DETAILED ACTION
This action is in response to the reply filed 19 February 2026.
Claims 24 and 27–43 are pending. Claims 24 and 42 are independent.
Claims 24 and 27–43 are rejected.
Notice of Pre-AIA or AIA Status
The present application, filed on or after 16 March 2013, is being examined under the first inventor to file provisions of the AIA .
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.
Response to Arguments
Regarding the objection to the claims (remarks, p. 9), although the claims were not present at the time of original filing, all of claims 24–43 were added together in a single amendment, and therefore could have been numbered in the correct order. Although Applicant is not required to reorder the claims during prosecution, in the event that the claims are found allowable, the claims will be reordered by the examiner.
The interpretation of claims 24–40 under § 112(f) is withdrawn in light of the amendments and accompanying arguments (remarks, p. 9).
Applicant’s arguments filed 19 February 2026 have been fully considered, but are not persuasive.
Applicant argues that Martin does not teach calculating a number of execution cycles and estimating an execution time for neural networks or inference techniques (remarks, p. 11). However, Martin teaches that its method may be applied to any kind of software (Martin, ¶ 42). Neural networks, including those of Tseng, are software (Tseng, ¶ 128). Therefore, a person having ordinary skill in the art would have understood that the method of estimating an execution time, as taught by Martin, could have been applied to the software neural network of Tseng.
Applicant further argues that there is no motivation or suggestion to combine Martin and Tseng (remarks, p. 11). The examiner respectfully disagrees. Tseng teaches configuring a neural network to satisfy a CPU usage constraint, including determining a number of computations and determining the time required to pass data through the neural network (Tseng, ¶ 85). Martin teaches a method of estimating the execution time of software generally (Martin, ¶ 42). Therefore, it would have been obvious to apply the method of Martin to the neural network of Tseng in order to obtain a more accurate estimate of its execution time.
Claim Rejections—35 U.S.C. § 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.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 C.F.R. § 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. § 102(b)(2)(C) for any potential 35 U.S.C. § 102(a)(2) prior art against the later invention.
Claims 24, 27, 28, 30, 35, 36, and 41–43 are rejected under 35 U.S.C. § 103 as being unpatentable over Tseng et al. (US 2020/0302292 A1) in view of Martin et al. (US 2002/0166112 A1) [hereinafter Martin].
Regarding independent claim 24, Tseng teaches [a] neural network construction device, comprising: a processor; and a non-transitory memory storing a program, wherein the program, when executed by the processor, causes the processor to perform: obtaining a first condition and a second condition, the first condition being used to determine a candidate hyperparameter that is a candidate of a hyperparameter of a neural network to be constructed, the second condition being related to required performance of a model of the neural network; Hyperparameters are determined for a neural network based on one or more performance requirements [second condition] and one or more computing resource constraints [first condition] (Tseng, ¶¶ 13–16). first-determining the candidate hyperparameter using the first condition; Hyperparameters are determined based on the specified constraints (Tseng, ¶ 85). generating the model of the neural network using the candidate hyperparameter; and Neural networks are generated based on sets of hyperparameters meeting the constraints (Tseng, ¶¶ 82–86). second-determining whether or not the model generated meets the second condition, and output data based on a result of the determination, Model validation is carried out on the generated models in order to determine the accuracy thereof; models that meet minimum accuracy constraint are output as candidate models (Tseng, ¶¶ 92–99). the first-determining comprises calculating at least one of an upper limit or a lower limit of the candidate hyperparameter using the first condition, and determining the candidate hyperparameter based on the at least one of the upper limit or the lower limit calculated, the candidate hyperparameter being one or more candidate hyperparameters, Hyperparameters such as a number of layers or number of kernels may be chosen based on, e.g., a maximum amount of memory or maximum CPU usage constraints, such that the number chosen satisfies the constraint [upper limit] (Tseng, ¶ 85). the first condition includes a resource condition related to a computational resource of an embedded device, The neural network/hyperparameters are chosen based on a maximum level of computational resource (Tseng, ¶¶ 16, 54, 85). the first-determining comprises calculating the upper limit of the candidate hyperparameter based on the resource condition, and determining, as the candidate hyperparameter, at least one of hyperparameters less than or equal to the upper limit, The computational resource may be those of a resource-constrained device, such as an embedded device (Tseng, ¶ 51). the second condition includes a temporal condition related to a reference duration of an inference process in which the model of the neural network is used, The constraints may include a maximum acceptable latency [temporal condition] (Tseng, ¶ 83). the generating comprises calculating, based on the resource condition, a duration of an inference process in which the model generated is used, the second-determining comprises determining, by comparing the duration calculated and the reference duration, whether or not the model generated meets the second condition, The type of processor of the device [resource condition] may be used to determine the time required to pass data through the neural network [duration of an inference process], an indication of latency (Tseng, ¶¶ 61, 85). the resource condition further includes information of an operating frequency of an arithmetic processing device of the embedded device, the generating comprises obtaining a total number of execution cycles for a portion corresponding to the inference process of the model generated, and calculating the duration using the total number of execution cycles and the operating frequency, and The number of computations that will be performed by the neural network may be determined based on, e.g., the number of layers (Tseng, ¶ 61). The type of processor [operating frequency] of the device may be used with the number of computations to determine a time required to pass data through the neural network (Tseng, ¶ 85).
Tseng teaches testing the prediction accuracy of generated models, but does not expressly teach doing so using a generated source code. However, Martin teaches: the generating comprises generating a first source code for the portion corresponding to the inference process of the model, and obtaining the total number of execution cycles using an intermediate code obtained by compiling the first source code, the first source code being written in a language dependent on the arithmetic processing device. A number of cycles needed for executing a software program is estimated (Martin, ¶ 38). The number of cycles is estimated based on a number of virtual instructions [intermediate code] generated by compiling the software code using a virtual compiler (Martin, ¶¶ 44–49). The software code may be C code [a language dependent on the arithmetic processing device1] (Martin, ¶ 42).
It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to combine the teachings of Tseng with those of Martin. One would have been motivated to do so in order to increase the accuracy of the estimated time required for executing the neural network (Martin, ¶ 10).
Regarding dependent claim 27, the rejection of claim 24 is incorporated and Tseng/Martin further teaches: wherein: the obtaining further comprises obtaining learning data on the neural network, The models are trained using training data [learning data] retrieved from a larger set of data (Tseng, ¶ 88). the second-determining comprises outputting data indicating a model generated in the generating and determined as meeting the second condition, Models are generated based on the model constraints (Tseng, ¶¶ 82–86). the executed program further causes the processor to perform, using the learning data, learning of the model indicated in the data output in the second-determining, and The models are trained using the training data (Tseng, ¶¶ 87–91). the executed program further causes the processor to perform outputting at least a part of the model that has already been learned. The trained models may be validated to determine their accuracy, and models meeting the minimum required accuracy are added [output] to a list of candidate models (Tseng, ¶¶ 92–100).
Regarding dependent claim 28, the rejection of claim 27 is incorporated and Tseng/Martin further teaches: wherein the learning comprises performing prediction accuracy evaluation of the model that has already been learned, and generating data related to the prediction accuracy evaluation that has been performed. The accuracy of each model is stored in the information about the model (Tseng, ¶ 61). The accuracy is determined using a validation [evaluation] process on the trained [learned] models (Tseng, ¶¶ 93–95).
Regarding dependent claim 30, the rejection of claim 24 is incorporated and Tseng/Martin further teaches: wherein: the resource condition includes information of a memory size of the embedded device, and The resource of the embedded device may be memory, including memory usage when executing the neural network, or memory usage when storing the model information defining the neural network (Tseng, ¶ 16). the first-determining comprises calculating, as the upper limit of the candidate hyperparameter, an upper limit of the hyperparameter of the neural network that fits within the memory size, and determining, as the candidate hyperparameter, at least one of hyperparameters less than or equal to the upper limit. The constraint may indicate a maximum memory that is allowed to be used to store the model, such that hyperparameters including the number of layers, number of kernels per layer, etc. are chosen to fit within the memory [upper limit of the candidate hyperparameter] (Tseng, ¶ 85).
Regarding dependent claim 35, the rejection of claim 24 is incorporated and Tseng/Martin further teaches: wherein: the first condition includes a target of accuracy of inference obtained using the model of the neural network, and The performance requirement may be a minimum level of accuracy (Tseng, ¶¶ 16, 54). The accuracy may be inference accuracy (Tseng, ¶ 35). the first-determining comprises calculating the lower limit of the candidate hyperparameter using the target of accuracy, and determining, as the one or more candidate hyperparameters, at least one of hyperparameters greater than or equal to the lower limit calculated. The model may be chosen based on one or more constraints, including a combination of a minimum acceptable accuracy [lower limit] and maximum use of computational resources (Tseng, ¶ 83).
Regarding dependent claim 36, the rejection of claim 28 is incorporated and Tseng/Martin further teaches: wherein: the data related to the prediction accuracy evaluation is data in an evaluated model list indicating a model on which the prediction accuracy evaluation has already been performed, and The accuracy information from the validation process may be added to the model information (Tseng, ¶ 98). The model information may be stored as a model grid [model list] (Tseng, ¶ 86). the generating, the second-determining, or the learning comprises excluding, from a processing subject, a model generated using a plurality of hyperparameters that are a combination identical to hyperparameters used for any model indicated in the evaluated model list. Each set of stored model information represents a different neural network model having a different set of hyperparameters [i.e., no stored model has a combination of hyperparameters identical to another] (Tseng, ¶¶ 7, 13–15).
Regarding dependent claim 41, Tseng/Martin further teaches [a]n information processing device, comprising: the processor; and A processing apparatus (Tseng, ¶¶ 113, 118). a storage, wherein: A memory (Tseng, ¶¶ 113, 118). the storage stores a model generated by the neural network construction device according to the storage stores a model generated by the neural network construction device according to the processor reads the model from the storage and implements the model. The neural network models are stored in a memory (Tseng, ¶ 118).
Regarding independent claim 42, this claim recites limitations similar to those of claim 24, and is rejected for the same reasons.
Regarding dependent claim 43, Tseng/Martin further teaches [a] non-transitory computer-readable recording medium having recorded thereon a program for causing the arithmetic processing device to execute the neural network construction method according to claim 42. The code for initializing, training, and validating the neural networks may be stored in memory (Tseng, ¶ 113).
Claims 31–32 are rejected under 35 U.S.C. § 103 as being unpatentable over Tseng et al. (US 2020/0302292 A1) [hereinafter Tseng] in view of Martin et al. (US 2002/0166112 A1) [hereinafter Martin], further in view of Choi (US 2016/0155049 A1).
Regarding dependent claim 31, the rejection of claim 24 is incorporated. Tseng/Martin teaches tuning hyperparameters based on constraints, but does not expressly teach using the size of input or output data. However, Choi teaches: wherein: the first condition includes information of at least one of a size of input data or a size of output data, the input data being data input to the neural network, the output data being data output from the neural network, and A neural network is configured based on, e.g., the input dimension [size of input data] (Choi, ¶ 142). the first-determining comprises calculating the upper limit of the candidate hyperparameter based on the at least one of the size of the input data or the size of the output data that is included in the first condition, and determining, as the one or more candidate hyperparameters, at least one of hyperparameters less than or equal to the upper limit calculated. A number of nodes [hyperparameter] included in the input layer is determined based on the input dimension (Choi, ¶ 142).
It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to combine the teachings of Tseng/Martin with those of Choi. Doing so would have been a matter of applying a known technique (the constraints or “first condition”, and the hyperparameters, of Choi) to a known device ready for improvement (the hyperparameter tuning device of Tseng) to yield a predictable result (a hyperparameter tuning device that uses the size of input data as a constraint).
Regarding dependent claim 32, the rejection of claim 31 is incorporated and Tseng/Martin/Choi further teaches: wherein: the size of the input data is dimensionality of the input data, A neural network is configured based on, e.g., the input dimension [dimensionality of the input data]. the size of the output data is dimensionality of the output data, and A number of nodes included in the input layer is determined based on the input dimension, a number of nodes included in the output layer is determined based on the number in the input layer, and the number of nodes in the hidden layer is preset [i.e., the total number depends on the input dimension] (Choi, ¶ 142). the one or more candidate hyperparameters include both a total number of layers in the neural network and a total number of nodes in the neural network. The hyperparameters may include a number of layers in the model (Tseng, ¶ 52).
Regarding dependent claim 33, the rejection of claim 31 is incorporated and Tseng/Martin/Choi further teaches: wherein the first condition further includes information indicating that the neural network is a convolutional neural network. The neural networks have hidden layers implementing convolution kernels [convolutional neural networks] (Tseng, ¶ 1).
Claims 29 and 37 are rejected under 35 U.S.C. § 103 as being unpatentable over Tseng et al. (US 2020/0302292 A1) [hereinafter Tseng] in view of Martin et al. (US 2002/0166112 A1) [hereinafter Martin], further in view of Ar et al. (US 2018/0307978 A1) [hereinafter Ar].
Regarding dependent claim 29, the rejection of claim 28 is incorporated. Tseng/Martin teaches testing the prediction accuracy of generated models, but does not expressly teach doing so using generated source code. However, Ar teaches: wherein the learning comprises generating a second source code for a portion corresponding to an inference process of the model that has already been learned, and performing the prediction accuracy evaluation using the second source code, the second source code being written in a language dependent on an arithmetic processing device. A deep learning model generator compiles intermediate representations of models into source code [generates second source code] and generates suggestions pertaining to the correctness [prediction accuracy] of the network (Ar, ¶ 31). The generated source code is execution-ready [dependent on an arithmetic processing device] (Ar, ¶ 24). The languages used may include C (Ar, ¶ 67).
It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to combine the teachings of Tseng/Martin/Choi with those of Ar. One would have been motivated to do so in order to make it easier for the user to design a model that can be used on different platforms (Ar, ¶¶ 21-23).
Regarding dependent claim 37, the rejection of claim 27 is incorporated. Tseng/Martin teaches generating models and outputting them, but does not expressly teach outputting them in source code in a language “dependent on an arithmetic processing device”. However, Ar teaches: wherein the outputting comprises outputting the model in a format of a source code in a language dependent on an arithmetic processing device. A deep learning model generator compiles intermediate representations of models into source code; the generated source code is execution-ready [dependent on an arithmetic processing device] (Ar, ¶ 24). The languages used may include C (Ar, ¶ 67).
It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to combine the teachings of Tseng/Martin with those of Ar. One would have been motivated to do so in order to make it easier for the user to design a model that can be used on different platforms (Ar, ¶¶ 21–23).
Claim 34 is rejected under 35 U.S.C. § 103 as being unpatentable over Tseng et al. (US 2020/0302292 A1) [hereinafter Tseng] in view of Martin et al. (US 2002/0166112 A1) [hereinafter Martin] and Choi (US 2016/0155049 A1), further in view of Ar et al. (US 2018/0307978 A1) [hereinafter Ar] and Talathi et al. (US 2016/0224903 A1) [hereinafter Talathi].
Regarding dependent claim 34, the rejection of claim 33 is incorporated and Tseng/Martin/Choi further teaches: wherein: the input data is image data, The input data may be image data (Tseng, ¶ 36). […] the one or more candidate hyperparameters include at least one of a total number of layers in the convolutional neural network, a size of a kernel, a depth of the kernel, a size of a feature map, a window size of a pooling layer, an amount of padding, or an amount of stride. The hyperparameters may include a number of layers and a number of kernels per layer (Tseng, ¶ 13).
Tseng/Martin teaches input data being images, and output data being a classification of the image (Tseng, ¶ 36), and Choi teaches input data being image data (Choi, ¶¶ 5, 6, 154), but Tseng/Martin/Choi does not teach a “first condition” being a total number of pixels or total number of classes. However, Ar teaches: the size of the output data is a total number of classes into which the image data is classified, and A deep learning network [neural network] is constructed; the input may be images (Ar, ¶ 21). The last layer [output layer] has a number of nodes based on the number of classes [total number of classes into which the image data is classified] (Ar, ¶ 43).
It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to combine the teachings of Tseng/Martin/Choi with those of Ar. Doing so would have been a matter of applying a known technique (the constraints or “first condition”, and the hyperparameters, of Ar) to a known device ready for improvement (the hyperparameter tuning device of Tseng) to yield a predictable result (a hyperparameter tuning device that uses the number of output classes as a constraint).
Tseng/Martin/Choi/Ar teaches using a number of classes as a constraint, and a dimensionality of image data as a constraint, but does not expressly teach using a size of images in pixels. However, Talathi teaches: the size of the input data is a total number of pixels in the image data, A convolutional neural network classifies scenes based on visual input; the input images have a particular size, e.g., 224x224 pixels, which is mapped to the dimensions of the neural network (Talathi, ¶ 93). The neural network is selected from a number of neural networks in a database having different specified hyperparameters (Talathi, ¶¶ 94, 97).
It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to combine the teachings of Tseng/Martin/Choi/Ar with those of Talathi. Doing so would have been a matter of applying a known technique (the constraints or “first condition”, and the hyperparameters, of Talathi) to a known device ready for improvement (the hyperparameter tuning device of Tseng) to yield a predictable result (a hyperparameter tuning device that uses the number of pixels in the input image as a constraint).
Claim 38 is rejected under 35 U.S.C. § 103 as being unpatentable over Tseng et al. (US 2020/0302292 A1) [hereinafter Tseng] in view of Martin et al. (US 2002/0166112 A1) [hereinafter Martin], further in view of Lin et al. (US 2018/0114117 A1) [hereinafter Lin].
Regarding dependent claim 38, the rejection of claim 27 is incorporated. Tseng/Martin teaches generating neural networks, but does not expressly teach outputting neural networks in HDL. However, Lin teaches: wherein the outputting comprises outputting the model in a format of a hardware description language. One or more hardware description language source files are generated for a DNN [deep neural network] (Lin, ¶¶ 7, 10, 11, 27).
It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to combine the teachings of Tseng/Martin with those of Lin. One would have been motivated to do so in order to make it easier for users to generate neural networks for FPGAs (Lin, ¶¶ 3–5).
Claims 39 and 40 are rejected under 35 U.S.C. § 103 as being unpatentable over Tseng et al. (US 2020/0302292 A1) [hereinafter Tseng] in view of Martin et al. (US 2002/0166112 A1) [hereinafter Martin] and Ar et al. (US 2018/0307978 A1) [hereinafter Ar], further in view of Baluja et al. (US 9,026,479 B1) [hereinafter Baluja].
Regarding dependent claim 39, the rejection of claim 29 is incorporated. Tseng/Martin/Ar teaches generating a number of neural networks based on different hyperparameters, but does not expressly teach a stopping condition based on the prediction accuracy. However, Baluja teaches: wherein the second-determining comprises stopping the generating the model of the neural network when a grade of the prediction accuracy evaluation that has been performed meets a predetermined condition. Parameters for a category prediction model are determined, based on iteratively adjusting the prediction parameters of the model until a termination event, such as a minimum error threshold [grade of the prediction accuracy evaluation] being reached (Baluja, col. 8 ll. 50–60). At least two sets of parameters are stored [models generated] based on the best parameters and the current iteration (Baluja, col. 10 ll. 10–45).
It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to combine the teachings of Tseng/Martin/Ar with those of Baluja. One would have been motivated to do so in order to prevent the program from looping/iterating infinitely and, e.g., crashing.
Regarding dependent claim 40, the rejection of claim 29 is incorporated. Tseng/Martin/Ar teaches generating a number of neural networks based on different hyperparameters, but does not expressly teach a stopping condition based on the prediction accuracy. However, Baluja teaches: wherein: the obtaining comprises obtaining a target of accuracy indicating a predetermined level of accuracy of the model of the neural network, and the predetermined condition is that grades of the prediction accuracy evaluation of at least a predetermined number of models that are continuous in order of generation fail to reach the target of accuracy. The termination event may be a combination of achieving the minimum error threshold, or achieving an iteration limit [a number of models that are continuous in order of generation failing to reach the accuracy target] (Baluja, col. 10 ll. 40–45, ll. 55–60).
It would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to combine the teachings of Tseng/Martin/Ar with those of Baluja. One would have been motivated to do so in order to prevent the program from looping/iterating infinitely and, e.g., crashing.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 C.F.R. § 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 C.F.R. § 1.17(a)) pursuant to 37 C.F.R. § 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.
The prior art made of record and not relied upon is considered pertinent to Applicant's disclosure.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Tyler Schallhorn whose telephone number is 571-270-3178. The examiner can normally be reached Monday through Friday, 8:30 a.m. to 6 p.m. (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, Tamara Kyle can be reached on 571-272-4241. 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 the USA or Canada) or 571-272-1000.
/Tyler Schallhorn/Examiner, Art Unit 2144
/TAMARA T KYLE/Supervisory Patent Examiner, Art Unit 2144
1 Applicant states that C is an example of a “language that is highly dependent on the arithmetic processing device” (specification, p. 37).