DETAILED ACTION
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
This final office action is in response to the amendment filed 30 June 2026.
Claims 1-10 are pending. Claim 1 and 10 are independent claims.
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.
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 CFR 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 1, 4, and 6-10 are rejected under 35 U.S.C. 103 as being unpatentable over Hua et al. (US 11087201, patented 10 August 2021, hereafter Hua) and further in view of Ru et al. (Interpretable Neural Architecture Search Via Bayesian Optimisation with Weisfeiler-Lehman Kernels, published 19 February 2021, hereafter Ru), and further in view of Yadav et al. (US 2021/0019309, published 21 January 2021, hereafter Yadav) and further in view of Liu (Hierarchical Representations For Efficient Architecture Search, published 22 February 2018, hereafter Liu), and further in view of Moloney et al. (WO 2020/092810, published 7 May 2020, hereafter Moloney).
As per independent claim 1, Hua discloses a method for determining an optimal architecture of a neural network for a given data set including training data and validation data, the method comprising the following steps:
defining a search space which characterizes possible architecture of the neural network (column 3, lines 25-44: Here, a search space is defined by the neural architecture search)
drawing a plurality of candidate architectures (column 4, lines 53-63: Here, as part of determining the architecture for the cell, which is a fully convolutional neural network configured to receive cell input and generate an output (column 3, lines 45-47), the neural architecture search (column 3, lines 19-24) obtains data that specifies a current set of candidate cells/CNN. This includes obtaining cells from a previous iteration and adding a respective one or more operation blocks to each of the previous cells)
training neural networks with the candidate architecture on the training data, and validating the trained neural networks on the validation data (column 5, lines 42-60: Here, the training engine receives data for training the instances in a variety of ways. This is performed by receiving training data for training instances on the particular machine learning task and a validation set for evaluating the performance of the trained instances of the task neural network on the particular machine learning task)
adapting the process such that given the candidate architectures, the process predicts the validation achieved with the candidate architectures (column 4, line 64- column 5, line 2: Here, based on receiving a candidate cell in the current set, the predictor receives data specifying the cell and processes the data using the performance prediction neural network in accordance with the current values of the prediction parameters go generate a performance prediction for each candidate cell)
repeating steps i-iii two or more times (column 6, lines 61-64):
determining a next candidate architecture to be evaluated depending on an acquisition function (column 5, lines 3-12: Here, the neural architecture search (NAS) generates an updated set of cells (candidates) to be evaluated and prunes candidates based upon performance prediction)
training a further neural network with the candidate architecture to be evaluated on the training data, and validating the further, trained neural network on the validation data (column 5, lines 42-60: Here, the training engine receives data for training the instances in a variety of ways. This is performed by receiving training data for training instances on the particular machine learning task and a validation set for evaluating the performance of the trained instances of the task neural network on the particular machine learning task)
adapting the process such that given previously used candidate architectures, the process predicts the validation achieved with the previously used candidate architectures (column 6, lines 4-25: Here, the predictor generates a performance prediction for each of the candidates within the current set of the NAS. These predictions are based upon the small number of trained cells (previously used candidate architectures))
outputting the candidate architecture that achieved a best performance on the validation data (column 6, line 65- column 7, line 3: Here, the candidate cell that hast e best performance as the output cell is determined)
Hua fails to specifically disclose:
defining a search space which characterizes possible architecture of the neural network using a context-free grammar
wherein the context-free grammar characterizes a plurality of hierarchies of levels
wherein a lowest level of each hierarchy defines a plurality of operations
wherein levels above the lowest level define at least one rule, levels except the lowest level can be combined with one another
randomly drawing a candidate architecture according to the context-free grammar
wherein the randomly drawing includes generating a string configured to be translated into a syntax tree according to the context-free grammar, wherein an edge-attributed graph representing at least one of the candidate architectures is generated based on the syntax tree
initializing a Gaussian process, wherein the Gaussian process includes a Weisfeiler-Lehman graph kernel
an acquisition function that depends on the Gaussian process, wherein the acquisition function is optimized using an evolutionary algorithm
However, Ru, which is analogous to the claimed invention because it is directed toward neural architecture search, discloses:
randomly drawing a candidate architecture (page 3, paragraph 1: Here, for architecture generation, new candidates are selected via random sampling)
initializing a Gaussian process, wherein the Gaussian process includes a Weisfeiler-Lehman graph kernel (page 1, paragraph 1; page 3, paragraph 1: Here, a neural architecture search combining a Gaussian process with a Weisfeiler-Leman subtree graph kernel (GPWL) is used for performing searching within the neural architecture space to optimize generation/selection of candidate architectures at each iteration)
an acquisition function that depends on the Gaussian process, wherein the acquisition function is optimized using an evolutionary algorithm (page 2, paragraph 6; page 3, paragraph 2 – page 4, paragraph 2: Here, a Gaussian process is used to compare candidates and perform iterations until an optimized selection of features is achieved for the NAS)
It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Ru with Hua, with a reasonable expectation of success, as it would have allowed for implementing a scalable model to achieving better prediction performance (Ru: page 2, paragraph 2).
Further, Yadav, which is analogous to the claimed invention because it is directed toward identifying grammars based upon a translated string, discloses:
wherein the drawing includes generating a string configured to be translated into a syntax tree according to the context-free grammar (paragraph 0036: Here, a translation layer is implemented to take a strong of text and translate it into natural language syntax data in the form of syntax tree data (paragraph 0103)), wherein an edge-attributed graph is generated based on the syntax tree (paragraph 0103: Here, once the input string is received, the string is compared to known tokens. Each string fragment may matching multiple tokens and the input string is input into a natural language processing system to generate a syntax tree)
It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Yadav with Hua-Ru, with a reasonable expectation of success, as it would have allowed for generating a tree graph based upon an input string (Yadav: paragraph 0103).
Additionally, Liu, which is analogous to the claimed invention because it is directed toward hierarchical representations for architecture search, discloses:
a plurality of hierarchies of levels (page 3, Section 2.2: Here, a hierarchy is described in which the lowest level of the hierarchy L are a set of primitive operations and the highest level l contains a single motif corresponding to the full architecture)
wherein a lowest level of each hierarchy defines a plurality of operations (page 3, Section 2.2: Here, a hierarchy is described in which the lowest level of the hierarchy L are a set of primitive operations and the highest level l contains a single motif corresponding to the full architecture)
wherein levels above the lowest level define at least one rule, levels except the lowest level can be combined with one another (page 3, Section 2.2: Here, a hierarchy is described in which the lowest level of the hierarchy L are a set of primitive operations and the highest level l contains a single motif corresponding to the full architecture. In between the lowest and highest levels are additional levels of motifs used for defining the highest level motif)
It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Liu with Hua-Yadav-Ru, with a reasonable expectation of success, as it would have allowed for improved search using a hierarchical structure to implement hierarchical modules able to process changes across the motif and propagate them across the whole network immediately (Liu: page 2, paragraph 1).
Finally, Moloney, which is analogous to the claimed invention because it is directed toward searching a neural network architecture, discloses using a context-free grammar to search generate a query (paragraph 0027: Here, the neural network generator system utilizes context free grammars to facilitate evolution of parameters associated with specific layers and optimize the process). It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Komenda with Hua-Ru-Yadav-Liu, with a reasonable expectation of success, as it would have allowed for improvement of the neural network using derivations and evolutions facilitated by the context-free grammar (Moloney: paragraph 0027).
As per dependent claim 4, Hua, Ru, Yadav, Liu, and Moloney disclose the limitations similar to those in claim 1, and the same rejection is incorporated herein. Ru discloses wherein the acquisition function is a guided acquisition function, wherein the acquisition function is evaluated using a guided evolutionary algorithm (page 3, paragraph 1: Here, a mutation algorithm is used to generate new candidates via random sampling. This process is guided via a Gaussian process is used to compare candidates and perform iterations until an optimized selection of features is achieved for the NAS((page 2, paragraph 6; page 3, paragraph 2 – page 4, paragraph 2)). It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Ru with Hua, with a reasonable expectation of success, as it would have allowed for implementing a scalable model to achieving better prediction performance (Ru: page 2, paragraph 2).
Hua fails to specifically disclose the use of a grammar. However, Moloney, which is analogous to the claimed invention because it is directed toward searching a neural network architecture, discloses using a context-free grammar to search generate a query (paragraph 0027: Here, the neural network generator system utilizes context free grammars to facilitate evolution of parameters associated with specific layers and optimize the process). It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Komenda with Hua-Ru-Yadav-Liu, with a reasonable expectation of success, as it would have allowed for improvement of the neural network using derivations and evolutions facilitated by the context-free grammar (Moloney: paragraph 0027).
As per dependent claim 6, Hua, Ru, Yadav, Liu, and Moloney disclose the limitations similar to those in claim 1, and the same rejection is incorporated herein. Liu further discloses wherein the context-free grammar additionally includes secondary conditions that characterize properties of the architecture (page 3, Section 2.2: Here, a hierarchy is described in which the lowest level of the hierarchy L are a set of primitive operations and the highest level l contains a single motif corresponding to the full architecture. In between the lowest and highest levels are additional levels of motifs used for defining the highest level motif. These motifs characterize the properties of the architecture)
It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Liu with Hua-Ru-Yadav, with a reasonable expectation of success, as it would have allowed for improved search using a hierarchical structure to implement hierarchical modules able to process changes across the motif and propagate them across the whole network immediately (Liu: page 2, paragraph 1).
As per dependent claim 7, Hua, Ru, Yadav, Liu, and Moloney disclose the limitations similar to those in claim 1, and the same rejection is incorporated herein. Hua further discloses wherein the input variables are image and the machine learning system is an image classifier (column 2, lines 54-64).
With respect to independent claim 8, the claim recites the limitations substantially similar to those in claim 1. Claim 8 is rejected under similar rationale.
Additionally, Hua discloses a device (column 11, lines 37-40).
As per dependent claim 9, Hua, Ru, Yadav, Liu, and Moloney disclose the limitations similar to those in claim 8, and the same rejection is incorporated herein. Hua discloses wherein the device is a training device (column 9, lines 33-34: Here, the system trains the instance to perform the particular learning task. Further, the system may include devices such as a clients and servers (column 12, lines 44-55)).
With respect to independent claim 10, the claim recites the limitations substantially similar to those in claim 1. Claim 10 is rejected under similar rationale.
Additionally, Hua discloses a non-transitory machine-readable storage medium (column 11, lines 58-64).
Claim 2 is rejected under 35 U.S.C. 103 as being unpatentable over Hua, Ru, Yadav, Liu, and Moloney and further in view of Tan et al. (WO 2021/170215, published 2 September 2021, hereafter Tan).
As per dependent claim 2, Hua, Ru, Yadav, Liu, and Moloney disclose the limitations similar to those in claim 1, and the same rejection is incorporated herein. Moloney discloses use of a context free grammar (paragraph 0027). Additionally, Liu further discloses wherein the evolutionary algorithm applies a mutation, wherein the mutation and crossover are applied to a syntax tree characterizing the candidate architecture, wherein a new syntax tree obtained by the mutation is tested (page 5, paragraphs 2-3: Here, an evolutionary search algorithm picks a promising genotype from the population and places mutated offspring back into the population. Repeating this operation increases the quality of the population via refinement over a period of time with the genotype with the highest fitness selected as the final output over a period of time).
It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Liu with Hua-Ru-Yadav-Liu-Moloney, with a reasonable expectation of success, as it would have allowed for improved search using a hierarchical structure to utilize mutated offspring to improve fitness selection over a period of time (Liu: page 5, paragraphs 2-3).
Hua fails to specifically disclose wherein the evolutionary algorithm applies a crossover, wherein the crossover are applied to a syntax tree characterizing the candidate architecture, wherein a new syntax tree obtained by the crossover is tested. However, Tan, which is analogous to the claimed invention because it is directed toward neural architecture search, discloses wherein the evolutionary algorithm applies a crossover, wherein the crossover are applied to a syntax tree characterizing the candidate architecture, wherein a new syntax tree obtained by the crossover is tested (paragraph 0051-0053: Here, a genetic mutation operator and a cross-over are used to construct a new network editing tree (syntax tree). These trees are tested, and the most fit tree in the final generation provides the searched neural architecture).
It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Tan with Hua-Ru-Liu-Moloney, with a reasonable expectation of success, as it would have allowed for selecting a searched neural architecture by implementing mutations and cross-overs to diversify the population and improve the selection (Tan: paragraph 0051).
Claim 3 is rejected under 35 U.S.C. 103 as being unpatentable over Hua, Ru, Yadav, Liu, and Moloney and further in view of Pal et al. (Self-crossover – a new genetic operator and its application to feature selection, 1998, hereafter Pal).
As per dependent claim 3, Hua, Ru, Yadav, Liu, and Moloney disclose the limitations similar to those in claim 1, and the same rejection is incorporated herein. Moloney discloses use of a context free grammar (paragraph 0027). Additionally, Liu further discloses wherein the evolutionary algorithm applies a mutation, wherein the mutation and crossover are applied to a syntax tree characterizing the candidate architecture, wherein a new syntax tree obtained by the mutation is tested (page 5, paragraphs 2-3: Here, an evolutionary search algorithm picks a promising genotype from the population and places mutated offspring back into the population. Repeating this operation increases the quality of the population via refinement over a period of time with the genotype with the highest fitness selected as the final output over a period of time).
It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Liu with Hua-Ru-Yadav-Liu-Moloney, with a reasonable expectation of success, as it would have allowed for improved search using a hierarchical structure to utilize mutated offspring to improve fitness selection over a period of time (Liu: page 5, paragraphs 2-3).
Hua fails to specifically disclose wherein the evolutionary algorithm applies a self-crossover, wherein the self-crossover are applied to a syntax tree characterizing the architecture, wherein a new syntax tree obtained by the crossover is tested, wherein the self-crossover is carried out randomly, wherein with the self-crossover, branches are swapped in the tree. However, Pal, which is analogous to the claimed invention because it is directed toward neural architecture search, discloses wherein the evolutionary algorithm applies a self-crossover, wherein the self-crossover are applied to a syntax tree characterizing the architecture, wherein a new syntax obtained by the self-crossover is tested, wherein the self-crossover is carried out randomly, wherein with the self-crossover, branches are swapped (Section 4: Here, a self-crossover is applied by selecting a random position within a string identifying a splitting position. The string is then split into two substrings and two additional random positions are determined for splitting the substrings. These four substrings are recombined as the first and fourth substrings and the third and second substrings. This constitutes swapping branches of the substrings. Further, self-crossover are one mutation operation used in optimization algorithms (Section 2)).
It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Pal with Hua-Ru-Liu-Moloney, with a reasonable expectation of success, as it would have allowed for applying mutations where the genetic information within a single string/node is selected randomly from the pool of candidates to produce mutated offspring (Pal: Section 4, paragraph 1).
Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over Hua, Ru, Yadav, Liu, and Moloney and further in view of Elliot et al. (US 2021/0142525, published 13 May 2021, hereafter Elliot).
As per dependent claim 5, Hua, Ru, Yadav, Liu, and Moloney disclose the limitations similar to those in claim 1, and the same rejection is incorporated herein. Liu discloses wherein a lowest level of each hierarchy defines a plurality of operations (page 3, Section 2.2: Here, a hierarchy is described in which the lowest level of the hierarchy L are a set of primitive operations and the highest level l contains a single motif corresponding to the full architecture).
It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Liu with Hua-Ru-Yadav, with a reasonable expectation of success, as it would have allowed for improved search using a hierarchical structure to implement hierarchical modules able to process changes across the motif and propagate them across the whole network immediately (Liu: page 2, paragraph 1).
Hua fails to specifically disclose wherein the operation is a down sampling operation. However, Elliot, which is analogous to the claimed invention because it is directed toward a neural network having a plurality layers implementing operations, discloses a down sampling operation (paragraph 0118: Here, a neural network may comprises a plurality of layers of operations, including a down sampling operation layer). It would have been obvious to one of ordinary skill in the art at the time of the applicant’s effective filing date to have combined Elliot with Hua-Ru-Yadav-Liu-Moloney, with a reasonable expectation of success, as it would have allowed for performing down sampling operations within a neural network (Elliot: paragraph 0118).
Response to Arguments
Applicant’s arguments with respect to the rejection of claims under 35 USC 103 with respect to the newly amended limitations have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Hua, Ru, Yadav, Liu, and Moloney.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
Secker-Walker et al. (US 9093061): Discloses string processing and creation of a candidate graph (column 21, line 63- column 22, line 12; claim 11)
Xu et al. (US 2006/0062468): Discloses a tree structure of sub-expressions based upon parsing text strings (paragraph 0169)
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 KYLE R STORK whose telephone number is (571)272-4130. The examiner can normally be reached 8am - 2pm; 4pm - 6pm.
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, Omar Fernandez Rivas can be reached at 571/272-2589. 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.
/KYLE R STORK/Primary Examiner, Art Unit 2128
Prosecution Insights