DETAILED ACTION
Claims 1 and 3-21 are presented for examination. Claims 1, 3, 12, 17, and 18 stand currently amended. Claim 21 is new.
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Finality of Office Action
The following is a brief summary description of new ground(s) of rejection (if any) and the reason why those new ground(s) are made necessary by this amendment:
New claim 21 is newly rejected.
This office action adds extensive discussion regarding interpretation of claim language “1D box kernel.” However, merely adding clarification on claim interpretation, even with citing additional evidentiary dictionary references to support the interpretation, does not constitute a new ground of rejection. See MPEP §706.07(a) and §1207.03 (“not including a newly relied upon dictionary definition”).
Response to Arguments
Applicant's remarks filed 20 August 2026 have been fully considered and Examiner’s response is as follows:
Applicant remarks page 8 argues:
The Office acknowledges that "Chaudhuri does not explicitly disclose a box kernel," and turns to Rackauckas, asserting that "A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions." Id. The Applicant respectfully disagrees.
This is merely a summary of Applicant’s arguments to follow. However, Examiner notes that Applicant nowhere defines what a “1D box kernel” is. Because Applicant does not dispute Examiner’s interpretation of what the metes and bounds of what a “1D box kernel” is, it is generally unpersuasive to argue that the teachings which Examiner has mapped to the box kernel are, or are not, a 1D box kernel.
For further clarification on Examiner’s claim interpretation, Examiner has added a “Claim Interpretation” section below. This claim interpretation provides useful context in discussing the details of Applicant’s arguments below.
Applicant remarks page 8 further argues:
Rackauckas does not teach or suggest a kernel. The passage the Office relies upon describes the design of a reservoir for a surrogate model, in which "A 10x10 cellular network is randomly initialized and excited with the training inputs, and a neighborhood of 3x3 was chosen." Rackauckas, [0123].
This argument is unpersuasive.
The neighborhood in a 3x3 shape corresponds with the stencil of the kernel. Compare with claim interpretation section below.
Applicant remarks page 8 further argues:
The 3 x 3 neighborhood is the coupling neighborhood of that grid. "Each cell also interacts with its neighbors via two template variables A and B," and the size of the neighborhood "controls the size of the two template variables." Rackauckas, [0123]. The 3 x3 neighborhood therefore sizes the template variables that govern coupling between the cells of a simulated circuit. It describes the topology of the modeled network, and it is not a kernel applied to a computing function.
This argument is unpersuasive.
The term ‘template’ corresponds with a stencil for the kernel.
Rackauckas paragraph 123 teaches “The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling.” The discontinuity corresponds with a discontinuity of a computing function. Thus, the Latin Hypercube sampling which defines the surrogate to control the discontinuity relates to application of the kernel to the computing function. A surrogate corresponds with the resulting model after the kernel has already been applied.
Furthermore, the template controlling the size of corresponds with a “distance” which controls the 1D box kernel.” Compare with claim 21. Accordingly, the existence of a size control being taught by Rackauckas is consistent with the scope of the claims in which the kernel is also claimed as being optionally defined by a distance in dependent claim 21.
Applicant remarks page 8 further argues:
Neither the word kernel nor any description of convolution or pre-filtering appears anywhere in Rackauckas.
This argument is unpersuasive.
Literal recitation of the verbatim claim terminology is not required for a prima facie rejection. Applicant is applying an incorrect legal standard here and therefore is unpersuasive. Furthermore, Applicant is making arguments against Rackauckas individually while Examiner’s rejection is based upon a combination of references.
Applicant remarks page 9 further argues:
A technique for choosing which sets of system parameters to simulate does not model a discontinuity, and it does not supply the 1D box kernel by which the discontinuity is modeled as recited in Claims 1, 11, and 17.
This argument is unpersuasive.
In response to applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).
Here, Examiner has cited Chaudhuri regarding modeling sharp discontinuities using Gaussian smoothing. Rackauckas is not required to cure any area of Chaudhuri where Chaudhuri is not deficient.
Chaudhuri page 3 section 2.2 second paragraph discloses “The smoothed semantics of P can now be defined as a smooth approximation of
P
in terms of
f
μ
,
σ
.” Chaudhuri page 3 right column last paragraph discloses “We use the phrase “Gaussian smoothing of P (with respect to ß)” to refer to the interpretation of P according to the smoothed semantics.” The smoothed semantics or Gaussian smoothing of P corresponds with a generated data model of the program based on the Gaussian sampling.
Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution.” A 1D distribution corresponds with a one dimensional kernel.
Chaudhuri page 4 section 2.3 second paragraph item 2 teaches “Observe, however, that YX is not required to be Gaussian.”
Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
The neighborhood shape corresponds with the 1D box shape for the neighborhood recited by the claim language “1D box kernel.”
Applicant remarks page 9 further argues:
Latin Hypercube sampling is a stratified sampling design, in which the sampled range is divided into equally probable intervals along each dimension and one sample is placed in each interval of each dimension. Applied to a three by three arrangement, it therefore selects three points occupying three different rows and three different columns, and it designates no center point from which a pair of samples is taken to either side.
This argument is unpersuasive. If there are three points in a line (1D), with equal intervals between them, then the middle point is a center point because it is in the center.
Applicant remarks page 9 further argues:
As described in the subject Application, the claimed 1D box kernel is a prefilter applied to the computing function along the sampling axis. The modeled discontinuity is generated by "differentiating the computing function ... after prefiltering the function with a 1D box kernel," and "[t]he distance between the two evaluation sites ... defines a size of the box kernel." Application, [0044]. No such pre-filter is described in either reference.
This argument is unpersuasive.
The description of Specification [0044] is language not recited in the claim. Dependent claim 21 recites “pre-filtering the computing function.” Pre-filtering does not specifically require differentiating the computing function.
Chaudhuri page 3 section 2.2 second paragraph discloses “The smoothed semantics of P can now be defined as a smooth approximation of
P
in terms of
f
μ
,
σ
.” Chaudhuri page 3 right column last paragraph discloses “We use the phrase “Gaussian smoothing of P (with respect to ß)” to refer to the interpretation of P according to the smoothed semantics.” The smoothed semantics or Gaussian smoothing of P corresponds with a generated data model of the program based on the Gaussian sampling. Smoothing of the corresponding program corresponds with a pre-filtering of the computing function of the program.
Applicant remarks page 9 further argues:
To the extent the Office looks to FIG. 10 of the subject Application to supply the reading it gives to Rackauckas, that reliance is misplaced.
Examiner does not cite Applicant’s figure 10 in the rejection of any claim.
Applicant remarks page 10 further argues:
The three by three arrangement of sample locations depicted in FIG. 10 is therefore the two dimensional alternative that the subject Application distinguishes from the claimed one dimensional kernel
Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution.” A 1D distribution corresponds with a one-dimensional kernel. Because Chaudhuri teaches 1D, Rackauckas is not required to also teach 1D. The close correspondence between the 3x3 neighborhood taught by Chaudhuri and the specific three by three arrangement of figure 10 shows that the teaching of Chaudhuri corresponds with at least the specific embodiment disclosed by Applicant’s figure 10.
Applicant remarks page 10 further argues:
The nine cells of the neighborhood described in Rackauckas are moreover nine distinct circuits, each having its own randomly initialized resistance and capacitance and its own state variable, and they are not evaluations of a computing function at locations offset from a target point along an axis. Rackauckas, [0123].
This argument is unpersuasive.
In response to applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).
Applicant remarks page 12 further argues:
In rejecting these features as originally filed, the Office asserts Chaudhuri. Office Action, p. 4. Specifically, the Office asserts that Chaudhuri page 3 section 2.2 discloses "a 1-D Gaussian distribution .... We assume that Y has K components (i.e., Y ranges over states of P), and that these components are independent variables." Id. The Office concludes that "The ID independent variables Y correspond with an identified axis," and that "The Gaussian distribution corresponds with a sampling along this axis." Id The Applicant respectfully disagrees.
The passage the Office relies upon defines Y as the state of the program under analysis. Chaudhuri states that "Y ranges over states of P." Chaudhuri, p. 3. A variable that ranges over the states of the program being analyzed is internal to that program. It is not an input received by a processing device, and it does not identify an axis.
This argument is unpersuasive.
Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution …. We assume that Y has K components (i.e., Y ranges over states of P), and that these components are independent variables.” The 1D independent variables Y, i.e. one component, correspond with an identified axis. Each component corresponds with an identified axis. Thus Chaudhuri teaches at least one identified axis. The Gaussian distribution corresponds with a sampling along this axis. See further Chaudhuri page 4 right column item 1 regarding Gaussian distribution of YX.
Applicant remarks page 12 further argues:
Claim 1 recites two distinct acts. First, receiving an input identifying an axis, and second generating a plurality of samples oriented along that axis. The Office maps the axis onto the independent variables of Chaudhuri, but identifies no disclosure in Chaudhuri of receiving an input that identifies an axis. Treating an assumption about the components of a random variable as the receipt of an axis identifying input gives no effect to the receiving act as recited in Claim 1.
This argument is unpersuasive.
Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution …. We assume that Y has K components (i.e., Y ranges over states of P), and that these components are independent variables.” These independent variables correspond with inputs of the approximation function and is thus received by the subsequent calculations.
Furthermore, the rejection is based upon a combination of references and Rackauckas teaches the respective processor which receives the data to be calculated thereon.
Applicant remarks page 13 further argues:
The stated rationale does not produce the feature for which Rackauckas is cited. Substituting Latin Hypercube sampling for Gaussian sampling in Chaudhuri would yield a different scheme for selecting the parameter sets at which a program is evaluated. It would not yield a 1D box kernel by which a discontinuity is modeled.
… and, for the reasons set out in section (I) above, it is not a kernel.
This is a repeat argument of the above argument that a 1D box kernel is allegedly not taught. Examiner has already responded to this argument above.
Applicant remarks page 14 further argues:
The reason offered for selecting a box shaped kernel is found in the claims of the subject Application, which is hindsight.
This argument is unpersuasive.
Examiner’s rejection clearly does not cite the claims or other disclosure of the instant application for motivation to combine references. Applicant is reminded of the MPEP guidance on hindsight:
Applicants may argue that the examiner’s conclusion of obviousness is based on improper hindsight reasoning. However, "[a]ny judgment on obviousness is in a sense necessarily a reconstruction based on hindsight reasoning, but so long as it takes into account only knowledge which was within the level of ordinary skill in the art at the time the claimed invention was made and does not include knowledge gleaned only from applicant’s disclosure, such a reconstruction is proper." In re McLaughlin, 443 F.2d 1392, 1395, 170 USPQ 209, 212 (CCPA 1971).
MPEP §2145(X)(A).
Claim Rejections - 35 USC § 112
Claims 3 and 12 have been appropriately corrected. Accordingly, Examiner's rejection of claims 3 and 12 under § 112 is withdrawn.
Claim Interpretation “1D Box Kernel”
Claim 1 recites “using a 1D box kernel.” Interpretation of this limitation follows the guidelines of MPEP §2111.
No Special Definition Found in Disclosure
A 1D box kernel is not explicitly defined within the Specification. Specification figures 7 and 8 provide examples:
FIG. 7 depicts an example implementation of generating the modeled discontinuity as a box kernel approximation using a 1D box kernel by the discontinuity modeling module.
FIG. 8 depicts an example implementation of generating the modeled discontinuity as a box kernel approximation using a 1D box kernel.
Figure 7 does not particularly identify which portion is the model of the modeled discontinuity and which portion represents the use of the box kernel itself (if represented at all).
Specification figure 7 shows:
PNG
media_image1.png
68
408
media_image1.png
Greyscale
and
PNG
media_image2.png
347
800
media_image2.png
Greyscale
Specification paragraph 43 states “generating the modeled discontinuity 122 as a box kernel approximation 318 using a 1D box kernel.” Neither (122) nor (318) are depicted within figure 7. Figure 3 shows (318) as a black box labeled “Box Kernel Approximation” which fails to illustrate an example of the box kernel itself.
Specification paragraph 43 further states “a corresponding gradient 710 are depicted.” The line on the right-hand-side or figure 7 corresponds with a representation of the gradient. Claim 6 recites “wherein the modeling of the discontinuity estimates one or more gradients.” Claim 1 is not limited to the particulars of any dependent claim, but even claim 6 does not state that the data model is itself a gradient approximation because the claim language of claim 6 allows for the model of the discontinuity to only be indirectly uses an estimate of one or more gradients. Accordingly, neither figure 7 nor corresponding description within the Specification provide any clear definition for the 1D box kernel.
Figure 8 appears to be a 2D version of figure 7 but figure 8 and corresponding description of Specification paragraph 45 fail to specifically define the 1D box kernel.
Therefore, Examiner finds that Applicant does not have a lexicographic definition for “1D box kernel”. Thus, the phrase “1D box kernel” is interpreted according to the plain meaning. “The plain meaning of a term means the ordinary and customary meaning given to the term by those of ordinary skill in the art at the relevant time.” MPEP §2111.01.
Ordinary and Customary Meaning
The phrase “1D box kernel” has more than one distinct definition within the art. The claims must be given their broadest reasonable interpretation consistent with the Specification. MPEP §2111. There are two usages within the art which are consistent with the claims and Specification for discussion here.
First interpretation Matlab’s “Box Kernel” function:
Mathworks, MATLAB & Simulink “Kernel Distribution” (2026) [herein “Matlab”] pages 3-4 states “You can choose one of several options for the kernel smoothing function. This plot shows the shapes of the available smoothing functions.” The plot shows a “box” name and a green dashed line “box” graph. The box graph shown therein looks similar to the gradient line approximation of the instant application’s figure 7.
The online documentation for Matlab does not establish a prior art date before the effective filing date for the instant application. However, Luengo, C. "What is Matlab's 'box' Interpolation Kernel" Stack Overflow, comment (July 2018) [herein “Luengo”] title clearly shows the existence of Matlab’s ‘box’ kernel at least as early as 2018. Luengo further describes “A ‘box’ kernel is an averaging kernel with uniform weights. If it is an interpolation kernel, then it corresponds to nearest neighbor interpolation (it always takes the average of one input sample).” Luengo in Li’s comment further discloses a “specified window size” for the box interpolation kernel.
This description is consistent with the language of being a 1D ‘box’ kernel because the kernel is explicitly described as ‘box’ and is 1D.
Second interpretation Stencil-based kernels:
Kamil, S., et al. “Implicit and Explicit Optimizations for Stencil Computations” Association for Computing Machinery, MSPC’06 (2006) [herein “Kamil”] abstract which begins “Stencil-based kernels.” A kernel being stencil-based is a species of kernel. Kamil section 1 first paragraph teaches:
Partial differential equation (PDE) solvers constitute a large fraction of scientific applications in such diverse areas as heat diffusion, electromagnetics, and fluid dynamics. These applications are often implemented using iterative finite-difference techniques, which sweep over a spatial grid, performing nearest neighbor computations called stencils. In a stencil operation, each point in a multidimensional grid is updated with weighted contributions from a subset of its neighbors in both time and space— thereby representing the coefficients of the PDE for that data element. These operations are then used to build solvers
This description is similar to the first in that both describe correspondence with nearest neighbor interpolation. Examiner views Kamil description as a generalized version of Matlab’s description in that the description of Kamil section 1 allows for a multitude of stencil shapes and any number of dimensions.
In Kamil, when the computational grid is 1D and the stencil shape is that of a ‘box’ then the stencil-based kernel may reasonably be interpreted as corresponding with a 1D box kernel. In such a case the stencil-based kernel description of Kamil matches exactly the nearest neighbor computations described in regard to Matlab’s ‘box’ kernel.
Defining “1D Box Kernel”:
From the above prior art disclosure, and consistent with the disclosure of the Specification, Examiner finds the limitation “1D Box Kernel” includes the following characteristics:
It is one-dimensional.
It corresponds with performs nearest neighbor sampling on a 1D ‘box’. While a geometric box is inherently 2D, a person of ordinary skill in the art would understand sampling a 1D box to correspond with sampling two points on either side, such as depicted in instant applicant’s figure 10(a) and 10(b).
Consistency with Dependent claim 21:
Claim 21 recites “wherein the modeling the discontinuity of the computing function includes pre-filtering the computing function with the 1D box kernel, and wherein a size of the 1D box kernel is defined by a distance between a pair of samples taken from the plurality of samples.” As discussed above, Luengo in Li’s comment further discloses a “specified window size” for the box interpolation kernel. The specified window size corresponds with a distance between a pair of samples. Accordingly, the first interpretation of Matlab’s description is consistent with and not in conflict with the recited claim language of dependent claim 21. Kamil’s description is broader than Matlab’s and thus also consistent and compatible with the further limitations of dependent claim 21.
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.
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, 3-19, and 21
Claims 1, 3-19, and 21 are rejected under 35 U.S.C. 103 as being unpatentable over Chaudhuri, S. & Solar-Lezama, A. “Smooth Interpretation” ACM SIGPLAN Notices, vol. 45, issue 6, pp. 279-291 (2010) [herein “Chaudhuri”] in view of US 2024/0086598 A1 Rackauckas, et al. [herein “Rackauckas”].
Claim 1 recites “1. A method comprising: receiving, …, a program having a computing function that includes a discontinuity.” Chaudhuri abstract discloses “Programs from many application domains make frequent use of discontinuous control flow constructs, and consequently, encode functions with highly discontinuous and irregular landscapes.”
Claim 1 further recites “by a processing device” and a plurality of times “by the processing device.” Chaudhuri abstract discloses “programs.” Chaudhuri section 1 first sentence discloses “software systems are inherently discontinuous.” Software and programs are related to computers.
But Chaudhuri does not explicitly disclose a processing device; however, in analogous art of training with surrogate models, Rackauckas paragraph 28 teaches “The computer system includes a memory and a processor.” A processor is a processing device.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use a computer into the system of smooth interpretation of program control flow constructs for art recognized suitable system for implementing software and programs including surrogates in a computing language. See Rackauckas paragraph 3 and MPEP §2144.07.
Claim 1 further recites “receiving, …, an input identifying an axis; generating, …, a plurality of samples oriented along the axis based on an output of the program.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution …. We assume that Y has K components (i.e., Y ranges over states of P), and that these components are independent variables.” The 1D independent variables Y, i.e. one component, correspond with an identified axis. Each component corresponds with an identified axis. Thus Chaudhuri teaches at least one identified axis. The Gaussian distribution corresponds with a sampling along this axis. See further Chaudhuri page 4 right column item 1 regarding Gaussian distribution of YX.
Claim 1 further recites “generating, …, a data model of the program based on the plurality of samples, the generating including modeling the discontinuity of the computing function based on the plurality of samples.” Chaudhuri page 3 section 2.2 second paragraph discloses “The smoothed semantics of P can now be defined as a smooth approximation of
P
in terms of
f
μ
,
σ
.” Chaudhuri page 3 right column last paragraph discloses “We use the phrase “Gaussian smoothing of P (with respect to ß)” to refer to the interpretation of P according to the smoothed semantics.” The smoothed semantics or Gaussian smoothing of P corresponds with a generated data model of the program based on the Gaussian sampling.
Claim 1 further recites “using a 1D box kernel.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution.” A 1D distribution corresponds with a one dimensional kernel.
Chaudhuri page 4 section 2.3 second paragraph item 2 teaches “Observe, however, that YX is not required to be Gaussian.”
Chaudhuri does not explicitly disclose a box kernel; however, in analogous art of training with surrogate models, Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side.
A 3x3 neighborhood sampling corresponds with sampling a literal box shape stencil and a nearest neighbor-type sampling. Accordingly, a 1D version corresponds with using a 1D box kernel. Compare with “Claim Interpretation” section above.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use Latin Hypercube sampling in a small neighborhood into the system of smooth interpretation of program control flow constructs for the advantageous purpose of an art recognized non-Gaussian sampling method. See Rackauckas paragraph 123 and Chaudhuri page 4 section 2.3 second paragraph item 2.
Claim 1 further recites “and outputting, …, the data model including the modeled discontinuity.” Chaudhuri page 4 right column item 2 discloses:
2. Apply the transformation
Y
X
'
=
P
Y
X
. Note that
Y
X
'
is a random variable; intuitively, it captures the output of the program P when executed on the input
Y
X
.
The transformed result of the Gaussian smoothing of P corresponds with the data model. Applying the transformation to a particular program creates and outputs the respective Gaussian smoothing approximation.
Claim 3 further recites “3. The method as described in claim 1, wherein the generating of the data model converts a sequence of operations of the computing function of the program into one or more gradients.” Chaudhuri page 3 right column section 2.2 fourth paragraph disclose “Smoothed semantics is defined not only at the level of the whole program, but also at the level of individual expressions. For example, the smoothed semantics of a boolean expression B is ….” Accordingly, the whole program is a combination of a plurality of individual expressions. Each individual expression of the whole program corresponds with a respective operation of the program.
Chaudhuri page 3 left column first paragraph discloses “Note that the sharp discontinuities and flat plateaus of Figure 1-(c) are now gone, and any point reasonably close to the quadrant … has a nonzero downhill gradient. Hence, the search method is now able to converge to a value close to the global minimum for a much larger range of initial points.” Having a downhill gradient corresponds with the data model having respective gradient(s).
Claim 4 further recites “4. The method as described in claim 1, wherein the generating the data model is performed automatically and without user intervention.” Chaudhuri page 8 left column second paragraph discloses “we aim to synthesize these values automatically.”
Claim 5 further recites “5. The method as described in claim 1, wherein the modeling the discontinuity is based on pairs of samples taken from the plurality of samples.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution.” A 1D distribution corresponds with a one dimensional kernel.
Chaudhuri page 4 section 2.3 second paragraph item 2 teaches “Observe, however, that YX is not required to be Gaussian.”
Chaudhuri does not explicitly disclose a box kernel; however, in analogous art of training with surrogate models, Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use Latin Hypercube sampling in a small neighborhood into the system of smooth interpretation of program control flow constructs for the advantageous purpose of an art recognized non-Gaussian sampling method. See Rackauckas paragraph 123 and Chaudhuri page 4 section 2.3 second paragraph item 2.
Claim 6 further recites “6. The method as described in claim 5, wherein the modeling of the discontinuity estimates one or more gradients at locations between the pairs of samples.” Chaudhuri page 3 section 2.2 fourth paragraph discloses “The function [[P]](x) is the convolution of the function [[P]] and a K-D Gaussian with mean 0 and standard deviation.” The mean of zero at x is a centering of the estimate between the respective sampling.
Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side.
Claim 7 further recites “7. The method as described in claim 6, wherein the one or more gradients are generated as one or more derivatives of the computing function.” Chaudhuri page 3 left column first paragraph discloses “Note that the sharp discontinuities and flat plateaus of Figure 1-(c) are now gone, and any point reasonably close to the quadrant … has a nonzero downhill gradient. Hence, the search method is now able to converge to a value close to the global minimum for a much larger range of initial points.” Having a downhill gradient corresponds with the data model having respective gradient(s). These gradients correspond with derivatives of the data model which is of the computing function(s).
Claim 8 further recites “8. The method as described in claim 1, wherein the discontinuity is caused by an if/else branch in the program.” Chaudhuri abstract discloses “discontinuous control flow constructs.” Chaudhuri page 2 right column second paragraph discloses “conditional branches.” Control flow conditional branches include if/else branches in the program. Chaudhuri page 4 example 2 shows an if/else part of a program.
Claim 9 further recites “9. The method as described in claim 1, further comprising executing a machine learning model using the data model.” Chaudhuri page 13 left column second paragraph discloses “The idea of using smooth approximations to improve the performance of numerical methods like gradient descent is well-known in the domain of neural networks [3].” Improving the performance of a gradient descent in the domain of neural networking corresponds with executing a machine learning model using the data model. The neural network corresponds with a machine learning model.
Claim 10 further recites “10. The method as described in claim 9, wherein the executing of the machine learning model uses the data model as part of a backpropagation operation.” Chaudhuri page 13 left column second paragraph discloses “The idea of using smooth approximations to improve the performance of numerical methods like gradient descent is well-known in the domain of neural networks [3].” Improving the performance of a gradient descent in the domain of neural networking corresponds with executing a machine learning model using the data model. The neural network corresponds with a machine learning model. Gradient descent is a backpropagation.
Claim 11 recites “11. A system comprising: an axis sampling module … to generate a plurality of samples from a program that includes a computing function having a discontinuity.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution …. We assume that Y has K components (i.e., Y ranges over states of P), and that these components are independent variables.” The 1D independent variables Y correspond with an identified axis. The Gaussian distribution corresponds with a sampling along this axis. See further Chaudhuri page 4 right column item 1 regarding Gaussian distribution of YX.
Claim 11 further recites “implemented by a processing device” and “implemented by the processing device.” Chaudhuri abstract discloses “programs.” Chaudhuri section 1 first sentence discloses “software systems are inherently discontinuous.” Software and programs are related to computers.
But Chaudhuri does not explicitly disclose a processing device; however, in analogous art of training with surrogate models, Rackauckas paragraph 28 teaches “The computer system includes a memory and a processor.” A processor is a processing device.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use a computer into the system of smooth interpretation of program control flow constructs for art recognized suitable system for implementing software and programs including surrogates in a computing language. See Rackauckas paragraph 3 and MPEP §2144.07.
Claim 11 further recites “and a data model generation module … to generate a data model, automatically and without user intervention, based on the plurality of samples as one or more gradients having the discontinuity modeled.” Chaudhuri page 3 section 2.2 second paragraph discloses “The smoothed semantics of P can now be defined as a smooth approximation of
P
in terms of
f
μ
,
σ
.” Chaudhuri page 3 right column last paragraph discloses “We use the phrase “Gaussian smoothing of P (with respect to ß)” to refer to the interpretation of P according to the smoothed semantics.” The smoothed semantics or Gaussian smoothing of P corresponds with a generated data model of the program based on the Gaussian sampling.
Chaudhuri page 8 left column second paragraph discloses “we aim to synthesize these values automatically.”
Chaudhuri page 3 left column first paragraph discloses “Note that the sharp discontinuities and flat plateaus of Figure 1-(c) are now gone, and any point reasonably close to the quadrant … has a nonzero downhill gradient. Hence, the search method is now able to converge to a value close to the global minimum for a much larger range of initial points.” Having a downhill gradient corresponds with the data model having respective gradient(s).
Claim 11 further recites “using a 1D box kernel.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution.” A 1D distribution corresponds with a one dimensional kernel.
Chaudhuri page 4 section 2.3 second paragraph item 2 teaches “Observe, however, that YX is not required to be Gaussian.”
Chaudhuri does not explicitly disclose a box kernel; however, in analogous art of training with surrogate models, Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side.
A 3x3 neighborhood sampling corresponds with sampling a literal box shape stencil and a nearest neighbor-type sampling. Accordingly, a 1D version corresponds with using a 1D box kernel. Compare with “Claim Interpretation” section above.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use Latin Hypercube sampling in a small neighborhood into the system of smooth interpretation of program control flow constructs for the advantageous purpose of an art recognized non-Gaussian sampling method. See Rackauckas paragraph 123 and Chaudhuri page 4 section 2.3 second paragraph item 2.
Claim 12 further recites “12. The system as described in claim 11, wherein the data model includes a sequence of operations converted from the computing function of the program based on the samples.” Chaudhuri page 3 right column section 2.2 fourth paragraph disclose “Smoothed semantics is defined not only at the level of the whole program, but also at the level of individual expressions. For example, the smoothed semantics of a boolean expression B is ….” Accordingly, the whole program is a combination of a plurality of individual expressions. Each individual expression of the whole program corresponds with a respective operation of the program.
Claim 12 further recites “and the discontinuity is caused by an if/else branch in the program.” Chaudhuri abstract discloses “discontinuous control flow constructs.” Chaudhuri page 2 right column second paragraph discloses “conditional branches.” Control flow conditional branches include if/else branches in the program. Chaudhuri page 4 example 2 shows an if/else part of a program.
Claim 13 further recites “13. The system as described in claim 11, wherein the one or more gradients are generated as one or more derivatives of the computing function.” Chaudhuri page 3 left column first paragraph discloses “Note that the sharp discontinuities and flat plateaus of Figure 1-(c) are now gone, and any point reasonably close to the quadrant … has a nonzero downhill gradient. Hence, the search method is now able to converge to a value close to the global minimum for a much larger range of initial points.” Having a downhill gradient corresponds with the data model having respective gradient(s). These gradients correspond with derivatives of the data model which is of the computing function(s).
Claim 14 further recites “14. The system as described in claim 11, wherein the data model generation module is configured to model the discontinuity based on pairs of samples taken from the plurality of samples.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution.” A 1D distribution corresponds with a one dimensional kernel.
Chaudhuri page 4 section 2.3 second paragraph item 2 teaches “Observe, however, that YX is not required to be Gaussian.”
Chaudhuri does not explicitly disclose a box kernel; however, in analogous art of training with surrogate models, Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use Latin Hypercube sampling in a small neighborhood into the system of smooth interpretation of program control flow constructs for the advantageous purpose of an art recognized non-Gaussian sampling method. See Rackauckas paragraph 123 and Chaudhuri page 4 section 2.3 second paragraph item 2.
Claim 15 further recites “15. The system as described in claim 14, wherein the data model generation module is configured to model the discontinuity by estimating gradients at locations between the pairs of samples.” Chaudhuri page 3 section 2.2 fourth paragraph discloses “The function [[P]](x) is the convolution of the function [[P]] and a K-D Gaussian with mean 0 and standard deviation.” The mean of zero at x is a centering of the estimate between the respective sampling.
Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side.
Claim 16 further recites “16. The system as described in claim 11, further comprising executing a data model execution system configured to execute a machine learning model using the data model as part of a backpropagation operation between layers of the machine learning model.”
Chaudhuri page 13 left column second paragraph discloses “The idea of using smooth approximations to improve the performance of numerical methods like gradient descent is well-known in the domain of neural networks [3].” Improving the performance of a gradient descent in the domain of neural networking corresponds with executing a machine learning model using the data model. The neural network corresponds with a machine learning model. Gradient descent is a backpropagation.
Claim 17 further recites “17. A computing device comprising: a processing device; and a computer-readable storage medium storing instructions that, responsive to execution by the processing device, causes the processing device to perform operations.” Chaudhuri abstract discloses “programs.” Chaudhuri section 1 first sentence discloses “software systems are inherently discontinuous.” Software and programs are related to computers and corresponds with instructions.
But Chaudhuri does not explicitly disclose a processing device; however, in analogous art of training with surrogate models, Rackauckas paragraph 28 teaches “The computer system includes a memory and a processor.” A processor is a processing device. A memory corresponds with computer-readable storage medium.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use a computer into the system of smooth interpretation of program control flow constructs for art recognized suitable system for implementing software and programs including surrogates in a computing language. See Rackauckas paragraph 3 and MPEP §2144.07.
Claim 17 further recites “including: generating samples oriented along an axis based on an output of a program having a computing function that includes a discontinuity.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution …. We assume that Y has K components (i.e., Y ranges over states of P), and that these components are independent variables.” The 1D independent variables Y correspond with an identified axis. The Gaussian distribution corresponds with a sampling along this axis. See further Chaudhuri page 4 right column item 1 regarding Gaussian distribution of YX.
Claim 17 further recites “generating one or more gradients that model the discontinuity of the computing function based on the samples.” Chaudhuri page 3 section 2.2 second paragraph discloses “The smoothed semantics of P can now be defined as a smooth approximation of
P
in terms of
f
μ
,
σ
.” Chaudhuri page 3 right column last paragraph discloses “We use the phrase “Gaussian smoothing of P (with respect to ß)” to refer to the interpretation of P according to the smoothed semantics.” The smoothed semantics or Gaussian smoothing of P corresponds with a generated data model of the program based on the Gaussian sampling.
Chaudhuri page 3 left column first paragraph discloses “Note that the sharp discontinuities and flat plateaus of Figure 1-(c) are now gone, and any point reasonably close to the quadrant … has a nonzero downhill gradient. Hence, the search method is now able to converge to a value close to the global minimum for a much larger range of initial points.” Having a downhill gradient corresponds with the data model having respective gradient(s).
Claim 17 further recites “using a 1D box kernel.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution.” A 1D distribution corresponds with a one dimensional kernel.
Chaudhuri page 4 section 2.3 second paragraph item 2 teaches “Observe, however, that YX is not required to be Gaussian.”
Chaudhuri does not explicitly disclose a box kernel; however, in analogous art of training with surrogate models, Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side.
A 3x3 neighborhood sampling corresponds with sampling a literal box shape stencil and a nearest neighbor-type sampling. Accordingly, a 1D version corresponds with using a 1D box kernel. Compare with “Claim Interpretation” section above.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use Latin Hypercube sampling in a small neighborhood into the system of smooth interpretation of program control flow constructs for the advantageous purpose of an art recognized non-Gaussian sampling method. See Rackauckas paragraph 123 and Chaudhuri page 4 section 2.3 second paragraph item 2.
Claim 17 further recites “and backpropagating the one or more gradients between layers as part of executing a machine learning model.” Chaudhuri page 13 left column second paragraph discloses “The idea of using smooth approximations to improve the performance of numerical methods like gradient descent is well-known in the domain of neural networks [3].” Improving the performance of a gradient descent in the domain of neural networking corresponds with executing a machine learning model using the data model. The neural network corresponds with a machine learning model. Gradient descent is a backpropagation.
Claim 18 further recites “18. The computing device as described in claim 17, further comprising receiving an input identifying the axis.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution …. We assume that Y has K components (i.e., Y ranges over states of P), and that these components are independent variables.” The 1D independent variables Y correspond with an identified axis. The Gaussian distribution corresponds with a sampling along this axis. See further Chaudhuri page 4 right column item 1 regarding Gaussian distribution of YX.
Claim 19 further recites “19. The computing device as described in claim 17, wherein the generating of the one or more gradients models the discontinuity based on pairs of the samples.” Chaudhuri page 3 section 2.2 disclose “a 1-D Gaussian distribution.” A 1D distribution corresponds with a one dimensional kernel.
Chaudhuri page 4 section 2.3 second paragraph item 2 teaches “Observe, however, that YX is not required to be Gaussian.”
Chaudhuri does not explicitly disclose a box kernel; however, in analogous art of training with surrogate models, Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use Latin Hypercube sampling in a small neighborhood into the system of smooth interpretation of program control flow constructs for the advantageous purpose of an art recognized non-Gaussian sampling method. See Rackauckas paragraph 123 and Chaudhuri page 4 section 2.3 second paragraph item 2.
Claim 21 further recites “21. The method as described in claim 1, wherein the modeling the discontinuity of the computing function includes pre-filtering the computing function with the 1D box kernel.” Chaudhuri page 3 section 2.2 second paragraph discloses “The smoothed semantics of P can now be defined as a smooth approximation of
P
in terms of
f
μ
,
σ
.” Chaudhuri page 3 right column last paragraph discloses “We use the phrase “Gaussian smoothing of P (with respect to ß)” to refer to the interpretation of P according to the smoothed semantics.” The smoothed semantics or Gaussian smoothing of P corresponds with a generated data model of the program based on the Gaussian sampling. Smoothing of the corresponding program corresponds with a pre-filtering of the computing function of the program.
Chaudhuri does not explicitly disclose a box kernel; however, in analogous art of training with surrogate models, Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side.
A 3x3 neighborhood sampling corresponds with sampling a literal box shape stencil and a nearest neighbor-type sampling. Accordingly, a 1D version corresponds with using a 1D box kernel. Compare with “Claim Interpretation” section above.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri and Rackauckas. One having ordinary skill in the art would have found motivation to use Latin Hypercube sampling in a small neighborhood into the system of smooth interpretation of program control flow constructs for the advantageous purpose of an art recognized non-Gaussian sampling method. See Rackauckas paragraph 123 and Chaudhuri page 4 section 2.3 second paragraph item 2.
Claim 21 further recites “and wherein a size of the 1D box kernel is defined by a distance between a pair of samples taken from the plurality of samples.” Chaudhuri page 4 right column section 2.2 definition 2 disclose “The constant ß is said to be the smoothing parameter.” This smoothing parameter corresponds with a distance defining the smoothing function. See further Chaudhuri figure 2.
Rackauckas paragraph 123 teaches:
The location of the discontinuity is controlled by inverter's system parameters, and the surrogate is trained on multiple sets of system parameters sampled using Latin Hypercube sampling. … a neighborhood of 3x3 was chosen.
A Latin Hypercube sampling of a 3x3 neighborhood corresponds with a box sampling in two dimensions. Latin Hypercube sampling of a 3x3 neighborhood in one dimension corresponds with a 3 neighborhood which is the target point with a pair of samples to either side. That pair of samples corresponds with a pair of samples taken from the plurality of possible samples and has a distance between them.
Dependent Claim 20
Claim 20 is rejected under 35 U.S.C. 103 as being unpatentable over Chaudhuri and Rackauckas as applied to claim 17 above, and further in view of US patent 11,727,274 B1 Forristal, et al. [herein “Forristal”].
Claim 20 further recites “20. The computing device as described in claim 17, wherein the machine learning model is configured to perform an image processing operation.” The configuration of the machine learning model is an intended use of the machine learning model. See MPEP §2111.
Chaudhuri pages 1-2 at the page-break discloses “Our definition is inspired by the literature on computer vision and signal processing, where numerical optimization is routinely used on noisy real-world signals, and often runs into problems.” Computer vision and signal processing correspond with image processing operation.
Chaudhuri page 13 left column second paragraph discloses “The idea of using smooth approximations to improve the performance of numerical methods like gradient descent is well-known in the domain of neural networks [3].” Neural networks correspond with a machine learning model.
But Chaudhuri does not explicitly disclose the neural network performing image processing; however, in analogous art of training neural networks with gradients, Forristal column 29 lines 14-19 teaches:
There are applications for neural network model training application 122 and prediction application 522 in many areas such as process control and equipment health monitoring, image processing and classification, data segmentation, data analysis, voice processing and recognition, language translation, etc.
An image processing application for a neural network corresponds with a machine learning model being applied to perform image processing operations.
Forristal abstract teaches “A computer trains a neural network. A neural network is executed with a weight vector to compute a gradient vector using a batch of observation vectors.” This shows that respective neural networks including those which perform image processing have configurations which involve training with gradients.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chaudhuri, Rackauckas, and Forristal. One having ordinary skill in the art would have found motivation to use neural network for image processing into the system of smooth interpretation of program control flow constructs for the advantageous purpose of “improved accuracies with fewer iterations resulting in significantly faster computing times.” See Forristal column 29 lines 20-21.
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 Jay B Hann whose telephone number is (571)272-3330. The examiner can normally be reached M-F 10am-7pm EDT.
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, Renee Chavez can be reached at (571) 270-1104. 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.
/Jay Hann/Primary Examiner, Art Unit 2186 3 October 2026
Prosecution Insights