclosed, because Prendki selects and retains training records based on their positive usefulness values derived from training behavior, which correspond to datapoints associated with computational pathways)
generate a second reduced training dataset comprising the second set of the one or more datapoints (Prendki, Par. [0058], “At block 160, the process executes computer instructions that are programmed to filter the second dataset of prospective training data using the data filter, and to output a refined training dataset comprising fewer records than the second dataset, the refined training dataset comprising only records of the second dataset having the usefulness value greater than a specified threshold”, thus generating a second reduced training dataset comprising the second set of the one or more datapoints is disclosed)
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.
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. US20210295175A1 is pertinent to applicant’s disclosure because it teaches training a neural network by monitoring the model’s internal behavior during training and repeatedly updating weights and biases across many iterations to reach a desired output behavior. It also teaches applying constraint operations during training that identify which connections are conforming versus non-conforming (based on importance measures like weight magnitude or sensitivity), then progressively shrinking the non-conforming connections toward zero over time to enforce restrictions such as limiting disallowed feature interactions and reducing improper bias effects. Since the applicant also focuses on controlling undesirable model behavior (including bias-related behavior) by adjusting internal pathways/parameters during training and corrective procedures, the reference is relevant to the claimed invention.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MAHLIET ADMASU whose telephone number is (571)272-0034. The examiner can normally be reached Mon-Fri, 8am-5pm.
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, Alexey Shmatov can be reached at (571)270-3428. 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.
/M.T.A./Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights