DETAILED ACTION
This action is responsive to the application filed on 02/02/2026. Claims 1-16, 18-22 are pending and have been examined. This action is Final.
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 .
Priority
Applicant’s claim for the benefit of a prior-filed application under 35 U.S.C. 119(e) or under 35 U.S.C.
120, 121, 365(c), or 386(c) is acknowledged.
Response to Argument
Argument 1: Applicant argues that the amended claims are patent eligible under all three parts of the 101 framework because the added limitations require a specific technical process for neural network search, not an abstract mental process or generic computer implementation. Specifically, Applicant contends that selecting a mutated neural network before training and testing based on structural superiority criteria, such as greater channel quantity or a greater number of convolutional layers, cannot practically be performed in the human mind and is instead a technical operation on neural network architectures under Step 2A, Prong 1. Applicant further argues under Step 2A, Prong 2 that these limitations are integrated into a practical application because they improve neural architecture search efficiency by pruning inferior networks before training and testing, thereby reducing the search space and conserving computing resources. Finally, under Step 2B, Applicant argues that the ordered combination of pre-training structural selection followed by post-training evaluation using P evaluation parameters is unconventional and amounts to significantly more than an abstract idea.
Response to Argument 1: The examiner has considered the argument set forth above, however they are not persuasive. Regarding Step 2A, Prong 1, although Applicant argues that the amended claims recite specific technical operations that cannot be performed mentally, the updated rejection explains that the claim still recites mutating a neural network structure, selecting a neural network structure, and determining whether one structure is “superior” by comparing numerical quantities, including whether one channel quantity is greater than another channel quantity or whether one quantity of convolutional layers is greater than another quantity of convolutional layers; these are rule-based evaluations, judgments, and mathematical comparisons, and therefore remain directed to a mathematical concept and/or mental process. Regarding Step 2A, Prong 2, although Applicant argues that the claims improve neural network search efficiency, the additional claim elements merely apply the abstract selection and comparison rules using a generic computing device, dataset, and neural networks; obtaining a dataset and neural networks is insignificant extra-solution activity, and performing the recited steps “by a computing device” amounts to no more than instructions to apply the abstract idea on a computer. Regarding Step 2B, the claim elements, individually and as an ordered combination, do not amount to significantly more because the claim relies on conventional computer implementation and routine data handling, including obtaining data, evaluating parameters, and selecting neural networks according to recited rules, without reciting a specific improvement to computer functionality or a particular technological mechanism beyond the abstract architecture-search process itself. Claim 19 remains non-eligible for similar reasons because the object-recognition steps merely obtain an image, input the image to a neural network, and output an object type, which are insignificant extra-solution input/output activities, and the remaining neural-network search limitations are analogous to claim. Claims 9, 18, and 22 are likewise analogous to claims 1 and 19 and remain rejected for the same reasons.
Argument 2: Applicant argues that Real does not teach the amended limitation requiring selection “from the mutated neural network and before training and testing the mutated neural network.” Applicant asserts that Real trains and evaluates the child architecture after mutation and only then uses validation fitness for selection. Applicant further argues that Bartz-Beielstein does not cure this deficiency because Bartz-Beielstein’s selection is based on “function value” or “fitness value,” which Applicant characterizes as performance-based evaluation rather than pre-training structural selection based on channel quantity or convolutional-layer quantity.
Response to Argument 2: Applicant’s argument is not persuasive because the rejection does not rely on Real alone, or on Bartz-Beielstein alone, to teach the entire amended limitation. Real is relied upon for the evolutionary neural architecture search framework, including obtaining a dataset and neural networks, performing evolutions, mutating a prior architecture, and teaching structural scaling through output filters and normal cells per stack. Bartz-Beielstein is relied upon for the parent-offspring selection framework, namely selecting the better of the offspring and parent so that the selected network advances through the evolutionary process. The new reference, Amizadeh, is relied upon specifically for the timing limitation because Amizadeh teaches selecting a neural network architecture before the architecture is trained and avoiding training fewer desirable architectures. Thus, Real in view of Bartz-Beielstein and further in view of Amizadeh teaches or suggests the amended pre-training selection limitation.
Argument 3: Applicant argues that the cited references do not teach filtering mutated networks based on structural superiority before training to reduce the search space. Specifically, Applicant contends that the prior art does not teach using channel quantity or convolutional-layer quantity as structural criteria during an evolutionary neural network search process before training and testing.
Response to Argument 3: Applicant’s argument is not persuasive because Real teaches that architecture size and accuracy may be altered using the number of output filters of convolution operations and the number of normal cells per stack. The number of output filters corresponds to channel quantity, and increasing the number of normal cells per stack corresponds to increasing the effective depth or quantity of convolutional layers. Real therefore teaches structural characteristics that correspond to the claimed channel-quantity and convolutional-layer-quantity criteria. Amizadeh further teaches selecting architectures before training to avoid training fewer desirable architectures. Accordingly, it would have been obvious to apply Amizadeh’s pre-training architecture selection in Real’s evolutionary neural architecture search process, as modified by Bartz-Beielstein, to reduce unnecessary training and testing of less desirable candidate networks. Claim 19 remains unpatentable for the same reasons because it recites materially similar neural network search limitations in the object-recognition context, and the remaining unchanged or analogous claims stand with claims 1 and 19.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition
of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the
conditions and requirements of this title.
Claims 1-16, 18-22 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1,Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
Step 2A Prong 1:
“mutating, … a network structure of a neural network obtained through an(i-1)th evolution, to obtain a mutated neural network;” -- The limitation is directed to mutating a network structure through an evolution to obtain a mutated type of a neural network. The limitation is directed to a mathematical calculation/operation, and thus is directed to math.
“selecting... and before training and testing the mutated neural network, a neural network whose network structure is superior to that of the neural network obtained through the (i-1)th evolution, to obtain a candidate neural network; and selecting,...from a set of the neural network obtained through the (i-1)th evolution and the candidate neural network and based on P evaluation parameters corresponding to each neural network in the set, a neural network obtained through the ith evolution, wherein the P evaluation parameters are for evaluating performance of a neural network obtained after each neural network in the set is trained and tested by using the dataset, i and P are positive integers, and 1≤i≤K.” -- The limitation recites selecting a neural network, before the training and testing of the mutated NN, structure that is superior to a neural network that was obtained through the evolution, to obtain a candidate neural network, and from the set of the neural network obtained by the evolution, the candidate neural network, and the P evaluation parameters, a neural network is then obtained once more, the parameters are for evaluating performance and training. The limitation is directed to a process that can be performed in the human mind using evaluation, observation, and judgement, and thus the limitation is directed to a mental process.
“wherein the network structure of the neural network is superior to that of the neural network obtained through the (i-1)1h evolution when at least one of the following conditions is met: a channel quantity of the neural network is greater than a channel quantity of the neural network obtained through the (i-l)th evolution, or a quantity of convolutional layers in the neural network is greater than a quantity of convolutional layers in the neural network obtained through the (i-1)th evolution” -- The limitation is directed to a mental process and/or mathematical concept because it determines whether one neural network structure is “superior” by comparing numerical quantities. Specifically, the limitation compares whether one channel quantity is greater than another channel quantity, or whether one quantity of convolutional layers is greater than another quantity of convolutional layers. Such greater-than comparisons are evaluations or judgments that can be performed mentally or with pen and paper, and also constitute mathematical comparisons of numerical values.
Step 2A Prong 2 and Step 2B:
“by a computing device...by the computing device from the mutated neural network” -- The limitation recites that tasks will be performed by the computing device. The limitation amounts to no more than mere instructions to apply onto a computer, and thus it does not integrate to a practical application, nor does it provide significantly more than the judicial exception (see MPEP 2106.05(f)).
“obtaining...a dataset and N neural networks, wherein N is a positive integer;” -- The limitation recites obtaining a dataset and neural network, to which N would be a positive integer amount of neural networks that are obtained. The limitation is directed to an insignificant, extra-solution activity that cannot be integrated to a practical application (see MPEP 2106.05(g)). Furthermore, under Step 2B, the act of obtaining data/a dataset is a well-understood, routine, and conventional activity that cannot provide significantly more than the judicial exception (see MPEP 2106.05(d)(II)).
Thus, claim 1 is non-patent eligible. Claim 9 and 18 is analogous to claim 1 (aside from claim type/ slight differences like “0th evolution” and “Kth evolution”, claim 9’s process/memories and the CRSM from claim 18, all of which are small and don’t change the scope of the claim), and thus would face the same rejection as set forth above.
Regarding claim 2,
Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
Step 2A Prong 1:
“The method according to claim 1, wherein the mutating, by the computing device, [[a]]the network structure of a neural network obtained through the (i-1)th evolution comprises at least one of the following steps: swapping locations of two convolutional layers in one or more neural networks of the neural network obtained through the (i-1)th evolution; doubling a channel quantity of one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution; doubling a step size of a convolution kernel of the one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)'h evolution;” - This recites rule-based manipulations of a model’s architecture (layer order, channel counts, stride values, presence/absence of layers). Such operations are mathematical relationships/operations applied to an abstract model graph and thus the limitation is directed to math.
Step 2A Prong 2 and Step 2B:
“inserting the one or more convolutional layers into the one or more neural networks of the neural network obtained through the (i-1 evolution; deleting the one or more convolutional layers from the one or more neural networks of the neural network obtained through the (i-I) evolution…inserting one or more pooling layers into the one or more neural networks of the neural network obtained through the (i-I) evolution; or deleting the one or more pooling layers from the one or more neural networks of the neural network obtained through the (i-l)' evolution.” -- The limitation recites inserting layers to neural networks obtained by the evolution, deleting layers from the neural works, and same for the pooling layer that are also obtained by the evolution. The limitation recites no more than mere instructions to apply onto a computer, and does not integrate to a practical application, nor does it provide significantly more than the judicial exception (see MPEP 2106.05(f)).
Thus, claim 2 is non-patent eligible. Claim 10 (difference: apparatus, mutation unit operations) and claim 11 (difference: apparatus, deep residual mutation operations) are analogous to claim 2 and thus will face the same rejection as set forth above.
Regarding claim 3,
Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
Step 2A Prong 1:
“doubling a channel quantity of one or more residual units in the one or more neural networks of the neural network obtained through the (i-1) evolution; doubling a step size of a convolution kernel of the one or more residual units in the one or more neural networks of the neural network obtained through the (i-1)th evolution;” -- The limitation is directed to doubling a channel quantity of the units of the neural network that is obtained through an i-1 evolution, and the same for the step size. The limitation is directed to a mathematical operation/calculation/relationship, and thus the limitation is directed to math.
Step 2A Prong 2 and Step 2B:
“The method according to claim 1, wherein the neural network obtained through the (i-1)* evolution is a deep residual network, and the mutating, by the computing device, [[a]]the network structure of a neural network obtained through the (i-1)* evolution comprises at least one of the following steps: swapping locations of two residual units in one or more neural networks of the neural network obtained through the (i-1) evolution; doubling a channel quantity of one or more residual units in the one or more neural networks of the neural network obtained through the (i-1) evolution; doubling a step size of a convolution kernel of the one or more residual units in the one or more neural networks of the neural network obtained through the (i-1)th evolution; inserting one or more residual units into the one or more neural networks of the neural network obtained through the (i-1)th evolution; or deleting the one or more residual units from the one or more neural networks of the neural network obtained through the (i-)th evolution.” -- The limitation recites the same limitations that were recited in claim 2 but further to include residual units vs convolutional layers. The limitation amounts to no more than mere instructions to apply onto a computer, and thus it does not integrate to a practical application, not does it provide significantly more than the juridical exception (see MPEP 2106.05(f)).
Thus, claim 3 is non-patent eligible. Claim 11 is analogous to claim 3 and will face the same rejection as set forth above.
Regarding claim 4,
Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
Step 2A Prong 1:
“The method according to claim 1, wherein the selecting, by the computing device from the mutated neural network, a neural network whose network structure is superior to that of the neural network obtained through the (i-1)th evolution, to obtain a candidate neural network comprises: selecting, by the computing device from a neural network obtained by mutating a first neural network, a neural network whose network structure is superior to that of the first neural network, wherein the candidate neural network comprises the neural network that is of the neural network obtained by mutating the first neural network and whose network structure is superior to that of the first neural network, and the first neural network is any neural network of the neural network obtained through the (i-1)th evolution.” -- The limitation recites selecting a neural network, by the computing device, whose structure is superior to a network that was obtained by the evolution for which will further comprise of selecting a neural network that is superior to the evolution from the device that was from the mutated neural network. The limitation is directed to a process that can be completed in the human mind using evaluation, observation, and judgment (with aid of pen and paper), and thus the limitation is directed to a mental process.
There are no elements to be evaluated under Step 2A Prong 2 and Step 2B.
Thus, claim 4 is non-patent eligible. Claim 12 is analogous to claim 4 and will face the same rejection as set forth above.
Regarding claim 5,
Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
There are no elements to be evaluated under Step 2A Prong 1.
Step 2A Prong 2 and Step 2B:
“The method according to claim 4, wherein when at least one of the following conditions is met, a network structure of the neural network obtained by mutating the first neural network is superior to the network structure of the first neural network: a channel quantity of the neural network obtained by mutating the first neural network is greater than a channel quantity of the first neural network; or a quantity of convolutional layers in the neural network obtained by mutating the first neural network is greater than a quantity of convolutional layers in the first neural network.” -- The limitation recites that when a condition is met, then a neural network obtained by a mutating first NN is superior to the network structure of the neural network, where a channel quantity obtained by the mutated NN is greater than the channel quantity from the first NNs, or the same is assessed in relation to the quantity of convolution layers. The limitation is directed to a process that can be completed in the human mind using evaluation, observation, and judgment (with aid of pen and paper), and thus the limitation is directed to a mental process.
There are no elements to be evaluated under Step 2A Prong 2 and Step 2B.
Thus, claim 5 is non-patent eligible. Claim 13 is analogous to claim 5 and will face the same rejection as set forth above.
Regarding claim 6,
Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
Step 2A Prong 1:
“ The method according to claim 1, wherein the selecting, by the computing device from a set of the neural network obtained through the (i-1)th evolution and the candidate neural network and based on P evaluation parameters corresponding to each neural network in the set, a neural network obtained through the ith evolution specifically comprises: determining,...that the neural network obtained through the in evolution is a neural network that is not dominated in the set, wherein a second neural network and a third neural network are two neural networks in the set, and if the second neural network is not inferior to the third neural network in terms of each of the P evaluation parameters, and the second neural network is superior to the third neural network in terms of at least one of the P evaluation parameters, the second neural network dominates the third neural network.” -- The limitation is directed to determining if a neural network obtained through evolution is dominated, with the other neural networks in the set, to also determine if the second NN is not inferior to the third NN based on the evaluation parameters/superior to the NN. The limitation is directed to a process that can be completed in the human mind using evaluation, observation, and judgment (with aid of pen and paper), and thus the limitation is directed to a mental process.
“performing...non-dominated sorting on the neural networks in the set based on the P evaluation parameters corresponding to each neural network in the set;” -- This recites a mathematical optimization procedure (multi-objective sorting over numerical metrics). It is a mathematical relationship/operation, and thus the limitation is directed to math.
There are no elements to be evaluated under Step 2A Prong 2 and Step 2B.
Thus, claim 6 is non-patent eligible. Claim 14 is analogous to claim 6 and will face the same rejection as set forth above.
Regarding claim 7,
Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
Step 2A Prong 1:
“and selecting, by the computing device, N neural networks from the M neural networks based on the P evaluation parameters corresponding to each of the M neural networks, wherein N is not greater than M.” -- The limitation is directed to selecting N networks from M networks based on the parameters that correspond to each M network. The limitation is directed to a process that can be completed in the human mind using evaluation, observation, and judgment (with aid of pen and paper), and thus the limitation is directed to a mental process.
Step 2A Prong 2 and Step 2B:
“The method according to claim 1, wherein the obtaining, by a computing device, N neural networks specifically comprises: randomly generating, by the computing device, neural networks, wherein M is a positive integer; training and testing, by the computing device, each of the M neural networks by using the dataset, to obtain P evaluation parameters corresponding to each of the M neural networks” -- The limitation recites obtaining neural networks by a computing device will comprise of random generations of the neural networks, training/testing of the networks using a dataset to obtain parameters,. The limitation is directed to an insignificant, extra-solution activity that cannot be integrated to a practical application (see MPEP 2106.05(g)). Furthermore, under Step 2B, the act of obtaining data using gathered datasets from the neural networks is a well-understood, routine, and conventional activity, and thus the limitation does not provide significantly more than the judicial exception (see MPEP 2106.05(d)(II)).
Thus, claim 7 is non-patent eligible. Claim 15 is analogous to claim 7 and will face the same rejection as set forth above.
Regarding claim 8,
Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
There are no elements to be evaluated under Step 2A Prong 1.
Step 2A Prong 2 and Step 2B:
“The method according to claim 1, wherein the P evaluation parameters comprise at least one of a running time, accuracy, and a parameter quantity.” --The limitation recites that the P evaluations parameters will further comprise a running time, accuracy and parameter quantity. The limitation merely further limits the claim to a field of use/environment, and thus it does not integrate to a practical application, nor does it provide significantly more than the judicial exception (see MPEP 2106.05(h)).
Thus, claim 8 is non-patent eligible. Claim 16 is analogous to claim 8 and will face the same rejection as set forth above.
Regarding claim 19, majority of claim 19 is analogous to claim 1, 9 and/or 18, aside from below:
Step 2A Prong 2 and Step 2B:
“An object recognition method, comprising: obtaining a to-be-recognized image; and inputting the to-be-recognized image to an object recognition neural network, to obtain an object type corresponding to the to-be-recognized image” -- The limitation recites a method that comprises obtaining an image, inputting the image to a neural network to be recognized, and obtain an object type that corresponds to the image. The limitation is directed to an insignificant, extra-solution activity that cannot be integrated to a practical application (see MPEP 2106.05(g)). Furthermore, under Step 2B, the act of inputting/outputting gather data over a network to obtain other data is a well-understood, routine, and conventional activity (WURC), that cannot provide significantly more than the judicial exception (see MPEP 2106.05(d)(II)).
“wherein the object recognition neural network is a target neural network determined from a search space by using a neural network search method...wherein the search space is built by using a basic unit and a parameter of the basic unit,” -- The limitation recites using a neural network search method to determine a search space, and building a search spaces using a unit and parameters of that unit. The limitation amounts to no more than mere instructions to apply onto a computer, and thus the limitation does not integrate to a practical application, nor does it provide significantly more than the judicial exception (see MPEP 2106.05(d)(II)).
Thus, claim 19 is non-patent eligible. Claim 22 is analogous to claim 19 and will face the same rejection as set forth above.
Regarding claim 20,
Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
There are no elements to be evaluated under Step 2A Prong 1.
Step 2A Prong 2 and Step 2B:
“The method according to claim 19, wherein the parameter of the basic unit comprises at least one of a type, a channel quantity parameter, and a size parameter of the basic unit.” -- The limitation recites the basic unit will further comprise at least one type, channel parameter and size parameter for the unit. The limitation merely further limits the claim to a field of use/environment, and thus it does not integrate to a practical application, nor does it provide significantly more than the judicial exception (see MPEP 2106.05(h)).
Thus, claim 20 is non-patent eligible.
Regarding claim 21,
Step 1: The claim is directed to a method, which falls under the category of a process. The claim satisfies step 1.
Step 2A Prong 1:
“The method according to claim 19, wherein the basic unit is configured to perform a first operation and a second operation on a feature map input to the basic unit, and wherein the feature map input is a feature map of the to-be-recognized image, the first operation is for doubling or maintaining a quantity of feature maps input to the basic unit, the second operation is for changing a size of the feature map input to the basic unit from an original first size to a second size or maintain a first size, and the first size is greater than the second size.” -- The limitation recites that the unit is configured to perform operations on a feature map input to the unit, and that the input is a feature map of the recognized image, and further the first operation is for doubling/maintaining feature maps, the second operations is for changing the size of the feature map, and that the first size is greater than the second size. The limitation is directed to mathematical operation/relationship, and thus the limitation is directed the math.
There are no elements to be evaluated over Step 2A Prong 2 and Step 2B.
Thus, claim 21 is non-patent eligible.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this
Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not
identically disclosed as set forth in section 102, if the differences between the claimed invention and the
prior art are such that the claimed invention as a whole would have been obvious before the effective filing
date of the claimed invention to a person having ordinary skill in the art to which the claimed invention
pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are
summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1, 4-7, 8-9, 12-16, 18-22 are rejected under 35 U.S.C. 103 as being unpatentable over Real in view of NPL reference “Evolutionary Algorithms” by Bartz-Beielstein et. al. (referred herein as Bartz-Beielstein) further in view of , by Amizadeh et. al. (referred herein as Amizadeh).
Regarding claim 1, Real teaches:
A neural network search method, comprising: obtaining, by a computing device, a dataset and N neural networks, wherein N is a positive integer, ([Real, page 4] “In particular, all methods used the same computer code for network construction, training and evaluation. Experiments always searched on the CIFAR-10 dataset” and [Real, page 3] “It keeps a population of P trained models throughout the experiment. The population is initialized with models with random architectures”, wherein the examiner interprets experiments always searched on the CIFAR-10 dataset to be the same as obtaining, by a computing device, a dataset because they are both the explicit acquisition and use of data for training and evaluation by the computing device. The examiner further interprets a population of P trained models to be the same as N neural networks, wherein N is a positive integer because they are both a finite, countable set of neural networks maintained for the search.)
performing, by the computing device, K evolutions on the N neural networks to obtain a neural network obtained through a Kth evolution, wherein K is a positive integer, ([Real, page 3, Algorithm 1] “Algorithm 1: … while |history| < C do . Evolve for C cycles.” and “After this, evolution improves the initial population in cycles”, wherein the examiner interprets evolve for C cycles / improves...in cycles to be the same as performing K evolutions...to obtain a neural network obtained through a K-th evolution, wherein K is a positive integer because they are both iterative, indexed evolutionary rounds executed by the computing device that yield a network at the K-th (final) cycle).
and an ith evolution of the K evolutions comprises: mutating, by the computing device, a network structure of a neural network obtained through an (i-1)th evolution, to obtain a mutated neural network; ([Real, page 3, Algorithm 1] “parent ← highest-accuracy model in sample ... child.arch ← MUTATE(parent.arch)” and “A mutation causes a simple and random modification of the architecture ... Only one of these mutations is applied in each cycle, choosing between them at random”, wherein the examiner interprets mutating a parent architecture to create a child and modifying the architecture in each cycle to be the same as mutating, by the computing device, a network structure of a neural network obtained through an (i-1)th evolution, to obtain a mutated neural network because they are both structural changes applied at the ith step to the previously obtained network to produce the mutated network.)
wherein the network structure of the neural network is superior to that of the neural network obtained through the (i-1)th evolution ([Real, page 3, sec Evolutionary Algorithm], “After this, evolution improves the initial population in cycles (“while |history|” in Algorithm 1). At each cycle, it samples S random models from the population, each drawn uniformly at random with replacement. The model with the highest validation fitness within this sample is selected as the parent.”, and [Real, page 3, Algorithm 1] “child.arch ← MUTATE(parent.arch),”wherein the examiner interprets the parent selected from the population during a given evolution cycle to be the same as the neural network obtained through the (i-1)th evolution because both are the prior evolutionary network used as the baseline for generating the next network. The examiner further interprets the child architecture generated by mutating the parent architecture to be the same as the neural network whose structure is compared against the neural network obtained through the (i-1)th evolution because both are directed to a next candidate network produced from the prior evolutionary baseline.)
when at least one of the following conditions is met: a channel quantity of the neural network is greater than a channel quantity of the neural network obtained through the (i-1)th evolution, ([Real, page 1] “Scaled to larger size, AmoebaNet-A sets a new state-of-the-art 83.9% top-1 / 96.6% top-5 ImageNet accuracy,” and [Real, page 3, sec Methods] “Once the architecture is specified, the model still has two free parameters that can be used to alter its size (and its accuracy): the number of normal cells per stack (N) and the number of output filters of the convolution ops (F),” wherein the examiner interprets increasing the number of output filters of the convolution operations to be the same as the channel quantity of the neural network being greater than the channel quantity of the neural network obtained through the (i-1)th evolution because output filters correspond to convolutional channels. The examiner further interprets scaling the architecture to a larger size improves accuracy to be the same as the larger-width neural network having a superior network structure relative to the prior evolutionary baseline.)
or a quantity of convolutional lavers in the neural network is greater than a quantity of convolutional layers in the neural network obtained through the (i-1)th evolution, ([Real, page 2, sec Methods] “normal cells are arranged in three stacks of N cells”, and [Real, page 3, sec Methods] “Once the architecture is specified, the model still has two free parameters that can be used to alter its size (and its accuracy): the number of normal cells per stack (N) and the number of output filters of the convolution ops (F), wherein the examiner interprets increasing the number of normal cells per stack to be the same as increasing a quantity of convolutional layers in the neural network relative to the neural network obtained through the (i-1)th evolution because the normal cells contain convolutional operations and adding more cells increases the depth of the convolutional architecture. The examiner further interprets increasing the depth of the architecture as a structurally superior network when compared to the prior evolutionary baseline because architecture’s size and accuracy are altered through the number of normal cells per stack.)
wherein the P evaluation parameters are for evaluating performance of a neural network obtained after each neural network in the set is trained and tested by using the dataset, ([Real, page 2, sec Methods] “Once the child architecture is constructed, it is then trained, evaluated...The model with the highest validation fitness within this sample is selected as the parent,” wherein the examiner interprets training and evaluating the child architecture and selecting based on highest validation fitness to be the same as using P evaluation parameters to evaluate performance of a neural network after training and testing by using the dataset because both are directed to computing performance measures after the neural network is trained and evaluated on data. The examiner further interprets validation fitness to be an evaluation parameter because it is a quantitative performance measure to help selection.)
i and P are positive integers, and 1<= i <= K. ([Real, page 3, Algorithm 1] “while |history| < C do . Evolve for C cycles” and [Real, page 3, sec Evolutionary Algorithm] “It keeps a population of P trained models throughout the experiment,” wherein the examiner interprets Real’s C evolution cycles to be the same as K evolutions with an index i satisfying 1 <= i <= K because both describe a positive integer-indexed evolutionary process. The examiner further interprets Real’s population of P trained models to be the same as P being a positive integer because both identify a countable number of models maintained during the search.)
Real does not teach selecting, by the computing device from the mutated neural network and before training and testing the mutated neural network, a neural network whose network structure is superior to that of the neural network obtained through the (i-1)th evolution, to obtain a candidate neural network …selecting, by the computing device from a set of the neural network obtained through the (i-1)th evolution and the candidate neural network and based on P evaluation parameters corresponding to each neural network in the set, a neural network obtained through the ith evolution.
Bartz-Beielstein teaches:
selecting, by the computing device from the mutated neural network…a neural network whose network structure is superior to that of the neural network obtained through the (i-1)th evolution, to obtain a candidate neural network, ([Bartz-Beielstein, page 5] “If the offspring solution is better (has a better function value) than the parent, take it as the new parent, otherwise retain the parent.”, wherein the examiner interprets taking the better offspring as the new parent to be the same as selecting, by the computing device from the mutated neural network, a neural network whose network structure is superior to that of the neural network obtained through the (i-1)th evolution, to obtain a candidate neural network because they are both a head-to-head comparison of offspring and its predecessor that keeps the superior network as the selected candidate)
and selecting, by the computing device from a set of the neural network obtained through the (i-1)th evolution and the candidate neural network and based on P evaluation parameters corresponding to each neural network in the set, a neural network obtained through the ith evolution, ([Bartz-Beielstein, page 5] “If the offspring solution is better (has a better function value) than the parent, take it as the new parent, otherwise retain the parent,” wherein the examiner interprets the parent to be the same as the neural network obtained through the (i-1)th evolution because both are the prior evolutionary network used as the baseline for the current selection step. The examiner further interprets the offspring solution to be the same as the candidate neural network because both are the newly generated candidate produced from the prior evolutionary baseline. The examiner further interprets the “function value” used to determine whether the offspring is better than the parent to be the same as P evaluation parameters corresponding to each neural network in the set because both are one or more objective performance-based values used to compare candidate solutions. Thus, selecting the better of the offspring and parent based on function value is the same as selecting, based on P evaluation parameters, between the candidate neural network and the prior-evolution neural network to obtain the neural network through the ith evolution.)
Real and Bartz-Beielstein does not teach…and before training and testing the mutated neural network.
Amizadeh teaches …and before training and testing the mutated neural network. ([Amizadeh, col 2-3, line 67, lines 1-3] “solving the technical problem of selecting, from a set of neural network architectures, a neural network architecture for solving a given machine learning problem before the neural network architecture is trained. Advantageously, as a result of some aspects, a high-performing neural network architecture is trained to solve the given machine learning problem, and less desirable architectures are not trained”; [Amizadeh, col 8, lines 29-33] “trains the architecture(s) ... and evaluates their performance”, wherein the examiner interprets selecting an architecture before the architecture is trained, and before later evaluation on a holdout set, to be the same as selecting before training and testing the mutated neural network because both are directed to screening or selecting a candidate neural network architecture before the selected candidate proceeds to training and performance evaluation/testing.)
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to neural network architecture search methods that generate, evaluate, and select candidate neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the neural network search method disclosed by Real to include the child-parent selection technique disclosed by Bartz-Beielstein. One would be motivated to do so to efficiently select the better candidate network during each evolutionary step and avoid advancing a lower-quality network, as suggested by Bartz-Beielstein ([Bartz-Beielstein, page 5] “If the offspring solution is better ... than the parent, take it as the new parent, otherwise retain the parent.”).
It would have also been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the neural network search method disclosed by Real to include the pre-training selection technique disclosed by Amizadeh. One would be motivated to do so to efficiently avoid training and testing less desirable candidate architectures and reduce unnecessary computational cost, as suggested by Amizadeh ([Amizadeh, col. 2-3, line 67, lines 1-3] “less desirable architectures are not trained.”). Claims 9, 18, 19, and 22 are analogous to claim 1, aside from claim type and small differences for which do not drastically change the scope, and thus would face the same rejection as set forth above.
Regarding claim 4, Real, Bartz-Beielstein, and Amizadeh teaches The method according to claim 1, (see rejection of claim 1).
Real further teaches wherein the selecting, by the computing device from the mutated neural network, a neural network whose network structure is superior to that of the neural network obtained through the (i-1)th evolution, to obtain a candidate neural network comprises: selecting, by the computing device from a neural network obtained by mutating a first neural network,.., wherein the candidate neural network comprises the neural network that is of the neural network obtained by mutating the first neural network and whose network structure is superior to that of the first neural network, and the first neural network is any neural network of the neural network obtained through the (i-1)th evolution. ([Real, page 3] “samples S random models from the population... The model with the highest validation fitness within this sample is selected as the parent. A new architecture, called the child, is constructed from the parent by the application of a transformation called a mutation.”, wherein the examiner interprets sampling random models from the population each cycle and selecting the highest-fitness parent and then constructing a child by a mutation to be the same as selecting from neural networks obtained through the (i-1)th evolution and obtaining a neural network by mutating a first neural network because they are both directed to choosing candidates from the previously evolved set and producing a mutated candidate architecture that is then evaluated/selected based on superiority (fitness/performance).)
Bartz-Beielstein further teaches a neural network whose network structure is superior to that of the first neural network ([Bartz-Beielstein, page 6] “If the offspring solution is better (has a better function value) than the parent, take it as the new parent, otherwise retain the parent.”, wherein the examiner interprets taking the better offspring as the new parent to be the same as selecting a neural network whose network structure is superior to that of the first neural network because they are both directed to the evolutionary rule that the mutated candidate replaces the parent when it outperforms it (i.e., is superior in objective/fitness).)
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to evolutionary neural architecture search.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the evolutionary architecture search over neural networks disclosed by Real to include the offspring architecture disclosed by Bartz-Beielstein. One would be motivated to do so to effectively favor higher-performing offspring and accelerate convergence of the search, as suggested by Bartz-Beielstein ([Bartz-Beielstein, page 6] “If the offspring solution is better (has a better function value) than the parent, take it as the new parent, otherwise retain the parent.”).
Regarding claim 5, Real, Bartz-Beielstein, and Amizadeh teaches The method according to claim 1, (see rejection of claim 1).
Real further teaches wherein when at least one of the following conditions is met, a network structure of the neural network obtained by mutating the first neural network is superior to the network structure of the first neural network: a channel quantity of the neural network obtained by mutating the first neural network is greater than a channel quantity of the first neural network; or a quantity of convolutional layers in the neural network obtained by mutating the first neural network is greater than a quantity of convolutional layers in the first neural network.; ([Real, page 1] “Scaled to larger size, AmoebaNet-A sets a new state-of-the-art 83.9% top-1 / 96.6% top-5 ImageNet accuracy.”, and [Real, page 3] “Once the architecture is specified, the model still has two free parameters that can be used to alter its size (and its accuracy): the number of normal cells per stack (N) and the number of output filters of the convolution ops (F).”, wherein the examiner interprets scaled to larger size and altering accuracy via the number of output filters (F) and the number of normal cells per stack (N) to be the same as increasing channel quantity and increasing the quantity of convolutional layers because they are both directed to superiority (higher accuracy) being achieved when either the number of filters/channels is greater or the effective depth (more cells/layers) is greater.)
Regarding claim 6, Real, Bartz-Beielstein, and Amizadeh teaches The method according to claim 1, (see rejection of claim 1).
Bartz-Beielstein further teaches wherein the selecting, by the computing device from a set of the neural network obtained through the (i-1)th evolution and the candidate neural network and based on P evaluation parameters corresponding to each neural network in the set, a neural network obtained through the i-th evolution specifically comprises: performing, by the computing device, non-dominated sorting on the neural networks in the set based on the P evaluation parameters corresponding to each neural network in the set; and determining, by the computing device, that the neural network obtained through the i-th evolution is a neural network that is not dominated in the set, wherein a second neural network and a third neural network are two neural networks in the set, and if the second neural network is not inferior to the third neural network in terms of each of the P evaluation parameters, and the second neural network is superior to the third neural network in terms of at least one of the P evaluation parameters, the second neural network dominates the third neural network. ([Bartz-Beielstein, page 12] “The ranking procedure in NSGA-II is called ‘non-dominated ranking’. The procedure determines all non-dominated solutions and assigns them to the first (best) class. Then, it iteratively removes these solutions from the population, again determines all non-dominated solutions, and assigns them the next best class, until the population is empty” and [Bartz-Beielstein, page 10] “An important concept in case of multiple objectives is that of Pareto dominance: a solution x dominates another solution y ... iff x is better in at least one objective and not worse in all the others.”, wherein the examiner interprets determining all non-dominated solutions and assigning them to the first (best) class, iteratively, to be the same as performing non-dominated sorting on the neural networks in the set based on the P evaluation parameters and determining that the neural network obtained through the i-th evolution is a neural network that is not dominated in the set, because they are both ranking a set of candidate solutions across multiple objectives (the P evaluation parameters) via Pareto dominance and then selecting from the non-dominated (first-front) set. The examiner further interprets, “better in at least one objective and not worse in all the others” to be the same as the second neural network is superior to the third neural network in terms of at least one of the P evaluation parameters and the second neural network is not inferior to the third neural network in terms of each of the P evaluation parameters, because they are both defining dominance as strict improvement in at least one criterion while being no worse in every remaining criterion.)
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to selecting among candidate neural networks using evolutionary procedures.
It would have been obvious to a person of ordinary skill in the art before the effective filing date to modify the evolutionary architecture-search process disclosed by Real to include the non-dominated sorting and Pareto-dominance selection disclosed by Bartz-Beielstein. One would be motivated to do so to balance Real’s multiple evaluation parameters with a known multi-objective survivor-selection technique, thereby enforcing advancement of non-dominated candidates and improving search efficiency consistent as suggested by Real (Real, page 6] “evolution reaches half-maximum accuracy in roughly half the time.”)
Regarding claim 7, Real, Bartz-Beielstein, and Amizadeh teaches The method according to claim 1, (see rejection of claim 1).
Real further teaches wherein the obtaining, by a computing device, N neural networks comprises: randomly generating, by the computing device, M neural networks, wherein M is a positive integer; training and testing, by the computing device, each of the M neural networks by using the dataset, to obtain P evaluation parameters corresponding to each of the M neural networks; and selecting, by the computing device, N neural networks from the M neural networks based on the P evaluation parameters corresponding to each of the M neural networks, wherein N is not greater than M. ([Real, page 3] “It keeps a population of P trained models throughout the experiment. The population is initialized with models with random architectures (“while j-population < P” in Algorithm 1). All architectures that conform to the search space described are possible and equally likely.”, [Real, page 3] “At each cycle, it samples S random models from the population... The model with the highest validation fitness within this sample is selected as the parent....Once the child architecture is constructed, it is then trained, evaluated, and added to the population.”; wherein the examiner interprets initialized with models with random architectures to be the same as randomly generating M neural networks because they are both directed to creating candidate networks by random construction rather than by a fixed, deterministic design. The examiner further interprets, highest validation fitness within this sample is selected as the parent to be the same as selecting ... based on the P evaluation parameters because they are both directed to choosing networks using measured performance criteria obtained after training and testing. Finally, the examiner interprets, trained, evaluated... assigns a quality measure or fitness value... decision... based on fitness values to be the same as training and testing each of the M neural networks by using the dataset to obtain P evaluation parameters and selecting N neural networks from the M neural networks based on the P evaluation parameters, wherein N is not greater than M because they are both directed to (i) computing per-model evaluation metrics after training/testing and (ii) selecting a subset no larger than the original set according to those metrics (fitness/evaluation parameters).)
Regarding claim 8, Real, Bartz-Beielstein, and Amizadeh teaches The method according to claim 1, (see rejection of claim 1).
Real further teaches wherein the P evaluation parameters comprise at least one of a running time, accuracy, and a parameter quantity. ([Real, page 5] “It shows test accuracy and model compute cost. The latter is measured in FLOPs, by which we mean the total count of operations in the forward pass, so lower is better. ... Number of parameters showed similar behavior to FLOPs.”, wherein the examiner interprets “test accuracy” and “model compute cost” (measured in FLOPs) to be the same as running time because they are both directed to accuracy and computational effort for executing the model (lower FLOPs reflecting faster/less time to run). The examiner further interprets “Number of parameters”... to be the same as a parameter quantity because they are both counts of learned weights that quantify model size.)
Regarding claim 12, Real, Bartz-Beielstein, and Amizadeh teaches The apparatus according to claim 9, (see rejection of claim 9).
Bartz-Beielstein further teaches:
wherein the computer-executable instructions, when executed by the at least one processor, cause the apparatus to: select, from a neural network obtained by mutating a first neural network, a neural network whose network structure is superior to that of the first neural network, wherein the candidate neural network comprises the neural network that is of ([Bartz-Beielstein, page 4-5] “offspringi = mutate(parenti);” AND “population = {population} ∪ {offspring}; population = environmentalSelection(population)” and “The decision which individuals to include in the next generation is usually based on fitness values.”, wherein the examiner interprets mutate(parenti) (offspring produced from a selected parent) to be the same as a neural network obtained by mutating a first neural network; interprets environmentalSelection(population) with the decision... based on fitness values to be the same as select ... a neural network whose network structure is superior to that of the first neural network (selection of the better-performing architecture), and interprets parenti ∈ population (the parent is any member of the current population/generation) to be the same as the first neural network is any neural network of the neural network obtained through the (i-1)th evolution, because they are both directed to (i) generating a mutated child from an arbitrary prior-generation network, and (ii) choosing the candidate child when it is superior according to evaluation (fitness) criteria. )
the neural network obtained by mutating the first neural network and whose network structure is superior to that of the first neural network, and the first neural network is any neural network of the neural network obtained through the (i-1)th evolution. ([Real, page 3] “The evolutionary method we used is summarized in Algorithm 1. It keeps a population of P trained models throughout the experiment. The population is initialized with models with random architectures (“while population” in Algorithm 1). All architectures that conform to the search space described are possible and equally likely.”,
PNG
media_image1.png
424
362
media_image1.png
Greyscale
wherein the examiner interprets MUTATE(parent.arch) to be the same as a neural network obtained by mutating a first neural network, and highest-accuracy model to be the same as a neural network whose network structure is superior (as measured by accuracy), because they are both directed to producing a mutated candidate from a prior network and selecting the better architecture based on performance evaluation.)
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to evolutionary selection of mutated neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the evolutionary programming framework that mutates a parent disclosed by Bartz-Beielstein to include the mutated model selection process disclosed by Real. One would be motivated to do so to efficiently prioritize architectures with demonstrably better validation performance for advancement, as suggested by Real (Real, page 3] “The model with the highest validation fitness within this sample is selected as the parent.”).
Regarding claim 13, Real, Bartz-Beielstein, and Amizadeh teaches The apparatus according to claim 12, (see rejection of claim 12).
Real further teaches wherein when at least one of the following conditions is met, a network structure of the neural network obtained by mutating the first neural network is superior to the network structure of the first neural network: a channel quantity of the neural network obtained by mutating the first neural network is greater than a channel quantity of the first neural network; ([Real, page 3] “the model still has two free parameters that can be used to alter its size (and its accuracy): the number of normal cells per stack (N) and the number of output filters of the convolution ops (F).” AND [Real, page 1] “Scaled to larger size, AmoebaNet-A sets a new state-of-the-art 83.9% top-1 / 96.6% top-5 ImageNet accuracy.”, wherein the examiner interprets the number of output filters of the convolution ops (F) to be the same as a channel quantity and interprets “scaled to larger size” achieving higher accuracy to be the same as the mutated network’s channel quantity being greater than the first network’s channel quantity and being superior, because they are both directed to increasing the per-layer channel count (F) as a way to enlarge the model and improve accuracy, which is a form of superiority in performance.)
or a quantity of convolutional layers in the neural network obtained by mutating the first neural network is greater than a quantity of convolutional layers in the first neural network. ([Real, page 2] “normal cells are arranged in three stacks of N cells.” AND [Real, page 3] “two free parameters...the number of normal cells per stack (N)...used to alter its size (and its accuracy)”, wherein the examiner interprets increasing the number of normal cells per stack (N) to be the same as increasing a quantity of convolutional layers in the mutated neural network relative to the first neural network and being superior, because they are both directed to adding more cell blocks that contain convolution operations, thereby increasing the total number of convolutional layers and, per Real, altering size in a way that affects (improves) accuracy.)
Regarding claim 14, Real, Bartz-Beielstein, and Amizadeh teaches The apparatus according to claim 9, (see rejection of claim 9).
Bartz-Beielstein further teaches:
wherein computer-executable instructions, when executed by the at least one processor, cause the apparatus to: perform non-dominated sorting on the neural networks in the set based on the P evaluation parameters corresponding to each neural network in the set; ([Bartz-Beielstein, page 11] “The ranking procedure in NSGA-II is called ‘non-dominated ranking’. The procedure determines all non-dominated solutions and assigns them to the first (best) class. Then, it iteratively removes these solutions from the population, again determines all non-dominated solutions, and assigns them the next best class, until the population is empty.” AND [Bartz-Beielstein, page 10] “An important concept in case of multiple objectives is that of Pareto dominance...”, wherein the examiner interprets non-dominated ranking over multiple objectives to be the same as perform non-dominated sorting on the neural networks in the set based on the P evaluation parameters corresponding to each neural network in the set because they are both ranking a set of candidates by Pareto comparison across multiple objectives (the P evaluation parameters) to form fronts of non-dominated solutions.)
and determine that the neural network obtained through the i-th evolution is a neural network that is not dominated in the set, ([Bartz-Beielstein, page 11] “determines all non-dominated solutions and assigns them to the first (best) class...”, wherein the examiner interprets identifying and assigning solutions to the first (best) non-dominated class to be the same as determine that the neural network obtained through the i-th evolution is a neural network that is not dominated in the set because they are both selecting from the current set the solution belonging to the non-dominated front.)
wherein a second neural network and a third neural network are two neural networks in the set, ([Bartz-Beielstein, page 10] “Because EAs work with a population of solutions, they can be used to generate a set of solutions”, wherein the examiner interprets population of solutions to be the same as the set of neural networks because they are both the collection from which pairwise dominance comparisons (e.g., between a second and a third neural network) are made.)
and if the second neural network is not inferior to the third neural network in terms of each of the P evaluation parameters, and the second neural network is superior to the third neural network in terms of at least one of the P evaluation parameters, the second neural network dominates the third neural network. ([Bartz-Beielstein, page 10] “a solution x dominates another solution y...iff x is better in at least one objective and not worse in all the others.”, wherein the examiner interprets better in at least one objective and not worse in all the others to be the same as the second neural network is superior...in terms of at least one of the P evaluation parameters and not inferior ... in terms of each of the P evaluation parameters, because they are both the standard Pareto dominance rule applied to the P criteria.)
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to selecting superior neural-network candidates to advance non-dominated architectures.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the evolutionary selection framework disclosed by Bartz-Beielstein to include the new model construction process disclosed by Real. One would be motivated to do so to efficiently instantiate and hence evaluate concrete neural-network candidates within each generation so that non-dominated (superior) networks can be identified under the P-parameter criteria, as suggested by Real ([Real, page 3-5] “Once the child architecture is constructed, it is then trained, evaluated, and added to the population.”)
Regarding claim 15, Real, Bartz-Beielstein, and Amizadeh teaches The apparatus according to claim 9, (see rejection of claim 9).
Real further teaches to obtain P evaluation parameters...and select N neural networks from the M neural networks based on the P evaluation parameters corresponding to each of the M neural networks, wherein N is not greater than M. ([Real, page 3] “The evolutionary method we used is summarized in Algorithm 1. It keeps a population of P trained models through out the experiment. The population is initialized with models with random architectures (“while population” in Algorithm 1). All architectures that conform to the search space described are possible and equally likely.”
PNG
media_image2.png
374
323
media_image2.png
Greyscale
wherein the examiner interprets initialized...with random architectures to be the same as randomly generate M neural networks because they are both random construction of the candidate set; interprets TRAINANDEVAL to be the same as train and test...to obtain P evaluation parameters because they are both evaluating models after training; and interprets highest-accuracy model (and tournament selection described nearby) to be the same as select N neural networks...based on the P evaluation parameters because they are both selection decisions driven by measured performance metrics.)
Bartz-Beielstein further teaches wherein the computer-executable instructions, when executed by the at least one processor, cause the apparatus to: randomly generate M neural networks, wherein M is a positive integer; train and test each of the M neural networks by using the dataset,...corresponding to each of the M neural networks; ([Bartz-Beielstein, page 3-4] “first, individuals, which form the population, are randomly generated (line 1)...evaluatePopulation(population)...the environmental (or survivor) selection is performed (line 12)...The decision which individuals to include in the next generation is usually based on fitness values.”
PNG
media_image3.png
238
521
media_image3.png
Greyscale
wherein the examiner interprets individuals...randomly generated to be the same as randomly generate M neural networks because they are both initializing a set of candidates stochastically. The examiner further interprets evaluatePopulation(population) assigning fitness to be the same as train and test each of the M neural networks by using the dataset, to obtain P evaluation parameters because they are both computing per-model performance measures; and interprets environmental...selection...based on fitness values to be the same as select N neural networks from the M neural networks based on the P evaluation parameters...wherein N is not greater than M because they are both choosing a subset from the original set using the measured criteria, yielding at most the original M items.)
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to initializing a population of neural networks and selecting a subset of networks based on parameters to advance across generations.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the evolutionary programming workflow disclosed by Bartz-Beielstein to include the architecture selection process based on model accuracy disclosed by Real. One would be motivated to do so to efficiently drive selection by explicit post-training evaluation metrics that correlate with improved recognition outcomes, as suggested by Real (Real, page 1] “Scaled to larger size, AmoebaNet-A sets a new state-of-the-art 83.9% top-1 / 96.6% top-5 ImageNet accuracy.”)
Regarding claim 16, Real, Bartz-Beielstein, and Amizadeh teaches The apparatus according to claim 9, (see rejection of claim 9).
Real further teaches wherein the P evaluation parameters comprise at least one of a running time, accuracy, and a parameter quantity. ([Real, page 5-7] “It shows test accuracy and model compute cost. The latter is measured in FLOPs, by which we mean the total count of operations in the forward pass, so lower is better…Number of parameters showed similar behavior to FLOPs…Algorithm speed…evolution reaches half-maximum accuracy in roughly half the time…Model speed. The speed of individual models produced is also relevant. Figure 4 demonstrated that evolved models are faster (lower FLOPs).”, wherein the examiner interprets test accuracy to be the same as accuracy because they are both explicit measures of model correctness and “model compute cost…measured in FLOPs…lower is better” to be the same as running time because they are both directed to computational effort/speed of executing the model (fewer operations → faster runtime); and interprets Number of parameters to be the same as a parameter quantity because they are both counts of learned weights that quantify model size. The examiner further interprets “half…the time” and “faster (lower FLOPs)” to be the same as running time because they are both time-based efficiency measures used as evaluation criteria and per-model execution speed as an evaluation parameter.).
Claims 2,3, and 10-11 are rejected under 35 U.S.C. 103 as being unpatentable over Real in view of Bartz-Beielstein in view of Amizadeh further in view of NPL reference “Variational discriminator bottleneck: Improving imitation learning, inverse rl, and gans by constraining information flow.”, by Peng et. al. (referred herein as Peng).
Regarding claim 2, Real, Bartz-Beielstein, and Amizadeh teaches The method according to claim 1, (see rejection of claim 1).
Real further teaches:
wherein the mutating, by the computing device, a the network structure of a neural network obtained through the (i-1)th evolution comprises at least one of the following steps: ([Real, page 3] “A mutation causes a simple and random modification of the architecture” and “New models are constructed by applying a mutation to existing models, transforming their architectures in random ways.”, wherein the examiner interprets a mutation causing a simple and random modification of the architecture and transforming architectures in random ways to be the same as mutating the network structure in the i-th evolution because they are both computer-executed architectural (structural) changes applied during an evolution step).
swapping locations of two convolutional layers in one or more neural networks of the neural network obtained through the (i-1) evolution; ([Real, page 3] “The chosen element has one hidden state. This hidden state is now replaced with another hidden state from within the cell...no loops are formed (to keep the feed-forward nature of the convnet).”, wherein the examiner interprets replacing the source hidden state of an operation with another hidden state within the cell to be the same as swapping locations of two convolutional layers because they are both rewiring which intermediate feature feeds a convolution, thereby exchanging the relative placement/order of convolutional operations within the cell graph).
doubling a step size of a convolution kernel of the one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Real, page 2] “every application of the reduction cell is followed by a stride of 2 that reduces the image size, whereas normal cells preserve the image size...Ops belong to a fixed set of common convnet operations such as convolutions and pooling layers.”, wherein the examiner interprets a stride of 2 that reduces the image size within convnet operations such as convolutions to be the same as “doubling a step size of a convolution kernel of the one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution” because they are both increasing the convolution stride from the usual 1 (size-preserving) to 2 (downsampling), i.e., doubling the step between sampled positions in the convolutional layer.)
inserting the one or more convolutional layers into the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Real, page 3] “The op mutation...modifies the opage..by replacing the existing op with a random choice from a fixed list of ops (see Methods Details).” and [Real, page 2] “Ops belong to a fixed set of common convnet operations such as convolutions “, wherein the examiner interprets replacing an existing op with a convolutional op from the fixed list to be the same as inserting one or more convolutional layers because they are both adding a convolution operation at the mutated edge of the architecture).
deleting the one or more convolutional layers from the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Real, page 3] the op set illustrated includes “none”, and the text states “replacing the existing op with a random choice from a fixed list of ops”, wherein the examiner interprets choosing none as the replacement op to be the same as deleting the convolutional layer because they are both removing that convolution operation from the network at the mutated connection).
inserting one or more pooling layers into the one or more neural networks of the neural network obtained through the (i-1)th evolution; or (([Real, page 2] “Ops belong to a fixed set of common convnet operations such as convolutions and pooling layers.” and Figure examples include “avg 3x3” and “max 3x3”, wherein the examiner interprets selecting a pooling op via op-mutation to be the same as inserting one or more pooling layers because they are both adding a pooling operation at that location in the architecture).
deleting the one or more pooling layers from the one or more neural networks of the neural network obtained through the (i-1)th evolution. ([Real, page 3] the op set includes “none” and the text states “replacing the existing op with a random choice from a fixed list of ops”, wherein the examiner interprets changing a pooling op to none via the op-mutation to be the same as deleting the pooling layer because they are both removing that pooling operation from the network at the mutated edge).
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to computer-implemented evolutionary neural-network architecture search that mutates network structure (e.g., rewiring sources, changing stride, inserting/deleting convolutional or pooling layers) during an i-th evolution step.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the evolutionary architecture-search procedure disclosed by Real to include the “offspringi = mutate(parenti)” disclosed by Bartz-Beielstein. One would be motivated to do so to efficiently organize and govern Real’s specific mutation operations within a standard evolutionary loop that reliably advances higher-quality variants, as suggested by Bartz-Beielstein (Bartz-Beielstein, [page 4-5] “offspringi = mutate(parenti)”.)
Real, Bartz-Beielstein, and Amizadeh do not teach doubling a channel quantity of one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution;.
Peng teaches doubling a channel quantity of one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Peng, page 22] “Resnet-block 64×32×32 32→64 ... Resnet-block 256×8×8 128→256 ... 1×1 Conv2D 2·256 ... 256→ 2·256”), wherein the examiner interprets 32→64 / 128→256 / 256→2·256 to be the same as “doubling a channel quantity of one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution” because they are both increasing the number of channels (i.e., feature-map width) of convolutional/residual layers by a factor of two.)
Real, Bartz-Beielstein, Amizadeh, Peng, and the instant application are analogous art because they are all directed to evolutionary modification of convolutional/residual network architectures to achieve superior recognition performance.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the evolutionary architecture search over convolutional/residual networks disclosed by Real to include the Resnet-block architecture disclosed by Peng. One would be motivated to do so to efficiently increase representational capacity and throughput of residual pathways, as suggested by Peng ([Peng, page 22] “Resnet-block 64×32×32 32→64...[i.e. doubling channel quantity]”).
Regarding claim 3, Real, Bartz-Beielstein, and Amizadeh teaches The method according to claim 1, (see rejection of claim 1).
Real further teaches wherein the neural network obtained through the (i-1)th evolution is a deep residual network, and the mutating, by the computing device, a the network structure of a neural network obtained through the (i-1)th evolution comprises at least one of the following steps: swapping locations of two residual units in one or more neural networks of the neural network obtained through the (i-1)th evolution; ... doubling a step size of a convolution kernel of the one or more residual units in the one or more neural networks of the neural network obtained through the (i-1)th evolution; inserting one or more residual units into the one or more neural networks of the neural network obtained through the (i-1)th evolution; or deleting the one or more residual units from the one or more neural networks of the neural network obtained through the (i-1)th evolution. ([Real, page 3] “every application of the reduction cell is followed by a stride of 2 that reduces the image size”, and [Real, page 2] “A mutation causes a simple and random modification of the architecture...our mutation rules only alter architectures by randomly reconnecting the origin of edges to different vertices and by randomly relabeling the edges...The hidden state mutation ... the chosen element has one hidden state. This hidden state is now replaced with another hidden state from within the cell...”, and cells include “a skip input from the cell before it”“, and Fig. 1 op set includes “none”, wherein the examiner interprets (i) “stride of 2 that reduces the image size” to be the same as doubling a step size of a convolution kernel because they are both directed to increasing the convolution stride from 1 to 2; (ii) “randomly reconnecting” edges and “randomly relabeling” edges to be the same as inserting one or more residual units because they are both directed to adding new residual-style pathways/blocks into the network graph; and (iii) the presence of the op “none” within the mutation’s op choices to be the same as deleting the one or more residual units because they are both directed to removing an operation/path in the residual-style module. The examiner further interprets, “replaced with another hidden state” within a cell that uses skip connections to be the same as swapping locations of two residual units because they are both directed to reassigning which residual-style block (formed by summing branches/skip) occupies a given position in the cell’s topology (i.e., exchanging sources so the units trade positions in the dataflow).
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to modifying deep residual network structures during an evolutionary search, like swapping residual units, changing stride, inserting/deleting residual units to improve model performance.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the evolutionary residual-network editing disclosed by Real to include the “population = environmentalSelection(population)” disclosed by Bartz-Beielstein. One would be motivated to do so to efficiently advance higher-performing residual-unit configurations and avoid regressions via survivor selection, as suggested by Bartz-Beielstein (Bartz-Beielstein, page 5] “The decision which individuals to include in the next generation is usually based on fitness values.”).
Real, Bartz-Beielstein, Amizadeh do not teach doubling a channel quantity of one or more residual units in the one or more neural networks of the neural network obtained through the (i-1)th evolution;
Peng teaches doubling a channel quantity of one or more residual units in the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Peng, page 22] Table 4: “Resnet-block 64×32×32 32→64 ... “Resnet-block 256×8×8 128→256 ... “1×1 Conv2D 2·256 ... 256→ 2·256”, wherein the examiner interprets “Resnet-block ... 32→64 / 128→256” and “1×1 Conv2D ... 256→ 2·256” to be the same as doubling a channel quantity of one or more residual units because they are both directed to increasing the number of channels (width) in residual units or immediately-adjacent layers by a factor of two in a deep residual network.)
Real, Bartz-Beielstein, Amizadeh, Peng, and the instant application are analogous art because they are all directed to modifying deep residual network structures to improve model performance.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the evolutionary architecture search disclosed by Real to include the Resnet-block architecture disclosed by Peng. One would be motivated to do so to efficiently increase representational capacity and throughput of residual pathways, as suggested by Peng ([Peng, page 22] “Resnet-block 64×32×32 32→64 ... [i.e. doubling throughput]”.)
Regarding claim 10, Real, Bartz-Beielstein, and Amizadeh teaches The apparatus according to claim 9, (see rejection of claim 9).
Real further teaches:
wherein computer-executable instructions, when executed by the at least one processor, cause the apparatus to perform at least one of the following steps: swapping locations of two convolutional layers in one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Real, page 1] “Our mutation rules only alter architectures by randomly reconnecting the origin of edges to different vertices and by randomly relabeling the edges, covering the full search space.”; and [Real, page 3] “This hidden state is now replaced with another hidden state from within the cell...”, wherein the examiner interprets applying a mutation to transform architectures to be the same as computer-executable instructions performing steps that modify network structure. The examiner further interprets randomly reconnecting the origin of edges to different vertices and replacing a hidden state within the cell to be the same as swapping locations of two convolutional layers because they are both re-attaching the inputs/positions of convolutional operations in the directed graph so that the locations of two convolutions are exchanged.)
doubling a step size of a convolution kernel of the one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Real, page 2] “every application of the reduction cell is followed by a stride of 2 that reduces the image size, whereas normal cells preserve the image size.”, wherein the examiner interprets stride of 2 that reduces the image size to be the same as doubling a step size of a convolution kernel because they are both increasing the convolution stride from 1 to 2 (i.e., doubling the step between sampled positions).)
inserting the one or more convolutional layers into the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Real, page 4] “It does this by replacing the existing op with a random choice from a fixed list of ops (see Methods Details).”; and [Real, Methods → Search Space] “Ops belong to a fixed set of common convnet operations such as convolutions and pooling layers.”, wherein the examiner interprets replacing the existing op with a random choice from a fixed list of ops that includes convolutions to be the same as inserting the one or more convolutional layers because they are both adding a convolutional operation at that site by operation substitution.)
deleting the one or more convolutional layers from the one or more neural networks of the neural network obtained through the (i-1)th evolution; inserting one or more pooling layers into the one or more neural networks of the neural network obtained through the (i-1)th evolution; or deleting the one or more pooling layers from the one or more neural networks of the neural network obtained through the (i-1)th evolution. ([Real, page 4] “...none...replacing the existing op with a random choice from a fixed list of ops...”, [Real, page 1] “Ops belong to a fixed set of common convnet operations (such as convolutions and pooling layers.)”; wherein the examiner interprets choosing “none” as the replacement op to be the same as deleting the one or more convolutional and pooling layers because they are both removing an operation at that position so that the convolutional and pooling layer is absent. The examiner further interprets selecting a pooling operation from the fixed list to be the same as inserting one or more pooling layers because they are both adding a pooling layer at the chosen graph position by operation substitution.)
Real, Bartz-Beielstein, and Amizadeh do not teach doubling a channel quantity of one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution;
Peng teaches doubling a channel quantity of one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Peng, page 22] “Resnet-block 64×32×32 32→64 ... Resnet-block 256×8×8 128→256 ... 1×1 Conv2D 2·256 ... 256→ 2·256”), wherein the examiner interprets 32→64 / 128→256 / 256→2·256 to be the same as “doubling a channel quantity of one or more convolutional layers in the one or more neural networks of the neural network obtained through the (i-1)th evolution” because they are both increasing the number of channels (i.e., feature-map width) of convolutional/residual layers by a factor of two.)
Real, Bartz-Beielstein, Amizadeh, Peng, and the instant application are analogous art because they are all directed to evolutionary modification of convolutional/residual network architectures to achieve superior performance.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the apparatus of claim 9 disclosed by Real, Bartz-Beielstein, and Amizadeh to include the Resnet-block architecture disclosed by Peng. One would be motivated to increase representational capacity and throughput in residual pathways, as suggested by Peng ([Peng, page 22] “Resnet-block 64×32×32 32→64 ... [i.e. doubling channel quantity] … 1x1 Conv2D … 256 → 2x256”).
Regarding claim 11, Real, Bartz-Beielstein, and Amizadeh teaches The apparatus according to claim 9, (see rejection of claim 9).
Real further teaches:
wherein the neural network obtained through the (i-1)th evolution is a deep residual network, and computer-executable instructions, when executed by the at least one processor, cause the apparatus to perform at least one of the following steps: ([Real, page 2] “Each cell receives a direct input from the previous cell ... and a skip input from the cell before it.” and “applying an operation (or op) to an existing hidden state, applying another op to another existing hidden state, and adding the results to produce a new hidden state.”, wherein the examiner interprets skip input and adding the results to be the same as a deep residual network because they are both directed to residual-style skip connections and additive combination across stacked units.)
swapping locations of two residual units in one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Real, page 1] “Our mutation rules only alter architectures by randomly reconnecting the origin of edges to different vertices” and [Real, page 3] “This hidden state is now replaced with another hidden state from within the cell”, wherein the examiner interprets randomly reconnecting the origin of edges to different vertices and replacing a hidden state within the cell to be the same as swapping locations of two residual units because they are both re-attaching which predecessors feed a unit so that the positions/attachments of two residual blocks are exchanged in the graph.)
doubling a step size of a convolution kernel of the one or more residual units in the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Real, page 3] “the number of output filters of the convolution ops (F).” and [Real, page 2] “every application of the reduction cell is followed by a stride of 2 that reduces the image size, whereas normal cells preserve the image size.”, wherein the examiner interprets the number of output filters of the convolution ops (F) to be the same as a channel quantity of residual units because they are both per-layer channel counts; increasing F (including by a factor of two) increases the channel quantity in the residual block’s convolutions. The examiner further interprets “stride of 2” to be the same as doubling a step size of a convolution kernel because they are both increasing the convolution stride from 1 to 2, i.e., doubling the sampling step in a residual unit. )
inserting one or more residual units into the one or more neural networks of the neural network obtained through the (i-1)th evolution; or deleting the one or more residual units from the one or more neural networks of the neural network obtained through the (i-1)th evolution. ([Real, page 4] “It does this by replacing the existing op with a random choice from a fixed list of ops (see Methods Details).”, [Real, page 2] “applying an op ... to an existing hidden state, applying another op ... and adding the results to produce a new hidden state.”, and [Real, page 4] “Possible ops: none (identity); 3x3, 5x5 and 7x7 separable convolutions ... 3x3 average pool; 3x3 max pool; ...”, wherein the examiner interprets replacing the existing op with a choice that yields the two-branch add to be the same as inserting one or more residual units because they are both adding a residual-style computation (two transformed paths combined by addition) at that position via operation substitution. The examiner further interprets choosing none (identity) as the replacement op to be the same as deleting the one or more residual units because they are both removing the residual computation at that location so the unit is absent/effectively bypassed.)
Real, Bartz-Beielstein, and Amizadeh do not teach doubling a channel quantity of one or more residual units in the one or more neural networks of the neural network obtained through the (i-1)th evolution;
Peng teaches doubling a channel quantity of one or more residual units in the one or more neural networks of the neural network obtained through the (i-1)th evolution; ([Peng, page 22] Table 4: “Resnet-block 64×32×32 32→64 ... “Resnet-block 256×8×8 128→256 ... “1×1 Conv2D 2·256 ... 256→ 2·256”, wherein the examiner interprets “Resnet-block ... 32→64 / 128→256” and “1×1 Conv2D ... 256→ 2·256” to be the same as doubling a channel quantity of one or more residual units because they are both directed to increasing the number of channels (width) in residual units or immediately-adjacent layers by a factor of two in a deep residual network.)
Real, Bartz-Beielstein, Peng, and the instant application are analogous art because they are all directed to modifying deep residual network structures (via mutation or block design) to improve model performance.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the apparatus of claim 9 disclosed by Real, Bartz-Beielstein, Amizadeh to include the Resnet-style residual blocks with width-doubling disclosed by Peng. One would be motivated to do so to efficiently increase representational capacity and throughput of residual pathways, as suggested by Peng ([Peng, page 22] “Resnet-block 64×32×32 32→64 … Resnet-block 256×8×8 128→256 … 1×1 Conv2D 2x256 …. 256→2x256”).
Regarding claim 20, Real, Bartz-Beielstein, and Amizadeh teaches The method according to claim 19, (see rejection of claim 19).
Real further teaches wherein the parameter of the basic unit comprises at least one of a type, a channel quantity parameter, and a size parameter of the basic unit. ([Real, page 2-3] “Ops belong to a fixed set of common convnet operations such as convolutions and pooling layers...Once the architecture is specified, the model still has two free parameters...the number of normal cells per stack (N) and the number of output filters of the convolution ops (F).”, [Real, page 2] “In the cell example of Figure 1 (right), the first pairwise combination applies a 3x3 average pool op [operation] to hidden state 0 and a 3x3 max pool op to hidden state 1, in order to produce hidden state 2. The next pairwise combination can now choose from hidden states 0, 1, and 2 to produce hidden state 3 (chose 0 and 1 in Figure 1), and so on.” and [Real, page 4] “replacing the existing op with a random choice from a fixed list of ops”, wherein the examiner interprets the fixed list of ops (e.g., convolutions and pooling) to be the same as a type of the basic unit, the number of output filters of the convolution ops (F) to be the same as a channel quantity parameter, and the op labels “3x3/5x5/7x7” to be the same as a size parameter of the basic unit because they are both defining, for each repeated cell (the basic unit), (i) which operation type is used, (ii) how many output channels it has, and (iii) the spatial size (e.g., kernel size) that configures that operation.)
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to configuring and selecting neural network architectures using parameters.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the cell-based architecture search with typed ops and width/size parameters disclosed by Real to include the fitness value calculation process disclosed by Bartz-Beielstein. One would be motivated to do so to efficiently select parameter settings (type, channel quantity, size) that improve model performance under the stated evaluation criteria, as suggested by Bartz-Beielstein ([Bartz-Beielstein, page 3] “The decision which individuals to include in the next generation is usually based on fitness values.”).
Regarding claim 21, Real, Bartz-Beielstein, and Amizadeh teaches The method according to claim 19, (see rejection of claim 19).
Real further teaches wherein the basic unit is configured to perform a first operation and a second operation on a feature map input to the basic unit, and wherein the feature map input is a feature map of the to-be-recognized image, ([Real, page 2] “applying an operation (or op) to an existing hidden state, applying another op to another existing hidden state, and adding the results to produce a new hidden state.” and ([Real, page 1, 5] “Regularized Evolution for Image Classifier Architecture Search” AND “Experiments always searched on the CIFAR-10 dataset.”, wherein the examiner interprets applying an operation ...and applying another o...to existing hidden state(s) to be the same as “the basic unit is configured to perform a first operation and a second operation on a feature map input to the basic unit” and image classifier...CIFAR-10 dataset to be the same as “the feature map input is a feature map of the to-be-recognized image” because they are both describing two ops acting on intermediate feature maps derived from input images in an image-recognition network.)
the first operation is for doubling or maintaining a quantity of feature maps input to the basic unit, ([Real, page 3] “the number of output filters of the convolution ops (F).”, wherein the examiner interprets the number of output filters ... (F) to be the same as “a quantity of feature maps input to the basic unit” and hence “the first operation is for doubling or maintaining a quantity of feature maps input to the basic unit” because they are both controlling per-layer channel count (feature-map quantity), which can be held constant or increased (e.g., doubled) by choosing a larger F.)
the second operation is for changing a size of the feature map input to the basic unit from an original first size to a second size or maintain the first size, and the first size is greater than the second size. ([Real, page 2] “every application of the reduction cell is followed by a stride of 2 that reduces the image size, whereas normal cells preserve the image size.”, wherein the examiner interprets stride of 2...reduces the image size and normal cells preserve the image size to be the same as “the second operation is for changing a size of the feature map input...from an original first size to a second size or maintain a first size, and the first size is greater than the second size because they are both” specifying that one operation downsamples (first size => smaller second size) while another keeps the spatial size unchanged.)
Real, Bartz-Beielstein, Amizadeh, and the instant application are analogous art because they are all directed to configuring and evolving neural-network building blocks that apply paired operations.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the evolutionary optimization over candidate architectures disclosed by Bartz-Beielstein to include the process of free parameter calculation disclosed by Real. One would be motivated to do so to efficiently control feature-map quantity (via free parameter F) and network capacity (via free parameter N) to enhance accuracy, as suggested by Real ([Real, page 3] “alter its size (and its accuracy)”).
Conclusion
THIS ACTION IS MADE FINAL. 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 DEVAN KAPOOR whose telephone number is (703)756-1434. The examiner can normally be reached Monday - Friday: 9:00AM - 5:00 PM EST (times may vary).
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, David Yi can be reached at (571) 270-7519. 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.
/DEVAN KAPOOR/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights