Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 10/28/2025 has been entered.
Response to Amendment
The office action is responsive to the amendment filed on 10/28/2025. As directed by the amendments claims 1, 21, and 22 are amended. Claims 1, 4-18 and 21-22 are pending for examination.
Response to Arguments
Regarding Rejections Under 35 US.C. § 101:
Applicant’s arguments, see pg. 9-13, filed 10/28/2025, with respect to claims 1,4-18 and 21-22 have been fully considered and are persuasive. The rejection of claims 1,4-18 and 21-22 under 35 US.C. § 101 has been withdrawn.
Regarding Rejections Under 35 US.C. § 103:
Applicant's arguments filed pg. 13-17 filled on 10/28/2025 have been fully considered but they are not persuasive.
APPLICANT ARGUMENT #1:
Applicant argues, amended claim 1 is non-obvious in view of the cited documents. Similar independent claims 21 and 22 are non-obvious for at least the same reasons as claim 1. In addition, applicant argues, claims 4-18 depend on claim 1, therefore are non-obvious for at least the same reason as claim 1.
EXAMINER RESPONSE #1: Applicant’s arguments with respect to claims 1, 4-18 and 21-22 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
Claims 1, 4-18 and 21-22 rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.
Independent claim 1, recites the limitation “testing, using at least one processor, the trained neural networks in the set to identify a plurality of functional blocks, each identified functional block being common to at least some of the trained neural networks in the set, wherein testing the trained neural networks in the set comprises...” (emphasis added). However, the specification does not teach or suggest the trained neural network being tested in order to identify the plurality of functional block. Rather, as can be seen in paragraphs, [0005], [0007], [0016-17] of the instant application, the trained neural networks are being inspected not tested. As would be familiar to one skilled in the art, testing a model and inspecting in the context of machine learning differ, as testing refers to assess the model performance on test data while inspecting involves analyzing the internal working, structure or component of the model. Therefore, the specification, does not provide support for the limitation of “testing” the trained neural networks as presented in amended claim 1.
Independent claims 21 and 22 recites similar limitation to those of claim 1, thus are rejected for reasons set forth in the rejection of claim 1.
Claims 4-18 are dependent on claim 1, and thus are rejected for reasons set forth in the rejection of claim 1.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 13-16, 18 and 21-22 are rejected under 35 U.S.C. 103 as being unpatentable over YU et al. US 2022/0405579 A1 (hereinafter Yu) in view of KIM et al. US 2021/0264237 A1 (hereinafter Kim) in further view of Gyungmin KIM US 2021/0182670 A1 (hereinafter Gyungmin).
Regarding claim 1:
Yu teaches A computer-implemented method comprising: (Yu [0018] teaches a computer implemented method).
obtaining a set of trained neural networks for performing a common task and test data for evaluating the performance of the trained neural networks in the set when performing said task; (Yu [0021] teaches obtaining training data (test data) that includes a “set of neural network inputs and, for each network input, a respective target output that should be generated by the neural network to perform the particular task”. Furthermore, Yu [0024] teaches a plurality of neural networks (trained neural networks) and [0026] teaches a system can determine a neural network that performs the machine learning task within a specified set of resource constraints).
testing, using at least one processor, the trained neural networks in the set to identify a plurality of functional blocks, each identified functional block being common to at least some of the trained neural networks in the set, wherein testing the trained neural networks in the set comprises: ( Yu [0053] teaches training/testing the set of neural networks. Further, Yu [0057] teaches a plurality of neural networks were “each neural network in the subset has a respective architecture selected from a proper subset of the search space of different architectures” and [0024] teaches “each of the neural networks has parameters (functional blocks) that are a subset of the shared set, with different neural networks having different subsets of the shared set”).
for each identified functional block: extracting a respective network component for implementing the identified functional block within each of at least some of the trained neural networks; and (Yu [0052] teaches each set of parameters (functional blocks) are a subset of a shared set of parameters (network component) and [0056] teaches “each of the plurality of neural networks has parameters that are a subset of the shared set and each of the plurality of neural networks has a respective architecture selected from a search space of different architectures that is defined by a respective set of possible values for each of a plurality of architectural dimensions”).
for each extracted network component: measuring performance characteristics of the network component when processing the test data; and ( Yu [0070] teaches in order to “determine the respective performance benchmark for a neural network, the system can determine the accuracy or other appropriate performance measure for the machine learning task on a data set”).
receiving, via an interface, a request to synthesize a neural network for performing said task subject to a given set of constraints, wherein the given set of constraints is based at least in part on specific hardware of a given device; and ( Yu [0021] teaches training data included a set of neural network input and [0022] teaches receiving training data (i.e., input) from a “remote user of the system over a data communication network, e.g., using an application programming interface (API) made available by the system”. A person skilled in the relevant art will recognize API are a type of software interface. Furthermore, Yu [0027] teaches the system can receiving an input (request) from a user of the system that specifies the set of resource constraints (i.e., “specify constraints on how many computational resources are consumed by the neural network when performing the task when deployed on a target set of hardware devices” see [0026]) or can automatically determine the set of resource constraints based on computational resources that are available to the system. In addition, Yu [0032] teaches the system can select a neural network that satisfies the constrains as the neural network to be used for performing the task).
composing a plurality of network components and in dependence on the performance data and the given set of constraints, thereby to synthesize a neural network for performing said task with the given device (Yu [0032] teaches selecting a neural network that satisfies the constrains as the neural network to be used for performing the task . Further, Yu [0059] teaches “the system can select a proper subset of the possible values for each of the architectural dimensions and then determines a performance benchmark for each combination of the proper subsets that yields a neural network that satisfies the constraints” to therefore “identify a single architecture or a range of architectures that can be deployed effectively on edge devices for any given machine learning task” ([0008])).
Yu does not specifically teaches processing, using the at least one processor, a common test data item with a plurality of trained neural networks in the set; recording, using the at least one processor during the processing of the common test data item, activations of network layers within the plurality of trained neural networks; comparing the recorded activations of network layers between the trained neural networks when processing the common test data item; and identifying, based on said comparing of the recorded activations, a functional block as corresponding to a contiguous groups of layers within at least some of the trained neural networks having consistently alike input activations and output activations to one another; for each extracted network component: storing performance data indicating the measured performance characteristics of the network component when processing the test data; storing configuration data indicating a configuration of the identified plurality of common functional blocks within said plurality of the trained neural networks; composing a plurality of network components in accordance with the stored configuration data...
Nevertheless, Kim teaches the following:
comparing the activations of network layers between the trained neural networks when processing the common test data item; and
PNG
media_image1.png
649
960
media_image1.png
Greyscale
(Kim [0058] teaches “the neural network analysis module may determine similar layer groups by directly comparing the layer groups” such that it can identify layers having similar structures by considering types and the number of respective layers of the layer groups, a kernel size, the number of input samples, the number of output samples, and the like”. Furthermore, to clarify, [0061] teaches “the neural network analysis module 143 of FIG. 3 analyzes the first to third neural networks NN1 to NN3” and teaches Fig. 5 the neural networks (NN1 to NN3) processing input sample (common test data item). See Fig. 5 above, with emphasize on “input sample”).
identifying, based on said comparing of the activations, a functional block as corresponding to a contiguous groups of layers within at least some of the trained neural networks having consistently alike input activations and output activations to one another; (To clarify, Kim teaches neural network having input activations and output activations, specifically [0023] teaches the artificial neural network (ANN) having a structure in which artificial neurons are connected to process received signal and transmit the signal to other neurons, where the output of the neuron is referend to an activation. Thus, a person skilled in the relevant art will recognize that the neural network described consist of both input and output activation since the output of the neuros “activation” is a result of processing input activation through the layers of the ANN. Furthermore, Kim [0062] and FIG. 5 teaches the “neural network analysis module” element 143 can group layers having structures that are “similar (alike) to those of reference neural networks from among the layers of the first neural network NN1, and thus may determine layer groups of the first neural network NN1”.)
storing performance data indicating said performance of the network component when processing the test data; ; (Kim [0039] teaches storing the sharing layer group in Random Access Memory (RAM) or in the memory and [0060] teaches “the neural network analysis module” can determine the sharing layer group having an excellent performance such as “a layer group capable of performing an operation with a smaller number of nodes or layers”. Since the layer group are stored in memory, the performance data determined could also be store in memory).
storing configuration data indicating a configuration of the identified plurality of common functional blocks within said plurality of the trained neural networks; (Kim [0039] teaches storing the sharing layer group in (RAM) or in the memory. This implies that the configuration data associated with the layer group will be also stored in RAM).
composing a plurality of network components in accordance with the stored configuration data and in dependence on the performance data and the given set of constraints, thereby to synthesize a neural network for performing said task with the given device (Kim [0039] teaches how some layer groups are stored in RAM or in memory. This implies that their performance data and configuration data will also be store within. Furthermore, Kim [0051] teaches when the neural network device receives an operation request, it identify sharing layer group forming neural network required for the operation).
Kim is also in the same field of endeavor as Yu (neural networks architectures). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of neural networks components, as being disclosed and taught by Kim, in the system taught by Yu to yield the predictable results of improve the efficiency of the storage space of an electronic device (see Kim [0080]).
Neither Yu or Kim teaches processing, using the at least one processor, a common test data item with a plurality of trained neural networks in the set; recording, using the at least one processor during the processing of the common test data item, activations of network layers within the plurality of trained neural networks; comparing the recorded activations of network layers between the trained neural networks when processing the common test data item;
However, Gyungmin teaches the following:
processing, using the at least one processor, a common test data item with a plurality of trained neural networks in the set; ( Gyungmin [0101] teaches a first module that implements a “first neural network” and a second module that implements a “second neural network”, further [0113] teaches providing to the first and second module the same test data ( i.e., common test data) for processing).
recording, using the at least one processor during the processing of the common test data item, activations of network layers within the plurality of trained neural networks; (Gyungmin [0004] teaches a neural network device is used to implement the neural networks, and [0059] teaches how for example, a neural network that is being implemented by the “neural network device” can generate output activation (i.e. activations of network layers). Moreover, Fig. 11 and [0128] teaches a memory (element 1110) that is nested within the neural network device (element 1100) stored various pieces of data that is being processed by the neural network device, thus suggesting the various pieces of data such as output activations can be recorded in memory for later retrieval to perform for example, activations comparations as described in [0096]).
comparing the recorded activations of network layers between the trained neural networks when processing the common test data item; and (Gyungmin [0090-91] teaches obtaining a “first data” generated by the first module which might include “first output data generated as a result of the operation of the layer of the first neural network “and [0093-94] teaches obtaining a “second data” generated by the second module which might include “second output data generated as a result of the operation of the layer of the second neural network”. Further, [0096] specifically teaches comparing “...the first input data of the first module with the second input data of the second module, compare the first output data with the second output data...”).
...based on said comparing of the recorded activations,... ( Gyungmin [0096] teaches comparing the activations).
Gyungmin is also in the same field of endeavor as Yu and Kim (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of processing test data by the multiple neural networks, recording various pieces of data and comparing the data generated by the neural networks that include output data generated as result of the operation of the layer of the neural networks, as being disclosed and taught by Gyungmin, in the system taught by Yu and Kim to yield the predictable results of “...provide various processing functions, such as performing data augmentation on input data provided to the neural network, generating the neural network, training the neural network, quantizing parameters of the neural network, or performing optimization to tune training parameters of the network” (Gyungmin [0075]).
Regarding claim 13:
Yu, Kim and Gyungmin teach The computer-implemented method of claim 1. Yu specifically teach wherein composing the plurality of network components comprises selecting a plurality of the extracted network components, using the stored performance data, for compliance with the given set of constraints (Yu [0032] teaches the system selects a neural network that has a proper subset of the shared set of parameters (network components) and that satisfies the constraints as the neural network to be used for performing the task”).
Regarding claim 14:
Yu, Kim and Gyungmin teach The computer-implemented method of claim 1. Yu specifically teaches wherein the given set of constraints includes at least one of an accuracy constraint, a memory constraint, a processing operation constraint, an execution time constraint, a latency constraint, and an energy consumption constraint (Yu [0026] teaches where the given constrain can include “resource constraints” that specify how many computational resources will be consumed by the neural network when performing the task when deployed on a target set of hardware devices).
Regarding claim 15:
Yu, Kim and Gyungmin teach The computer-implemented method of claim 1. Yu specifically, teaches wherein the request indicates an order of priority for the given set of constraints; and (Yu [0026] teaches where the given constrain can include “resource constraints” that “specify how many computational resources will be consumed by the neural network” by specifying the computation resources it’s possible to indicate an order of priority).
the composing of the plurality of network components is dependent on the indicated order of priority for the given set of constraints ( Yu [0032] teaches the system can select the neural network that has a proper subset of the shared set of parameters (network components) and that satisfied the constraints as the neural network to be used for performing the task).
Regarding claim 16:
Yu, Kim and Gyungmin teach The computer-implemented method of claim 1. Kim specifically teaches wherein the set of trained neural networks is a first set, the method further comprising: (Kim [0042] teaches a set of neural networks).
obtaining one or more further sets of trained neural networks, the trained neural networks in each further set configured to performing a respective further common task; and
(Kim [0023] teaches artificial neural networks that can learn to perform tasks according to predefines conditions and [0042] teaches obtaining sets of trained neural networks NN1, NN2, ... , and NNn).
inspecting the trained neural networks in the one or more further sets to identify that at least some of the plurality of functional blocks are common to at least some of the trained neural networks in the first set and the one or more further sets, (Kim [0036] teaches “network reconstruction module may receive and analyze the neural networks and thus determine a layer group...that the neural networks may commonly use”).
wherein the composed plurality of network components includes at least one network component derived from the trained neural networks in the one or more further sets (Kim [0051] teaches the relearning model can store the neural networks and the layer group such that when it receive an operation request it can identify layer groups forming neural networks requires for operation).
Regarding claim 18:
Yu, Kim and Gyungmin teach The computer-implemented method of claim 1. Yu specifically teaches further comprising training the synthesized neural network using machine learning (Yu [0052] teaches training a neural network that satisfies a specific set of constrains such that it can be selected and deployed).
Regarding claim 21: is rejected under the same rational of claim 1. Claim 21 only recites the additional elements of One or more non-transitory storage media comprising computer-readable instructions which, when executed by one or more processors, cause the one or more processors to carry out a method comprising... for which Yu [0093] teaches “one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible non-transitory storage medium for execution by, or to control the operation of, data processing apparatus”.
Regarding claim 22: is rejected under the same rational of claim 1. Claim 22 only recites the additional elements of A system comprising: at least one processor; and at least one non-transitory storage media comprising computer-readable instructions which, when executed by the at least one processor, cause the at least one processor to carry out operations comprising..., for which Kim Fig. 4 teaches a system element 300 with a processor element 310 and Yu [0093] teaches “one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible non-transitory storage medium for execution by, or to control the operation of, data processing apparatus”.
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Yu, Kim, Gyungmin in further view of Xue et al. US 20220198260 A1 (hereinafter Xue).
Regarding claim 4:
Yu, Kim and Gyungmin teach The computer-implemented method of claim 1.
Neither Yu, Kim and Gyungmin teach wherein comparing the activations of the network layers between the trained neural networks is performed on the basis of a random search or a grid search.
However, Xue teaches the following:
wherein comparing the activations of the network layers between the trained neural networks is performed on the basis of a random search or a grid search (Xue teaches [0041] using a “grid architecture search or a random architecture search” in order to in order to analyze the hyperparameters and compare the performance of different networks architectures during the use of neural architecture search (NAS)).
Xue is also in the same field of endeavor as Yu, Kim and Gyungmin (automated machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of random and grid search, in the system taught by Xue, in the system taught by Yu, Kim and Gyungmin to yield the predictable results of improve the speed and accuracy of mobile devices such that by using NAS it’s possible to “to build a CNN model to fit a particular problem defined by multiple objectives in a multi-level hierarchy based on hyperparameters for various conditions and/or constraints” ( Xue [0040]).
Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over Yu, Kim, Gyungmin in further view of Cao et al. WO 202202793 A1 (hereinafter Cao).
Regarding claim 5:
Yu, Kim and Gyungmin teach The computer-implemented method of claim 1. Yu, Kim and Gyungmin do not disclose wherein comparing the activations of the network layers between the trained neural networks uses meta-learning.
However, Cao teaches the following:
wherein comparing the activations of the network layers between the trained neural networks uses meta-learning (Cao pg. 5, Detail Description, 3rd paragraph teaches supervised and unsupervised meta-learning which can be used to compare the activation of the network layers).
Cao is also in the same field of endeavor as Yu, Kim and Gyungmin (neural architecture search). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of meta-learning, as being disclosed and taught by Cao, in the system taught by Yu, Kim and Gyungmin to yield the predictable results of optimize neural network architectures for performing a machine learning task (Cao pg. 3, Summary of Invention, 1st paragraph).
Claims 6-9 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Yu, Kim, Gyungmin and in further view of Dohan et al. US10,997,503 B2 (hereinafter Dohan).
Regarding claim 6:
Yu, Kim and Gyungmin The computer-implemented method of claim 1. Kim specifically teaches further comprising, for at least one identified functional block, processing the extracted network components for implementing the functional block, (Yu [0052] teaches how each set of parameters (functional blocks) are a subset of a shared set of parameters (network component) and [0056] teaches “each of the plurality of neural networks has parameters that are a subset of the shared set and each of the plurality of neural networks has a respective architecture selected from a search space of different architectures that is defined by a respective set of possible values for each of a plurality of architectural dimensions”).
Neither Yu, Kim or Gyungmin disclose using machine learning, to generate one or more further network components for implementing the functional block, wherein the composed plurality of network components includes at least one of the generated further network components.
However, Dohan teaches the following:
...using machine learning, to generate one or more further network components for implementing the functional block, (Dohan (col.6:16-20) teaches the “new architecture generator” that can generate a new candidate architecture based on the one or more selected candidate architectures and Dohan (col.12:3-10) teaches the candidate architectures includes a stack of cell (stack of operation blocks)).
wherein the composed plurality of network components includes at least one of the generated further network components (Dohan (col.6: 26-35) teaches the “new architecture generator” can randomly select a mutation from a set of mutations, and apply the randomly selected mutation to the selected candidate architecture and (col.6: 26-35) teaches the “mutations can include any of a variety of architecture modifications that represent the addition, removal, or modification of a component (e.g., a layer or a structure) from an architecture or a change in a hyper-parameter for the training of the neural network having the architecture”).
Dohan is also in the same field of endeavor as Yu, Kim and Gyungmin (neural network architecture search). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of new architecture search and mutations, as being disclosed and taught by Dohan, in the system taught by Yu, Kim and Gyungmin to yield the predictable results of optimize neural network architecture for performing the machine learning task such that “a system can enable a population of candidate architectures to improve over time, resulting in an optimized neural network having better performance (e.g., better accuracy) when performing the particular machine learning task compared to existing neural network architecture search methods” (Dohan col. 2:49-54).
Regarding claim 7:
Yu, Kim, Gyungmin and Dohan teach The computer-implemented method of claim 6. Yu specifically teach further comprising, for said at least one functional block (Yu [0052] teaches each set of parameters (functional blocks).
evaluating performance of the ( Yu [0070] teaches in order to “determine the respective performance benchmark for a neural network, the system can determine the accuracy or other appropriate performance measure for the machine learning task on a data set”).
wherein the composing of the plurality of network components is (Further, Yu [0027] teaches receiving an input (request) and [0032] teaches selecting a neural network that satisfies the constrains as the neural network to be used for performing the task . Further, Yu [0059] teaches “the system can select a proper subset of the possible values for each of the architectural dimensions and then determines a performance benchmark for each combination of the proper subsets that yields a neural network that satisfies the constraints).
Yu does not disclose storing further performance data indicating said performance of the
However, Kim teaches the following:
storing further performance data indicating said performance of the (Kim [0039] teaches storing the sharing layer group in Random Access Memory (RAM) for rapid calculation and the other layers group in the memory and [0060] teaches “the neural network analysis module” can determine the sharing layer group having an excellent performance such as “a layer group capable of performing an operation with a smaller number of nodes or layers”. Since the layer group are stored in memory, the performance data determined could also be store in memory).
While Yu, Kim and Gyungmin does not specifically evaluate performance of the one or more further network components.. storing further performance data indicating said performance of the one or more further network components.. wherein the composing of the plurality of network components is
However, Dohan teaches the following:
Dohan introduces mutations that enables adding new components into the candidate architecture. Specifically, Dohan (col. 6: 16-20) teaches “ new architecture generator” that can generate a new candidate architecture based on the one or more selected candidate architecture and (col.12: 29-33) teaches the system can applies mutations by replacing an existing component in the architecture with a new component and adding the new component into the architecture therefore creating further network components.
Regarding claim 8:
Yu, Kim, Gyungmin and Dohan teach The computer-implemented method of claim 6. Yu specifically teaches wherein generating the one or more further network components is performed in response to receiving the request for the neural network (Yu [0027] teaches the system can receiving an input (request) and [0032] teaches the system can select a neural network that satisfies the constrains as the neural network to be used for performing the task).
Regarding claim 9:
Yu, Kim, Gyungmin and Dohan teach The computer-implemented method of claim 6. Yu specifically teach wherein the processing of the extracted network components using machine learning uses neural architecture search ( Yu FIG. 1 shows an example neural architecture search system element 100).
Regarding claim 17:
Yu, Kim, and Gyungmin The computer-implemented method of claim 15, wherein the request for a neural network is a first request, the method further comprising. Yu specifically, teaches receiving a ( Yu [0027] teaches the system can receiving an input (request) from a user of the system that specifies the set of resource constraints or can automatically determine the set of resource constraints based on computational resources that are available to the system. Further, [0032] teaches the system can select a neural network that satisfies the constrains as the neural network to be used for performing the task).
composing a ( Yu [0027] teaches receiving an input (request) and [0032] teaches selecting a neural network that satisfies the constrains as the neural network to be used for performing the task).
Yu, Kim, and Gyungmin do not disclose the limitation of further request, further task, further given set of constrains, further plurality of network components. However, Dohan teaches the following:
Dohan (col. 5 :36-40) teaches in order to determine an optimized architecture, the optimization system is capable of performing multiple iterations of a search process to repeatedly update the candidate architectures in a population of candidate architecture. Since the optimization system enable to perform multiple iteration it’s possible to obtain further request, further task, further given set of constrains, further plurality of network components.
Claims 10 and 11 are rejected under 35 U.S.C. 103 as being unpatentable over Yu, Kim, Gyungmin, Dohan in further view of Changiz Rezaei et al. US 20230140142 A1 (hereinafter Changiz).
Regarding claim 10:
Yu, Kim, Gyungmin, and Dohan teach The computer-implemented method of claim 6.
Neither Yu, Kim, Gyungmin, and Dohan disclose wherein the processing of the extracted network components using machine learning comprises training a generative model to generate the further network components.
However, Changiz teaches the following:
wherein the processing of the extracted network components using machine learning comprises training a generative model to generate the further network components ( Changiz [0011] teaches implementing generative model such as generative adversarial network (GAN) or any other generator model such as variational auto encoders (VAEs) and [0020] teaches how the GAN implemented has a generator that can generate a plurality of neural network architectures (further network components).
Changiz is also in the same field of endeavor as Yu, Kim, Gyungmin, and Dohan (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of the generative adversarial network, as being disclosed and taught by Changiz, in the system taught by Yu, Kim, Gyungmin, and Dohan to yield the predictable results of improve the performance of a device such that “GA-NAS may improve the performance of smartphones and chipsets by identifying deep learning models to be used thereon. [Therefore, ] the optimized neural network models can boost the performance of smartphones and chipsets”( Changiz [0012]).
Regarding claim 11:
Yu, Kim, Gyungmin, Dohan and Changiz teach The computer-implemented method of claim 10, Changiz specifically teaches wherein said training comprises adversarial training (Changiz [0011] teaches using GAN which inheritably is trained using adversarial training).
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over Yu, Kim, Gyungmin, Dohan in further view of Cao.
Regarding claim 12:
Yu, Kim, Gyungmin, and Dohan teach The computer-implemented method of claim 6. Neither Yu, Kim, Gyungmin, or Dohan wherein said processing of the extracted network components using machine learning uses knowledge distillation or model compression.
However, Cao teaches the following:
wherein said processing of the extracted network components using machine learning uses knowledge distillation or model compression ( Cao pg. 4, 6th-10th paragraphs, teaches a neural network compression device that is “configured to predict the optimal network structure under the target constraint condition through the target element generation network, and obtain the compressed neural network model”).
Cao is also in the same field of endeavor as Yu, Kim, Gyungmin and Dohan (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of a neural network compression device, as being disclosed and taught by Cao, in the system taught by Yu, Kim, Gyungmin and Dohan to yield the predictable results of optimize neural network architectures for performing a machine learning task (Cao pg. 3, Summary of Invention, 1st paragraph).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GISEL G FACCENDA whose telephone number is (703)756-1919. The examiner can normally be reached Monday - Friday 8:00 am - 4:00 pm.
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, Abdullah Al Kawsar can be reached at (571) 270-3169. 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.
/G.G.F./Examiner, Art Unit 2127
/ABDULLAH AL KAWSAR/Supervisory Patent Examiner, Art Unit 2127
Prosecution Insights