Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Remarks
Claim Rejections – 35 U.S.C. 101
Applicant’s amendments have been fully considered and they are persuasive.
The rejection of claims 1-20 under 35 U.S.C. 101 has been withdrawn.
The foregoing applies to all independent claims and their dependent claims.
Claim Rejections – 35 U.S.C. 103
Applicant’s prior art arguments have been fully considered and they are persuasive.
Applicant argues (pgs. 9-10) that that Krasanakis adapts training sample weights, which are technologically different from hyperparameters. Applicant argues that “dynamically reweighting individual data samples (Krasanakis) provides no teaching or suggestion for dynamically updating evaluation metrics used to navigate a hyperparameter search space”. Applicant further argues that the cited references do not teach the amendments clarifying “allocating an initial computational training budget to each of the candidate combinations of hyperparameters; pruning a subset of the candidate combinations of hyperparameters based on the updated relative weighting; and allocating an increased computational training budget to a remaining subset of the candidate combinations of hyperparameters for a subsequent iteration”.
Examiner agrees. Accordingly, a new reference, Li et al. (“Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization”), has been added to the rejection, as further detailed below.
The foregoing applies to all independent claims and their dependent claims.
Claim Rejections – 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 3-5, 7-18 are rejected under 35 U.S.C. 103 as being unpatentable over Hu et al. (“FairNN - Conjoint Learning of Fair Representations for Fair Decisions”) hereinafter known as Hu in view of Li et al. (“Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization”) hereinafter known as Li in view of Krasanakis et al. (“Adaptive Sensitive Reweighting to Mitigate Bias in Fairness-aware Classification”) hereinafter known as Krasanakis.
Regarding independent claim 1, Hu teaches:
A method, comprising: receiving a fairness evaluation metric for evaluating a fairness of a machine learning model to be trained; (Hu [Page 5, Paragraph 3]: “In this work, we employ Equalized Odds … Eq.Odds ∈ [0, 2], with 0 indicating no discrimination and 2 indicating maximum discrimination” Hu teaches a fairness evaluation metric called equalized odds, where the objective is to minimize the metric, as a lower score means lower discrimination.)
receiving a performance metric for evaluating performance of the machine learning model to be trained; (Hu [Page 8, Paragraph 2]: “The Binary Cross Entropy is used as loss function to train the classifier … where cb is the true label and ˙cb is the predicted probability of the data point b having the label cb” Hu teaches the binary cross entropy as the performance metric to evaluate the performance of the machine learning model.)
…
…
…
…
…
and using the selected hyperparameter combination to train the machine learning model to balance the fairness of the machine learning model and the performance of the machine learning model. (Hu [Page 10, Paragraph 1]: “We train the auto-encoder and classifier simultaneously by minimizing the objective function Eq. 8. In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5].” Hu teaches training the machine learning model using the hyperparameter combination, of α – β, objective function.)
Hu does not explicitly teach:
automatically evaluating candidate combinations of hyperparameters of the machine learning model, the hyperparameters being different from model weights of the machine learning model and the hyperparameters including categorical hyperparameters, based at least in part on multi-objective optimization including scalarization and using the fairness evaluation metric and the performance metric to select a hyperparameter combination to utilize among the candidate combinations of hyperparameters, wherein evaluating the candidate combinations of hyperparameters of the machine learning model includes automatically and dynamically determining an updated relative weighting between the fairness evaluation metric and the performance metric for a current iteration, the updated relative weighting being determined based at least in part on an average fairness and an average performance of a set of candidate combinations of hyperparameters evaluated in one or more previous iterations;
allocating an initial computational training budget to each of the candidate combinations of hyperparameters;
automatically and dynamically determining an updated relative weighting between the fairness evaluation metric and the performance metric for a current iteration, the updated relative weighting being determined based at least in part on an average fairness and an average performance of a set of candidate combinations of hyperparameters evaluated in one or more previous iterations;
pruning a subset of the candidate combinations of hyperparameters based on the updated relative weighting;
and allocating an increased computational training budget to a remaining subset of the candidate combinations of hyperparameters for a subsequent iteration.
However, Li teaches:
automatically evaluating candidate combinations of hyperparameters of the machine learning model, the hyperparameters being different from model weights of the machine learning model and the hyperparameters including categorical hyperparameters, based at least in part on multi-objective optimization including scalarization and using the fairness evaluation metric and the performance metric to select a hyperparameter combination to utilize among the candidate combinations of hyperparameters, wherein evaluating the candidate combinations of hyperparameters of the machine learning model includes: (Li [Page 8, Algorithm 1]: “Configuration with the smallest intermediate loss seen so far.” Li teaches that the algorithm evaluates and finds the configuration of combination of hyperparameters with the smallest intermediate loss. Note that the combination with Hu provides the fairness and performance metric, as that is the loss function from Hu.)
allocating an initial computational training budget to each of the candidate combinations of hyperparameters; (Li [Page 6, Section 3.1]: “uniformly allocate a budget to a set of hyperparameter configurations, evaluate the performance of all configurations, throw out the worst half, and repeat until one configuration remains” Li teaches that the algorithm allocates a budget to a set of hyperparameter configurations.)
…
pruning a subset of the candidate combinations of hyperparameters based on the updated relative weighting; (Li [Page 6, Section 3.1]: “uniformly allocate a budget to a set of hyperparameter configurations, evaluate the performance of all configurations, throw out the worst half, and repeat until one configuration remains” Li teaches that the algorithm throws out, or prunes out, the worst half repeatedly until one configuration remains.)
and allocating an increased computational training budget to a remaining subset of the candidate combinations of hyperparameters for a subsequent iteration. (Li [Page 6, Section 3.1]: “The algorithm allocates exponentially more resources to more promising configurations” Li teaches that the algorithm allocates exponentially more resources in later iterations after it is determined that it is a more promising configuration.)
Hu and Li are in the same field of endeavor as the present invention, as the references are directed to iteratively improving the allocation of multiple objectives in a training function. It would have been obvious, before the effective filing date of the claimed invention, to a person of ordinary skill in the art, to combine balancing the fairness and performance of the machine learning model with a loss function as taught in Hu with automatically, iteratively selective hyperparameter combinations with feedback such as pruning and increasing allocation depending on the combination as taught in Li. Li provides this additional functionality. As such, it would have been obvious to one of ordinary skill in the art to modify the teachings of Hu to include teachings of Li because the combination would allow for the allocation of fairness and performance of the machine learning model to be determined using a iterative automatic algorithm that rewards good combination candidates while punishing/pruning bad candidates. This has the potential benefit of efficiently and accurately selecting an optimal hyperparameter allocation that maximizes fairness and performance.
Hu and Li do not explicitly teach:
automatically and dynamically determining an updated relative weighting between the fairness evaluation metric and the performance metric for a current iteration, the updated relative weighting being determined based at least in part on an average fairness and an average performance of a set of candidate combinations of hyperparameters evaluated in one or more previous iterations;
However, Krasanakis teaches:
automatically and dynamically determining an updated relative weighting between the fairness evaluation metric and the performance metric for a current iteration, the updated relative weighting being determined based at least in part on an average fairness and an average performance of a set of candidate combinations of hyperparameters evaluated in one or more previous iterations; (Krasanakis [Page 858, Equation 10]: Krasanakis teaches a equation that tunes two attributes using one variable, α, to scalarize the attributes on a gradient. Krasanakis explains the equation in more detail in the following. Krasanakis [Page 859, Column 2, Paragraph 2]: “Fairness-aware classifiers are usually able to train towards mitigating various fairness metrics. At the same time, they need to preserve the accuracy (acc) of the base classification model as much as possible. … it can employ either linear scalarization, where a linear trade-off is set between the objectives … it is easier to tune the parameters of Eq. 10 in a linear than in a constrained space.” Krasanakis provides the context for Equation 10 – that the equation aims at balancing the fairness as well as the accuracy/efficiency of the objective function using scalarization. Krasanakis [Page 860, Column 1, Paragraph 4]: “convergence of Algorithm 1 towards optimal weights and the impact on the objective function … measure the root mean square weight edits on each iteration of Algorithm 1” Krasanakis teaches that in determining the optimal weights and the convergence toward it, an iterative process adjusts toward the optimal weighting based on the root mean square of the weights of the fairness and performance. These weights can be mapped to hyperparameters, as they are not true weights of the model – instead, they are the weights of the objective function. Krasanakis [Page 858, col. 1, paragraph 4]: “The sensitive group S and the non-sensitive group S’ can adhere to different misclassification biases, which skew error in different ways”. Krasanakis teaches categorical hyperparameters, since in equation 10, as referenced above, the weight depends on whether the samples is in S or S’.)
Krasanakis are in the same field of endeavor as the present invention, as it is directed to combining multiple objective functions into one objective function to train a neural network. It would have been obvious, before the effective filing date of the claimed invention, to a person of ordinary skill in the art, to combine combining the fairness and the performance metrics of the objective function as taught in Hu as modified by Li with doing so in an iterative way that takes into account past iteratives as taught in Krasanakis. Krasanakis provides this additional functionality. As such, it would have been obvious to one of ordinary skill in the art to modify the teachings of Hu as modified by Li to include teachings of Krasanakis because the combination would allow for the convergence of the weighting of performance and fairness metrics to be done iteratively, adjusting hyperparameters, all while taking into account both fairness and performance. This has the potential benefit of more likely ensuring that the fairness of the objective function does not get overshadowed by the performance metric.
Regarding dependent claim 3, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the fairness evaluation metric includes a measure of at least one of: group fairness. (Hu [Page 5, Paragraph 3]: “In particular, let δFPR (δFNR) be the difference in false positive rates (false negative rates, respectively) between the protected and non-protected groups … The goal of Eq.Odds is to minimize both differences” Hu teaches that the goal of the fairness evaluation metric is to minimize the difference between the protected and non-protected groups, meaning that the evaluation includes group fairness.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 4, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the performance metric includes a measure of performance of a predictive task. (Hu [Page 8, Paragraph 2]: “The Binary Cross Entropy is used as loss function to train the classifier … where cb is the true label and ˙cb is the predicted probability of the data point b having the label cb” Hu teaches the binary cross entropy as the performance metric to evaluate the performance of the machine learning model. This metric provides the measure of performance of a predictive task, namely the accuracy of the model in predicting the data points having the correct label.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 5, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the selected hyperparameter combination is included in a Pareto frontier. (Hu [Page 10, Paragraph 1]: “In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5]. Finally, α = 0.9, β = 0.2 for the Adult Census Income Dataset and α = 0.8, β = 0.4 for the Bank Marketing Dataset are selected.” Hu teaches evaluating the candidate combinations of hyperparameters and choosing a combination of hyperparameters by determining a relative weighting between the fairness evaluation metric and the performance metric, denoted β. This is done using grid search, which automatically evaluates combinations of the possible hyperparameters and selects one of the set of best ones (Pareto frontier).)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 7, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
further comprising evaluating the fairness of the machine learning model to be trained according to the fairness evaluation metric, wherein the evaluation of the fairness of the machine learning model to be trained is based on the same predictions used to evaluate the performance of the machine learning model to be trained. (Hu [Page 10, Paragraph 1]: “In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5]. Finally, α = 0.9, β = 0.2 for the Adult Census Income Dataset and α = 0.8, β = 0.4 for the Bank Marketing Dataset are selected.” Hu teaches evaluating the candidate combinations of hyperparameters and choosing a combination of hyperparameters by determining a relative weighting between the fairness evaluation metric and the performance metric, denoted β. The evaluation of the fairness of the model is based on the evaluation of the performance, as they are inextricably linked by this β coefficient in the multi-objective function.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 8, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the dynamic determination of the relative weighting between the fairness evaluation metric and the performance metric is based on a user-defined fairness-performance trade-off. (Hu [Page 10, Paragraph 1]: “In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5]. Finally, α = 0.9, β = 0.2 for the Adult Census Income Dataset and α = 0.8, β = 0.4 for the Bank Marketing Dataset are selected.” Hu teaches evaluating the candidate combinations of hyperparameters and choosing a combination of hyperparameters by determining a relative weighting between the fairness evaluation metric and the performance metric, denoted β. What value that β turns out to be is based on the user-defined terms of the fairness and performance metric terms in the function. In a word, the weight of the fairness evaluation metric relative to the performance evaluation metric is based on the characteristics of how both metrics are originally constructed, which has the effect of being a user-defined trade-off.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 9, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
further comprising outputting a sorted set of one or more machine learning models trained using the selected hyperparameter combination. (Hu [Page 10, Paragraph 1]: “In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5]. Finally, α = 0.9, β = 0.2 for the Adult Census Income Dataset.” Hu teaches that for the Adult Census Income Dataset, a set of one trained machine learning model, with hyperparameters α = 0.9, β = 0.2, is outputted.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 10, Hu and Krasanakis teach:
The method of claim 9,
Hu teaches:
wherein the set of one or more machine learning models are output to a graphical user interface including by at least one of: displaying an associated fairness and performance for each of the one or more machine learning models; (Hu [Page 11, Figure 3]: Hu teaches displaying an associated fairness and performance for each of the machine learning models that were created with different methods. This graph can be interacted with by a user.)
or displaying at least one comparison between machine learning models in the set of one or more machine learning models, wherein the machine learning models meet at least one Pareto criterion. (Hu [Page 11, Figure 3]: Hu teaches displaying an associated fairness and performance for each of the chosen (Pareto optimal) machine learning models that were created with different methods. The format is a bar graph such that the comparison between the models can be shown. This graph can be interacted with by a user, as it had to have been at one point in order to be generated.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 11, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the dynamic determination of the relative weighting between the fairness evaluation metric and the performance metric is performed automatically and does not require specific domain knowledge. (Hu [Page 10, Paragraph 1]: “In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5]. Finally, α = 0.9, β = 0.2 for the Adult Census Income Dataset and α = 0.8, β = 0.4 for the Bank Marketing Dataset are selected.” Hu teaches evaluating the candidate combinations of hyperparameters and choosing a combination of hyperparameters by determining a relative weighting between the fairness evaluation metric and the performance metric, denoted β. This is done using grid search, which automatically evaluates combinations of the possible hyperparameters and is not specific to this domain nor requires specific domain knowledge.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 12, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the dynamic determination of the relative weighting between the fairness evaluation metric and the performance metric includes guiding a search towards minimizing a difference between the average fairness and the average performance. (Hu [Page 10, Paragraph 1]: “In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5]. Finally, α = 0.9, β = 0.2 for the Adult Census Income Dataset and α = 0.8, β = 0.4 for the Bank Marketing Dataset are selected.” Hu teaches evaluating the candidate combinations of hyperparameters and choosing a combination of hyperparameters by determining a relative weighting between the fairness evaluation metric and the performance metric, denoted β. This is done using grid search, which automatically evaluates combinations of the possible hyperparameters.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 13, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the dynamic determination of the relative weighting between the fairness evaluation metric and the performance metric includes guiding a search towards regions of higher fairness if current candidate combinations of hyperparameters correspond to machine learning model performance above a first threshold and machine learning model fairness below a second threshold. (Hu [Page 13, Paragraph 2]: “perform ablation studies to evaluate how different parts influence the predictive and fairness performance of our method. In Fig. 4, α = 0 represents the outcome without KL-Divergence regularization and β = 0 without Eq.Odds regularization respectively” Hu teaches that when β = 0, then the fairness metric is not taken into consideration. In the opposite side of the spectrum, when β = 1, the determination of the weighting go to a region of higher fairness. The threshold depends on β.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 14, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the dynamic determination of the relative weighting between the fairness evaluation metric and the performance metric includes determining an updated weighting for the fairness evaluation metric and an updated weighting for the performance metric in each iteration. (Hu [Page 10, Paragraph 1]: “In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5]. Finally, α = 0.9, β = 0.2 for the Adult Census Income Dataset and α = 0.8, β = 0.4 for the Bank Marketing Dataset are selected.” Hu teaches evaluating the candidate combinations of hyperparameters and choosing a combination of hyperparameters by determining a relative weighting between the fairness evaluation metric and the performance metric, denoted β. The evaluation of the fairness of the model is updated when the weighting is changed by changing the coefficient in each iteration.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 15, Hu and Krasanakis teach:
The method of claim 14,
Hu teaches:
wherein at least one of: the updated weighting of the fairness evaluation metric at a current iteration is increased relative to that of a previous iteration if at least one of: the average fairness of the evaluated candidate combinations of hyperparameters decreased in the previous iteration, or the average performance of the evaluated candidate combinations of hyperparameters increased in the previous iteration; (Hu [Page 10, Paragraph 1]: “We train the auto-encoder and classifier simultaneously by minimizing the objective function Eq. 8. For training, we use the Adam optimization method” Hu teaches using the Adam optimization method, which is a modified form of the popular gradient descent for multiple gradients (because of multiple objectives). That the fairness metric is increased in an iteration where the previous iteration the fairness decreased is the definition of gradient descent, as the solution tends toward the direction perpendicular to the gradient.)
or the updated weighting of the performance metric at a current iteration is increased relative to that of a previous iteration if at least one of: the average fairness of the evaluated candidate combinations of hyperparameters increased in the previous iteration, or the average performance of the evaluated candidate combinations of hyperparameters decreased in the previous iteration. (Hu [Page 10, Paragraph 1]: “We train the auto-encoder and classifier simultaneously by minimizing the objective function Eq. 8. For training, we use the Adam optimization method” Hu teaches using the Adam optimization method, which is a modified form of the popular gradient descent for multiple gradients (because of multiple objectives). That the performance metric is increased in an iteration where the previous iteration the fairness decreased is the definition of gradient descent, as the solution tends toward the direction perpendicular to the gradient.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 16, Hu and Krasanakis teach:
The method of claim 14,
Hu teaches:
wherein the updated weighting of at least one of the fairness evaluation metric and the updated weighting of the performance metric at a current iteration is determined based least in part on the average fairness and the average performance of already trained hyperparameter combinations. (Hu [Page 10, Paragraph 1]: “We train the auto-encoder and classifier simultaneously by minimizing the objective function Eq. 8. For training, we use the Adam optimization method” Hu teaches using the Adam optimization method, which is a modified form of the popular gradient descent for multiple gradients (because of multiple objectives). Gradient descent in this context involves keeping an average of the direction of the hyperparameter combinations to apply the change at the end.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 17, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the dynamic determination of the relative weighting between the fairness evaluation metric and the performance metric includes determining a weighting for the fairness evaluation metric based on an associated range of values for the fairness evaluation metric and a weighting for the performance metric based on an associated range of values for the performance metric. (Hu [Page 10, Paragraph 1]: “In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5]. Finally, α = 0.9, β = 0.2 for the Adult Census Income Dataset and α = 0.8, β = 0.4 for the Bank Marketing Dataset are selected.” Hu teaches evaluating the candidate combinations of hyperparameters and choosing a combination of hyperparameters by determining a relative weighting between the fairness evaluation metric and the performance metric, denoted β. Hu teaches an associated range of values for the fairness and the performance metrics by specifying the range of β. Note that the fairness metric coefficient is β and the fairness is 1-β, so specifying β specifies both.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 18, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
wherein the use of the selected hyperparameter combination to optimize the machine learning model to be trained includes selecting the machine learning model based at least in part on the average fairness and performance of all sampled hyperparameter combinations. (Hu [Page 10, Paragraph 1]: “In order to get the best α − β combination (see Eq. (5) and 7), grid search is operated within α ∈ [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] and β ∈ [0.1, 0.2, 0.3, 0.4, 0.5]. Finally, α = 0.9, β = 0.2 for the Adult Census Income Dataset and α = 0.8, β = 0.4 for the Bank Marketing Dataset are selected.” Hu teaches evaluating the candidate combinations of hyperparameters and choosing a combination of hyperparameters by determining a relative weighting between the fairness evaluation metric and the performance metric. If the average fairness and performance of all sampled hyperparameter combinations is lower than that of a particular combination, and that no other combination has a greater score, then that particular combination can be chosen. In this case, the selected hyperparameter combination may be chosen in part based on the average of all sampled hyperparameter combinations.)
The reasons to combine are substantially similar to those of claim 1.
Claim 2, 19, 20, 21 are rejected under 35 U.S.C. 103 as being unpatentable over Hu in view of Li in view of Krasanakis in view of Denolf et al. (US 20200104715 A1) hereinafter known as Denolf.
Regarding dependent claim 2, Hu and Krasanakis teach:
The method of claim 1,
Denolf teaches:
wherein the scalarization reduces objectives of the multi- objective optimization to a single scalar output and includes a weighted lp-norm. (Denolf ¶[0040]: “Linear scalarization, g=Σw.sub.if.sub.i(x), where w.sub.i>0 is a weight associated with each objective function; and L.sub.p norm, g=∥f−z∥.sub.p, where f={f.sub.1(x), f.sub.2(x), . . . , f.sub.k(x)}, and z∈R.sup.k is a vector of ideal cost values.” Denolf teaches using a weighted lp norm in the scalarization of the weights.)
Denolf is in the same field of endeavor as the present invention, since it is directed to combining multiple objective functions into one objective function to train a neural network. It would have been obvious, before the effective filing date of the claimed invention, to a person of ordinary skill in the art, to combine combining a fairness objective function with a performance objective function as taught in Hu as modified by Krasanakis with combining different objective functions using scalarization as taught in Denolf. Denolf provides this additional functionality. As such, it would have been obvious to one of ordinary skill in the art to modify the teachings of Hu as modified by Krasanakis to include teachings of Denolf because the combination would allow for different combinations of varying relative strengths of the fairness objective and the performance objective. This has the potential benefit of choosing the multiple objective function with the relative strength of fairness to performance that results in the most accurate and fair model.
Claim 19 is substantially similar to claim 1, but has the following additional elements:
Regarding independent claim 19, Denolf teaches:
A system, comprising: a processor configured to: (Denolf ¶[0031]: “the CPU 206 can be any type of general-purpose central processing unit (CPU), such as an x86-based processor, ARM®-based processor, or the like” Denolf teaches a system comprising a processor, of different possible types.)
and a memory coupled to the processor and configured to provide the processor with instructions. (Denolf ¶[0032]: “The system memory 208 can store data 226 and program code (“code 228”) processed and executed by the CPU 206 to implement the software platform 204.” Denolf teaches memory coupled to the CPU that provides the processor with instructions to be executed.)
The reasons to combine are substantially similar to those of claim 2.
Claim 20 is substantially similar to claim 1, but has the following additional elements:
Regarding independent claim 20, Denolf teaches:
A computer program product embodied in a non-transitory computer readable medium and comprising computer instructions for: (Denolf ¶[0007]: “a non-transitory computer readable medium comprising instructions, which when executed in a computer system, causes the computer system to carry out a method of implementing a neural network includes” Denolf teaches a computer product embodied in a non-transitory computer readable medium comprising instructions.)
The reasons to combine are substantially similar to those of claim 2.
Regarding dependent claim 21, Hu and Krasanakis teach:
The method of claim 1,
Denolf teaches:
further comprising determining candidate combinations of hyperparameters of the machine learning model including by applying at least one of the following hyperparameter tuners: Random Search, Tree Parzen Estimator, or bandit-based hyperparameter tuner. (Denolf ¶[0051]: “A random search is conceptually similar to a grid search, except that a random search picks random values from a specified range for each hyperparameter, rather than selecting them from a grid.”)
The reasons to combine are substantially similar to those of claim 2.
Claims 6 is rejected under 35 U.S.C. 103 as being unpatentable over Hu in view of Li in view of Krasanakis in view of Elbsat (US 20200301408 A1) hereinafter known as Elbsat.
Regarding dependent claim 6, Hu and Krasanakis teach:
The method of claim 1,
Hu teaches:
… the updated relative weighting between the fairness evaluation metric and the performance metric includes a fairness evaluation metric weight and a performance metric weight … and sums to 1. (Hu [Page 9, Paragraph 1]: “Therefore, Equalized Odds (Eq.Odds) is used as the constraint term and added to the classification loss … where β ∈ [0, 1), is a balancing coefficient between the classification loss Lbce and the Eq.Odds fairness regularization.” Hu teaches that the terms of the multi-objective loss function are weighted such that the sum is 1.)
Hu and Krasanakis do not explicitly teach:
… that are inversely proportional …
However, Elbsat teaches:
… that are inversely proportional … (Elbsat ¶[0335]: “Due to the inverse relationship between φ.sub.1′ and COP, input weighter 1508 can generate a weighting function that assigns weights that are inversely proportional to the value of φ.sub.1′.” Elbsat teaches a weighting of a metric that is inversely proportional.)
Elbsat is in the same field as the present invention, since it is directed to weighting metrics in an inversely proportional manner. It would have been obvious, before the effective filing date of the claimed invention, to a person of ordinary skill in the art, to combine generating a multiple objective function with the both fairness and performance metrics represented in the objective function as taught in Hu as modified by Krasanakis with weighting the metrics inversely proportionally as taught in Elbsat. Elbsat provides this additional functionality. As such, it would have been obvious to one of ordinary skill in the art to modify the teachings of Hu as modified by Krasanakis to include teachings of Elbsat because the combination would allow for the weighting of the fairness evaluation metric to be inversely proportional. This has the potential benefit of weighting the fairness metric in the direction that maximizes it, rather than minimizes it, in the multi-objective function.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to KYU HYUNG HAN whose telephone number is (703) 756-5529. The examiner can normally be reached on MF 9-5.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Alexey Shmatov can be reached on (571) 270-3428. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Kyu Hyung Han/
Examiner
Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights