DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on January 27, 2023; October 2, 2025; December 10, 2025; February 5, 2026; and April 9, 2026 have been considered by the examiner.
Specification
The specification is objected to as failing to provide proper antecedent basis for the claimed subject matter. See 37 CFR 1.75(d)(1) and MPEP § 608.01(o). Correction of the following is required:
Claim 21 recites as the limitation regarding “The method of claim 1, wherein each of the bias feature heads are configured to actively introduce one or more biases into a first set of the plurality of output labels generated by the bias feature heads,…” where the term ‘actively’ is not clearly defined in the specification.
Similarly, claim 22 recites as the limitation regarding “The method of claim 1, wherein a first output label of the output labels generated through a first bias feature head of the bias feature heads is a first prediction in which a first bias specified by the first bias feature head is intentionally formed, the first prediction being formed to intentionally include the first bias when a first intermediate output of the intermediate outputs is consumed by the first bias feature head,” where the term ‘intentionally’ is not clearly defined in the specification.
Response to Amendment
The Amendment filed March 26th, 2026 has been entered. Claims 1-18, 21 and 22, remain pending in the application. Applicant’s amendments to the Specification, Drawings, and Claims have overcome each and every objection and 112(b) rejections previously set forth in the Non-Final Office Action mailed December 29, 2025.
The objection to paragraph [0021] in the specification has been considered from 3/26/2026 applicant remarks.
Response to Arguments:
Applicant’s arguments filed on 03-05-2026 have been fully considered.
In reference to Applicant’s arguments:
-Claim rejections under 35 U.S.C. 101.
Examiner’s response:
From considering the applicant remarks on page 12-13 regarding an improvement to the functioning of the claimed invention, examiner is unable to find support from the specification regarding such improvement, other than the invention can reduce latent bias from specification paragraph [0017] noting “By performing the combined training and untraining, trained inference models may be less likely to exhibit latent bias or may exhibit latent bias to reduced degrees,” and the mentioning of improving the system located in paragraph [0169] noting “In various embodiments, to enable a thinner and lighter system design as well as to improve system responsiveness, this mass storage may be implemented via a solid state device (SSD),” which was not recited in the amended claims on 03/26/2026. Furthermore, the remarks on page 12 citing paragraphs [0015]-[0020], and [0033]-[0046], where the specification in [0035] describes “training data may include a correlation that is not obvious but that may result in latent bias being introduced into inference models trained using training data. If consumed by computer-implemented services, these inaccurate or otherwise undesirable inferences may negatively impact the computer-implemented services.” This mentioning of unwanted latent bias that negatively impact computer-implemented services, does not mention the improvement to the claimed invention. Therefore, the 35 U.S.C. 101 rejection is maintained.
In reference to Applicant’s arguments:
-Claim rejections under 35 U.S.C. 103.
Examiner’s response:
Applicant’s arguments are directed to the amended limitations as well as to the applicant’s arguments of:
Regarding applicant’s remarks on paragraphs 2-3, page 14, noting “it should be clear that the claimed multiheaded inference model is not just any mere parallel processing capable model (as the Office has mapped the claimed multiheaded inference model to) but rather a specifically designed model having the particular neural network structure as claimed in the above-referenced limitation of the amended independent claims,” the examiner notes that the applicant’s arguments are fully considered, but are moot in view of new grounds of rejection because the arguments do not apply to the combination of references used in the current rejection, in regards to the amended independent claims 1, 8, and 15.
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.
Claim 1:
Regarding claim 1, in step 1 of the 101-analysis set forth in MPEP 2106, the claim recites
“A method for configuring and training inference models to reduce an amount of latent bias exhibited by the inference models when generating inferences, the method comprising:… ,” and a method is one of the four statutory categories of invention. In step 2A prong 1 of the 101-analysis set forth in the MPEP 2106, the examiner has determined that the following limitations recite a process that, under the broadest reasonable interpretation, covers a mental process but for recitation of generic computer components:
“identifying, for the trained multiheaded inference model, predictive power levels for labels and predictive power levels for bias features;” (mental process, a person can mentally evaluate a numeric quantity for predictive power level for labels and a numeric quantity for a predictive power levels for bias features, see (MPEP 2106))
“making a determination, based on the predictive power levels for the labels and the predictive power levels for the bias features, regarding whether the trained multiheaded inference model meets a predetermined criterion;” (mental process, a person can mentally evaluate and then judge if a trained inference model meets a preset criterion based on the values from the predictive power levels for labels and the values from the predictive power levels for bias features, see (MPEP 2106))
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
In step 2A prong 2 of the 101-analysis set forth in MPEP 2106, the examiner has determined that the following additional elements do not integrate this judicial exception into a practical application:
“A method for configuring and training inference models to reduce an amount of latent bias exhibited by the inference models when generating inferences, the method comprising: obtaining inference goals for label inferences and bias feature inferences for a multiheaded inference model of the inference models, (In step 2A, prong2, this recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),).
the multiheaded inference model being an inference model comprising a multi-path neural network that comprises multiple label prediction heads and multiple bias feature heads through which intermediate outputs generated by a shared body of the multi-path neural network are consumed to generate a plurality of output labels;” (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f))
“obtaining learning rates for the label prediction heads and the bias feature heads of the multiheaded inference model using the inference goals;” (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f))
“training the multiheaded inference model based on the learning rate to obtain a trained multiheaded inference model,” (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f))
“and in an instance of the determination where the multiheaded inference model meets a predetermined criterion: providing computer implemented services using first inferences generated by the trained multiheaded inference model, the first inferences being ones of the inferences,” (This recites generally linking the use of the judicial exception to a particular technological environment or field of use, in this case, linking use of the exception to generic computer to provide computer related services – see MPEP 2106.05 (h)).
Since the claim as a whole, looking at the additional elements individually and in combination, does not contain any other additional elements that are indicative of integration into a practical application, the claim is “directed” to an abstract idea.
In step 2B of the 101-analysis set forth in the 2019 PEG, the examiner has determined that the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional element iii recites “a method for configuring and training inference models, to reduce an amount of latent bias exhibited by the inference models when generating inferences, the method comprising: obtaining inference goals for label inferences and bias feature inferences for a multiheaded inference model of the inference models”, (in step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes obtaining or receiving data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016) – see MPEP 2106.05(d) (II)(i)).
As discussed above, additional element iv, v, and vi recites mere instructions to apply the judicial exception, which is not indicative of significantly more. The additional element vii recites linking the use of the judicial exception to a particular technological environment or field of use. Considering the additional elements individually and in combination, and the claim as a whole, the additional element does not provide significantly more than the abstract idea. Therefore, the claim is not patent eligible.
Claim 2:
Regarding claim 2, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1.
Further, claim 2 recites an additional element, “The method of claim 1, wherein the inference goals specify: a minimum predictive power level for the labels; and maximum predictive power levels for the bias features.” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 3:
Regarding claim 3, it is dependent upon claim 2, and thereby incorporates the limitations of, and corresponding analysis applied to claim 2. Further, claim 3 recites an abstract idea in:
“The method of claim 2, wherein obtaining the learning rates comprises: establishing a first learning rate based on the minimum predictive power level for the labels, and establishing second learning rate based on the acceptable minimum and maximum predictive power levels for the bias features.” (Math concept, in this case, the learning rate is considered a math calculation shown from the specification in paragraph 134 describing the “learning rates may be based on the ratios of the predictive power levels for the bias features to the predictive power level for the features”, see MPEP 2106).
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mathematical concept but for the recitation of generic computer components, then it falls within the mathematical concept grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 4:
Regarding claim 4, it is dependent upon claim 3, and thereby incorporates the limitations of, and corresponding analysis applied to claim 3. Further, claim 4 recites the following additional elements:
“wherein training the multiheaded inference model comprises: performing a first number of training cycles based on the first learning rate;” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor and memory to train the multiheaded inference model based on a first learning rate– see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
“and performing a second number of untraining cycles based on the second learning rate.” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor and memory to train the multiheaded inference model based on a second learning rate see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 5:
Regarding claim 5, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 5 recites the following additional elements:
“obtaining the inference goals comprises: presenting, to a user, a graphical user interface comprising: a first inference goal control corresponding to a label of the labels, and a second inference goal control corresponding to a bias feature of the bias features;” (In step 2A, prong2, obtaining inference goals recites mere presentation of data, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes presenting offers and gathering statistics, OIP Techs., 788 F.3d at 1362-63, 115 USPQ2d at 1092-93; – see MPEP 2106.05(d) (II)(iv)),
“and obtaining, from the user and via the graphical user interface: first user input that indicates a first inference goal of the inference goals, and second user input that indicates a second inference goal of the inference goals.” (In step 2A, prong2, obtaining user inputs recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016), – see MPEP 2106.05(d) (II)(i)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 6:
Regarding claim 6, it is dependent upon claim 5, and thereby incorporates the limitations of, and corresponding analysis applied to claim 5. Further, claim 6 recites the following additional elements:
“wherein the first inference goal control comprises: a slider that the user may actuate along a path to provide the first user input, and a position of the slider in the path defining a range for the predictive power level for label.” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor, memory, with an interface slider icon to retrieve user inputs to generate inference goals, see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 7:
Regarding claim 7, it is dependent upon claim 6, and thereby incorporates the limitations of, and corresponding analysis applied to claim 6. Further, claim 7 recites the following additional elements:
“The method of claim 6, wherein making the determination comprises: instantiating, in the graphical user interface and to obtain an updated graphical user interface, a performance indicator along the path,” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor, memory, with a graphic user interface to run the method, see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
“the performance indicator indicating an actual predictive power level for the label by the trained multiheaded inference model;” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor, memory, with an interface to run the method, see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
“obtaining, via the updated graphical user interface, second user input from the user, the second user input indicating a level of acceptability of the actual predictive power level for the label;” (In step 2A, prong2, obtaining a user input recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016), – see MPEP 2106.05(d) (II)(i)).
“and in a first instance of the second user input indicating that the level of acceptability is accepted by the user: determining that the trained multiheaded inference model meets a predetermined criterion,” (mental process- a person can evaluate and then judge to decide if a first instance of the second user input shows a level of acceptability is accepted by a user, and then judge to decide if a trained multiheaded inference model meets a preset criterion, see MPEP 2106.04(a)(2)(III)),
“and in a second instance of the second user input indicating that the level of acceptability is not accepted by the user: determining that the trained multiheaded inference model fails to meet the predetermined criterion.” (mental process- a person can evaluate and then judge to decide if a second instance of the second user input shows the user did not accept a level of acceptability, and then judge to decide if a trained multiheaded inference model does not meet a preset criterion, see MPEP 2106.04(a)(2)(III)),
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 8:
Regarding claim 8, in step 1 of the 101-analysis set forth in MPEP 2106, the claim recites
“a non-transitory machine-readable medium having instructions stored therein, which when executed by a processor, cause the processor to perform operations for configuring and training inference models to reduce an amount of latent bias exhibited by the inference models when generating inferences, …”, and a non-transitory machine-readable medium is considered a machine and is one of the four statutory categories of invention.
In step 2A prong 1 of the 101-analysis set forth in the MPEP 2106, the examiner has determined that the following limitations recite a process that, under the broadest reasonable interpretation, covers a mental process but for recitation of generic computer components:
“identifying, for the trained multiheaded inference model, predictive power levels for labels and predictive power levels for bias features;” (mental process, a person can mentally find and evaluate a numeric quantity for predictive power level for labels and a numeric quantity for a predictive power levels for bias features, see MPEP 2106.04(a)(2)(III)),
“making a determination, based on the predictive power levels for the labels and the predictive power levels for the bias features, regarding whether the trained multiheaded inference model meets a predetermined criterion,” (mental process, a person can mentally evaluate and then judge if a trained inference model meets a preset criterion based on the values from the predictive power levels for labels and the values from the predictive power levels for bias features, see MPEP 2106.04(a)(2)(III)),
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
In step 2A prong 2 of the 101-analysis set forth in MPEP 2106, the examiner has determined that the following additional elements do not integrate this judicial exception into a practical application:
“A non-transitory machine-readable medium having instructions stored therein, which when executed by a processor, cause the processor to perform operations for configuring and training inference models to reduce an amount of latent bias exhibited by the inference models when generating inferences,” (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
“obtaining inference goals for label inferences and bias feature inferences for a multiheaded inference model of the inference models,” (In step 2A, prong2, obtaining inference goals recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),).
“the multiheaded inference model being an inference model comprising a multi-path neural network that comprises multiple label prediction heads and multiple bias feature heads through which intermediate outputs generated by a shared body of the multi-path neural network are consumed to generate a plurality of output labels;” (This recites mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
“obtaining learning rates for label prediction heads and bias feature heads of the multiheaded inference model using the inference goals,” (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f))
“training the multiheaded inference model based on the learning rate to obtain a trained multiheaded inference model,” (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f))
“and in an instance of the determination where the multiheaded inference meets a predetermined criterion: providing computer implemented services using first inferences generated by the trained multiheaded inference model, the first inferences being ones of the inferences,” (This recites generally linking the use of the judicial exception to a particular technological environment or field of use, in this case, linking use of the exception to generic computer to provide computer related services – see MPEP 2106.05 (h)).
Since the claim as a whole, looking at the additional elements individually and in combination, does not contain any other additional elements that are indicative of integration into a practical application, the claim is “directed” to an abstract idea.
In step 2B of the 101-analysis set forth in the 2019 PEG, the examiner has determined that the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional element iv recites “ obtaining inference goals for label inferences and bias feature inferences for a multiheaded inference model of the inference models”, (in step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes obtaining or receiving data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016) – see MPEP 2106.05(d) (II)(i)).
As discussed above, additional elements iii, v, vi, and vii recite mere instructions to apply the judicial exception, which are not indicative of significantly more. The additional element viii recites linking the use of the judicial exception to a particular technological environment or field of use. Considering the additional elements individually and in combination, and the claim as a whole, the additional element does not provide significantly more than the abstract idea. Therefore, the claim is not patent eligible.
Claim 9:
Regarding claim 9, it is dependent upon claim 8, and thereby incorporates the limitations of, and corresponding analysis applied to claim 8.
Further, claim 9 recites an additional element, “The non-transitory machine-readable medium of claim 8, wherein the inference goals specify: a minimum predictive power level for the labels; and maximum predictive power levels for the bias features.” (In step 2A, prong 2, the non-transitory medium and inference goals are considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 10:
Regarding claim 10, it is dependent upon claim 9, and thereby incorporates the limitations of, and corresponding analysis applied to claim 9. Further, claim 10 recites the following abstract idea in:
“wherein obtaining the learning rates comprises: establishing a first learning rate based on the minimum predictive power level for the labels, and establishing second learning rate based on the minimum and maximum predictive power levels for the bias features.” (Math concept, in this case, the learning rate is considered a math calculation shown from the specification in paragraph 134 describing the “learning rates may be based on the ratios of the predictive power levels for the bias features to the predictive power level for the features”, see MPEP 2106).
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mathematical concept but for the recitation of generic computer components, then it falls within the mathematical concept grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
Further, claim 10 recites the additional elements, “the non-transitory machine-readable medium” (In step 2A, prong 2, this element is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 11:
Regarding claim 11, it is dependent upon claim 10, and thereby incorporates the limitations of, and corresponding analysis applied to claim 10. Further, claim 11 recites the following additional elements:
“wherein training the multiheaded inference model comprises: performing a first number of training cycles based on the first learning rate;” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor and memory to train the multiheaded inference model based on a first learning rate– see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
“and performing a second number of untraining cycles based on the second learning rate.” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor and memory to train the multiheaded inference model based on a second learning rate see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Further, claim 11 recites the additional elements, “the non-transitory machine-readable medium” (In step 2A, prong 2, this element is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 12:
Regarding claim 12, it is dependent upon claim 8, and thereby incorporates the limitations of, and corresponding analysis applied to claim 8. Further, claim 12 recites the following additional elements:
“obtaining the inference goals comprises: presenting, to a user, a graphical user interface comprising: a first inference goal control corresponding to a label of the labels, and a second inference goal control corresponding to a bias feature of the bias features;” (In step 2A, prong2, obtaining inference goals and inference goal controls recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes presenting offers and gathering statistics, OIP Techs., 788 F.3d at 1362-63, 115 USPQ2d at 1092-93; – see MPEP 2106.05(d) (II)(iv)).
“and obtaining, from the user and via the graphical user interface: first user input that indicates a first inference goal of the inference goals, and second user input that indicates a second inference goal of the inference goals.” (In step 2A, prong2, obtaining user inputs recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016), – see MPEP 2106.05(d) (II)(i)).
Further, claim 12 recites the additional element, “the non-transitory machine-readable medium” (In step 2A, prong 2, this element is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 13:
Regarding claim 13, it is dependent upon claim 12, and thereby incorporates the limitations of, and corresponding analysis applied to claim 12. Further, claim 13 recites the following additional elements:
“wherein the first inference goal control comprises: a slider that the user may actuate along a path to provide the first user input, and a position of the slider in the path defining a range for the predictive power level for label.” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor, memory, with an interface slider icon to retrieve user inputs to generate inference goals, see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Further, claim 13 recites the additional elements, “the non-transitory machine-readable medium” (In step 2A, prong 2, this element is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 14:
Regarding claim 14, it is dependent upon claim 13, and thereby incorporates the limitations of, and corresponding analysis applied to claim 13. Further, claim 14 recites the following additional elements:
“making the determination comprises: instantiating, in the graphical user interface and to obtain an updated graphical user interface, a performance indicator along the path,” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor, memory, with a graphic user interface to run the method, see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
“the performance indicator indicating an actual predictive power level for the label by the trained multiheaded inference model;” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor, memory, with an interface to run the method, see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
“obtaining, via the updated graphical user interface, second user input from the user, the second user input indicating a level of acceptability of the actual predictive power level for the label;” (In step 2A, prong2, obtaining a user input recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),). In step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016), – see MPEP 2106.05(d) (II)(i)).
“and in a first instance of the second user input indicating that the level of acceptability is accepted by the user: determining that the trained multiheaded inference model meets a predetermined criterion;” (mental process- a person can evaluate and then judge to decide if a first instance of the second user input shows a high level of acceptability, and then judge to decide if a trained multiheaded inference model is acceptable – see MPEP 2106.04(a)(2)(III)),
“and in a second instance of the second user input indicating that the level of acceptability is not accepted by the user: determining that the trained multiheaded inference model fails to meet a predetermined criterion.” (mental process- a person can evaluate and then judge to decide if a second instance of the second user input shows a low level of acceptability, and then judge to decide if a trained multiheaded inference model is unacceptable – see MPEP 2106.04(a)(2)(III)),
Further, claim 14 recites the additional element, “the non-transitory machine-readable medium” (In step 2A, prong 2, this element is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 15:
Regarding claim 15, in step 1 of the 101-analysis set forth in MPEP 2106, the claim recites “a data processing system, comprising: a processor; and a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations for configuring and training inference models to reduce an amount of latent bias exhibited by the inference models when generating inferences,…” which is considered a machine and is one of the four statutory categories of invention.
In step 2A prong 1 of the 101-analysis set forth in the MPEP 2106, the examiner has determined that the following limitations recite a process that, under the broadest reasonable interpretation, covers a mental process but for recitation of generic computer components:
“identifying, for the trained multiheaded inference model, predictive power levels for labels and predictive power levels for bias features;” (mental process, a person can mentally find and evaluate a numeric quantity for predictive power level for labels and a numeric quantity for a predictive power levels for bias features, see MPEP 2106.04(a)(2)(III)),
“making a determination, based on the predictive power levels for the labels and the predictive power levels for the bias features, regarding whether the trained multiheaded inference model meets a predetermined criterion,” (mental process, a person can mentally evaluate and then judge if a trained inference model meets a preset criterion based on the values from the predictive power levels for labels and the values from the predictive power levels for bias features, see MPEP 2106.04(a)(2)(III)),
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
In step 2A prong 2 of the 101-analysis set forth in MPEP 2106, the examiner has determined that the following additional elements do not integrate this judicial exception into a practical application:
“a data processing system, comprising: a processor; and a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations for managing inference models, the operations comprising:… (This recites mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
“obtaining inference goals for label inferences and bias feature inferences for a multiheaded inference model of the inference models,” (In step 2A, prong2, obtaining inference goals recites mere data gathering, which is considered insignificant extra-solution activity – see MPEP 2106.05(g),).
“the multiheaded inference model being an inference model comprising a multi-path neural network that comprises multiple label prediction heads and multiple bias feature heads through which intermediate outputs generated by a shared body of the multi-path neural network are consumed to generate a plurality of output labels;” (This recites mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
“obtaining learning rates for label prediction heads and bias feature heads of the multiheaded inference model using the inference goals,” (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f))
“training the multiheaded inference model based on the learning rate to obtain a trained multiheaded inference model,” (Mere instructions to apply an exception using generic computer – see MPEP 2106.05(f))
“and in an instance of the determination where the multiheaded inference meets a predetermined criterion: providing computer implemented services using first inferences generated by the trained multiheaded inference model, the first inferences being ones of the inferences,” (This recites generally linking the use of the judicial exception to a particular technological environment or field of use, in this case, linking use of the exception to generic computer to provide computer related services – see MPEP 2106.05 (h)).
Since the claim as a whole, looking at the additional elements individually and in combination, does not contain any other additional elements that are indicative of integration into a practical application, the claim is “directed” to an abstract idea.
In step 2B of the 101-analysis set forth in the 2019 PEG, the examiner has determined that the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional element iv recites “ obtaining inference goals for label inferences and bias feature inferences for a multiheaded inference model of the inference models”, (in step 2B, this insignificant extra-solution activity is well understood routine and conventional activity which includes obtaining or receiving data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016) – see MPEP 2106.05(d) (II)(i)).
As discussed above, additional elements iii, v, vi, and vii recite mere instructions to apply the judicial exception, which are not indicative of significantly more. The additional element viii recites linking the use of the judicial exception to a particular technological environment or field of use. Considering the additional elements individually and in combination, and the claim as a whole, the additional element does not provide significantly more than the abstract idea. Therefore, the claim is not patent eligible.
Claim 16:
Regarding claim 16, it is dependent upon claim 15, and thereby incorporates the limitations of, and corresponding analysis applied to claim 15.
Further, claim 16 recites the additional elements, “the data processing system”, and “the inference goals specify: a minimum predictive power level for the labels; and maximum predictive power levels for the bias features.” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 17:
Regarding claim 17, it is dependent upon claim 16, and thereby incorporates the limitations of, and corresponding analysis applied to claim 16. Further, claim 17 recites the following abstract idea in:
“wherein obtaining the learning rates comprises: establishing a first learning rate based on the minimum predictive power level for the labels, and establishing second learning rate based on the minimum and maximum predictive power levels for the bias features.” (Math concept, in this case, the learning rate is considered a math calculation shown from the specification in paragraph 134 describing the “learning rates may be based on the ratios of the predictive power levels for the bias features to the predictive power level for the features”, see MPEP 2106).
If claim limitations, under their broadest reasonable interpretation, covers performance of the limitations as a mathematical concept but for the recitation of generic computer components, then it falls within the mathematical concept grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
Further, claim 17 recites the additional elements, “the data processing system” (In step 2A, prong 2, this element is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 18:
Regarding claim 18, it is dependent upon claim 17, and thereby incorporates the limitations of, and corresponding analysis applied to claim 17. Further, claim 18 recites the following additional elements:
“wherein training the multiheaded inference model comprises: performing a first number of training cycles based on the first learning rate;” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor and memory to train the multiheaded inference model based on a first learning rate– see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
“and performing a second number of untraining cycles based on the second learning rate.” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer with a processor and memory to train the multiheaded inference model based on a second learning rate see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Further, claim 18 recites the additional elements, “the data processing system” (In step 2A, prong 2, this element is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 21:
Regarding claim 21, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 21 recites the following additional elements:
“The method of claim 1, wherein each of the bias feature heads are configured to actively introduce one or more biases into a first set of the plurality of output labels generated by the bias feature heads, and each of the label prediction heads are not configured to actively introduce the one or more biases into a second set of the plurality of output labels generated by the label prediction heads.” (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer, see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 22:
Regarding claim 22, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 22 recites the following additional elements:
The method of claim 1, wherein a first output label of the output labels generated through a first bias feature head of the bias feature heads is a first prediction in which a first bias specified by the first bias feature head is intentionally formed, the first prediction being formed to intentionally include the first bias when a first intermediate output of the intermediate outputs is consumed by the first bias feature head, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer, see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
and a second output label of the output labels generated through a first label prediction head of the label prediction heads comprises a second prediction in which the first bias of is not intentionally introduced by the first label prediction head when a second intermediate output of the intermediate outputs is consumed by the first label prediction head. (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer, see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1, 5, 8, 12, 15, 19, 21, and 22 are rejected under 35 U.S.C. 103 as being unpatentable over Settipalli V. et al. "Reducing Unintended bias in Text Classification using Multitask learning”, available at https://www.diva-portal.org/smash/get/diva2:1534152/FULLTEXT02 on January, 2021, (hereafter, Settipalli), in view of Hu. Z. et al. (Pub. No. CN 110516541 A), published on November, 29, 2019 (hereafter, Hu), further in view of Bucklin N. et al. (Pub. No. WO2022240860 A1), published on November 17, 2022, (hereafter, Bucklin), and further in view of Oneto, L. et al., in “Taking advantage of multitask learning for fair classification,” published on January 27–28, 2019, available at https://dl.acm.org/doi/pdf/10.1145/3306618.3314255, (hereafter, Oneto).
Claim 1:
Referring to claim 1, Settipalli teaches “A method for configuring and training inference models…, …obtaining inference goals for label inferences and bias feature inferences for a multiheaded inference model of the inference models,”
(See Settipalli on page 27, chapter 4 opening paragraphs, where Settipalli describes “in this experiment, we will compare BERT based models i.e.. BERT-base and DistilBERT with the Bi-LSTM model proposed in [53]. In this comparative analysis, we are considering three models namely two variants of BERT i.e. BERT-base, Distil- BERT and the Bi-LSTM model proposed by Vaidya et al. [53]. Multitask learning is used to train these models so that they can perform two tasks namely identity task and the toxicity task in parallel. Identity task refers to predicting the presence of the identity in the given sentence. Toxicity task refers to predicting the toxicity score for the given sentence. First, we will train all the three models and then we will test these models to compare the bias mitigation performance.” Here, Settipalli teaches using multitask learning to train models (i.e. a multiheaded inference model of inference models) to perform two tasks, which are identity task and the toxicity task in parallel (i.e. inference goals for label inferences and bias feature inferences). Settipalli also describes training of the models.
See Settipalli on page 35, section 4.6.1. classification performance, "to measure the overall classification performance of the models we have used the metrics f1 score and Overall-AUC i.e. AUC [21] calculated on the entire test set. The test set is imbalanced i.e. only 7.9% of instances belong to toxic class and the remaining 82.1 belong to non-toxic class. So we had chosen the metrics overall-AUC and f1-score as they are not effected by the unbalanced class distribution. The overall-AUC is a threshold agnostic metric so we can apply it directly on the model’s prediction." Here, Settipalli teaches and describes measuring the prediction performance for label inferences or classification performance with AUC scores or F1 scores as obtaining inference goals for label inferences. Further, on page 35, section 4.6.1., Settipalli discusses that the study has “chosen the metrics overall-AUC and f1-score as they are not effected by the unbalanced class distribution. The overall-AUC is a threshold agnostic metric so we can apply it directly on the model’s prediction.” Here, Settipalli describes the label inference goals using the metrics overall-AUC and F1 since they measure the model’s prediction in labeling.
On pages 35-36, in section 4.6.2 Bias mitigation performance section, Settipalli discusses “for measuring the bias mitigation performance over different identity subgroups we have used the AUC based metrics that were selected in the literature review i.e. subgroup-AUC, BPSN-AUC, BNSP-AUC and Generalized mean AUC. Subgroup-AUC: For calculating this measure, a sample set which contains an equal number of toxic and non-toxic examples corresponding to particular identity subgroup is selected from the test set. Calculating the AUC metric on this sample set will give us the get subgroup AUC for the particular identity. The subgroup-AUC will tell [how] well the model is able to distinguish positive and negative sentences corresponding to a particular identity subgroup”. Here, Settipalli describes using a specific type of AUC score called sub-group AUC score to measure bias mitigation performance to determine the success of the model in distinguishing prediction inferences of labeling if a sentence has positive or negative sentiment that relate to a specific identity subgroup or an attribute that is frequently biased. Settipalli here also teaches getting the metrics including generalized mean AUC scores of various subgroups to measure bias mitigation performance, which corresponds to obtaining inference goals for bias feature inference. See Settipalli also on page 25, section 3.6 metrics selection for additional information. See page 34 where Settipalli discusses various models such as BERT base, BERT large, and Bi-LSTM on table 4.4. The models here represent multi-headed inference models.
Further, Settipalli teaches “obtaining learning rates for the label prediction heads and the bias feature heads of the multiheaded inference model using the inference goals,” (see Settipalli in abstract on page 3, where Settipalli mentions "the main aim of this research is to use multitask learning to fine tune the BERT models namely BERT-base and Distil BERT", and that these models "can jointly predict the toxicity of the text comment and the presence of identity in that comment in order to reduce bias over frequently attacked identity terms. The proposed BERT models are also compared with the previously proposed Bi-LSTM model in terms of bias mitigation and classification performance." Here, Settipalli describes that the models are used to perform both classification and bias mitigation tasks, where tasks correspond to heads in the claim, and the models are connected to the classification prediction task (i.e. label prediction heads) and bias reduction task (i.e. bias feature heads). Additionally, on page 34, Settipalli shows table 4.4 where the learning rates are used for the prediction models are listed, where each model also performs both classification prediction and bias reduction tasks. Here, Settipalli shows using learning rates for performing the prediction and bias mitigation tasks.)
Further, Settipalli teaches “training the multiheaded inference model based on the learning rate to obtain a trained multiheaded inference model,”
(See Settipalli in page 34, where Settipalli describes that a "learning rate: …is a hyper-parameter which tells the extent to which the model parameters need to be changed while back-propagating the loss through the network during the training process," Here, Settipalli shows that the learning rate is part of the training process of the multi-task model. Settipalli also talks about using a learning rate for each of the BERT models, and described how to finely tune the learning rate as "the value of the learning rate should be tuned carefully because using a very small rate will slow down the training process whereas selecting a large learning rate will result in a bad parameter tuning while back-propagating the loss through the network." This disclosure shows how a multiheaded inference model is trained based on the learning rate to get a trained multiheaded inference model.)
Further, Settipalli teaches “identifying, for the trained multiheaded inference model, predictive power levels for labels and predictive power levels for bias features,”
(See Settipalli in page 35 section 4.6.1 where Settipalli discusses using the metrics AUC and F1 score for both predictions in "to measure the overall classification performance of the models we have used the metrics f1 score and Overall-AUC", and Settipalli in page 25, section 3.6 uses generalised mean AUC scores in identifying bias subgroups in "use the per identity AUC metrics i.e. subgroup-AUC, BPSN-AUC and BNSP-AUC to measure the bias mitigation performance of the model over individual identity subgroup and the metric generalised mean AUC to measure the overall bias mitigation performance". Later, Settipalli on pages 39-41 in figures 5.1- 5.3, and tables 5.1-5.3 displays scores of AUC and F1 scores for all models including subgroups, which corresponds to the recited claim limitation in identifying, for the trained multiheaded inference model, predictive power levels for labels, and predictive power levels for bias features. The AUC overall and F1 scores measure predictive power level for labels, and the sub-group AUC score measure the predictive power levels for bias features. By displaying the AUC overall and F1 scores for model performance and the sub-group AUC for bias features, this relates to identifying the predictive power levels for labels and predictive power levels for bias features from the claim limitation. Note: For the terms predictive power levels for labels, and predictive power levels for bias features the examiner has construed the meanings for the terms to be performance measures or metrics for the prediction tasks, and performance metrics for the bias task, respectively.
Further, Settipalli teaches “making a determination, based on the predictive power levels for the labels and the predictive power levels for the bias features, …,”
(See Settipalli on page 54, section 6.3 Discussion, where Settipalli discusses that "the higher value of Generalised mean AUC indicates that the overall bias mitigation performance of BERT based models is better when compared to the previous Bi-LSTM model proposed in the paper," and this concept applies to AUC values for classification tasks as well. Here, Settipalli teaches the higher the AUC value, the better the model is in implementing the prediction task. This means the higher the AUC score, the model performs well. On page 53, section 6.2.1, Settipalli mentions that "it is evident that BERT-base has the highest classification performance followed by Distill-BERT, then followed by Bi-LSTM in terms of the all tow metrics f1-score and overall-AUC." On page 53, section 6.2.2., Settipalli further discusses "BERT-base and DistilBERT have a higher subgroup-AUC, BPSN-AUC and BSPN-AUC for all the nine identity subgroups when compared to Bi-LSTM model." This shows that the BERT-base and DistilBERT models perform better and is considered to be an acceptable inference model compared to the Bi-LSTM model.) Note: For this limitation, the examiner has construed the word acceptable to mean best model performance to satisfy a condition.
However, Settipalli fail to teach:
A method … to reduce an amount of latent bias exhibited by the inference models when generating inferences, the method comprising: … the multiheaded inference model being an inference model comprising a multi-path neural network that comprises multiple label prediction heads and multiple bias feature heads through which intermediate outputs generated by a shared body of the multi-path neural network are consumed to generate a plurality of output labels;
“making a determination, …, regarding whether the trained multiheaded inference model meets a predetermined criterion;”
“ and in an instance of the determination where the multiheaded inference model meets a predetermined criterion: providing computer implemented services using first inferences generated by the trained multiheaded inference model, the first inferences being ones of the inferences.”
Yet, in an analogous system, Hu teaches “making a determination, …, regarding whether the trained multiheaded inference model meets a predetermined criterion;”
See Hu in paragraph [0104], where Hu describes that "the predicted position of the text in the invoice image samples is determined according to the training position distribution map and the training boundary image, the loss value between the predicted position and the reference label is calculated, and the parameters in the multi-task network model are adjusted by the loss value until the predicted position output by the multi-task network model after parameter adjustment meets the position conditions.” Here, Hu discloses using a trained multi-task model, once the model satisfies the position conditions from training, to help locate text in the invoice to prevent adjacent text boxes from sticking together. This correlates multi-task network model (i.e. multi headed inference model) meets the position conditions (i.e. meets a predetermined criterion).
Further, Hu teaches “in an instance of the determination where the multiheaded inference model meets a predetermined criterion: providing computer implemented services using first inferences generated by the trained multiheaded inference model ...”
See Hu in paragraph [0104], where Hu describes that "the predicted position of the text in the invoice image samples is determined according to the training position distribution map and the training boundary image, the loss value between the predicted position and the reference label is calculated, and the parameters in the multi-task network model are adjusted by the loss value until the predicted position output by the multi-task network model after parameter adjustment meets the position conditions. Thus, a multi-task network model that can locate the text in the invoice can be obtained. By locating the invoice image using this multi-task network model, the situation of adjacent text boxes sticking together can be avoided." Here, Hu discloses using a multi-task model, once the model satisfies the position conditions from training, to help locate text in the invoice to prevent adjacent text boxes from sticking together. This correlates multi-task network model (i.e. multi headed inference model) meets the position conditions (i.e. meets a predetermined criterion) and locate the text in the invoice corresponds with providing computer implemented services in the claim limitation.
Further, see Hu in [0139] describe “determining the text box boundaries of the text features through the boundary determination network; classifying the text features according to the classification network to obtain text features belonging to each text type;” Here, Hu describes the model performing classification tasks on text features, which is viewed as generating an inference. Later, see Hu in [0012-0013] describe “In one embodiment, before determining the position of the text in the invoice image based on the location distribution map and the boundary image, the method further includes: adjusting the size of the feature map corresponding to the text feature according to a preset size to obtain an adjusted feature map;” Further, see Hu in [0027] describe “calculating the loss value between the predicted position and the reference label, and adjusting the parameters in the multi-task network model through the loss value until the predicted position output by the multi-task network model after parameter adjustment meets the position conditions.” Here, Hu in [0027] shows if a multi-task model meets position conditions that was set by a preset size from [0012-0013].
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli and incorporate into the teachings of Hu because both references teach using multiheaded models for inference and determining whether the models provide computer services.
One of ordinary skill in the art would be motivated to do so because after determining the multiheaded inference model is acceptable, computer-implemented services would be subsequently provided can “reduce the amount of calculations and improve the generalization ability of multi-task network models” (Hu, [0072]).
However, Settipalli in view of Hu, did not teach
A method … to reduce an amount of latent bias exhibited by the inference models when generating inferences, the method comprising: … the multiheaded inference model being an inference model comprising a multi-path neural network that comprises multiple label prediction heads and multiple bias feature heads through which intermediate outputs generated by a shared body of the multi-path neural network are consumed to generate a plurality of output labels;
providing computer implemented services using first inferences generated by the trained multiheaded inference model, the first inferences being ones of the inferences”
In an analogous art, Bucklin teaches “to reduce an amount of latent bias exhibited by the inference models when generating inferences, … the multiheaded inference model being an inference model comprising a multi-path neural network that comprises multiple label prediction heads and multiple bias feature heads through which intermediate outputs generated by a shared body of the multi-path neural network are consumed to generate a plurality of output labels;”
See Bucklin in [0385] describe "In some cases (generally referred to as “supervised learning”), a training data set used to train a machine learning model can include known outcomes (e.g., labels). In alternative cases (generally referred to as “unsupervised learning”), a training data set does not include known outcomes." Bucklin here describes generating outcomes in form of labels (i.e. generate a plurality of output labels) .
Further, see Bucklin in [0045] describe “In some configurations, the feature may have multiple categories and the output generated by the model (that uses that feature) may have multiple types. For instance, a model may use a feature (e.g., gender that has a first category of men and a second category of women) to analyze applicant data and predict whether they would default on a loan (e.g., the output is the default prediction, which has a first type indicating a default and a second type of output indicating a no default). In some other configurations, the system may only mitigate bias associated with binary classes/values." Here, Bucklin shows a prediction task such as predicting whether an applicant default on a loan, and Bucklin indicates the output may have multiple types of categories or tasks, which relate to multiple label prediction tasks.
Later, see Bucklin describe in [0111] for more details “the system may revise the blueprint and add a new task corresponding to bias mitigation, which includes reweighting of the model. The new task can be placed directly after the categorical input node and before any categorical preprocessing/featurization tasks." Here, Bucklin shows multiple tasks regarding bias mitigation, which relates to multiple bias features. See Bucklin further describe in [0042] “the user may use the methods and systems discussed herein to determine whether a model is biased towards a particular class of individuals (e.g., whether a decision to grant loans to applicants is biased based on the applicants’ gender or race). The user may then request the system to mitigate the bias (if any), such that the model’ bias toward a class of individuals is reduced or the model is no longer biased towards that class of individuals (e.g., the bias is within a tolerable threshold).” where Bucklin shows using methods to detect if a model is biased towards a particular class of individuals then mitigate the bias (i.e. reducing an amount of latent bias) from the model itself.
See Bucklin also in [0161] describe “there are many different ways to implement the user interface for specifying the arrangement of the task hierarchy. But from a logical perspective, a task that is not at the leaf-level may include a directed graph of sub-tasks. At each of the top and intermediate levels of this hierarchy, there may be one starting sub-task whose input is from the parent task in the hierarchy (or the parent modeling technique at the top level of the hierarchy). There may also be one ending sub-task whose output is to the parent task in the hierarchy (or the parent modeling technique at the top". Here, Bucklin mentions the outputs from the model’s sub-tasks or intermediate levels, which relate to intermediate outputs generated by a multi-headed neural network. See Bucklin for details in [0287].
Bucklin further teaches “providing computer implemented services using first inferences generated by the trained multiheaded inference model, the first inferences being ones of the inferences”
See Bucklin in [0053] describe "Therefore, the page 300 may display additional inferences regarding the dataset 200 and may include summary statistics of the dataset 200. The page 300 may include the column 302 that provides a list of feature names, may indicate a value type for each feature (304), and other summary statistical information 306." Here, Bucklin describes displaying additional inferences in addition to an initial inference via page 300, which is part of a system that delivers computer services. See Bucklin in [0116] for more details.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli and Hu, and incorporate into the teachings of Bucklin because the references teach using multi-headed models to retrieve inference goals from user inputs for both prediction task and bias feature mitigation task.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli and Hu of inference goals for prediction and bias feature mitigation with Bucklin’s interface design and user inputs to provide “systems and methods of this technical solution can systematically and cost- effectively evaluate the space of potential predictive modeling techniques for prediction problems. This technical solution can utilize statistical learning techniques to systematically and cost-effectively evaluate the space of potential predictive modeling solutions for prediction problems” (Bucklin, [0132]).
However, Settipalli in view of Hu, further in view of Bucklin did not teach “… the multiheaded inference model … generated by a shared body …”
In an analogous method, Oneto teaches “the multiheaded inference model…generated by a shared body..”
See Oneto describe in second paragraph, page 228, “a particular instance of MTL which jointly learns a shared model between the groups as well as a specific model per group. We show how fairness constraints, measured with Equalized Odds or Equal Opportunities introduced in [20], can be built in MTL directly during the training phase.” Here, Oneto in page 228 describes creating a multi-task learning model (MTL) that learns a shared model that includes the fairness constraints, which are the bias terms of the bias features, as well as generating a model between the groups, which are part of the overall multi-task learning model.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli, Hu, and Bucklin, and incorporate into the teachings of Oneto because the references teach using multi-headed models to retrieve inference goals from user inputs for both prediction task and bias feature mitigation task.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli, Hu, and Bucklin of inference goals for prediction and bias feature mitigation with Oneto’s concept of a shared body to “show that MTL can be effectively used even when the sensitive feature is not available during testing by predicting the sensitive feature based on the non-sensitive ones.” (See Oneto in page section 5. Experiments, page 230).
Claim 5:
Regarding claim 5, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, teach the limitations in claim 1.
Further, Bucklin teaches
“obtaining the inference goals comprises: presenting, to a user, a graphical user interface comprising: a first inference goal control corresponding to a label of the labels, and a second inference goal control corresponding to a bias feature of the bias features,”
See Bucklin in paragraph [0029] where Bucklin shows “figures 3 A-3P illustrate different graphical user interfaces displayed within a bias and fairness evaluation system” and illustrates in paragraph [0086] that the interface has “may provide several input elements that can modify the display, allowing the user to focus on information of particular interest. For instance, the interactive element 343a may allow the user to revise the prediction threshold. The prediction threshold may be the dividing line for interpreting results in binary classification models. The system may use a default threshold of 0.5 (e.g., every prediction above this dividing line has a positive class label). However, this threshold can be revised by the user.” Here, Bucklin teaches using user input to navigate the interactive elements like the prediction threshold to connect to a first inference goal control corresponding to a label of the labels. This relates to obtaining, from the user and via the graphical user interface: first user input that indicates a first inference goal of the inference goals.
Further, Bucklin teaches in paragraph [0042] that “using the method 100, the system may display one or more GUIs on a user computer device, such as a computer operated by a user. As used herein, the user may be a user or a customer utilizing services associated with the system. For instance, the user may be a subscriber of the services rendered by the system and may utilize the system and its various models to generate decisions or receive predicted outputs. For instance, the user may access an electronic platform (e.g., website) associated with the system and interact with various GUIs and features discussed herein to evaluate how a model treats a particular feature. For instance, the user may use the methods and systems discussed herein to determine whether a model is biased towards a particular class of individuals (e.g., whether a decision to grant loans to applicants is biased based on the applicants' gender or race). The user may then request the system to mitigate the bias (if any), such that the model' bias toward a class of individuals is reduced or the model is no longer biased towards that class of individuals (e.g., the bias is within a tolerable threshold).” Here, Bucklin describes that the graphic user interface that contains the multi-task model with multiple user inputs performs two functions, one to generate decisions or receive predicted outputs that correspond to a first inference goal control, and two is to determine whether a model is biased towards a particular class of individuals that relates to the second inference goal control. For more information, see Bucklin in paragraphs [0043, 0045].
Bucklin also teaches “obtaining, from the user and via the graphical user interface: first user input that indicates a first inference goal of the inference goals, and second user input that indicates a second inference goal of the inference goals.”
See Bucklin mentions in paragraph [0090] that “using the input elements depicted within the graphical element 354, the user may select a protected feature and two class values of that feature to measure for data disparities. For instance, the user may select “gender,” “male,” and “female.” The system may then present the chart 356 that depicts data disparity vs feature importance. The chart 356 can be used to perform root-cause analysis of the model's bias for the selected classes (e.g., the data disparity vs feature importance chart can be used to identify which features in the dataset impact bias the most). The chart 356 may also detail where the bias exists within the feature.” Here, Bucklin shows that the interface requires the user to provide at least two inputs of a class value (i.e. first inference goal of the inference goals) and a protected feature (i.e. second inference goal of the inference goals). See Bucklin in paragraphs [0093, 0098, 0124] for more information.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli and Hu, and incorporate into the teachings of Bucklin because the references teach using multi-headed models to retrieve inference goals from user inputs for both prediction task and bias feature mitigation task.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli and Hu of inference goals for prediction and bias feature mitigation with Bucklin’s interface design and user inputs to provide “systems and methods of this technical solution can systematically and cost- effectively evaluate the space of potential predictive modeling techniques for prediction problems. This technical solution can utilize statistical learning techniques to systematically and cost-effectively evaluate the space of potential predictive modeling solutions for prediction problems” (Bucklin, [0132]).
Claim 8:
Regarding claim 8, Settipalli teaches “… when executed by a processor, cause the processor to perform operations for managing inference models, the operations comprising: obtaining inference goals for label inferences and bias feature inferences for a multiheaded inference model of the inference models;
See Settipalli teaches a processor in page 28, using a “GPU (Graphics Processing Unit) based compute engine from google cloud to train these models” along with a memory of 16gb vRAM from table 4.1, where Settipalli describes using a processing system.
PNG
media_image1.png
318
1090
media_image1.png
Greyscale
See Settipalli teaches obtaining inference goals for label inferences on page 35, section 4.6.1. classification performance, "to measure the overall classification performance of the models we have used the metrics f1 score and Overall-AUC i.e. AUC [21] calculated on the entire test set… The overall-AUC is a threshold agnostic metric so we can apply it directly on the model’s prediction." Here, Settipalli teaches and describes measuring the prediction performance for label inferences or classification performance with AUC scores or F1 scores as obtaining inference goals for label inferences.
See Settipalli also on page 25, section 3.6 metrics selection, discussing inference goals for bias feature inferences, in "after studying various metrics used to measure the biases in text classification we decided to use the per identity AUC metrics i.e. subgroup-AUC, BPSN-AUC and BNSP-AUC to measure the bias mitigation performance of the model over individual identity subgroup and the metric generalised mean AUC to measure the overall bias mitigation performance," Here, Settipalli teaches getting the metrics including generalised mean AUC scores of various subgroups to measure bias mitigation performance, which corresponds to obtaining inference goals for bias feature inference. See Settipalli on page 27, chapter 4 opening paragraphs, stating “multitask learning is used to train these models so that they can perform two tasks namely identity task and the toxicity task in parallel” for more information on the models used for multi-task learning).
However, Settipalli in view of Hu did not teach “a non-transitory machine-readable medium having instructions stored therein, …”
In an analogous art, Bucklin teaches “a non-transitory machine-readable medium having instructions stored therein, …”
See Bucklin describe in paragraph [0024] that “another aspect is directed towards another computer system that comprises a server comprising a processor and a non-transitory computer-readable medium containing instructions that when executed by the processor causes the processor to perform operations comprising receiving a feature of a plurality of features used by a model to generate output, wherein the feature comprises a plurality of categories, and the output comprises a plurality of types. The instruction may also cause the processor to identify a metric used to evaluate a performance of the model and a threshold for the metric. …The instruction may also cause the processor to generate a notification indicating the performance of the model responsive to a comparison of the value for the metric with the threshold for the metric.” Here, Bucklin teaches that a non-transitory computer-readable medium is part of the system that helps obtain inference goals for multiheaded models.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli and Hu, and incorporate into the teachings of Bucklin because the references teach using multi-headed models to retrieve inference goals from user inputs for both prediction task and bias feature mitigation task.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli and Hu of inference goals for prediction and bias feature mitigation with Bucklin’s interface design and user inputs that run on a non-transitory computer-readable medium to provide “systems and methods of this technical solution can systematically and cost- effectively evaluate the space of potential predictive modeling techniques for prediction problems. This technical solution can utilize statistical learning techniques to systematically and cost-effectively evaluate the space of potential predictive modeling solutions for prediction problems” (Bucklin, [0132]).
Regarding claim 8, the claim recites similar limitations as corresponding independent claim 1 and is rejected for similar reasons as claim 1 using similar teachings and rationale.
Claim 12:
Regarding claim 12, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, teach the limitations of claim 8. Further, claim 12 comprises of similar additional limitations as claim 5, and is rejected under the same rationale.
Claim 15:
Settipalli in view of Hu, did not teach a processing system specifically for data with “a data processing system, comprising: a processor; and a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations for managing inference models,”
In an analogous art, Bucklin teaches “a data processing system, comprising: a processor; and a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations for managing inference models,”
See Bucklin in paragraph [0006] that “the method may include receiving, by a data processing system comprising one or more processors and memory, a feature of a plurality of features used by a model to generate output,... The method may include identifying, by the data processing system, a metric used to evaluate the performance of the model and a threshold for the metric. …The method may include generating, by the data processing system, a notification indicating the performance of the model responsive to a comparison of the value for the metric with the threshold for the metric.” Here, Bucklin teaches that a data processing system, which includes processors and a memory, is part of the system that helps manage inference goals for multiheaded models.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli and Hu, and incorporate into the teachings of Bucklin because the references teach using multi-headed models to retrieve inference goals from user inputs for both prediction task and bias feature mitigation task.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli and Hu of inference goals for prediction and bias feature mitigation with Bucklin’s interface design and user inputs that run on a data processing system to provide “systems and methods of this technical solution can systematically and cost- effectively evaluate the space of potential predictive modeling techniques for prediction problems. This technical solution can utilize statistical learning techniques to systematically and cost-effectively evaluate the space of potential predictive modeling solutions for prediction problems” (Bucklin, [0132]).
Regarding claim 15, the claim recites similar limitations as corresponding independent claim 1 and is rejected for similar reasons as claim 1 using similar teachings and rationale.
Claim 19:
Regarding claim 19, Settipalli in view of Hu, and further in view of Bucklin teaches the limitations of claim 15. Further, claim 19 comprises of similar additional limitations as claim 5, and is rejected under the same rationale.
Claim 21:
Regarding claim 21, Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, teach the limitations in claim 1.
Further, Bucklin teaches “The method of claim 1, wherein each of the bias feature heads are configured to actively introduce one or more biases into a first set of the plurality of output labels generated by the bias feature heads, and each of the label prediction heads are not configured to actively introduce the one or more biases into a second set of the plurality of output labels generated by the label prediction heads.”
See Bucklin in [0059] describe “Accordingly, the favorable outcome may not always be the same as the assigned positive class. For example, when predicting whether or not an applicant will default on their loan. The positive class could be 1 (or "will default"), whereas the favorable target outcome would be 0 (or "will not default"). The favorable target outcome refers to the outcome that the protected individual would prefer to receive.” Bucklin here describes the prediction task of the multi-task model, where the prediction is either 1 with default on loan, or 0 with not default on loan.
Bucklin in [0060-0062] shows “The input element 314 requests the user to input a primary fairness metric. A fairness metric may refer to a metric that indicates how fair or biased a model is behaving towards a particular feature (protected feature).
[0061] Fairness metrics may refer to statistical measures of parity constraints used to assess the fairness of a model. The system may calculate the fairness metric in two steps. First, the system may calculate a fairness score for each protected class of the model's protected feature (e.g., the feature received from the user). The fairness score may refer to a numerical computation of model fairness against the protected class, based on the underlying fairness metric. Second, the system may normalize the fairness scores against the highest fairness score for the protected feature. This may be referred to herein as the relative score.
[0062] Metrics that measure “fairness by error” evaluate whether the model's error rate is equivalent across each protected class. These metrics may be best suited when the user does not have control over the outcome or wishes to conform to the ground truth, and simply desires a model to be equally right between each protected group. Metrics that measure “fairness by representation” evaluate whether the model's predictions are equivalent across each protected class. These metrics are best suited when the user has control over the target outcome or is willing to depart from ground truth in order for a model's predictions to exhibit more equal representation between protected groups, regardless of the target distribution in the training data.” Bucklin mentions here that the user can adjust settings (i.e. actively introduce one or more biases) regarding a fairness metric for each protected class (or group that can be biased against), and the user can also choose to not set the metric from ground truth (i.e. not actively introduce one or more biases).
See Bucklin in [0111] describe “In another example, the system may revise the blueprint and add a new task corresponding to bias mitigation, which includes reweighting of the model. The new task can be placed directly after the categorical input node and before any categorical preprocessing/featurization tasks. The new task may calculate a set of mitigation row weights using the target and the bias mitigation feature.” Bucklin shows that this can be done in another task, or bias feature head, different from the label prediction task or label prediction head. See Bucklin for details in [0208, 0385].
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli and Hu, and incorporate into the teachings of Bucklin because the references teach using multi-headed models to retrieve inference goals from user inputs for both prediction task and bias feature mitigation task.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli and Hu of inference goals for prediction and bias feature mitigation with Bucklin’s interface design and user inputs to provide “systems and methods of this technical solution can systematically and cost- effectively evaluate the space of potential predictive modeling techniques for prediction problems. This technical solution can utilize statistical learning techniques to systematically and cost-effectively evaluate the space of potential predictive modeling solutions for prediction problems” (Bucklin, [0132]).
Claim 22:
Regarding claim 22, Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, teach the limitations in claim 1.
Further, Bucklin teaches “22. The method of claim 1, wherein a first output label of the output labels generated through a first bias feature head of the bias feature heads is a first prediction in which a first bias specified by the first bias feature head is intentionally formed, the first prediction being formed to intentionally include the first bias when a first intermediate output of the intermediate outputs is consumed by the first bias feature head,”
See Bucklin in [0059] describe “... For example, when predicting whether or not an applicant will default on their loan. The positive class could be 1 (or "will default"), whereas the favorable target outcome would be 0 (or "will not default"). The favorable target outcome refers to the outcome that the protected individual would prefer to receive.” Bucklin here describes the prediction task of the multi-task model, where the prediction is either 1 with default on loan, or 0 with not default on loan.
Further, Bucklin in [0060-0062] shows “The input element 314 requests the user to input a primary fairness metric. A fairness metric may refer to a metric that indicates how fair or biased a model is behaving towards a particular feature (protected feature).
[0061] Fairness metrics may refer to statistical measures of parity constraints used to assess the fairness of a model. The system may calculate the fairness metric in two steps. First, the system may calculate a fairness score for each protected class of the model's protected feature (e.g., the feature received from the user). The fairness score may refer to a numerical computation of model fairness against the protected class, based on the underlying fairness metric. Second, the system may normalize the fairness scores against the highest fairness score for the protected feature. This may be referred to herein as the relative score.
[0062] Metrics that measure “fairness by error” evaluate whether the model's error rate is equivalent across each protected class. These metrics may be best suited when the user does not have control over the outcome or wishes to conform to the ground truth, and simply desires a model to be equally right between each protected group. Metrics that measure “fairness by representation” evaluate whether the model's predictions are equivalent across each protected class. These metrics are best suited when the user has control over the target outcome or is willing to depart from ground truth in order for a model's predictions to exhibit more equal representation between protected groups, regardless of the target distribution in the training data.” Bucklin mentions here that the user can adjust settings (i.e. intentionally form one or more biases) regarding a fairness metric for each protected class (or group that can be biased against), and the user can also choose to not set the metric from ground truth (i.e. not actively introduce one or more biases).
Further, see Bucklin in [0161] mention “At each of the top and intermediate levels of this hierarchy, there may be one starting sub-task whose input is from the parent task in the hierarchy (or the parent modeling technique at the top level of the hierarchy).” Here, Bucklin shows incorporating an intermediate output among the tasks.
See Bucklin in [0111] describe “In another example, the system may revise the blueprint and add a new task corresponding to bias mitigation, which includes reweighting of the model. The new task can be placed directly after the categorical input node and before any categorical preprocessing/featurization tasks. The new task may calculate a set of mitigation row weights using the target and the bias mitigation feature.” Bucklin shows that this can be done in another task, or bias feature head, different from the first label prediction task or label prediction head. See Bucklin for details in [0208, 0385].
Further, Bucklin teaches “and a second output label of the output labels generated through a first label prediction head of the label prediction heads comprises a second prediction in which the first bias of is not intentionally introduced by the first label prediction head when a second intermediate output of the intermediate outputs is consumed by the first label prediction head.”
See Bucklin in [0059] describe “... For example, when predicting whether or not an applicant will default on their loan. The positive class could be 1 (or "will default"), whereas the favorable target outcome would be 0 (or "will not default"). The favorable target outcome refers to the outcome that the protected individual would prefer to receive.” Bucklin here describes the prediction task of the multi-task model, where the prediction is either 1 with default on loan, or 0 with not default on loan.
Further, Bucklin in [0060-0062] shows “The input element 314 requests the user to input a primary fairness metric. A fairness metric may refer to a metric that indicates how fair or biased a model is behaving towards a particular feature (protected feature).
[0061] Fairness metrics may refer to statistical measures of parity constraints used to assess the fairness of a model. The system may calculate the fairness metric in two steps. First, the system may calculate a fairness score for each protected class of the model's protected feature (e.g., the feature received from the user). The fairness score may refer to a numerical computation of model fairness against the protected class, based on the underlying fairness metric. Second, the system may normalize the fairness scores against the highest fairness score for the protected feature. This may be referred to herein as the relative score.
[0062] Metrics that measure “fairness by error” evaluate whether the model's error rate is equivalent across each protected class. These metrics may be best suited when the user does not have control over the outcome or wishes to conform to the ground truth, and simply desires a model to be equally right between each protected group. Metrics that measure “fairness by representation” evaluate whether the model's predictions are equivalent across each protected class. These metrics are best suited when the user has control over the target outcome or is willing to depart from ground truth in order for a model's predictions to exhibit more equal representation between protected groups, regardless of the target distribution in the training data.” Bucklin mentions here that the user can adjust settings (i.e. intentionally form one or more biases) regarding a fairness metric for each protected class (or group that can be biased against), and the user can also choose to not set the metric from ground truth (i.e. not actively introduce one or more biases).
Further, see Bucklin in [0161] mention “At each of the top and intermediate levels of this hierarchy, there may be one starting sub-task whose input is from the parent task in the hierarchy (or the parent modeling technique at the top level of the hierarchy).” Here, Bucklin shows incorporating an intermediate output among the tasks.
See Bucklin in [0111] describe “In another example, the system may revise the blueprint and add a new task corresponding to bias mitigation, which includes reweighting of the model. The new task can be placed directly after the categorical input node and before any categorical preprocessing/featurization tasks. The new task may calculate a set of mitigation row weights using the target and the bias mitigation feature.” Bucklin shows that this can be done in another task, or bias feature head, different from the label prediction task or label prediction head. See Bucklin for details in [0208, 0385].
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli and Hu, and incorporate into the teachings of Bucklin because the references teach using multi-headed models to retrieve inference goals from user inputs for both prediction task and bias feature mitigation task.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli and Hu of inference goals for prediction and bias feature mitigation with Bucklin’s interface design and user inputs to provide “systems and methods of this technical solution can systematically and cost- effectively evaluate the space of potential predictive modeling techniques for prediction problems. This technical solution can utilize statistical learning techniques to systematically and cost-effectively evaluate the space of potential predictive modeling solutions for prediction problems” (Bucklin, [0132]).
Claims 2 and 3 are rejected under 35 U.S.C. 103 as being unpatentable over Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, further in view of Li K. et al., (Pub. No. CN 114463063), published on May 10, 2022 (hereafter, Li K.), and further in view of Ganguly D. (Pub. No. US 2022/0198297 A1), published on June 23, 2022 (hereafter, Ganguly).
Claim 2:
Regarding claim 2, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, teach the limitations in claim 1. Further, Settipalli teaches both the “predictive power level for the labels” and “predictive power levels for the bias features” in page 35 section 4.6.1 where Settipalli discusses using the metrics AUC and F1 score for both predictions in "to measure the overall classification performance of the models we have used the metrics f1 score and Overall-AUC", and Settipalli in page 25, section 3.6 Metrics Selection, uses generalised mean AUC scores in identifying bias subgroups in "use the per identity AUC metrics i.e. subgroup-AUC, BPSN-AUC and BNSP-AUC to measure the bias mitigation performance of the model over individual identity subgroup and the metric generalised mean AUC to measure the overall bias mitigation performance".
Further, on page 35, section 4.6.1., Settipalli discusses that the study has “chosen the metrics overall-AUC and f1-score as they are not effected by the unbalanced class distribution. The overall-AUC is a threshold agnostic metric so we can apply it directly on the model’s prediction. But to calculate the metric f1-score we need to model’s predictions into binary format and the corresponding ground truth values into 0 or 1 using a threshold of 0.5”. Here, Settipalli describes the label inference goals using the metrics overall-AUC and F1 since they measure the model’s prediction in labeling. On pages 35-36, in section 4.6.2, Settipalli discusses “for measuring the bias mitigation performance over different identity subgroups we have used the AUC based metrics that were selected in the literature review i.e. subgroup-AUC, BPSN-AUC, BNSP-AUC and Generalized mean AUC. Subgroup-AUC: For calculating this measure, a sample set which contains an equal number of toxic and non-toxic examples corresponding to particular identity subgroup is selected from the test set. Calculating the AUC metric on this sample set will give us the get subgroup AUC for the particular identity. The subgroup-AUC will tell [how] well the model is able to distinguish positive and negative sentences corresponding to a particular identity subgroup”. Here, Settipalli describes using a specific type of AUC score called sub-group AUC score to measure bias mitigation performance to determine the success of the model in distinguishing prediction inferences of labeling if a sentence has positive or negative sentiment that relate to a specific identity subgroup or an attribute that is frequently biased. See page 34 where Settipalli discusses various models such as BERT base, BERT large, and Bi-LSTM on table 4.4. The models here represent multi-headed inference models.
Further, see Settipalli also on page 25, section 3.6 metrics selection, Settipalli teaches getting the metrics including generalised mean AUC scores of various subgroups to measure bias mitigation performance, in "after studying various metrics used to measure the biases in text classification we decided to use the per identity AUC metrics i.e. subgroup-AUC, BPSN-AUC and BNSP-AUC to measure the bias mitigation performance of the model over individual identity subgroup and the metric generalised mean AUC to measure the overall bias mitigation performance," which corresponds to obtaining inference goals for bias feature inferences.
Later, Settipalli on pages 39-41 in figures 5.1- 5.3 and tables 5.1- 5.3 displays scores of AUC and F1 scores for all models including subgroups, which corresponds to the recited claim limitation in identifying, for the trained multiheaded inference model, predictive power levels for labels, and predictive power levels for bias features.
However, Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, did not explicitly teach
a minimum predictive power level for labels
or a maximum predictive power levels for the bias features
In an analogous system, Li K. teaches “a minimum predictive power level for labels”, See paragraph [0124], in figure 5, where Li K. describes "after the multi-task model training, the computer device also can test the trained multi-task model to obtain the model evaluation index value of the trained multi-task model, for example, the model evaluation index value of the multi-task model may include the area under the ROC curve (Area Under Curve, AUC), wherein the AUC can be used for evaluating the training effect of the multi-task model, AUC= 1; It is believed that the multitask model is a perfect multitasking model (in the ideal state). if the model evaluation index value of the trained multi-task model satisfies the model evaluation index condition, then the computer device can call the trained multi-task model to predict one or more task index corresponding to the target advertisement data provided by the first service party. wherein the model evaluation index condition can include index threshold value, if the trained multi-task model of the model evaluation index value (e.g., 0.9) is greater than or equal to the index threshold (0.88), it can determine the model evaluation index value of the trained multi-task model satisfies the model evaluation index condition." This is applied later for encryption process of data owned by business organizations. Here, Li K. describes a quantity for a minimum threshold value being greater than or equal to 0.88, that the multi-task model can take AUC as a model evaluation index value (i.e. minimum predictive power level for labels).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, and Oneto and incorporate into the teachings of Li K. because these references teach specifying an inference goal for a minimum metric for a labeling prediction.
One of ordinary skill in the art would be motivated to do so because making such a combination helps “the trained multi-task model” to “predict multiple task indicators, thereby improving the accuracy of the predicted task indicators” (Li K., paragraph [0051]).
However, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, in further view of Li K. did not teach a “maximum predictive power levels for the bias features.”
In an analogous system, Ganguly describes in paragraph [0033] from figure 3 that a “neural network architecture 300” is responsible “for jointly learning to perform well in a primary task and poorly in a pseudo-bias task.” Here, Ganguly teaches two tasks performed on the same multi-objective model, one to improve prediction from the primary task, and one to perform poorly for identifying bias features. Ganguly also mentions these two tasks are learned within the same model. Additionally, Ganguly describes that jointly learning implies having two learning rates on the same multi-objective model. Ganguly also mentions in paragraph [0031] about a multi-objective loss function
PNG
media_image2.png
5
166
media_image2.png
Greyscale
“which learns to perform above a first threshold on a primary task and perform below a second threshold (note the − sign in the second half of the function) on a pseudo-bias task. In various embodiments, the performance is effective based on a determination that the likelihood is maximized and/or the performance is poor based on an inverse determination that the likelihood is maximized. In another embodiment, the performance is effective based on a determination that the likelihood is at or above a defined threshold.” Here, the loss function defines what is acceptable or not for both prediction (i.e. predictive power level for labels) and bias features (i.e. predictive power levels for bias features). In particular, the loss function includes a maximum cutoff with stating that the inverse determination that the likelihood is maximized on the bias features from the equation in paragraph [0031], and that the likelihood is at or above a defined threshold (i.e. maximum predictive power levels). Ganguly shows the term ‘perform above a first threshold’ to relate to a maximum value for model performance. This connects to each of the learning rates for the respective two tasks, of prediction performance and identifying bias features, performed by the multi-objective model.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, and Oneto, Li K., and incorporate into the teachings of Ganguly because these references teach establishing learning of the multiheaded inference model as well as teaching a range for label predictive power levels and bias feature predictive power levels.
One of ordinary skill in the art would be motivated to do so because an “advantage of such a system is that it can both identify and mitigate bias from the predictions generated by a machine-learning model without the need of either human supervision or annotations for secondary-identity attributes” (Ganguly, [0005]).
Claim 3:
Regarding claim 3, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, and in further view of Li K., and further in view of Ganguly, teaches the limitations in claim 2.
Further, Ganguly teaches
“establishing a first learning rate based on minimum predictive power level for the labels and establishing second learning rate based on the minimum and maximum predictive power levels for the bias features”
See Ganguly in paragraph [0039] describes that a “system trains (e.g., via a multi-objective learning component 140 operatively coupled to a processor 110) the machine-learning model 150 to mitigate bias from predictions,” which shows a multi-objective learning component (i.e. multiheaded inference model). Further, Ganguly describes in paragraph [0033] from figure 3 that a “neural network architecture 300” is responsible “for jointly learning to perform well in a primary task and poorly in a pseudo-bias task.” Here, Ganguly teaches two tasks performed on the same multi-objective model, one to improve prediction from the primary task, and one to perform poorly for identifying bias features. Ganguly also mentions these two tasks are learned within the same model. Additionally, Ganguly describes that jointly learning implies having two learning rates on the same multi-objective model.
Further, Ganguly also mentions in paragraph [0031] about a loss function “which learns to perform above a first threshold on a primary task and perform below a second threshold (note the − sign in the second half of the function) on a pseudo-bias task. In various embodiments, the performance is effective based on a determination that the likelihood is maximized and/or the performance is poor based on an inverse determination that the likelihood is maximized. In another embodiment, the performance is effective based on a determination that the likelihood is at or above a defined threshold.” Here, the loss function defines a range for both prediction (i.e. predictive power level for labels) and bias features (i.e. predictive power levels for bias features). Ganguly also mentions the model with loss function that perform above a first threshold on a primary task and perform below a second threshold on a pseudo-bias task, which relates to having minimum and maximum values since Ganguly is describing a range. This applies to each of the learning rates for the respective two tasks, of prediction performance and identifying bias features, performed by the multi-objective model. Ganguly also relates that performing above a first threshold corresponds to a minimum predictive power level for the labels and performing below a second threshold accounts for the maximum predictive power level for bias features.
Ganguly also introduces a multi-task model in paragraph [0013] that the diagram “FIG. 3 illustrates a non-limiting schematic diagram of a neural network architecture for jointly learning to perform at a first level in a primary task and perform at a second level that is less than the first level in a pseudo-bias task in accordance with one or more embodiments described herein.” Here, Ganguly shows that the model performs two functions, one primary task to identify the category label (i.e. establish predictive power level for the labels,), and a second task to perform debiasing on the category labels (i.e. establish predictive power levels for the bias features).
PNG
media_image3.png
518
538
media_image3.png
Greyscale
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, and Oneto, Li K., and incorporate into the teachings of Ganguly because these references teach establishing learning of the multiheaded inference model as well as what is acceptable for label predictive power levels and bias feature predictive power levels.
One of ordinary skill in the art would be motivated to do so because an “advantage of such a system is that it can both identify and mitigate bias from the predictions generated by a machine-learning model without the need of either human supervision or annotations for secondary-identity attributes” (Ganguly, [0005]).
Claims 9, 10, and 16-17 are rejected under 35 U.S.C. 103 as being unpatentable over Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, and further in view of Ganguly.
Claim 9:
Regarding claim 9, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, teach the limitations of claim 8.
Further, Bucklin teaches “The non-transitory machine-readable medium of claim 8, wherein the inference goals specify:”
See Bucklin in paragraph [0024] describing “another aspect is directed towards another computer system that comprises a server comprising a processor and a non-transitory computer-readable medium containing instructions that when executed by the processor causes the processor to perform operations comprising receiving a feature of a plurality of features used by a model to generate output, wherein the feature comprises a plurality of categories, and the output comprises a plurality of types. The instruction may also cause the processor to identify a metric used to evaluate a performance of the model and a threshold for the metric. The instruction may also cause the processor to determine a value for the metric for a category of the plurality of categories of the feature ... The instruction may also cause the processor to generate a notification indicating the performance of the model responsive to a comparison of the value for the metric with the threshold for the metric.” Bucklin discloses a non-transitory computer readable medium to implement computer services such as generating outputs of a machine learning model. Later, Bucklin in paragraph [0124] states “the cloud 575 may include back end platforms, e.g., servers 195, storage, server farms or data centers. For example, the cloud 575 can include or correspond to a server 195 or system remote from one or more clients 565 to provide third party control over a pool of shared services and resources. The computing environment 560 can provide resource pooling to serve multiple users via clients 565 through a multi-tenant environment or multi-tenant model with different physical and virtual resources dynamically assigned and reassigned responsive to different demands within the respective environment. The multi-tenant environment can include a system or architecture that can provide a single instance of software, an application or a software application to serve multiple users.” Here, Bucklin teaches a non-transitory computer-readable medium that serves multiple users and operates to run the multi-head inference models.
Further, Bucklin teaches “a minimum predictive power level for the labels,”
See Bucklin at paragraphs [0086- 0087] that “the interactive element 343a may allow the user to revise the prediction threshold. The prediction threshold may be the dividing line for interpreting results in binary classification models. The system may use a default threshold of 0.5 (e.g., every prediction above this dividing line has a positive class label). However, this threshold can be revised by the user. [0087] For imbalanced datasets, a threshold of 0.5 can result in a validation partition without any positive class predictions, preventing the calculation of fairness scores on the per-class bias tab. To recalculate and surface fairness scores, the system may receive a revised prediction threshold from the user and may resolve the dataset imbalance. All fairness metrics (except prediction balance) may use the model's prediction threshold when calculating fairness scores. Changing this value may cause the system to recalculate the fairness scores and update the chart to display the new values.” Here, Bucklin specifies a value of 0.5 for a prediction threshold metric or minimum acceptable predictive power level for the labels. Any value above 0.5 is labeled one class, and any value below 0.5 is labeled a different class.
However, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, did not explicitly teach: “and a maximum predictive power levels for the bias features,”
In an analogous system, Ganguly teaches “and a maximum predictive power levels for the bias features,”
See Ganguly describe in paragraph [0033] from figure 3 that a “neural network architecture 300” is responsible “for jointly learning to perform well in a primary task and poorly in a pseudo-bias task.” Here, Ganguly teaches two tasks performed on the same multi-objective model, one to improve prediction from the primary task, and one to perform poorly for identifying bias features. Ganguly also mentions these two tasks are learned within the same model. Additionally, Ganguly describes that jointly learning implies having two learning rates on the same multi-objective model. Ganguly also mentions in paragraph [0031] about a multi-objective loss function
PNG
media_image2.png
5
166
media_image2.png
Greyscale
“which learns to perform above a first threshold on a primary task and perform below a second threshold (note the − sign in the second half of the function) on a pseudo-bias task. In various embodiments, the performance is effective based on a determination that the likelihood is maximized and/or the performance is poor based on an inverse determination that the likelihood is maximized. In another embodiment, the performance is effective based on a determination that the likelihood is at or above a defined threshold.” Here, the loss function defines what is acceptable or not for both prediction (i.e. predictive power level for labels) and bias features (i.e. predictive power levels for bias features). In particular, the loss function includes a maximum cutoff with stating that the inverse determination that the likelihood is maximized on the bias features from the equation in paragraph [0031], and that the likelihood is at or above a defined threshold (i.e. maximum acceptable predictive power levels). This connects to each of the learning rates for the respective two tasks, of prediction performance and identifying bias features, performed by the multi-objective model.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, Oneto, and incorporate into the teachings of Ganguly because these references teach establishing learning of the multiheaded inference model as well as teaching a range for label predictive power levels and bias feature predictive power levels.
One of ordinary skill in the art would be motivated to do so because an “advantage of such a system is that it can both identify and mitigate bias from the predictions generated by a machine-learning model without the need of either human supervision or annotations for secondary-identity attributes” (Ganguly, [0005]).
Claim 10:
Regarding claim 10, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, and further in view of Ganguly, teach the limitations of claim 9.
Further, Ganguly teaches
“establishing a first learning rate based on minimum predictive power level for the labels,”
“establishing second learning rate based on the minimum and maximum predictive power levels for the bias features”
See Ganguly in paragraph [0039] using a system that “trains (e.g., via a multi-objective learning component 140 operatively coupled to a processor 110) the machine-learning model 150 to mitigate bias from predictions,” which shows a multi-objective learning component (i.e. multiheaded inference model). Further, Ganguly describes in paragraph [0033] from figure 3 that a “neural network architecture 300” is responsible “for jointly learning to perform well in a primary task and poorly in a pseudo-bias task.” Here, Ganguly teaches two tasks performed on the same multi-objective model, one to improve prediction from the primary task, and one to perform poorly for identifying bias features. Ganguly also mentions these two tasks are learned within the same model. Additionally, Ganguly describes that jointly learning implies having two learning rates on the same multi-objective model. See figure 10 in paragraph [0013] for more information.
Further, Ganguly also mentions in paragraph [0031] about a loss function “which learns to perform above a first threshold on a primary task and perform below a second threshold (note the − sign in the second half of the function) on a pseudo-bias task. In various embodiments, the performance is effective based on a determination that the likelihood is maximized and/or the performance is poor based on an inverse determination that the likelihood is maximized. In another embodiment, the performance is effective based on a determination that the likelihood is at or above a defined threshold.” Here, the loss function defines what is acceptable or not for both prediction (i.e. predictive power level for labels) and bias features (i.e. predictive power levels for bias features). This connects to each of the learning rates for the respective two tasks, of prediction performance and identifying bias features, performed by the multi-objective model. Ganguly also elaborates that the jointly learning tasks involve first primary task for prediction, and another second task for predictive power levels of bias features. This involves a first learning rate for a threshold level acceptable predictive power level for the labels, and a second learning rate for predictive power levels of bias features. Ganguly also relates that performing above a first threshold corresponds to a minimum predictive power level for the labels and performing below a second threshold accounts for the maximum predictive power level for bias features.
PNG
media_image3.png
518
538
media_image3.png
Greyscale
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, and Oneto, and incorporate into the teachings of Ganguly because these references teach establishing learning of the multiheaded inference model as well as what is acceptable for label predictive power levels and bias feature predictive power levels.
One of ordinary skill in the art would be motivated to do so because an “advantage of such a system is that it can both identify and mitigate bias from the predictions generated by a machine-learning model without the need of either human supervision or annotations for secondary-identity attributes” (Ganguly, [0005]).
Claim 16:
Regarding claim 16, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, teach the limitations of claim 15.
Further, Bucklin teaches “The data processing system of claim 15, wherein the inference goals specify:”
See Bucklin in paragraph [0024] describing “another aspect is directed towards another computer system that comprises a server comprising a processor and a non-transitory computer-readable medium containing instructions that when executed by the processor causes the processor to perform operations comprising receiving a feature of a plurality of features used by a model to generate output, …The instruction may also cause the processor to identify a metric used to evaluate a performance of the model and a threshold for the metric. The instruction may also cause the processor to determine a value for the metric for a category ... The instruction may also cause the processor to generate a notification indicating the performance of the model responsive to a comparison of the value for the metric with the threshold for the metric.” Further, Bucklin in paragraph [0006] mentions the “method may include receiving, by a data processing system comprising one or more processors and memory, a feature of a plurality of features used by a model to generate output, … The method may include identifying, by the data processing system, a metric used to evaluate the performance of the model and a threshold for the metric. The method may include determining, by the data processing system, a value for the metric for a category of the plurality of categories of the feature …The method may include generating, by the data processing system, a notification indicating the performance of the model responsive to a comparison of the value for the metric with the threshold for the metric.” Here, Bucklin discloses a data processor to implement computer services such as generating outputs of a machine learning model. For more information, see Bucklin in paragraph [0124].
Further, Bucklin teaches “a minimum predictive power level for the labels,”
See Bucklin at paragraphs [0086- 0087] that “the interactive element 343a may allow the user to revise the prediction threshold. The prediction threshold may be the dividing line for interpreting results in binary classification models. The system may use a default threshold of 0.5 (e.g., every prediction above this dividing line has a positive class label). However, this threshold can be revised by the user. [0087] For imbalanced datasets, a threshold of 0.5 can result in a validation partition without any positive class predictions, preventing the calculation of fairness scores on the per-class bias tab. To recalculate and surface fairness scores, the system may receive a revised prediction threshold from the user and may resolve the dataset imbalance. All fairness metrics (except prediction balance) may use the model's prediction threshold when calculating fairness scores. Changing this value may cause the system to recalculate the fairness scores and update the chart to display the new values.” Here, Bucklin specifies a value of 0.5 for a prediction threshold metric or minimum acceptable predictive power level for the labels. Any value above 0.5 is labeled one class, and any value below 0.5 is labeled a different class.
However, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, did not explicitly teach: “and maximum predictive power levels for the bias features,”
In an analogous system, Ganguly teaches “and maximum predictive power levels for the bias features,”
See Ganguly describes in paragraph [0033] from figure 3 that a “neural network architecture 300” is responsible “for jointly learning to perform well in a primary task and poorly in a pseudo-bias task.” Here, Ganguly teaches two tasks performed on the same multi-objective model, one to improve prediction from the primary task, and one to perform poorly for identifying bias features. Ganguly also mentions these two tasks are learned within the same model. Additionally, Ganguly describes that jointly learning implies having two learning rates on the same multi-objective model. Ganguly also mentions in paragraph [0031] about a multi-objective loss function
PNG
media_image2.png
5
166
media_image2.png
Greyscale
“which learns to perform above a first threshold on a primary task and perform below a second threshold (note the − sign in the second half of the function) on a pseudo-bias task. In various embodiments, the performance is effective based on a determination that the likelihood is maximized and/or the performance is poor based on an inverse determination that the likelihood is maximized. In another embodiment, the performance is effective based on a determination that the likelihood is at or above a defined threshold.” Here, the loss function defines what is acceptable or not for both prediction (i.e. predictive power level for labels) and bias features (i.e. predictive power levels for bias features). Ganguly also mentions the model with a loss function that perform above a first threshold on a primary task and perform below a second threshold on a pseudo-bias task, which relates to having minimum and maximum values since Ganguly is describing a range. This applies to each of the learning rates for the respective two tasks, of prediction performance and identifying bias features, performed by the multi-objective model.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, Oneto, and incorporate into the teachings of Ganguly because these references teach establishing learning of the multiheaded inference model as well as teaching a range for label predictive power levels and bias feature predictive power levels.
One of ordinary skill in the art would be motivated to do so because an “advantage of such a system is that it can both identify and mitigate bias from the predictions generated by a machine-learning model without the need of either human supervision or annotations for secondary-identity attributes” (Ganguly, [0005]).
Claim 17:
Regarding claim 17, Settipalli in view of Hu, and further in view of Bucklin, further in view of Oneto, and further in view of Ganguly, teaches the limitations of claim 16. Further, claim 17 comprises of similar additional limitations as claims 3 and 10, and is rejected under the same rationale.
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Settipalli, in view of Hu, further in view of Bucklin, further in view of Oneto, further in view of Li K., and further in view of Ganguly, and further in view of Ragonesi R. in “Addressing Dataset Bias in Deep Neural Networks”, published on February 25, 2022, available at https://tesidottorato.depositolegale.it/handle/20.500.14242/71496, (hereafter, Ragonesi).
Claim 4:
Regarding claim 4, Settipalli in view of Hu, in further view of Bucklin, in further view of Oneto, in further view of Li K., and in further view of Ganguly, teaches the limitations in claim 3.
Settipalli teaches “The method of claim 3, wherein training the multiheaded inference model comprises: performing … training … based on the …learning rate;” see page 34, where Settipalli shows table 4.4 where the learning rates are used for the prediction models are listed, where each model also performs both classification prediction and bias reduction tasks.
However, Settipalli in view of Hu, in further view of Bucklin, in further view of Oneto, and in further view of Li K., did not explicitly teach
“a first number of training cycles based on the first learning rate,”
“performing a second number of untraining cycles based on the second learning rate”
In an analogous art, Ganguly did teach “a first number of training cycles based on the first learning rate”, See paragraphs [0024-0026] where Ganguly shows that “data provided to the network of learner units can include a training set (e.g., a set of data with known classifications that is employed for the training process) that is employed at a beginning of the training process. Utilizing the training set, the network of learner units can perform iterative processing stages in which data generated during a particular processing stage is determined from data generated during one or more previous processing stages,” where the learner units used in the beginning of the training process means an initial learning rate and are used iteratively and that “once identified, the machine-learning model can be trained, via a multi-objective learning component operatively coupled to the processer, to mitigate bias from the one or more predictions.” Here, Ganguly shows that the iterative training cycles are repetitively performed to train the multi-objective model (i.e. multiheaded inference model) using an initial learning rate.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, Oneto, and Li K., and incorporate into the teachings of Ganguly because these references teach establishing learning rates and training cycles of the multiheaded inference model for label predictive power levels and bias feature predictive power levels.
One of ordinary skill in the art would be motivated to do so because an “advantage of such a system is that it can both identify and mitigate bias from the predictions generated by a machine-learning model without the need of either human supervision or annotations for secondary-identity attributes,” (Ganguly, [0005]).
However, Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, further in view of Li K., and in further view of Ganguly, did not teach “performing a second number of untraining cycles based on the second learning rate”.
In an analogous system, Ragonesi teaches this in page 36, section 4.2, where Ragonesi discusses "a similar formulation is related to the so-called “algorithmic fairness” (57). The problem here is learning representations that do not rely on sensitive attributes (such as, e.g., gender, age or ethnicity), in order to prevent from learning discriminant capabilities towards protected categories. Our methods can be applied in this setting, in order to minimize the mutual information between the learned representation and the sensitive attribute (whose distribution might be biased for what concerns the training set." Here, Ragonesi shows that prevent from learning means unlearning. Ragonesi also discusses unlearning iterations where the model is prevented from learning (i.e. cycles of untraining) the sensitive attributes (i.e. bias features). Further, on page 66, Ragonesi describes how to "devise an iterative method that computes pseudo-labels for unlabeled samples and then train a Generative Adversarial Network to generate new data points according to such noisy labels. Repeating this procedure in an iterative way, allows to refine the pseudo-labels, leading to improved accuracy on the target domain." Here, Ragonesi further describes an iterative method which correspond to the cycles of untraining, which improves training process to recognize noisy labels (i.e. labels for bias features).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, Oneto, Li K., and Ganguly, incorporate into the teachings of Ragonesi because these references teach establishing learning rates and training cycles of the multiheaded inference model for label predictive power levels and untraining cycles of the bias feature predictive power levels.
One of ordinary skill in the art would be motivated to make the combination because doing so helps “a bi-level optimization algorithm inspired by meta-learning to learn efficiently
from such data” (Ragonesi, page 56, section 5.3.2), and training with “such weights help the model to remove the bias from the data representation” (Ragonesi, page 14, section 2.2.2).
Claims 11 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Settipalli, in view of Hu, further in view of Bucklin, further in view of Oneto, further in view of Ganguly, and further in view of Ragonesi .
Claim 11:
Regarding claim 11, Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, and further in view of Ganguly, teach the limitations of claim 10.
Further, Settipalli teaches “The method of claim 3, wherein training the multiheaded inference model comprises: performing … training … based on the …learning rate;”
See page 34, where Settipalli shows table 4.4 where the learning rates are used for the prediction models are listed, where each model also performs both classification prediction and bias reduction tasks.
However, Settipalli in view of Hu, in further view of Bucklin, and in further view of Oneto, did not explicitly teach
“a first number of training cycles based on the first learning rate,”
“performing a second number of untraining cycles based on the second learning rate”
In an analogous art, Ganguly teaches “a first number of training cycles based on the first learning rate”, See paragraphs [0024-0026] where Ganguly shows that “data provided to the network of learner units can include a training set (e.g., a set of data with known classifications that is employed for the training process) that is employed at a beginning of the training process. Utilizing the training set, the network of learner units can perform iterative processing stages in which data generated during a particular processing stage is determined from data generated during one or more previous processing stages,” where the learner units used in the beginning of the training process means an initial learning rate and are used iteratively and that “once identified, the machine-learning model can be trained, via a multi-objective learning component operatively coupled to the processer, to mitigate bias from the one or more predictions.” Here, Ganguly shows that the iterative training cycles are repetitively performed to train the multi-objective model (i.e. multiheaded inference model) using an initial learning rate.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, Oneto, and incorporate into the teachings of Ganguly because these references teach establishing learning rates and training cycles of the multiheaded inference model for label predictive power levels and bias feature predictive power levels.
One of ordinary skill in the art would be motivated to do so because an “advantage of such a system is that it can both identify and mitigate bias from the predictions generated by a machine-learning model without the need of either human supervision or annotations for secondary-identity attributes,” (Ganguly, [0005]).
However, Settipalli in view of Hu, in view of Bucklin, in further view of Oneto, and in further view of Ganguly, did not teach “performing a second number of untraining cycles based on the second learning rate”.
In an analogous system, Ragonesi teaches this in page 36, section 4.2, where Ragonesi discusses "a similar formulation is related to the so-called “algorithmic fairness” (57). The problem here is learning representations that do not rely on sensitive attributes (such as, e.g., gender, age or ethnicity), in order to prevent from learning discriminant capabilities towards protected categories. Our methods can be applied in this setting, in order to minimize the mutual information between the learned representation and the sensitive attribute (whose distribution might be biased for what concerns the training set)." Here, Ragonesi shows that prevent from learning means unlearning. Ragonesi also discusses unlearning iterations where the model is prevented from learning (i.e. cycles of untraining) the sensitive attributes (i.e. bias features). Further, on page 66, Ragonesi describes how to "devise an iterative method that computes pseudo-labels for unlabeled samples and then train a Generative Adversarial Network to generate new data points according to such noisy labels. Repeating this procedure in an iterative way, allows to refine the pseudo-labels, leading to improved accuracy on the target domain." Here, Ragonesi further describes an iterative method which correspond to the cycles of untraining, which improves training process to recognize noisy labels (i.e. labels for bias features).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Settipalli, Hu, Bucklin, Oneto, and Ganguly, incorporate into the teachings of Ragonesi because these references teach establishing learning rates and training cycles of the multiheaded inference model for label predictive power levels and untraining cycles of the bias feature predictive power levels.
One of ordinary skill in the art would be motivated to make the combination because doing so helps “a bi-level optimization algorithm inspired by meta-learning to learn efficiently
from such data” (Ragonesi, page 56, section 5.3.2), and training with “such weights help the model to remove the bias from the data representation” (Ragonesi, page 14, section 2.2.2).
Claim 18:
Regarding claim 18, Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, and further in view of Ganguly teaches the limitations of claim 17. Further, claim 18 comprises of similar additional limitations as claim 11, and is rejected under the same rationale.
Claims 6 and 13 is rejected under 35 U.S.C. 103 as being unpatentable over Settipalli, in view of Hu, further in view of Bucklin, further in view of Oneto, further in view of Li, T., available on https://ieeexplore.ieee.org/abstract/document/10075647, on October 22, 2018, (hereafter, Li, T.)
Claim 6:
Regarding claim 6, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, teaches the limitations in claim 5.
Further, Settipalli teaches the “The method of claim 5, wherein the first inference goal control comprises: … for the predictive power level for label”,
See Settipalli in page 35, section 4.6.1 in "to measure the overall classification performance of the models we have used the metrics f1 score and Overall-AUC i.e. AUC [21] calculated on the entire test set ... The overall-AUC is a threshold agnostic metric so we can apply it directly on the model’s prediction." Here, Settipalli teaches and describes measuring the prediction performance for label inferences or classification performance with AUC scores or F1 scores as obtaining inference goals for label inferences. Settipalli here discusses the first inference goal corresponding to the prediction task (i.e. label of the labels) for the predictive power level for label. Here, Settipalli teaches the inference goal control for the predictive power level for label for multi-task models.
However, Settipalli in view of Hu did not teach “…a slider that the user may actuate along a path to provide the first user input, and a position of the slider in the path defining a range for the predictive power level for label.”
In an analogous system, Bucklin teaches “the method …, wherein the first inference goal control comprises: a slider that the user may actuate along a path to provide the first user input, ….”
See Bucklin in paragraph [0085] disclosing that “the system may allow the user to toggle between different protected classes and features. For instance, as depicted on page 348, when the system determines that the user has interacted with the interactive element 350 (e.g., the user has toggled to age bracket from gender), the system dynamically revises the chart 340 to the chart 352. The chart 352 displays similar bias evaluations as the chart 340 but for a different feature to be protected.” This shows that the toggle feature corresponds to a slider function of the interface that let user to actuate along a path for the first user input.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli and Hu, and incorporate into the teachings of Bucklin because the references teach using multi-headed models to retrieve inference goal control from user inputs for a task using a slider interface that the user adjusts to also define an acceptable model range for the prediction task.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli and Hu of inference goal control for a task with Bucklin’s slider interface providing user inputs to “systematically and cost- effectively evaluate the space of potential predictive modeling techniques for prediction problems. This technical solution can utilize statistical learning techniques to systematically and cost-effectively evaluate the space of potential predictive modeling solutions for prediction problems” (Bucklin, [0132]).
However, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto failed to teach “a position of the slider in the path defining a range for the predictive power level for label.”
In an analogous art, Li, T. teaches a slider “a position of the slider in the path defining a range for the predictive power level for label.”
See Li, T. in page 7, left column, second paragraph where Li, T. mentions “updated sliders in the parameter panel on the left: the lower and upper limit of each range (blue circles in each slider is automatically re-positioned to reflect the hyperparameter and performance metrics of the experiments selected.” Such disclosure of the performance metrics as well as lower and upper limit of each range correspond with an acceptable range for the predictive power level for label. Here, Li, T. describes a lower and upper limit that relates to a range for adjusting a slider for performance metric (i.e. defining a range for predictive power level).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli, Hu, Bucklin, and Oneto to incorporate into the teachings of Li, T. because the references teach prediction models with graphic user interfaces and slider icons to retrieve inference goals from user inputs for prediction task in labeling.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli and Hu of inference goal control for a task with Bucklin’s interface design for user inputs, and Oneto’s framework of shared body, along with Li. T’s slider interface to provide a system that is “generalizable to different models and dataset, and efficiently enhances the interpretability of a given model” (Li. T. in page 3, section 3.3, left column, third paragraph).
Claim 13:
Regarding claim 13, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, teach the limitations of claim 12. Further, claim 13, comprises of similar additional limitations as claim 6, and is rejected under the same rationale.
Claims 7 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Settipalli, in view of Hu, further in view of Bucklin, further in view of Oneto, further in view of Li, T., and further in view of Perrone, V. et al., (published patent: US11481659B1), published on October 25, 2022, (hereafter, Perrone).
Claim 7:
Regarding claim 7, Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, further in view of Li, T., teaches the limitations in claim 6.
However, Settipalli in view of Hu, did not teach
“instantiating, in the graphical user interface and to obtain an updated graphical user interface, a performance indicator along the path, the performance indicator indicating an actual predictive power level for the label by the trained multiheaded inference model,”
“obtaining, via the updated graphical user interface, second user input from the user, the second user input indicating a level of acceptability of the actual predictive power level for the label,”
“in a first instance of the second user input indicating that the level of acceptability is accepted by the user: determining that the trained multiheaded inference model meets a predetermined criterion;”
“and in a second instance of the second user input indicating that the level of acceptability is not accepted by the user: determining that the trained multiheaded inference model fails to meet a predetermined criterion.”
In an analogous field, Bucklin teaches “instantiating, in the graphical user interface and to obtain an updated graphical user interface, a performance indicator along the path, the performance indicator indicating an actual predictive power level for the label by the trained multiheaded inference model,”
(See Bucklin in paragraphs [0086-0087] mentions the system units “348 and 336 may provide several input elements that can modify the display, allowing the user to focus on information of particular interest. For instance, the interactive element 343a may allow the user to revise the prediction threshold. The prediction threshold may be the dividing line for interpreting results in binary classification models. The system may use a default threshold of 0.5 (e.g., every prediction above this dividing line has a positive class label). However, this threshold can be revised by the user... To recalculate and surface fairness scores, the system may receive a revised prediction threshold from the user and may resolve the dataset imbalance. All fairness metrics (except prediction balance) may use the model's prediction threshold when calculating fairness scores. Changing this value may cause the system to recalculate the fairness scores and update the chart to display the new values.” Here, the system receiving a revised threshold for model prediction according to the user corresponds to instantiating in the graphic user interface to get an updated graphic user interface. The interface system gets updated based on the user setting a value for a metric, in this case, is a threshold value. Here, Bucklin teaches having two graphical visual indicators that take two different user inputs to get an updated graphical user interface with a performance indicator for each task, and the performance indicator contain a metric that corresponds to the predictive power level for the labels.
Further, see Bucklin in paragraph [0029] describing “FIGS. 3 A- 3 P illustrate different graphical user interfaces displayed within a bias and fairness evaluation system in accordance with various embodiments. Bucklin in figure 3J illustrates an example of a graphic user interface that gets updated based on user inputs of prediction threshold, and metric scores across various feature variables. For additional information, see Bucklin in paragraphs [0138, 0245].
PNG
media_image4.png
438
674
media_image4.png
Greyscale
Further, Bucklin in paragraphs [0076-0077] describes “the system may use various fairness metric evaluation methods to calculate a fairness value or a fairness score for the model with respect to the feature to be protected. [0077] At step 108, the system may generate a notification indicating the performance of the model responsive to a comparison of the value for the metric with the threshold.” This notification is considered a performance indicator for a prediction of the trained inference model. For more information, see Bucklin in paragraphs [0013, 0046-0050].)
Further, Bucklin teaches “obtaining, via the updated graphical user interface, second user input from the user, the second user input indicating a level of acceptability of the actual predictive power level for the label,”
(See Bucklin in paragraphs [0072 –0074] where “When the system determines that the user has interacted with the input element 314, the system may display the drop-down menu 320 allowing the user to select a fairness metric. Alternatively, in embodiments where the user is not experienced in bias and fairness evaluations, the system may display a series of interactive interfaces that can help the user identify their desired fairness metric. The system may first direct the user (e.g., display a pop-up window or direct the user to a new page) to page 322. The page 322 requests that the user selects whether the user is interested in evaluating the model based on equal error or equal representation. The system may also display a description of each method for the user. [0073] When the system determines that the user has selected one of the options (e.g., the user has selected ‘equal error’ in the depicted embodiment), the system displays the second questions, as depicted on page 324. On the page 324, the system asks the user whether the favorable target outcome occurs for a very small percentage of the population. On the next page (page 326), the system displays another question (do you want to ensure the favorable outcome for an equal number of the equal relative percentage of rows for each ‘protected class.’ The system may display more/different questions that depicted on page 324 (or other figures). The system may retrieve a list of questions to be presented to the user from a pre-generated list of questions. Therefore, the depicted questions do not represent an exhaustive list of questions.
[0074] After displaying the iterative questions, the system may select a suitable fairness metric for the user. The system may apply a set of pre-generated rules to the responses received from the user and may recommend a fairness metric.” Later, Bucklin in paragraph [0075] describes “The fairness threshold may be used by the system to measure if a model performs within appropriate fairness bounds for each protected class and 15 does not affect the fairness score or performance of any protected class. In a non-limiting example of evaluating based on “gender,” if men receive a favorable outcome 60% of the time and women receive a favorable outcome 40% of the time, the system would scale and normalize the 60% of favorable outcome to 1. The system may then determine the women’s outcome in accordance with the scaled results for men. The fairness threshold may indicate a desired ratio of the normalized and scaled values for men and women.” Here, Bucklin shows the second user input to indicate a percentage of a favorable outcome or actual predictive power in labels).
Further, Bucklin also teaches “in a first instance of the second user input indicating that the level of acceptability is accepted by the user, determining that the trained multiheaded inference model meets a predetermined criterion”,
(See Bucklin in [0060-0061] describe “The input element 314 requests the user to input a primary fairness metric. A fairness metric may refer to a metric that indicates how fair or biased a model is behaving towards a particular feature (protected feature).
Fairness metrics may refer to statistical measures of parity constraints used to assess the fairness of a model. The system may calculate the fairness metric in two steps. First, the system may calculate a fairness score for each protected class of the model's protected feature (e.g., the feature received from the user)”, where Bucklin shows user input to enter a fairness metric to later be calculated as a fairness score. See additional details from Bucklin in [0075].
Further, see Bucklin in [0092-0097] describing “the system may use various visual methods to provide additional information regarding the points/features within the chart 356… In some embodiments, the training dataset may be separated into two sub-datasets based on the classes of the protected feature the user chooses. Then the system may calculate the PSI between all the features in these two datasets.
[0093] The points displayed within the chart 356 may be interactive. For instance, when the system identifies that the user has interacted with the point 358, the user may display the window 360 that indicates the importance value and data disparity value for that particular point. The chart 356 displays insights as to why a model is biased.
… When the system identifies that the user has interacted with the dropdown menu 364, the system displays the options 366 that correspond to a list of features used by the model.
[0095] Based on the user’s selection, the system may display the chart 368. For instance, the system determines that the user has selected “hours per week” worked by each individual within the dataset 200. As a result, the system displays the chart 368 with an X-axis based on different bins corresponding to different ranges of the “hours per week” associated with different individuals and the Y-axis corresponding to the percentage of records (both men and women) who fall within the binned ranges. …The page 370 may include a cross-class accuracy table (table 372) that depicts the model's accuracy performance for each protected class. The system may revise the table 372 if the user changes the protected feature using the dropdown 374. The table 372 identifies various accuracy metrics when the data is partitioned based on the protected feature. Therefore, the table 372 depicts how accurate the evaluated model is when predicting the target value while accounting for gender. The user can also use the input elements depicted herein to change the prediction threshold”). This shows that Bucklin describes a system where the second user input’s “importance value” (from [0093]) of features (i.e. is acceptable), can determine acceptability or how well the multi-headed inference model is performing with the “model's accuracy performance” (from paragraph [0097]. See Bucklin describe in [0123] “The computing environment 560 can provide resource pooling to serve multiple users via clients 565 through a multi-tenant environment or multi-tenant model with different physical and virtual resources dynamically assigned and reassigned responsive to different demands within the respective environment” about a multi-tenant model or multi-task inference model.
Further, Bucklin also teaches “in a second instance of the second user input indicating that the level of acceptability is not accepted by the user: determining that the trained multiheaded inference model fails to meet a predetermined criterion,”
See Bucklin in paragraphs [0103-0105], describes a situation where “in some configurations, the system may not allow the user to change the prediction threshold for a deployment. In some embodiments, the above metrics and thresholds may only be displayed within the page 400 while the user may change them using other input elements (e.g., input elements via the setting tab where those input elements may be clickable through and direct the user to the page 400). [0104] When the system receives an instruction from the user, the system displays the chart 406 that shows model performance for the time frame indicated using the input element 402 (similar to the visualizations depicted in FIGS. 3I-P). [0105] The system may revise the chart 406 in accordance with inputs received from the user. For instance, when the user revises the timeframe (410 depicted in page 408), the system displays the chart 412. As depicted, in the revised time frame, the model was not fair for gender or age bracket. However, in the time frame depicted in FIG. 4A, the model was fair for gender and not for age bracket.” This disclosure shows that the interface updates based on user inputs (which includes a second user input) to determine if it is fair or not fair (i.e. acceptable or not acceptable by a user) for different variables. See Bucklin in paragraph [0106] for more information.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli and Hu, and incorporate into the teachings of Bucklin because the references teach using inference models to retrieve inference goal control from user inputs with graphic user interfaces and a performance indicator that the user adjusts to also define an actual predictive power level for the label for the prediction task.
One of ordinary skill in the art would be motivated to do so because using the methods of Settipalli and Hu of inference goal control for a task with Bucklin’s slider interface providing user inputs to “systematically and cost- effectively evaluate the space of potential predictive modeling techniques for prediction problems. This technical solution can utilize statistical learning techniques to systematically and cost-effectively evaluate the space of potential predictive modeling solutions for prediction problems” (Bucklin, [0132]).
However, Settipalli in view of Hu, further in view of Bucklin, further in view of Oneto, and further in view of Li, T., did not teach
“in a first instance of the second user input indicating that the level of acceptability is accepted by the user: determining that the trained multiheaded inference model meets a predetermined criterion;”
“and in a second instance of the second user input indicating that the level of acceptability is not accepted by the user: determining that the trained multiheaded inference model fails to meet a predetermined criterion.”
In an analogous art, Perrone teaches “in a first instance of the second user input indicating that the level of acceptability is accepted by the user: determining that the trained multiheaded inference model meets a predetermined criterion;”
See Perrone in col. 9, lines 24-36 describe "In at least some embodiments, fairness constraint(s) may be user-defined. For example, a bias or fairness threshold range or maximum specification of risk for a particular type machine learning model application (e.g., classification labels) may be specified in model job tuning request 520 in some embodiments. In some embodiments, machine learning service 510 may implement validation or other approval techniques before accepting and tuning a machine learning model using a user-defined fairness constraint (e.g., by reviewing how respective attributes are treated by the fairness constraint to ensure that sensitive attributes are not intentionally (or unintentionally) skewed in an unfair manner by the constraint)". Perrone mentions that a user defines a fairness constraint, which is a level of acceptability for a particular feature variable.
Further, see Perrone describe in col. 6, lines 18-24 “A fairness constraint may be a threshold, range, or other criteria corresponding to a fairness definition, such as those examples discussed above, which may be evaluated according to a corresponding measure of bias for a model (e.g., satisfied constraint or not satisfied, minimizing a bias measure, etc.)” Here, Perrone mentions that the fairness constraint that is defined by the user includes a threshold that corresponds to a fairness definition, and can be evaluated as either satisfying constraint (i.e. meets a predetermined criteria) or not satisfy constraint.
Further, Perrone teaches “and in a second instance of the second user input indicating that the level of acceptability is not accepted by the user: determining that the trained multiheaded inference model fails to meet a predetermined criterion.”
See Perrone describe in col. 6, lines 18-24 “A fairness constraint may be a threshold, range, or other criteria corresponding to a fairness definition, such as those examples discussed above, which may be evaluated according to a corresponding measure of bias for a model (e.g., satisfied constraint or not satisfied, minimizing a bias measure, etc.)” Here, Perrone mentions that the fairness constraint that is defined by the user includes a threshold that corresponds to a fairness definition, and can be evaluated as either satisfying constraint (i.e. meets a predetermined criteria) or not satisfy constraint ( i.e. not meet a predetermined criteria).
Later, see Perrone in col. 8 lines 23-27, lines 33-35 describe "Once the probabilistic models are updated in step 340, a determination may be made whether a stop condition for tuning has been satisfied. In some embodiments, this determination may be made using an evaluation of the determined set of metrics ... If the stop condition is not satisfied, the optimization returns to step 320 to further refine the probabilistic models. Otherwise, the optimization proceeds to step 360." Perrone shows here that if the model fails to meet a stop condition or a preset criterion, then the optimization returns to a previous step to further refine the model so the model can satisfy the condition.
Further, see Perrone in col 9, lines 24-36, describe “fairness constraint(s) may be user-defined. For example, a bias or fairness threshold range or maximum specification of risk for a particular type machine learning model application (e.g., classification labels) may be specified in model job tuning request 520 in some embodiments. In some embodiments, machine learning service 510 may implement validation or other approval techniques before accepting and tuning a machine learning model using a user-defined fairness constraint (e.g., by reviewing how respective attributes are treated by the fairness constraint to ensure that sensitive attributes are not intentionally (or unintentionally) skewed.” Here, Perrone shows that user defines the fairness constraint (i.e. level of acceptability), since this is set by the user, the user can identify if a criteria corresponding to a fairness definition from col 6, lines 18-24, is satisfied or not, and relates to level of acceptability is not accepted by the user.
See Perrone in col 5, lines 58-63 describe "Optimizer 150 may then evaluate the constraints, as indicated at 152, with respect to the generated measures 145 to update the probabilistic models 155 and may further iterate on the above steps to generate an optimized set of hyperparameters 140 for the constraints 112." Here, Perrone describes that the process of optimization of model’s hyperparameters for constraints that was defined by the user is an iterative process.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the base reference of Settipalli, Hu, Bucklin, Oneto, and Li, T., and incorporate into the teachings of Perrone because the references teach using inference models to retrieve inference goal control from user inputs with level of acceptability satisfying a preset criterion or not.
One of ordinary skill in the art would be motivated to do so because using the method of Settipalli, Hu, Bucklin, Oneto, and Li, T., of inference goal control and incorporate into the framework of Perrone to achieve a method “ where a point of the probabilistic model of training accuracy may be identified that maximizes an expected improvement in accuracy of training and feasibly meets the fairness constraint according to the probabilistic model of bias”, (see Perrone in col. 8, lines 53-57).
Claim 14:
Regarding claim 14, Settipalli in view of Hu, further in view of Bucklin, and further in view of Oneto, and further in view of Li, T. teaches the limitations of claim 13. Further, claim 14, comprises of similar additional limitations as claim 7, and is rejected under the same rationale.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to WENWEI ZENG whose telephone number is (571)272-7111. The examiner can normally be reached Monday-Friday, 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, Usmaan Saeed can be reached at (571) 272-4046. 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.
/WenWei Zeng/Examiner, Art Unit 2146
/USMAAN SAEED/Supervisory Patent Examiner, Art Unit 2146
Prosecution Insights