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 .
Note to Applicant
In response to Applicant's remarks submitted on 4/28/26, the Examiner has vacated the previous Non-Final Action dated 1/30/26, in favor of a new Non-Final Action. Please see the rejection below.
Status of Claims
Applicant’s submission filed 4/28/26 has been entered. Claims 21, 22 are new. Claims 1-22 are presented for examination.
Claim Rejections - 35 USC § 101
Applicant' s arguments, filed 4/28/26, with respect to the 101 rejection, have been fully considered and are persuasive. The rejection of claims 1 -20 has been withdrawn.
.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 2, 3, 5, 8, 10, 11, 12, 13, 15, 18, 21 are rejected under 35 U.S.C. 103 as being unpatentable over Lenc et al. (US 20200234142 A1), in view of Szegedy et al. (US 10521718 B1).
Re-claim 1, Lenc et al. teach A method implemented by a computer device and comprising:
--obtaining, in an nth training round of iterative training on a neural network model, a first training data subset from a training data set based on an index table, wherein n is a positive integer;
[0050] The neural network training system 100 is a system that receives, i.e., from a user of the system, training data 102 for training a neural network 110 to perform a machine learning task and uses the training data 102 to train the neural network to determine trained values 150 of the weights of the neural network 110.)
--training the neural network model based on training data in the first training data subset;
[0060] A training engine 120 within the system 100 iteratively trains the neural network 110, i.e., to determine trained values of the differentiable weights and to determine final distribution parameters.
--obtaining gradient information corresponding to the neural network model; --evaluating the training data based on the gradient information, to obtain an evaluation result;
(see e.g. [0009] determining, for each of the training inputs, a gradient with respect to the differentiable weights of the objective function through backpropagation; and determining, from the gradients for each of the training inputs, the update to the current values of the differentiable weights in accordance with the gradient-based training technique.
[0085] Thus, the same objective function and the same training inputs are used to determine both (i) the update to the distribution parameters and (ii) the update to the differentiable weights.)
Although Lenc et al. anticipate --adjusting the index table based on the evaluation result, to obtain an adjusted index table;
(see e.g. [0123] Thus, for example, the index database can include multiple collections of data, each of which may be organized and accessed differently.
[0087] The system updates the current values of the differentiable weights and the distribution parameters using the respective updates (step 208).
[0010] apply the final update to the current distribution parameters.)
Lenc et al. do not teach -- using the adjusted index table to obtain a second training data subset in subset for an (n+1)th round of the iterative training.
However, Szegedy et al. teach ----adjusting the index table based on the evaluation result, to obtain an adjusted index table; and using the adjusted index table to obtain a second training data subset in subset for an (n+1)th round of the iterative training.
(see e.g. col 2, lines 43-50 --The neural network training system 100 trains a neural network 110 on training inputs from a training data repository 120 to determine trained values of the parameters of the neural network 110 from initial values of the parameters. The neural network 110 can be a feedforward deep neural network, e.g., a convolutional neural network, or a recurrent neural network, e.g., a long short term (LSTM) neural network.
--col. 4, lines 36-40-- The system trains the neural network on each of the multiple training inputs and, for each of the training inputs, an adversarial perturbation of the training input to determine trained values of the parameters of the neural network (step 204).
--col. 5, lines 39-42 --In some implementations, the system determines a gradient of the specified objective function with respect to the training input, e.g., using backpropagation, and modifies the training input using the gradient of the specified objective function to determine the adversarial perturbation of the training input.
--col. 4, lines 16-22 -In some implementations, once the neural network 110 has been trained to determine the trained values of the parameters, the neural network training system 100 stores the trained values of the parameters of the neural network 110 for use in instantiating a trained neural network or provides the trained values of the parameters to another system for use in instantiating a trained neural network.)
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Lenc et al., and include the steps cited above, as taught by Szegedy et al., so the performance of the neural network when trained can be improved (see e.g. col. 2, line 3).
Re-claim 2, Lenc et al. teach -- The method according to claim 1, wherein the evaluating the training data comprises:
--obtaining a preset evaluation rule; and --evaluating the training data in the first training data subset based on the preset evaluation rule.
(see e.g. [0096] The system can determine the update in accordance with an update rule employed by the gradient-based training technique that determines how the gradients for the training inputs are combined to generate an update to the differentiable weights.)
Re-claim 3, Lenc et al. teach-- The method of claim 1, wherein the evaluation result comprises an effect of the training data on model or a manner of processing the training data in a next training round.
(see e.g. [0094] The system evaluates, for each training input, the objective function that measures a quality of the network output for the training input (step 306). That is, the system determines a value that of the objective function that represents the quality of the network output relative to the corresponding target output.
[0084] Generally, the fitness of the neural network is based on the values of the objective function on a set of training inputs).
Re-claim 5, Lenc et al. teach ---The method according to claim 3, wherein the manner comprises deleting the training data, decreasing a weight of the training data, increasing the weight, or retaining the training data.
(see e.g. [0056] During the training, the neural network training system 100 maintains and repeatedly updates weight data 130.
[0112] FIG. 5 shows an example of an iteration of updating the distribution parameters and the differentiable weights when the non-differentiable weights are a sparsity mask. In the example of FIG. 5, the gradient-based training technique used to learn the differentiable weights is stochastic gradient descent (SGD).
[0116] As a result of the step of SGD, the system determines (i) a fitness f.sub.i and (ii) an update to the differentiable weights. The system averages the updates to the differentiable weights and then applies the averaged update to generate updated values of the differentiable weights.)
Re-claim 8, Lenc et al. teach The method of claim 1, wherein the neural network model comprises computing layers; and wherein the method further comprises further obtaining the gradient information for at least one of the computing layers.
(see e.g. [0095] The system determines, for each of the training inputs, a gradient with respect to the differentiable weights of the objective function (step 308). The system can determine this gradient using the gradient-based training technique, e.g., through backpropagation.
[0004] Some neural networks are recurrent neural networks. A recurrent neural network is a neural network that receives an input sequence and generates an output sequence from the input sequence. In particular, a recurrent neural network can use some or all of the internal state of the network from a previous time step in computing an output at a current time step.)
Re-claim 10, Lenc et al. teach-- The method of claim 1, further comprising receiving configuration information from a user and through an interface, wherein the configuration information comprises dynamic training information comprises information about the neural network model, information about the training data set, a running parameter for model training, or computing resource information for the model training.
(see e.g. [0055] The system 100 can receive the training data 102 in any of a variety of ways. For example, the system 100 can receive training data as an upload from a remote user of the system over a data communication network, e.g., using an application programming interface (API) made available by the system 100. As another example, the system 100 can receive an input from a user specifying which data that is already maintained by the system 100 should be used as the training data 102.
[[0050] The neural network training system 100 is a system that receives, i.e., from a user of the system, training data 102 for training a neural network 110 to perform a machine learning task and uses the training data 102 to train the neural network to determine trained values 150 of the weights of the neural network 110.)
Re-claim 21, Lenc et al. teach-- --The method of claim 1, further comprising further obtaining, through training the neural network model, the gradient information, wherein the gradient information is based on sequential calculation for each computing layer of the neural network model during back propagation.
(see e.g. [0095] The system determines, for each of the training inputs, a gradient with respect to the differentiable weights of the objective function (step 308). The system can determine this gradient using the gradient-based training technique, e.g., through backpropagation.
[0004] Some neural networks are recurrent neural networks. A recurrent neural network is a neural network that receives an input sequence and generates an output sequence from the input sequence. In particular, a recurrent neural network can use some or all of the internal state of the network from a previous time step in computing an output at a current time step.
[0057] As a particular example, for image processing tasks, the neural network can be a convolutional neural network. As another particular example, for sequence generation tasks, i.e., tasks that require the neural network to generate outputs sequentially, the neural network can be an auto-regressive neural network, e.g., a recurrent neural network, an auto-regressive convolutional neural network, or a self-attention neural network.)
Claim 11 recites similar limitations as claim 1 and is therefore rejected under the same arts and rationale.
Claim 12 recites similar limitations as claim 2 and is therefore rejected under the same arts and rationale.
Claim 13 recites similar limitations as claim 3 and is therefore rejected under the same arts and rationale.
Claim 15 recites similar limitations as claim 5 and is therefore rejected under the same arts and rationale.
Claim 18 recites similar limitations as claim 8 and is therefore rejected under the same arts and rationale.
Claims 4, 14 are rejected under 35 U.S.C. 103 as being unpatentable over Lenc et al. (US 20200234142 A1), in view of Szegedy et al. (US 10521718 B1), in further view of TAKEDA et al. (US 20170358045 A1).
Re-claim 4, Lenc et al., in view of Szegedy et al., do not teach the limitations as claimed.
However, TAKEDA et al. teach --The method according to claim 3, wherein the effect is "invalid," "inefficient," "efficient," or "indeterminate,"
(see e.g. [0024] The data analysis system evaluates the relation between data elements included in the training data and the classification information and evaluates the possibility of falling under invalid materials from a large amount of search object data (for example, unknown data such as patent documents and papers) by using the results of the above-described evaluation.
[0036] The relation evaluation unit 120 evaluates the relation between data elements included in the training data and the classification information. More specifically, the relation evaluation unit 120 evaluates the data elements extracted from the training data acquired by the data acquisition unit 110 in accordance with specified standards. In other words, the relation evaluation unit 120 can learn patterns (widely including abstract concepts and meanings and without limitation to so-called “specified patterns” [for example, specified design patterns or regularity]) included in the training data by evaluating the degree of contribution to combinations included in the training data set, which has been acquired by the data acquisition unit 110, by the data elements constituting at least part of the training data. Incidentally, the “specified standards” will be explained later.
[0088] The data evaluation unit 150 calculates the score indicating the relation between each piece of the partial unknown data and the training data on the basis of the evaluation results stored in the evaluation memory unit 220 in the memory unit 200 (S230). The evaluation integration unit 160 generates the integrated score with respect to each piece of the unknown data by integrating the scores calculated by the data evaluation unit 150 with respect to the partial unknown data obtained by breaking down the unknown data (S240).
[0081] The score indicating the relation to the training data with respect to the plurality of pieces of unknown data is calculated by using the evaluation results. As a result, it is possible to analyze the unknown data mechanically according to certain standards and support finding of data related to data in which specific ideas, events, etc. are described from a large amount of unknown data.)
With respect to the following limitation,” wherein "invalid" indicates that a contribution provided by the training data to training precision to be achieved by the model training is 0; wherein "inefficient" indicates that the contribution reaches a first contribution degree; wherein "efficient" indicates that the contribution reaches a second contribution degree that is greater than the first contribution degree; and wherein "indeterminate" indicates that the contribution is indeterminate.”
TAKEDA et al. teach [0034] Alternatively, the data acquisition unit 110 can acquire the training data from a storage device which is connected in a manner capable of communications. The classification information may be, by way of example and without limitation to, “1” assigned to the correct data and “−1” assigned to the incorrect data.
[0011] The data evaluation unit may calculate a score indicative of strength of a relation between the partial unknown data and the classification information so that when a relation between a data element included in the partial unknown data and the classification information is strong, a value of the score will become larger than a case where the relation is weak; and the evaluation integration unit may generate an integrated score as the integrated index by summing a specified number of the score, which is calculated by the data evaluation unit, in descending order.
***TAKEDA et al. assign scores to training data. It is considered an obvious variation of TAKEDA et al. label the contribution degree as claimed.
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Lenc et al., in view of Szegedy et al, and include the steps cited above, as taught by TAKEDA et al., in order to examine and judge the relevance of the data (see e.g. [0109]).
Claim 14 recites similar limitations as claim 4 and is therefore rejected under the same arts and rationale.
Claims 6, 7, 9, 16, 17, 19 are rejected under 35 U.S.C. 103 as being unpatentable over Lenc et al. (US 20200234142 A1), in view of Szegedy et al. (US 10521718 B1), in further view of Bhatia et al. (US 20200193531 A1).
Re-claims 6, 9, Lenc et al., in view of Szegedy et al, do not teach the limitations as claimed.
However, Bhatia t al. teach-- The method of claim 2, further comprising:
--testing the neural network model, using test data to obtain a test result; and updating the preset evaluation rule based on a preset target value and the test result.
9. The method of claim 6, further comprising further updating the preset evaluation rule based on a negative feedback mechanism when the test result does not reach the preset target value.
(see e.g. [0015] Responsive to the resulting property data not meeting target data, manufacturing parameters (e.g., hardware parameters, process parameters) of the manufacturing equipment may be updated to attempt to meet the target data.
[0036] The testing engine 186 may be capable of testing a trained machine learning model 190 using a corresponding set of features of a testing set from data set generator 172. For example, a first trained machine learning model 190 that was trained using a first set of features of the training set may be tested using the first set of features of the testing set.
[0053] In some embodiments, subsequent to generating a data set and training, validating, or testing machine learning model 190 using the data set, the machine learning model 190 may be further trained, validated, or tested (e.g., using updates to manufacturing parameters 154 and tested film property data 156 of FIG. 1) or adjusted (e.g., adjusting weights associated with input data of the machine learning model 190, such as connection weights in a neural network).
[0054] The system 300 may be a feedback system for determining updates to manufacturing parameters 354 to meet target data 352 (e.g., target data 152 of FIG. 1) based on historical or experimental data 342 (e.g., historical or experimental data 142 of FIG. 1).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Lenc et al., in view of Szegedy et al, and include the steps cited above, as taught by Bhatia et al. in order to , validate each of the trained models using a corresponding set of features of the validation set. (see e.g. [0057] ).
Re-claim 7, The method of claim 6, further comprising further updating the preset evaluation rule based on a positive feedback mechanism when the test result reaches or is better than the preset target value.
Bhatia teaches a feedback system for determining updates to manufacturing parameters (hardware parameters, process parameters), during testing of a model, and then adjust the parameters accordingly. Therefore, it is considered an obvious variation of Bhatia et al. to update the rules based on positive feedback.
Claim 16 recites similar limitations as claim 6 and is therefore rejected under the same arts and rationale.
Claim 17 recites similar limitations as claim 7 and is therefore rejected under the same arts and rationale.
Claim 19 recites similar limitations as claim 9 and is therefore rejected under the same arts and rationale.
Claim 22 is rejected under 35 U.S.C. 103 as being unpatentable over Lenc et al. (US 20200234142 A1), in view of Szegedy et al. (US 10521718 B1), and further in view of Cervantes Martín (US 20190340507 A1)
Re-claim 22, Lenc et al., in view of Szegedy et al, do not teach the limitations as claimed.
However, Cervantes Martín teaches --The method of claim 1, wherein the gradient information is based on a calculation using a chain rule and is of an objective function, the gradient information comprises a weight, or the gradient information determines an impact of the training data on model training convergence.
(see e.g. [0002] First, the forward pass, which computes all the activations of the neurons until reaching the output given input data. Second, the backward pass, which takes the output activations and computes the error compared to the desired prediction (annotation). This comparison and computation of the error is defined by the objective function, which models the problem to solve by the network. The error is used to calculate the gradients of the objective function with respect to each weight in the network following the chain rule. According to these gradients and using an optimization method all the weights between neurons are adjusted in order to do better predictions in the next forward pass.
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Lenc et al., in view of Szegedy et al., and include the steps cited above, as taught by Cervantes Martín, in order to allow the network to find its optimal configuration (see e.g. [0002]).
Response to Arguments
Applicant’s arguments filed 4/28/26, with respect to the rejections of claims 1-20 have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground of rejection is made in view of the newly found prior art references. A new search was performed based on the remarks and the new references cited teach the limitations as claimed.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to LUNA CHAMPAGNE whose telephone number is (571)272-7177. The examiner can normally be reached M-F 8:00-5:00.
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, Florian Zeender can be reached at 571 272-6790. 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.
/LUNA CHAMPAGNE/ Primary Examiner, Art Unit 3627
June 17, 2026
Prosecution Insights