DETAILED ACTION
Claims 1-20 are presented for examination.
This office action is in response to submission of application on 24-FEBRUARY-2026.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 30-DECEMBER-2022 is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Response to Amendment
The amendment filed 24-FEBRUARY-2026 in response to the non-final office action mailed 25-NOVEMBER-2026 has been entered. Claims 1-20 remain pending in the application.
With regards to the non-final office action’s rejection under 101, the amendments to the claims are not sufficient to overcome the original rejection with regards to the claims being directed towards an abstract idea, save for claim 17 which includes an improvement in the mitigation of the neural network’s biases.
With regards to the non-final office action’s rejections under 103, the newly added claims necessitated a new consideration of the art. After this consideration of claims 17-18, the examiner respectfully disagrees with the applicant’s arguments that the art referenced in the previous office action does not teach the newly added claims’ limitations. A new 103 rejection over the prior art has been provided.
With regards to the non-final office action’s rejection under 103, the newly added claims have overcome the original rejection. However, upon a new search for the amended limitations, a new 103 rejection over Nguyen in view of Baker in view of Townsend has been written for claims 16 and 20. In light of the new art, the arguments are moot.
Claim 19 has overcome the prior art, though it is still subject to a rejection under 101.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-16, 18-20 rejected under 35 U.S.C. 101 because the claimed invention is direction to an abstract idea without significantly more.
MPEP 2106.04(a)(2)(Ill) “Accordingly, the "mental processes" abstract idea grouping is defined as concepts performed in the human mind, and examples of mental processes include observations, evaluations, Judgments, and opinions.
Further, the MPEP recites “The courts do not distinguish between mental processes that are performed entirely in the human mind and mental processes that require a human to use a physical aid (e.g., pen and paper or a slide run) to perform the claim limitation.
MPEP 2106.04(a)(2)(I) “The mathematical concepts grouping is defined as mathematical
relationships, mathematical formulas or equations, and mathematical calculations.”
Regarding claim 1:
Step 2A, Prong 1 will now be evaluated for this claim:
A judicial exception is recited in this claim as it recites a mental process:
creating at least one decision tree, where each decision tree is trained to approximate the ANN and optimize a defined criterion
Creating a decision tree would be accomplishable by a human using pen and paper.
for each node of the decision tree a threshold value for the defined criterion being calculated to determine for which node of the ANN the input activations should be split between branches of the decision tree
Determining a threshold value would be part of the creation of a decision tree, which could be performed by a human using a pen and paper.
for each of the threshold value combinations, performing a selected rule extraction algorithm using the combination of threshold values to extract from the ANN at least one rule for explaining the output of the layer of the ANN
Performing a selected rule extraction algorithm may be a series of steps determining the format, nodes, or edges of a decision tree.
for each of the threshold value combinations, obtaining a fidelity metric for the at least one rule using the combination of threshold values, the fidelity metric indicating the accuracy of the rule with respect to the predictions of the ANN
The fidelity metric is a form of comparing the results of the rule to the predictions of the ANN, wherein comparison is a form of evaluation.
determining which of the combinations of threshold values yields the best fidelity metric
This describes comparison of the fidelity metrics, which is an evaluation.
using the selected rule extraction algorithm with the combination of threshold values determined to yield the best fidelity metric to extract at least one rule for explaining the output of the layer of the ANN
This would be an evaluation of the decision tree following a series of steps, which is an evaluation performable by a human with the aid of a pen and paper.
Step 2A, Prong 2 will now be evaluated for this claim:
Furthermore, MPEP 2106.05(g) Insignificant Extra-Solution Activity has found mere data gathering and post-solution activity to be insignificant extra-solution activity.
The following steps are mere data gathering:
recording node activations for each node in a layer of a trained artificial neural network - ANN - and predictions of the ANN, in respect of each item of training data used to train the ANN;
taking as input the recorded node activations and as targets the recorded predictions of the ANN
recording the threshold values associated with respective nodes of the ANN;
obtaining threshold value combinations, each combination comprising one of the threshold values obtained for respective nodes of the ANN
The additional elements have been considered both individually and as an ordered combination in order to determine whether they integrate the exception into a practical application. Therefore, no meaningful limits are imposed practicing the abstract idea.
Therefore, the claim is related to an abstract idea.
Step 2B will now be discussed with regards to this claim:
The claim does not provide an inventive concept. There is no additional Insignificant Extra- Solution Activity, as identified in Step 2A Prong Two, that provides an inventive concept.
Adding insignificant extra-solution activity to the judicial exception, e.g., mere data gathering in conjunction with a law of nature or abstract idea such as a step of obtaining information about credit card transactions so that the information can be analyzed by an abstract mental process, as discussed in CyberSource v. Retail Decisions, Inc., 654 F.3d 1366, 1375, 99 USPQ2d 1690, 1694 (Fed. Cir. 2011) (see MPEP § 2106.05(g)) does not overcome a rejection.
The additional elements have been considered both individually and as an ordered combination as to whether they whether they warrant significantly more consideration.
The claim is ineligible.
Regarding claim 2, which depends upon claim 1:
The following would be a mental process:
ranking the threshold values, for each node of the ANN, according to occurrence frequency and average depth of appearance in the decision tree
Ranking the threshold values via comparison by a certain metric to each other would be evaluation performable in the human mind with the aid of a pen and paper.
performing the selected rule extraction algorithm for each combination of threshold values includes performing the selected rule extraction algorithm first on that combination of threshold values which includes the threshold values occurring with the highest frequencies
Performing a selected rule extraction algorithm may be a series of steps determining the format, nodes, or edges of a decision tree.
This claim incorporates the deficiencies of its parent claim. Furthermore, it in and of itself does not overcome the parent claim’s rejection as an abstract idea.
The claim is ineligible.
Regarding claim 3, which depends upon claim 1:
This claim further limits the obtaining threshold values of claim 1. Further specifying the obtaining of threshold values in this manner does not overcome the parent claim’s rejection.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
Regarding claim 4, which depends upon claim 1:
This claim further limits the defined criterion of claim 1. Further specifying the defined criterion in this manner does not overcome the parent claim’s rejection.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
Regarding claim 5, which depends upon claim 1:
This claim further limits the recording of threshold values of claim 1. Further specifying the recording of threshold values in this manner does not overcome the parent claim’s rejection.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
Regarding claim 6, which depends upon claim 1:
This claim further limits the manner of creating decision trees of claim 1. Further specifying the manner of creating decision trees in this manner does not overcome the parent claim’s rejection as it only incorporates the use of a generic computer.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
Regarding claim 7, which depends upon claim 1:
This claim consists of mere instructions to apply the judicial exception to a particular field of use (MPEP 2106.05(h)), which does not integrate the judicial exception into a practical application.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
Regarding claim 8, which depends upon claim 1:
This claim consists of mere instructions to apply the judicial exception to a particular field of use (MPEP 2106.05(h)), which does not integrate the judicial exception into a practical application.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
Claim 9 recites a non-transitory computer readable storage medium that parallels the method of claim 1. Therefore, the analysis discussed above with respect to claim 1 also applies to claim 9. Accordingly, claim 9 is rejected based on substantially the same rationale as set forth above with respect to claim 1.
Claims 10-15 recite an apparatus that parallels the method of claims 1-6 respectively. Therefore, the analysis discussed above with respect to claims 1-6 also applies to claims 10-15 respectively. Accordingly, claims 10-15 are rejected based on substantially the same rationale as set forth above with respect to claims 1-6 respectively.
Regarding claim 16, which depends upon claim 1:
The following limitation would be a mathematical concept:
performing a guided search using a Bayesian approach to find the threshold values that result in the highest fidelity
The use of a Bayesian approach specifies that a specific, known, algorithm is applied in order to find the highest fidelity, which would be a mathematical process.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
Regarding claim 18, which depends upon claim 1:
The following is well known, routine, conventional activity in the art:
wherein creating the at least one decision tree includes knowledge distillation to approximate the ANN's behavior by taking filter or neuron activations as input and the ANN's predictions as targets.
“The focus of this paper is knowledge distillation, which has received increasing attention from the research community in recent years.” Gou et al. “Knowledge Distillation: A Survey”, May 2021. This paper addressed that knowledge distillation in and of itself is conventional activity, as it has a sufficient body of work to necessitate a review of the method, wherein knowledge distillation itself addresses the approximation of the ANN’s behavior by taking an input and a target. The specific values of neuron activations and predictions do not distinguish the overall method from this routine activity.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
Regarding claim 19, which depends upon claim 1:
The following would be a mental process:
determining an occurrence frequency for each unique threshold value in the decision tree and ranking the threshold values based on said frequency
This would be the same as counting each unique threshold value by occurrence and ranking them, which would be accomplishable by the human mind at a small enough scale.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
Regarding claim 20, which depends upon claim 1:
The following would be mathematical concept:
when a specific node of the ANN does not appear in the at least one decision tree, recording a per sample mean activation of that node as the threshold value
Recording a per sample mean activation would be a calculation of the mean activation.
This claim is rejected for incorporating the parent claim in full.
This claim is ineligible.
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, 3-10, 12-15, and 17-18 are rejected under 35 U.S.C. 103 as being unpatentable over Nguyen et al. (“Towards Interpretable ANNs: An Exact Transformation to Multi-Class Multivariate Decision Trees”, published November 17th 2021, hereinafter Nguyen) in view of Baker et al. (Pub. No. US 20200293897 A1, filed June 3rd 2020, hereinafter Baker).
Regarding claim 1:
Claim 1 recites:
A computer-implemented method comprising: recording node activations for each node in a layer of a trained artificial neural network - ANN - and predictions of the ANN, in respect of each item of training data used to train the ANN; taking as input the recorded node activations and as targets the recorded predictions of the ANN, creating at least one decision tree, where each decision tree is trained to approximate the ANN and optimize a defined criterion, for each node of the decision tree a threshold value for the defined criterion being calculated to determine for which node of the ANN the input activations should be split between branches of the decision tree; recording the threshold values associated with respective nodes of the ANN; obtaining threshold value combinations, each combination comprising one of the threshold values obtained for respective nodes of the ANN; for each of the threshold value combinations, performing a selected rule extraction algorithm using the combination of threshold values to extract from the ANN at least one rule for explaining the output of the layer of the ANN; for each of the threshold value combinations, obtaining a fidelity metric for the at least one rule using the combination of threshold values, the fidelity metric indicating the accuracy of the rule with respect to the predictions of the ANN; determining which of the combinations of threshold values yields the best fidelity metric; and using the selected rule extraction algorithm with the combination of threshold values determined to yield the best fidelity metric to extract at least one rule for explaining the output of the layer of the ANN
Nguyen discloses recording node activations for each node in a layer of a trained artificial neural network - ANN - and predictions of the ANN, in respect of each item of training data used to train the ANN:
Nguyen teaches the representation of an artificial neural network by a decision tree, where the activation of each of nodes, which would include each node in a layer (Page 3). Furthermore, the training data is used to train the ANN as well (Page 7).
Nguyen discloses taking as input the recorded node activations and as targets the recorded predictions of the ANN, creating at least one decision tree, where each decision tree is trained to approximate the ANN [and optimize a defined criterion]:
Nguyen teaches that in addition to the previously taught use of recorded node activations the decision trees are used to represent the input and output of the neural network (Page 3) wherein the output would be recorded predictions of the ANN as target. The creation of the decision tree is the representation of ANN, wherein a representation would be analogous to an approximation of the ANN.
Nguyen does not teach [optimizing a defined criterion], which is taught by Baker further below.
Nguyen discloses for each node of the decision tree a threshold value for the defined criterion being calculated to determine for which node of the ANN the input activations should be split between branches of the decision tree; recording the threshold values associated with respective nodes of the ANN:
Nguyen teaches that the decision tree’s node may be determined by the activation value, which would act as a threshold value as the threshold for the node to be activated acts as a split between branches of the decision tree (Page 13). In doing so, the decision tree records the threshold values associated with respective nodes of the ANN.
Nguyen discloses obtaining threshold value combinations, each combination comprising one of the threshold values obtained for respective nodes of the ANN; for each of the threshold value combinations, performing a selected rule extraction algorithm using the combination of threshold values to extract from the ANN at least one rule for explaining the output of the layer of the ANN:
Nguyen teaches that a decision tree can generate a rule expression in terms of a combination of multiple variables as inputs (Page 12), wherein the variables may include the previously taught threshold value, and in such a case may be a combination comprises one of the threshold values of the decision tree. Furthermore, as stated, Nguyen uses this combination to generate a rule expression which would analogous to performing a selected rule extraction algorithm using the combination of threshold values to extract from the ANN at least one rule for explaining the output of the layer of the ANN as the latter describes the process of generating a rule from the ANN.
Nguyen discloses for each of the threshold value combinations, obtaining a fidelity metric for the at least one rule using the combination of threshold values, the fidelity metric indicating the accuracy of the rule with respect to the predictions of the ANN:
Nguyen teaches a generation of a fidelity metric, wherein the fidelity metric indicates the accuracy of the rule with respect to the predictions of the ANN (Page 3-4). As the fidelity is of the rules specifically, the metric would be for at least one rule using the combination of threshold values, as the rules use the threshold values as previously taught.
Nguyen discloses determining which of the combinations of threshold values yields the best fidelity metric; and using the selected rule extraction algorithm with the combination of threshold values determined to yield the best fidelity metric to extract at least one rule for explaining the output of the layer of the ANN:
Nguyen teaches testing multiple rules, and with them combinations of threshold values, to determine which maintain high fidelity metrics (Page 30-31). As the decision trees are tested, the decision tree with the best fidelity metric has already used the selected rule extraction algorithm with the combination of threshold values determined to yield the best fidelity metric, and extracted as least one rule for explaining the output of the layer of the ANN.
Baker in the same field of endeavor of machine learning discloses optimize a defined criterion:
Baker teaches optimization of hyperparameters (Paragraph 30), which would be an example of optimizing a defined criterion.
Baker and the present application are analogous art because they are in the same field of endeavor of machine learning.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Nguyen and the teachings of Baker. This would have provided the advantage of tuning the learning process to work better with particular kinds of data (Baker, Paragraph 23).
Regarding claim 3, which depends upon claim 1:
Claim 3 recites:
The method as claimed in claim 1, wherein obtaining threshold value combinations comprises one of: obtaining all possible combinations of the threshold values; obtaining combinations using only a preset number of the most frequently-appearing threshold values for each node of the ANN; obtaining combinations using only a random subset of the threshold values for each node of the ANN; obtaining combinations of only threshold values for each node of the ANN which meet a user-defined metric
Nguyen in view of Baker disclose the method of claim 1 upon which claim 3 depends. Furthermore, Nguyen discloses the limitation of claim 3:
Nguyen teaches that the threshold values are decided based the activation of the node, wherein relevance to the activation would be a user-defined metric (Page 13). Therefore, only combinations of threshold values for each node of the ANN which meet a user-defined metric are obtained, which would be one of the above obtaining methods.
Regarding claim 4, which depends upon claim 1:
Claim 4 recites:
The method as claimed in claim 1, wherein the defined criterion to be optimized is entropy or Gini index
Nguyen in view of Baker disclose the method of claim 1 upon which claim 4 depends. Furthermore, Nguyen discloses the limitation of claim 4:
Nguyen teaches a maximization of the information gain ratio (Page 10), wherein this would be analogous to optimization of the entropy as highest information gain is synonymous with the lowest entropy, as discuss the present application’s specification.
Regarding claim 5, which depends upon claim 1:
Claim 5 recites:
The method as claimed in claim 1, wherein recording the threshold values associated with respective nodes of the ANN comprises, when there is no threshold value associated with a particular node of the ANN, recording as a threshold value for the node the per sample mean activation of the node
Nguyen in view of Baker disclose the method of claim 1 upon which claim 5 depends. Furthermore, Baker discloses the limitation of claim 5:
Baker teaches the calculation of an average activation value which may act as the threshold value of a particular node lacking a threshold value as a selected feature node is made to use this mean activation (Paragraph 93). Therefore, the selected feature node has no threshold value associated with it until the sample mean activation is recorded.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Nguyen and the teachings of Baker. This would have provided the advantage of tuning the learning process to work better with particular kinds of data (Baker, Paragraph 23).
Regarding claim 6, which depends upon claim 1:
Claim 6 recites:
The method as claimed in claim 1, wherein creating at least one decision tree comprises using a random forest generation algorithm to build a plurality of diverse decision trees
Nguyen in view of Baker disclose the method of claim 1 upon which claim 6 depends. Furthermore, Baker discloses the limitation of claim 6:
Baker teaches the use of a random forest generation algorithm (Paragraph 48), which are used to produce a plurality of diverse decision trees.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Nguyen and the teachings of Baker. This would have provided the advantage of tuning the learning process to work better with particular kinds of data (Baker, Paragraph 23).
Regarding claim 7, which depends upon claim 1:
Claim 7 recites:
Use of the method as claimed in claim 1 to extract at least one rule for an ANN for use with one of an autonomous driving algorithm and a healthcare algorithm
Nguyen in view of Baker disclose the method of claim 1 upon which claim 7 depends. Furthermore, Nguyen discloses the limitation of claim 7:
Nguyen teaches that its method may be applied as a form of control for unmanned ground vehicles (Page 24), which would be analogous to an autonomous driving algorithm (which would control unmanned ground vehicles) and be one of the above.
Regarding claim 8, which depends upon claim 1:
Claim 8 recites:
Use of the method as claimed in claim 1 to either: (i) extract the at least one rule for a CNN used in the control of an autonomous driving vehicle; or (ii) determine, using the extracted at least one rule, that the ANN is functioning correctly
Nguyen in view of Baker disclose the method of claim 1 upon which claim 8 depends. Furthermore, Nguyen discloses the limitation of claim 8:
Nguyen teaches that its method may be used to analyze the functioning of an ANN (Page 40), using the extracted rule set, in order to determine the correct functioning of the ANN.
Claim 9 recites a non-transitory computer readable storage medium that parallels the method of claim 1. Therefore, the analysis discussed above with respect to claim 1 also applies to claim 9. Accordingly, claim 9 is rejected based on substantially the same rationale as set forth above with respect to claim 1.
Claims 10, 12-15 recite an apparatus that parallels the method of claims 1, 3-6 respectively. Therefore, the analysis discussed above with respect to claims 1, 3-6 also applies to claims 10, 12-15 respectively. Accordingly, claims 10, 12-15 are rejected based on substantially the same rationale as set forth above with respect to claims 1, 3-6 respectively.
Regarding claim 17, which depends upon claim 1:
Claim 17 recites:
The method as claimed in claim 1, further comprising identifying hidden biases in the original ANN using high-fidelity extracted rules and re-training the ANN to mitigate the identified biases.
Nguyen in view of Baker disclose the method of claim 1 upon which claim 17 depends. However, Nguyen does not disclose the limitations of this claim. Instead, Baker discloses the limitation of claim 17:
Baker teaches correcting biases through the initialization of biases with a larger network i.e. an original neural network (Paragraph 121), wherein this network may be retrained including this step (Paragraph 117). This may be used in combination with the extracted rules previously taught in claim 1.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Nguyen and the teachings of Baker. This would have provided the advantage of tuning the learning process to work better with particular kinds of data (Baker, Paragraph 23).
Regarding claim 18, which depends upon claim 1:
Claim 18 recites:
The method as claimed in claim 1, wherein creating the at least one decision tree includes knowledge distillation to approximate the ANN's behavior by taking filter or neuron activations as input and the ANN's predictions as targets.
Nguyen in view of Baker disclose the method of claim 1 upon which claim 18 depends. Furthermore, Nguyen discloses approximate the ANN's behavior by taking filter or neuron activations as input and the ANN's predictions as targets:
Nguyen recites: “Each rule induced by each leaf of the resultant Extended C-Net tree can be traversed back to weights of the network to deduce the input-target relationships as represented by the network and its weights.” (Page 20-21)
Nguyen teaches approximating the neural networks behavior through the deduction of input-target relationships, wherein the previously taught neuron activations and predictions may act as those inputs and targets.
However, Nguyen does not disclose the use of knowledge distillation. This is disclosed instead by Baker:
Baker teaches transferring knowledge from one learning module to another (Paragraph 71) where some of these learning modules may be smaller than other (Paragraph 48). Transfer of knowledge from the larger learning modules to the smaller ones would be knowledge distillation.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Nguyen and the teachings of Baker. This would have provided the advantage of tuning the learning process to work better with particular kinds of data (Baker, Paragraph 23).
Claims 2 and 11 are rejected under 35 U.S.C. 103 as being unpatentable over Nguyen in view of Baker, further in view of Zoldi et al. (Pub. No. US 20190354853 A1, filed May 21st 2018, hereinafter Zoldi).
Regarding claim 2, which depends upon claim 1:
Claim 2 recites:
The method as claimed in claim 1, wherein recording the threshold values includes ranking the threshold values, for each node of the ANN, according to occurrence frequency and average depth of appearance in the decision tree, and performing the selected rule extraction algorithm for each combination of threshold values includes performing the selected rule extraction algorithm first on that combination of threshold values which includes the threshold values occurring with the highest frequencies
Nguyen in view of Baker disclose the method of claim 1 upon which claim 2 depends. Furthermore, Zoldi in the same field of endeavor of machine learning discloses wherein recording the threshold values includes ranking the threshold values, for each node of the ANN, according to occurrence frequency and [average depth of appearance in the decision tree]:
Zoldi teaches for the nodes of a neural network creating a distance matrix indicating how similar their activations are, wherein the activations would contain the threshold values that are recorded, and the more similar activations would be ranked by occurrence frequency, as more frequent activations would be more similar more frequently (Paragraph 60).
Furthermore, while Zoldi does not teach the average depth of appearance in the decision tree, Nguyen has previously taught a decision tree and with it the depth of node appearances within the tree.
Zoldi and the present application are analogous art because they are in the same field of endeavor of machine learning.
Zoldi further discloses [performing the selected rule extraction algorithm for each combination of threshold values includes performing the selected rule extraction algorithm first on] that combination of threshold values which includes the threshold values occurring with the highest frequencies:
Zoldi teaches that the above method is used in translating a neural network model in a simplified neural network (Paragraph 6) wherein the ranking of nodes is used to find similar nodes, which may be nodes include threshold values occurring with the highest frequencies, to cluster together as part of the simplification (Paragraph 60).
Zoldi does not teach performing the selected rule extraction algorithm for each combination of threshold values includes performing the selected rule extraction algorithm. However, this limitation has previously been taught by Nguyen as seen in claim 1, wherein the method of Nguyen may incorporate the nodes of Zoldi for the reasons discussed further below.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Nguyen in view of Baker and the teachings of Zoldi. This would have provided the advantage of better representing relationships between data learned by machine learning models (Zoldi, Paragraph 3).
Claim 11 recites an apparatus that parallels the method of claim 2. Therefore, the analysis discussed above with respect to claim 2 also applies to claim 11. Accordingly, claim 11 is rejected based on substantially the same rationale as set forth above with respect to claim 2.
Claims 16 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Nguyen in view of Baker, further in view of Townsend et al. (“ERIC: Extracting Relations Inferred from Convolutions”, published October 19th 2020, hereinafter Townsend).
Regarding claim 16, which depends upon claim 1:
Claim 16 recites:
The method as claimed in claim 1, wherein obtaining the threshold value combinations comprises performing a guided search using a Bayesian approach to find the threshold values that result in the highest fidelity.
Nguyen in view of Baker disclose the method of claim 1 upon which claim 16 depends. However, neither Nguyen nor Baker does not disclose the limitations of this claim. Instead, Townsend in the same field of explainable AI models discloses the limitation of claim 16:
Townsend recites: “The preference of fidelity and accuracy over explainability or vice-versa may depend on the nature of the task or a user’s preference” (Page 3).
Townsend herein teaches that a user’s preference e.g. through a guided search may steer the approximation method towards fidelity, i.e. values that result in the highest fidelity. In combination with the previously taught threshold values of Nguyen and Baker, those values may be what are altered by user preference.
Townsend and the present application are in the same field of endeavor since they are both
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Nguyen and the teachings of Baker further in view of Townsend. This would have provided the advantage of more flexible approximations of machine learning models (“Furthermore our method is entirely post-hoc and does not assume a convolutional architecture has been designed […] ERIC is not necessarily incompatible with such architectures either, allowing for flexible usage.”, Townsend page 2).
Regarding claim 20, which depends upon claim 1:
Claim 20 recites:
The method as claimed in claim 1, further comprising, when a specific node of the ANN does not appear in the at least one decision tree, recording a per sample mean activation of that node as the threshold value.
Nguyen in view of Baker disclose the method of claim 1 upon which claim 20 depends. However, neither Nguyen nor Baker does not disclose the limitations of this claim. Instead, Townsend in the same field of explainable AI models discloses the limitation of claim 20:
Townsend recites: “We define a kernel’s threshold as the mean norm of its activations with respect to the training data xtr” (Page 7).
Townsend teaches recording a per sample mean activation of that node as the threshold value, which may be used when a specific node of the ANN does not appear in the at least one decision tree.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to implement a methodology that utilized the teachings of Nguyen and the teachings of Baker further in view of Townsend. This would have provided the advantage of more flexible approximations of machine learning models (“Furthermore our method is entirely post-hoc and does not assume a convolutional architecture has been designed […] ERIC is not necessarily incompatible with such architectures either, allowing for flexible usage.”, Townsend page 2).
Allowable Subject Matter
The following is a statement of reasons for the indication of allowable subject matter for claim 19:
Claim 19 describes counting the frequency of unique activation thresholds within a decision tree. While Nguyen teaches that the decision tree’s node may be determined by the activation value, which would act as a threshold value as the threshold for the node (Page 13), it does not teach any counting or ranking of unique threshold values. Furthermore, the prior art is silent as to the counting and ranking of particular parameters of a decision tree, including these thresholds. For this reason, the examiner believes that claim 19 would be allowable under 103, through it is not wholly allowable due to the outstanding 101 rejection.
Response to Arguments
Applicant’s arguments filed 24-FEBRUARY-2026 have been fully considered, but the examiner believes that not all are fully persuasive.
Regarding the applicant’s remarks on non-final office action’s 101 rejection of the claims as an abstract idea, the applicant argues that regardless of the claims are not directed towards an abstract idea. The examiner respectfully requests the applicant’s consideration of the following:
The applicant argues that the claims are not mere mental processes because the approximation of an ANN requires “combinatorial evaluations, repeated numerical optimization across large datasets, and algorithmic recursion at a scale and precision that exceeds human cognitive capabilities”. However, the examiner believes that these aspects are not reflected in the claims. There is no discussion of a large dataset, or a particular minimum number of samples, that would reasonably make the claim impractical for the human mind. Exercises made y humans that depict neural networks as trees of nodes and activation functions are routinely used to teach the concept of a neural network at a basic level. This may be seen in applicant’s own figure 8, which demonstrates an approximating decision tree.
Furthermore, the applicant argues that a technical improvement is recited, in the form of improving “the fidelity (accuracy of representation) and interpretability of the extracted rules while reducing computational complexity”. However, this improvement would result solely from the limitations of the claim considered by the examiner to be a mental process. Since the judicial exception alone cannot provide the improvement (MPEP 2106.05(a)), the examiner does not consider these limitations to integrate the claim into a practical application.
The newly amended dependent claims 16-20 have likewise been analyzed under 101. While the examiner believes that 16 and 18-20 describe an abstract idea or insignificant extra-solution activity, claim 17 would be allowable under 101. Even if the examiner considered the identification of hidden biases as a mental process, the re-training to mitigate those biases would demonstrate that the re-training is not insignificant, as it provides an improvement to the art through that mitigation.
Regarding the applicant’s remarks on the non-final office action’s 103 rejection of the claims, the applicant argues that Nguyen in view of Baker do not teach the limitations of these claims. As such, the applicant argues that all claims dependent on the above would additionally not be obvious under 103. (However, the examiner believes that Nguyen in view of Baker does teach the limitations and respectfully requests applicant’s consideration of the following:
The applicant states that “to ‘quantize’ Nguyen’s exact model as the Office Action suggests would be antithetical to Nguyen’s stated purpose of exactness, as it would introduce approximation error’.
However, Nguyen addressed its decisions trees as an approximation, stating: “Decision Trees are interpretable representations able to approximate the underlying functions that ANNs represent” (Page 12). The applicant uses Nguyen’s description of the tree as ‘complete and sound’ as evidence against this, but that description is stated to refer to ‘high fidelity’ as the present claims do.
Furthermore, the applicant states that “the nodes of the tree [of Nguyen] all use the same threshold. Nguyen does not, therefore, disclose ‘for each node of the decision tree a threshold value for the defined criterion being calculated to determine for which node of the ANN the input activations should be split between branches of the decision tree’ or ‘recording the threshold values associated with the respective nodes of the ANN’”.
However, the claim does not provide further limitations on how the calculation is performed. As such, there is nothing that forbids each node having an associated threshold value of 0 as is taught in Fig. 2 of Nguyen, it merely acts as a specific case of the claimed limitations.
Furthermore, the applicant states that Nguyen does not teach ‘obtaining threshold value combinations, each combination comprising one of the threshold values obtained for respective nodes of the ANN’ since ‘in a multivariate decision tree [used by Nguyen] multiple variables may be used in each node, but the variables are not thresholds.’
However, multiple thresholds are explicitly not required for the threshold value combinations, as the combination comprises one of the threshold values. Therefore, the threshold and the other variables would be a threshold value combination as only one threshold is required.
Furthermore, the applicant argues “Nguyen teaches exact conversion […] which is fundamentally different from the Applicant’s quantization method”. However, while the specification discusses quantization, this specificity is not reflected in the claim language. Rather, the closest language describes ‘the approximation of an ANN’ using a decision tree. As discussed above, Nguyen explicitly addresses that its decision trees are also an approximation.
Finally, the applicant argues that “Baker is directed to hyperparameter optimization during training. This is not the same as optimizing a criterion for rule extraction after training”. However, in addition to its previously described teachings, Baker also teaches an embodiment wherein the learning coach controls optimized hyperparameters for retrained models (Paragraph 117-118) which would be after an initial training.
For these reasons, the examiner believes that the 103 rejection of claims 1 and 10 over Nguyen in view of Baker is proper, as well as the rejections of the previously presented claims.
Regarding the new claims, the examiner has addressed them in the above 103 section.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ALEXANDRIA JOSEPHINE MILLER whose telephone number is (703)756-5684. The examiner can normally be reached Monday-Thursday: 7:30 - 5:00 pm, every other Friday 7:30 - 4:00.
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, Mariela Reyes can be reached at (571) 270-1006. 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.
/A.J.M./Examiner, Art Unit 2142
/Mariela Reyes/Supervisory Patent Examiner, Art Unit 2142
