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 .
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.
Claim(s) 1, 2, 4, 11-14, 16, 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Leo Breiman et al. (hereinafter Breiman) (“Nonlinear Discriminant Analysis Via Scaling and ACE,” 1984), in view of Maya Gupta et al. (hereinafter Gupta) (“Deep Lattice Networks and Partial Monotonic Functions”, 09/19/2017), further in view of Choi Yong Jun et al. (hereinafter Choi) (KR 20220109258 A, 08/04/2022).
Regarding claim 1, Breiman teaches;
A method for training a [nonlinear discriminant analysis] ([Title] nonlinear discriminant analysis … [pg. 9] an efficient form of the algorithm is constructed for estimating [i.e. training] y(x) and the centers yj from the data)
receiving, ([pg. 9] estimating y(x) and the centers yj from the data) dataset ([pg. 1] the data is of the form (jn, xn), n = 1, …, N) comprising one or more features ([pg. 1] where … xn is the vector of measured variables on the [nth] case);
and training, ([pg. 9] an efficient form of the algorithm is constructed for estimating [i.e. training] y(x) and the centers yj from the data)
the ([pg. 2] study a procedure which constructs discriminant functions of the form ∑mφm(xm), where the φm are … functions … [pg. 3] The φm are not restricted to be of any fixed functional form…),
and training the ([pg. 9] estimating [i.e., training] y(x) and the centers yj from the data) defines a decision boundary ([pg. 4] Our procedure gives the boundaries graphed on Figure l [see fig. 1 below]) separating a first class of data points (circle class, see fig. 1 below) in the training dataset ([pg. 1] In a classification problem, the data is of the form (jn, xn), n = 1, …, N where jn is the class label of the nth case) from a second class of data points (star class, see fig. 1 below) in the training dataset ([pg. 1] In a classification problem, the data is of the form (jn, xn))
PNG
media_image1.png
894
751
media_image1.png
Greyscale
Breiman fails to explicitly teach but Gupta teaches;
Lattice ([Abstract] ensembles of lattices, and calibrators (piecewise linear functions) … [pg. 1] Calibrators are one-dimensional lattices)
selecting, ([pg. 8] we train a model … All 9 features are required to be monotonic) … training … using … the selected set of shape constraints ([pg.3 section 2] We also experimented with constraining all calibrators to be monotonic (even for non-monotonic inputs) for more stable/regularized training)
OBVIOUSNESS TO COMBINE GUPTA:
Gupta is analogous art to the present disclosure as it pertains to lattices and training using shape constraints.
Breiman already forms a nonlinear discriminant function from separate feature functions, y(x) = ∑mφm(xm), where each input feature xm is separately transformed by a corresponding function φm, which are combined additively, producing a more flexible decision boundary for classification. Gupta teaches a similar feature by feature transformation role using (monotonic) shape constrained one-dimensional lattice calibrator functions.
Gupta further states;
([pg. 1] if one is predicting whether to give someone else a loan, we expect and would like to constrain the prediction to be monotonically increasing with respect to the applicant’s income, if all other features are unchanged. Imposing monotonicity acts as a regularizer, improves generalization to test data, and makes the end-to-end model more interpretable, debuggable, and trustworthy … [pg. 3] We also experimented with constraining all calibrators to be monotonic (even for non-monotonic inputs) for more stable/regularized training)
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement Breiman’s respective one-dimensional feature transformations φm as Gupta’s monotonicity constrained one-dimensional lattice calibrators, because Gupta teaches that such calibrators provide constrained transformations of respective inputs and that imposing monotonicity provides more stable training, improved generalization, and improved interpretability. The modification would predictably retain Breiman’s feature-wise additive discriminant architecture while providing the known benefits of Gupta’s shape constrained one-dimensional lattice calibrators.
Accordingly, Breiman’s LDA-derived discriminant analysis architecture, modified to implement its respective feature transformations as Guptas one-dimensional lattice calibrators, teaches; lattice linear discriminant analysis model (Lattice-LDA)
Additionally, the discriminant function resulting from the combination would be the nonlinear discriminant function of Breiman (y(x) = ∑mφm(xm)) where the functions φm are the monotonicity constrained one-dimensional lattice calibrators of Gupta. The resulting nonlinear discriminant boundary is then defined by an additive combination of shape-constrained one-dimensional lattice functions, thereby generating a shape-restricted hyperplane as described by the applicant’s spec ([0005] developing a nonlinear discriminant hyperplane using an additive model of 1-D lattice functions). Accordingly, the combination of Breiman and Gupta teaches; Lattice-LDA generates a shape-restricted hyperplane
Additionally, when Breiman’s respective feature functions are implemented as Guputa’s monotonicity constrained one-dimensional lattice calibrators, the resulting lattice-LDA model is trained using both Breimn’s training data and Gupta’s selected shape constraints. Accordingly, the combination of Breiman and Gupta teaches; training, … the Lattice-LDA using the training dataset and the selected set of shape constraints, wherein:
Breiman and Gupta fail to teach but Choi teaches;
receiving, by communications hardware, a training dataset ([pg.5 3rd paragraph] The computing device 100 of the present disclosure may receive … information …, and build a learning data set based on the information)
training circuitry ([pg. 16 6th paragraph] The processor 130 performs classification by processing the training input data … of the first classification sub-model, and reflects the weights on values with incorrect predictions)
OBVIOUSNESS TO COMBINE CHOI:
Choi is analogous art to the present disclosure as it pertains to the present disclosure as it pertains to training classifiers including LDA models.
Breiman teaches the base discriminant analysis model training, Gupta teaches selecting shape constraints and utilizing shape constrained lattice functions in model training, while Choi teaches hardware capable of implementing and training various classification models such as discriminant analysis models.
It would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement the model-training method of Breiman as modified by Gupta using Choi’s computing device and associated processor, because Choi teaches using such conventional computing components to receive data for generating a classification model and to perform model learning based on training data. Such a modification would have predictably provided a computer implemented mechanism for receiving the training data and executing the model-training operations of Breiman and Gupta, including the shape constraint training operations taught by Gupta.
Regarding claim 2, Breiman fails to explicitly teach but Gupta teaches;
The method of claim 1, wherein selecting the set of shape constraints comprises: receiving
([pg.6 section 6] We present results on the same benchmark dataset (Adult) with the same monotonic features as in Canini et al. [3], and for three problems from a large internet services company where the monotonicity constraints were specified by product groups.)
NOTE: Gupta teaches receiving user input comprising a shape constraint selection (monotonicity constraints specified by product groups from the company) for a feature (monotonic features) in the training dataset (benchmark dataset, which includes training data, see table 2 below).
PNG
media_image2.png
189
770
media_image2.png
Greyscale
and selecting, ([pg. 6 section 6] the monotonicity constraints were specified by product groups)
OBVIOUSNESS:
Using the same reasoning from claim 1, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement Breiman’s respective one-dimensional feature transformations φm as Gupta’s monotonicity constrained one-dimensional lattice calibrators, because Gupta teaches that such calibrators provide constrained transformations of respective inputs and that imposing monotonicity provides more stable training, improved generalization, and improved interpretability. The modification would predictably retain Breiman’s feature-wise additive discriminant architecture while providing the known benefits of Gupta’s shape constrained one-dimensional lattice calibrators.
Gupta fails to teach but Choi teaches [in claim 1];
Communications hardware … Training circuitry
(using the same reasoning from claim 1)
OBVIOUSNESS:
Using the same reasoning from claim 1, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement the model-training method of Breiman as modified by Gupta using Choi’s computing device and associated processor, because Choi teaches using such conventional computing components to receive data for generating a classification model and to perform model learning based on training data. Such a modification would have predictably provided a computer implemented mechanism for receiving the training data and executing the model-training operations of Breiman and Gupta, including the shape constraint training operations taught by Gupta.
Regarding claim 4, Breiman teaches;
the training dataset
(using the same reasoning from claim 1)
Breiman fails to explicitly teach but Gupta teaches;
generating, ([pg. 1] Calibrators are one-dimensional lattices) based on the selected set of shape constraints ([pg. 1] calibrators, … which [are] trained discriminatively to optimize a structural risk objective and obey any given monotonicity constraints), wherein each lattice corresponds to one or more of the one or more features ([pg. 1] Calibrators are one-dimensional lattices, which nonlinearly transform a single input)
OBVIOUSNESS:
Using the same reasoning from claim 1, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement Breiman’s respective one-dimensional feature transformations φm as Gupta’s monotonicity constrained one-dimensional lattice calibrators, because Gupta teaches that such calibrators provide constrained transformations of respective inputs and that imposing monotonicity provides more stable training, improved generalization, and improved interpretability. The modification would predictably retain Breiman’s feature-wise additive discriminant architecture while providing the known benefits of Gupta’s shape constrained one-dimensional lattice calibrators.
Accordingly, using the same reasoning from claim 1, the combination of Breiman and Gupta teach;
generating, … one or more lattices based on the selected set of shape constraints (Gupta), wherein each lattice corresponds to one or more of the one or more features (Gupta) in the training dataset (Breiman), wherein the Lattice-LDA comprises the generated lattices (Breiman’s LDA based discriminant analysis model, as modified according to claim 1, comprises Gupta’s generated 1D lattice calibrators); and combining, … the generated lattices (Breiman purely additively combines the respective feature functions to form the discriminant function; after the modification of claim 1, those respective feature functions are Gupta’s generated 1D lattice calibrators) to generate the shape-restricted hyperplane (the shape-restricted hyperplane resulting from the discriminant function, using the same reasoning from claim 1)
Breiman and Gupta fail to teach but Choi teaches;
training circuitry
(using the same reasoning from claim 1)
OBVIOUSNESS:
Using the same reasoning from claim 1, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement the model-training method of Breiman as modified by Gupta using Choi’s computing device and associated processor, because Choi teaches using such conventional computing components to receive data for generating a classification model and to perform model learning based on training data. Such a modification would have predictably provided a computer implemented mechanism for receiving the training data and executing the model-training operations of Breiman and Gupta, including the shape constraint training operations taught by Gupta.
Regarding claim 11, Breiman teaches;
training an LDA based model
(using the same reasoning from claim 1)
Breiman fails to teach but Gupta teaches;
lattice
(using the same reasoning from claim 1)
OBVIOUSNESS:
Using the same reasoning from claim 1, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement Breiman’s respective one-dimensional feature transformations φm as Gupta’s monotonicity constrained one-dimensional lattice calibrators, because Gupta teaches that such calibrators provide constrained transformations of respective inputs and that imposing monotonicity provides more stable training, improved generalization, and improved interpretability. The modification would predictably retain Breiman’s feature-wise additive discriminant architecture while providing the known benefits of Gupta’s shape constrained one-dimensional lattice calibrators.
Using the same reasoning from claim 1, the combination of Breiman and Gupta yeilds; the trained Lattice-LDA
Breiman and Gupta fail to explicitly teach but Choi teaches;
Outputting, by the communications hardware, the trained [model]
PNG
media_image3.png
500
446
media_image3.png
Greyscale
([pg.9 7th paragraph] The processor 130 may read a computer program stored in the memory 120 to provide a classification model … the processor 130 may perform calculation for training the classification model.)
NOTE: The processor of the aforementioned communication hardware / computing device is capable of outputting / providing a trained classification model.
OBVIOUSNESS:
Using the same reasoning from claim 1, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement the model-training method of Breiman as modified by Gupta using Choi’s computing device and associated processor, because Choi teaches using such conventional computing components to receive data for generating a classification model and to perform model learning based on training data. Such a modification would have predictably provided a computer implemented mechanism for receiving the training data and executing the model-training operations of Breiman and Gupta, including the shape constraint training operations taught by Gupta.
Regarding claim 12, Breiman teaches;
receiving, ([pg. 4] test set of size 2000); classifying, ([pg. 2] a map y(x) [the classification rule applied using Breiman’s trained nonlinear mapping y(x), see pg. 3, 9] … put x into a class for which ||y(x)-yj||^2 is a minimum), the target data point as a first classification or a second classification ([pg. 3] we give three examples. The first is a two class problem... [pg. 4] Our procedure gives a misclassification rate of .20 on the same independent test set); and outputting, ([pg. 2] put x into a class for which ||y(x)-yj||^2 is a minimum)
Breiman and Gupta fail to teach but Choi teaches;
communications hardware ([pg.8 4th paragraph] the computing device 100 may include … the network unit ... The network unit 110 may transmit/receive data for performing a method for generating a classification model for performing classification
classifier circuitry ([pg. 16 6th paragraph] The processor 130 performs classification by processing the training input data … of the first classification sub-model, and reflects the weights on values with incorrect predictions)
OBVIOUSNESS:
Using the same reasoning from claim 1, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement the model-training method of Breiman as modified by Gupta using Choi’s computing device and associated processor, because Choi teaches using such conventional computing components to receive data for generating a classification model and to perform model learning based on training data. Such a modification would have predictably provided a computer implemented mechanism for receiving the training data and executing the model-training operations of Breiman and Gupta, including the shape constraint training operations taught by Gupta.
Regarding claim 13;
Claim 13 is a system claim that is substantially similar to method claim 1, and is rejected using the same reasoning.
Regarding claim 14;
Claim 14 is a system claim that is substantially similar to method claim 2, and is rejected using the same reasoning.
Regarding claim 16;
Claim 16 is a system claim that is substantially similar to method claim 4, and is rejected using the same reasoning.
Regarding claim 20;
Claim 20 is a computer program product claim that is substantially similar to method claim 1, with one added limitation, which Breiman and Gupta fail to explicitly teach, but Choi teaches;
the computer program product comprising at least one non-transitory computer-readable storage medium storing software instructions that, when executed, cause an apparatus to: ([pg. 20 paragraph 2] the present disclosure may be … implemented as a software module executed by hardware … A software module may include random access memory (RAM))
OBVIOUSNESS:
Using the same reasoning from claim 1, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement the model-training method of Breiman as modified by Gupta using Choi’s computing device and associated processor, because Choi teaches using such conventional computing components to receive data for generating a classification model and to perform model learning based on training data. Such a modification would have predictably provided a computer implemented mechanism for receiving the training data and executing the model-training operations of Breiman and Gupta, including the shape constraint training operations taught by Gupta.
Claim(s) 3, 15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Breiman (“Nonlinear Discriminant Analysis Via Scaling and ACE,” 1984), in view of Gupta (“Deep Lattice Networks and Partial Monotonic Functions”, 09/19/2017), further in view of Choi (KR 20220109258 A, 08/04/2022) as applied to claims 1, 13 above, further in view of Pya Natalya (Hereinafter Pya) (“Shape Constrained Additive Models”, 02/25/2014).
Regarding claim 3, Breiman teaches;
a linear shape constraint ([Abstract] linear discriminant analysis produces discriminant functions linear in x1, …, xM)
NOTE: In view of the applicant’s spec, features described by a linear model with no shape constraint at all are considered to have a linear shape constraint (“[0096] certain features with no shape constraint at all, described by a linear model for the shape constraint”).
Breiman, Gupta, and Choi fail to teach but Pya teaches;
wherein the set of shape constraints is selected from a set of candidate shape constraints, the set of candidate shape constraints comprising: …, a monotone increasing shape constraint, a monotone decreasing shape constraint, a convex shape constraint, a concave shape constraint, and a combination of two or more of the above.
([pg.546 section 2.2.2] monotonically increasing smooth can be extended to a variety of monotonic functions, including decreasing, convex/concave, increasing/decreasing and concave, increasing/ decreasing and convex, the difference between alternative shape constraints being the form of the matrices Σ and D)
OBVIOUSNESS TO COMBINE PYA:
Pya is analogous art to the present disclosure as it pertains to shape constrained additive models.
Breiman teaches discriminant analysis using an additive combination of feature-wise functions as its discriminant function, Gupta teaches shape monotonicity constrained feature-wise lattice functions, Choi teaches hardware capable of LDA training functions and communication, and Pya teaches multiple alternative shape constraints for functions of additive models.
Pya further provides motivation for utilizing different shape constraints for modeling different types of data:
([pg. 543] it is natural to assume that the relationship between a response variable and one or more predictors obeys certain shape restrictions. For example, the growth of children over time … are known to be monotonic. The relationships between … body mass index and incidence of heart disease are other examples requiring shape restrictions. Unconstrained models might be too flexible and give implausible or un-interpretable results)
Pya thus teaches that different predictor response relationships may appropriately be modeled using respective shape restrictions, and that imposing such restrictions can prevent overly flexible models from producing implausible or uninterpretable results.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to select feature specific shape constraints from the set of shape constraints taught by Pya and apply the selected shape constraints to the corresponding feature wise lattice functions of Breiman as modified by Gupta and Choi, because Pya teaches that predictor relationships may naturally obey particular shape restrictions and that imposing appropriate shape restrictions can prevent overly flexible models from producing implausible or uninterpretable results. The modification would predictably preserve Breiman’s additive discriminant architecture while providing Pya’s known benefits of shape constrained modeling.
Regarding claim 15;
Claim 15 is a system claim that is substantially similar to method claim 3 and is rejected using the same reasoning.
Claim(s) 5-10, 17-19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Breiman (“Nonlinear Discriminant Analysis Via Scaling and ACE,” 1984), in view of Gupta (“Deep Lattice Networks and Partial Monotonic Functions”, 09/19/2017), further in view of Choi (KR 20220109258 A, 08/04/2022) as applied to claims 1, 13 above, further in view of Taylor Darwin Berkeley Berg-Kirkpatrick (hereinafter Taylor) (US 20170352344 A1, 12/07/2017), further in view of Andrew Cotter et al. (hereinafter Cotter) (“Monotonic Calibrated Interpolated Look-Up Tables”, 07/16/2016).
Regarding claim 5, Breiman fails to teach but Gupta teaches;
generating the one or more lattices includes: ([pg. 3] we fix the input range for each calibrator to [amin,amax], and we fix the K keypoints a ∈ RK to be uniformly-spaced over [amin,amax])
NOTE: Gupta’s set of K keypoint locations ‘a’ correspond to the claimed candidate knot set
defining, ([pg.3 section 2] For monotonic inputs, we can constrain the calibrator functions to be monotonic by constraining the calibrator parameters b 2 [0; 1]K to be monotonic, by adding the linear inequality constraints)
OBVIOUSNESS:
Using the same reasoning from claim 1, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to implement Breiman’s respective one-dimensional feature transformations φm as Gupta’s monotonicity constrained one-dimensional lattice calibrators, because Gupta teaches that such calibrators provide constrained transformations of respective inputs and that imposing monotonicity provides more stable training, improved generalization, and improved interpretability. The modification would predictably retain Breiman’s feature-wise additive discriminant architecture while providing the known benefits of Gupta’s shape constrained one-dimensional lattice calibrators.
Breiman and Gupta fail to teach but Choi teaches;
Communications hardware
the training circuitry
(using the same reasoning and obviousness rational from claim 1)
Breiman, Gupta and Choi fail to teach but Taylor teaches;
wherein the candidate knot set is associated with a knot value vector;
([0078] The full vector of all knot heights Ξ and the full set of segment scores Ψ can be parameterized jointly as a function of the full input sequence x: (Ξ, Ψ)=h(θ, x), where h is a non-linear function parameterized by θ that maps the input x to knot heights Ξ and segment scores Ψ.)
OBVIOUSNESS TO COMBINE TAYLOR:
Taylor is analogous art as it pertains to machine learning models employing piecewise defined functions having knots and corresponding values.
Gupta teaches one-dimensional lattice calibrators having respective key points and corresponding learned output values, while Taylor teaches representing the values corresponding to a plurality of knots collectively as a full vector of knot heights and jointly parameterizing the knot values.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to associate the candidate knot set of the lattice calibrators of Breiman as modified by Gupta and Choi with a knot value vector, as taught by Taylor, by collectively representing the respective values associated with the knots as a vector. Such as modification would have provided an organized parameter representation permitting the values corresponding to the knot set to be jointly parameterized and processed during training, while each knot and its corresponding value continue to perform its known function. The modification therefore would have amounted to applying Taylor’s known vector representation of knot values to Gupta’s known set of lattice key points and corresponding values. Yielding the predictable result of a candidate knot set associated with a knot value vector.
Breiman, Gupta, Choi, and Taylor fail to teach but Cotter teaches;
and optimizing, ([pg.20 section 7.1] Cd changepoint locations (also called knots) … The changepoint values are then optimized jointly with the lattice parameters, detailed in Section 9.3.) [knot value vector taught by Taylor above] within constraints of the defined constraint function for the corresponding lattice to obtain an optimized knot value vector ([pg. 26 section 9.3] the parameters α(j)[d] are the values of the calibration function at the knots of the piecewise linear function … each row of ˜A similarly specifies a monotonicity constraint for a pair of adjacent calibration parameters for one of the piecewise linear calibration functions)
OBVIOUSNESS TO COMBINE COTTER:
Cotter is analogous art to the present disclosure as it pertains to training machine learning models employing lattice function subject to shape constraints.
Gupta teaches feature-wise one-dimensional lattice calibrators whose parameters are trained subject to monotonicity constraints, while Cotter teaches a closely related implementation in which knot locations define piecewise linear calibration functions, the corresponding knot values are jointly optimized with the lattice parameters, and the calibration parameter values at the knots are optimized subject to monotonicity constraints.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to optimize the knot value parameters of the feature wise lattice calibration functions of Breiman as modified by Gupta, Choi, and Taylor according to Cotters constrained optimization technique, such that the knot value vector is optimized subject to the selected shape constraints. The modification would have predictably provided Gupta’s feature wise lattice functions with optimized knot values satisfying the selected shape constraints, while preserving their one-dimensional lattice structure and Breimans’ additive discriminant architecture.
Regarding claim 6, Breiman and Gupta fail to teach but Choi teaches;
Training circuitry
(Using the same reasoning and obviousness rational from claim 1)
Breiman, Gupta, Choi, and Taylor fail to teach but Cotter teaches;
The method of claim 5, wherein selecting the candidate knot set includes: identifying,
([pg.20 section 7.1] we treat the number of changepoints Cd for the dth feature as a hyperparameter, and fix the Cd changepoint locations (also called knots) at equally-spaced quantiles of the feature values. The changepoint values are then optimized jointly with the lattice parameters, detailed in Section 9.3.)
OBVIOUSNESS:
Cotter explains;
([pg. 20 section 7.1] To simplify estimating the parameters, we treat the number of changepoints Cd for the dth feature as a hyperparameter, and fix the Cd changepoint locations (also called knots) at equally-spaced quantiles of the feature values)
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to select the candidate knot set of feature wise lattice functions of Breiman as modified by Gupta, Choi, and Taylor according to Cotter’s quantile based knot selection technique, by receiving an identified number Cd of quantiles and selecting the Cd knot locations at those quantiles of the corresponding feature values. A person of ordinary skill would have been motivated to make this modification because Cotter expressly teaches that fixing the knot locations in this manner simplifies parameter estimation. The modification would predictably provide a feature based set of knot locations for the lattice functions while preserving Gupta’s feature wise lattice structure and Breiman’s additive discriminant architecture.
Regarding claim 7, Breiman and Gupta fail to teach but Choi teaches;
Communications hardware … training circuitry
(Using the same reasoning and obviousness rational as claim 1)
Breiman, Gupta, Choi, and Taylor fail to teach but Cotter teaches;
The method of claim 5, wherein selecting the candidate knot set includes: receiving,
([pg.20 section 7.1] we treat the number of changepoints Cd for the dth feature as a hyperparameter, and fix the Cd changepoint locations (also called knots) at equally-spaced quantiles of the feature values.)
OBVIOUSNESS:
Using the same reasoning from claim 6, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to select the candidate knot set of feature wise lattice functions of Breiman as modified by Gupta, Choi, and Taylor according to Cotter’s quantile based knot selection technique, by receiving an identified number Cd of quantiles and selecting the Cd knot locations at those quantiles of the corresponding feature values. A person of ordinary skill would have been motivated to make this modification because Cotter expressly teaches that fixing the knot locations in this manner simplifies parameter estimation. The modification would predictably provide a feature based set of knot locations for the lattice functions while preserving Gupta’s feature wise lattice structure and Breiman’s additive discriminant architecture.
Regarding Claim 8 Gupta fails to teach but Choi teaches,
Communications hardware
(using the same reasoning and obvious rational from claim 1)
Gupta, Choi, and Taylor fail to teach but Cotter further teaches;
The method of claim 5, wherein selecting the candidate knot set includes: receiving,
([pg.20 section 7.1] we treat the number of changepoints Cd for the dth feature as a hyperparameter, and fix the Cd changepoint locations (also called knots) at equally-spaced quantiles of the feature values.)
NOTE: The number of knots (changepoints) is a hyperparameter. A hyperparameter is a user defined parameter.
OBVIOUSNESS:
Using the same reasoning from claim 6, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to select the candidate knot set of feature wise lattice functions of Breiman as modified by Gupta, Choi, and Taylor according to Cotter’s quantile based knot selection technique, by receiving an identified number Cd of quantiles and selecting the Cd knot locations at those quantiles of the corresponding feature values. A person of ordinary skill would have been motivated to make this modification because Cotter expressly teaches that fixing the knot locations in this manner simplifies parameter estimation. The modification would predictably provide a feature based set of knot locations for the lattice functions while preserving Gupta’s feature wise lattice structure and Breiman’s additive discriminant architecture.
Regarding claim 9, Breiman fails to teach but Gupta teaches;
by an Adaptive Moment Estimation (Adam) stochastic gradient descent algorithm ([pg. 6] We use the ADAM optimizer [16] and batched stochastic gradients to update model parameters)
OBVIOUSNESS: It would have been obvious to one of ordinary skill in the art, before the effective filing date, to utilize Gupta’s ADAM stochastic gradient optimization technique when optimizing the knot-value patameters of the lattice functions of Breiman as modified by Gupta and Cotter. Gupta expressly teaches using the ADAM optimizer with batched stochastic gradients to update parameters of monotonic lattice and calibrator models. Accordingly, applying Gupta’s known ADAM optimization technique to Cotter’s optimization of the corresponding knot value parameters would have constituted the use of a known optimization technique for the same type of trainable lattice parameters, yielding the predictable result of iteratively optimizing those knot values during training.
Breiman, Gupta, and Choi fail to teach but Taylor teaches;
knot value vector
(using the same reasoning and obviousness rational from claim 5)
Breiman, Gupta, Choi, and Taylor fail to teach but Cotter teaches;
minimizing … an objective function (see below)
[pg. 26] objective:
PNG
media_image4.png
124
1111
media_image4.png
Greyscale
an objective function (see above) that depends on each knot value ([pg. 26] parameters α(j)[d] are the values of the calibration function at the knots of the piecewise linear function) ([pg. 26] each row of ˜A similarly specifies a monotonicity constraint for a pair of adjacent calibration parameters for one of the piecewise linear calibration functions)
OBVIOUSNESS:
Using the same reasoning from claim 5, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to optimize the knot value parameters of the feature wise lattice calibration functions of Breiman as modified by Gupta according to Cotters constrained optimization technique, such that the knot value vector is optimized subject to the selected shape constraints. The modification would have predictably provided Gupta’s feature wise lattice functions with optimized knot values satisfying the selected shape constraints, while preserving their one-dimensional lattice structure and Breiman’s additive discriminant architecture.
Regarding claim 10, Breiman fails to tech but Gupta teaches;
The method of claim 9, further comprising: applying, ([pg.6 section 5] After each gradient update, we project parameters to satisfy their monotonicity)
OBVIOUSNESS:
Gupta teaches training shape-constrained lattice functions using the ADAM optimizer and batched stochastic gradient updates and, following each gradient update, projecting the updated parameters to satisfy their monotonicity constraints.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date, to apply Gupta’s shape-projections technique at each iteration of the stochastic gradient optimization of the constrained knot value parameters of Breiman as modified by Gupta, Choi, Taylor, and Cotter. Such a modification would predictably ensure that parameter updates produced during iterative optimization continue to satisfy the selected shape constraints, thereby maintaining the desired shape-constrained lattice function throughout training.
Gupta fails to teach but Choi teaches;
training circuitry
(using the same reasoning and obvious rational from claim 10)
Regarding claim 17;
Claim 17 is a system claim that is substantially similar to method claim 5 and is rejected using the same reasoning.
Regarding claim 18;
Claim 18 is a system claim that is substantially similar to method claim 6 and is rejected using the same reasoning.
Regarding claim 19;
Claim 19 is a system claim that is substantially similar to method claim 7 and is rejected using the same reasoning.
Response to Arguments
Applicant’s arguments, see page 1, filed 06/22/2026, with respect to 35 U.S.C. § 101 rejections have been fully considered and are persuasive. The 101 rejections of claims 9 and 20 have been withdrawn.
Applicant's arguments filed 06/22/2026 regarding the 35 U.S.C. § 103 rejections have been fully considered but they are not persuasive. Starting on page 2, the applicant remarks that; “The applied references (i.e., Gupta and Choi), singly or in any proper combination, do not disclose and would not have rendered obvious "[a] Lattice-LDA compris[ing] a discriminant function comprising a plurality of nonlinear functions combined purely additively," as recited in independent claim 1.” Applicant’s arguments are not persuasive. Amended claims 1, 13, and 20 are being rejected by Breiman in view of Gupta and Choi, as necessitated by the amendments. Breiman teaches a discriminant function comprising a plurality of feature wise functions combined purely additively, and Gupta teaches nonlinear feature wise lattice functions, where the combination of Breiman and Gupta teaches a Lattice-LDA comprising a discriminant function comprising a plurality of nonlinear functions combined purely additively (using the same reasoning from claim 1 of the current office action). The applicant further remarks; “Additionally, claim 1 recites a "shape-restricted hyperplane that defines a decision boundary." Choi discloses only the general concept of a hyperplane common to traditional LDA, and nothing in Choi or Gupta provides a teaching or motivation to develop the claimed shape- restricted hyperplane based on the general concept of an LDA hyperplane and the notion of shape restrictions from Gupta (Gupta has no teaching of how to apply the disclosed shape restrictions, which apply specifically in its deep lattice network framework, to an LDA analysis). The combination of Gupta and Choi do not disclose or render obvious the limitations of claims 1. The other applied references fail to cure these defects of Gupta and Choi. Accordingly, independent claim 1 is allowable over the applied art.” Applicant’s arguments are not persuasive. Amended claims 1, 13, and 20 are being rejected by Breiman in view of Gupta and Choi, as necessitated by the amendments. The rejection does not rely on Gupta alone to expressly instruct that its shape restricted lattice calibrators be incorporated into an LDA model. Rather, Breiman already teaches an LDA derived discriminant analysis architecture in which respective features are transformed by corresponding functions, and Gupta teaches shape constrained, feature wise lattice functions that would have been obvious to one of ordinary skill in the art, before the effective filing date, to be utilized in the discriminant analysis architecture of Breiman, as taught in the 103 rejection of amended claim 1 of the current office action. Accordingly (as reflected by the current office action), the discriminant function resulting from the combination of Breiman and Gupta would be the nonlinear discriminant function of Breiman (y(x) = ∑mφm(xm)) where the functions φm are the shape constrained one-dimensional lattice calibrators of Gupta. The resulting nonlinear discriminant boundary is then defined by an additive combination of shape-constrained one-dimensional lattice functions, thereby generating a shape-restricted hyperplane that defines a decision boundary as described by the applicant’s specification ([0005] developing a nonlinear discriminant hyperplane using an additive model of 1-D lattice functions).
Thus, the 103 rejections of the present office action for independent claims 1, 13 and 20, and associated dependent claims 2-12, 14-19, stand.
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 Matthew Alan Cady whose telephone number is (571) 272-7229. The examiner can normally be reached Monday - Friday, 7:30 am - 5:00 pm ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Cesar Paula can be reached on (571)272-4128. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC)
at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/MATTHEW ALAN CADY/ Examiner, Art Unit 2145
/CESAR B PAULA/ Supervisory Patent Examiner, Art Unit 2145
Prosecution Insights