DETAILED ACTION
Remarks
Claims 2-23 have been examined and rejected. This Office Action is responsive to the amendment filed on 12/19/2025, which has been entered in the above identified application.
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 .
Claims 2-23 are presented for examination.
Respond to Amendment
The amendment filed 12/19/2025 has been entered. Claims 2, 14 and 23 have been amended. Claim 3-13, 15-20 and 22 were previously presented. Claim 1 was canceled. Claims 2-23 are pending in the application.
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 2-23 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Independent claims
Step 1
Claim 2 is drawn to a method of determining an architecture for a task neural network, claim 14 is drawn to a system and claim 23 is drawn to a non-transitory computer storage media storing instructions executed by one or more computers to perform the method of claim 2. Therefore, each of these claim groups falls under one of four categories of statutory subject matter (process/method, machines/product/apparatus, manufactures, and composition of matter).
Step 2A – Prong 1
Claims 2, 14 and 23 are directed to a judicially recognized exception of an abstract idea without significantly more.
Claims 2, 14 and 23 recite a method of determining an architecture for a task neural network in a computationally efficient manner using a performance prediction neural network that under its broadest reasonable interpretation enumerates both mathematical concept and certain methods of organizing human activity. A performance prediction neural network is a mix of mathematical principles and organized human effort. Therefore, the step of determining an architecture for a task neural network using a performance prediction neural network is either a mathematical concept or certain methods of organizing human activity (MPEP 2106.04(a)(2)).
Claims 2, 14 and 23 recite a method of determining an architecture for a task neural network that is configured to perform a particular machine learning task that under its broadest reasonable interpretation enumerates a mental concept. A human can mentally perform, with the physical aid such as pen and paper, to determine an architecture for a neural network. Therefore, the step of determining an architecture for a task neural network is nothing more than an abstract mental concept (MPEP 2106.04(a)(2)(III)).
Claims 2, 14 and 23 recite further a method of obtaining data specifying a current set of candidate architectures for the task neural network that under its broadest reasonable interpretation enumerates a mental concept. A human can mentally perform, with the physical aid such as pen and paper, to obtain data. Therefore, the step of obtaining data specifying a current set of candidate architectures is nothing more than an abstract mental concept (MPEP 2106.04(a)(2)(III)).
Claims 2, 14 and 23 recite further methods of generating an input that specifies the candidate architecture; generating an updated set of candidate architectures by selecting one or more of the candidate architectures in the current set based on the performance predictions for the candidate architectures in the current set; generating an instance of the task neural network having the candidate architecture and generating a plurality of new candidate architectures from the candidate architecture by adding, for each new candidate architecture, a respective new operation block having respective hyperparameters to the candidate architecture those under their broadest reasonable interpretation enumerates a mental concept. A human can mentally perform, with the physical aid such as pen and paper, to generate an input, to generate an updated set of candidate architectures, to generate an instance and to generate a plurality of candidates. Therefore, the steps of generating an input, generating an updated set of candidates, generating an instance of the task neural network and generating a plurality of new candidate architectures are nothing more than an abstract mental concept (MPEP 2106.04(a)(2)(III)).
Step 2A – Prong 2
Claims 2, 14 and 23 recite an additional step of processing input that specifies the candidate architecture using a performance prediction neural network having a plurality of performance prediction parameters, wherein the performance prediction neural network is configured to process the input specifying the candidate architecture in accordance with current values of the performance prediction parameters to generate a performance prediction for the candidate architecture prior to instantiating or training a candidate task neural network having the candidate architecture on the particular machine learning task, the performance prediction characterizing how well a neural network having the candidate architecture would perform if the neural network having the candidate architecture were trained on the particular machine learning task that fails to integrate the abstract idea into a practical application. The step of processing the input that specifies the candidate architecture is a form of insignificant input and output solution activity, where processing the input is necessary for all uses of the judicial exception. This additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (MPEP 2106.05(g)).
Claims 2, 14 and 23 recite additional steps of training the instance to perform the particular machine learning task; evaluating a performance of the trained instance on the particular machine learning task to determine an actual performance of the trained instance; and using the actual performances for the trained instances to adjust the current values of the performance prediction parameters of the performance prediction neural network using supervised learning those fail to integrate the abstract idea into a practical application. The steps of training the instance, evaluating a performance and using the actual performances are forms of insignificant input and output solution activities, where training the instance, evaluating a performance and using the actual performances are necessary for all uses of the judicial exception. These additional elements do not integrate the abstract idea into a practical application because they do not impose any meaningful limits on practicing the abstract idea (MPEP 2106.05(g)).
Claims 2, 14 and 23 recite an additional step of for each new candidate architecture in the plurality of new candidate architectures, processing data specifying the new candidate architecture using the performance prediction neural network in accordance with the updated values of the performance prediction parameters to generate a performance prediction for the new candidate architecture that fails to integrate the abstract idea into a practical application. The step of processing data specifying the new candidate architecture is a form of insignificant input and output solution activity, where processing data specifying the new candidate architecture is necessary for all uses of the judicial exception. This additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (MPEP 2106.05(g)).
Step 2B
The additional elements in step 2A-Prong 2 those are forms of insignificant extra-solution activities, do not amount to significantly more than an abstract idea because the court decision have determined that these additional elements of processing the input, training the instance, evaluating a performance, using the actual performances and processing data specifying new candidate architecture to be well-understood, routine, and conventional when claimed in a merely generic manner (MPEP 2106.05(d)(II)).
As such, claims 2, 14 and 23 are not patent eligible.
Dependent claims
Claims 3-13 and 15-22 merely narrow the previously recited abstract idea limitations. For the reasons described above with respect to claims 2 and 14, this judicial exception is not meaningfully integrated into a practical application, or significantly more than the abstract idea. The claims disclose similar limitations described for the independent claims above and do not provide anything more than the mental processes that are practically capable of being performed in the human mind with the assistance of pen and paper and mathematical concepts that are achievable through mathematical computation. Therefore, claims 3-13 and 15-22 also recite abstract ideas that do not integrate into a practical application or amount to significantly more than the judicial exception, and are rejected under U.S.C. 101.
Step 1
Claims 3-13 are drawn to a method of determining an architecture for a task neural network and claims 15-22 are drawn to a system executed by one or more computers to perform the method of claims 3-13. Therefore, each of these claim groups falls under one of four categories of statutory subject matter (process/method, machines/product/apparatus, manufactures, and composition of matter).
Step 2A – Prong 1
Dependent claims 6 and 15 recite further the mental concepts by training a task neural network having the determined architecture; and using the trained task neural network having the determined architecture to perform the particular machine learning task on received network inputs those based on one or more features of the ML project (MPEP 2106.04(a)(2)(III)).
Dependent claims 7 and 16 recite further the mental concepts by generating a new set of candidate architectures by selecting one or more of the new candidate architectures based on the performance predictions for the new candidate architectures that is based on one or more features of the ML project (MPEP 2106.04(a)(2)(III)).
Dependent claims 8 and 17 recite further the mental concepts by selecting one of the new candidate architectures in the new set as the architecture for the task neural network those based on one or more features of the ML project (MPEP 2106.04(a)(2)(III)).
Dependent claims 9 and 18 recite further the mental concepts by generating an instance of the task neural network having the new candidate architecture; training the instance to perform the particular machine learning task; and evaluating a performance of the trained instance on the particular machine learning task to determine an actual performance of the trained instance; and selecting a new candidate architecture corresponding to the trained instance having the best actual performance as the architecture for the task neural network those based on one or more features of the ML project (MPEP 2106.04(a)(2)(III)).
Step 2A – Prong 2
Dependent claims 3-5 recite further the insignificant extra solution activities of the particular machine learning task comprises image processing, image or video classification, and speech recognition. These additional elements do not integrate the abstract idea into a practical application because they do not impose any meaningful limits on practicing the abstract idea (MPEP 2106.05(g)).
Dependent claims 10 and 19 recite further the insignificant extra solution activities of wherein the architecture for the task neural network comprises a plurality of convolutional cells that each share one or more hyperparameters, each of the plurality of convolutional cells comprising one or more operation blocks that each receive one or more respective input hidden states and generate a respective output hidden state, and wherein each candidate architecture and each new candidate architecture defines values for the hyperparameters that are shared by each convolutional cell. These additional elements do not integrate the abstract idea into a practical application because they do not impose any meaningful limits on practicing the abstract idea (MPEP 2106.05(g)).
Dependent claims 11 and 20 recite further the insignificant extra solution activities of wherein each candidate architecture defines an architecture for a convolutional cell having a first number of operation blocks. This additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (MPEP 2106.05(g)).
Dependent claims 12 and 21 recite further the insignificant extra solution activities of wherein the input specifying the candidate architecture is a sequence of embeddings that define the candidate architecture, and wherein the performance prediction neural network is a recurrent neural network. This additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (MPEP 2106.05(g)).
Dependent claims 13 and 22 recite further the insignificant extra solution activities of wherein the performance prediction is an output of the recurrent neural network after processing a last embedding in the sequence of embeddings. This additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (MPEP 2106.05(g)).
As such, dependent claims 3-13 and 15-22 are not 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.
Claims 2-11, 14-20 and 23 are rejected under 35 U.S.C. 103 as being unpatentable over Ravindran et al (US 20160259994 A1) hereafter Ravindran, further in view of Hester et al (US 11604941 B1) hereafter Hester, and further in view of Bowen et al (“Accelerating Neural Architecture Search Using Performance Prediction”) hereafter Bowen.
Bowen was filed on 05/30/2017.
With respect to claim 2, Ravindran teaches a method for determining an architecture for a task neural network in a computationally efficient manner using a (a method provided for building and tuning CNNs for an image processing system [par. 0010]), the method comprising: determining an architecture for a task neural network that is configured to perform a particular machine learning task (a candidate architecture is selected among potential candidate architectures during the iterative process [par. 0010]), comprising:
obtaining data specifying a current set of candidate architectures for the task neural network (a data store is provided with the image processing server to store training sets and validation sets that includes digital images. The iterative process may use these sets to evaluate the performance of CNN [par. 0011, 0021]);
for each candidate architecture in the current set:
generating an input that specifies the candidate architecture (a candidate architecture and candidate parameters are selected during the process, such that the candidate parameters may include some input properties like a learning rate, a batch size, an image size, etc. [par. 0010]);
processing the input that specifies the candidate architecture using a (the performance of an intermediate CNN may be evaluated based on the selected candidate parameters, and then a prediction for each intermediate CNN may be performed to classify an image or an object of an image [par. 0010, 0011]), wherein (the image preprocessor configured to crop and enhance the images from the training set and input those images to the CNN builder. The CNN builder may select some candidate architectures and candidate parameters to train the intermediate CNN. The validation circuit configured to flag an intermediate CNN whether it meets a designated threshold. If an intermediate CNN does meet, it would be selected to be one of ensemble CNNs. The predictions from the ensemble CNNs may be used to classify the images from the inputted digital image without having to train the CNN on the digital images again [par. 0011, 0019]), the performance prediction characterizing how well a neural network having the candidate architecture would perform if the neural network having the candidate architecture were trained on the particular machine learning task (at step 555, the classifier may be used to classify the extent of damage of an object image in a validation set. The predictions gathered from the flagged CNNs after being trained may have a much higher accuracy of 90% compared to other CNNs of approximately 80-85% [par. 0042, 0043 and FIGS. 4, 5]);
generating an updated set of candidate architectures by selecting one or more of the candidate architectures in the current set based on the performance predictions for the candidate architectures in the current set (an updated set of candidate architecture is selected based on the ensemble of the most accurate intermediate CNNs which has a combination of the most accurate predictions [par. 0011]); and
for each candidate architecture in the updated set,
generating an instance of the task neural network having the candidate architecture (a training set configured to create a training set of images that may be used to build a CNN model. The CNN model may be built with the selected candidate architecture and candidate parameters [par. 0025 and FIG. 4]);
training the instance to perform the particular machine learning task; and evaluating a performance of the trained instance on the particular machine learning task to determine an actual performance of the trained instance (the validation circuit at step 416 may be used to train and to evaluate the performance of the CNN model whether it meets a certain threshold [par. 0025 and FIG. 4]); and
for each candidate architecture in the updated set: generating a plurality of new candidate architectures from the candidate architecture by adding, for each new candidate architecture, a respective new operation block having respective hyperparameters to the candidate architecture (based on the performance of the image classification, the method may select candidate architectures those having candidate parameters for each candidate architecture through the iterative process to build, train and optimize a CNN [par. 0010, 0030-0034]); and
for each new candidate architecture in the plurality of new candidate architectures, processing data specifying the new candidate architecture using the performance prediction neural network and in accordance with the updated values of the performance prediction parameters to generate a performance prediction for the new candidate architecture (the intermediate CNN is evaluated based on the performance using the training set, whereas each CNN has different selected values of parameters to generate an ensemble which selects an algorithm to generate the predictions for each CNN. The predictions may be used in CNN builder to select many candidate architectures for the next image classification. The model builder, which is used to tune CNN model, may select candidate architecture from a plurality of candidate architectures [par. 0011, 0019, 0031-0034]).
However, Ravindran does not particularly disclose using the actual performances for the trained instances to adjust the current values of the performance prediction parameters of the performance prediction neural network using supervised learning.
In the same field of endeavor, Hester teaches using the actual performances for the trained instances to adjust the current values of the performance prediction parameters of the performance prediction neural network using supervised learning (a supervised learning technique may be used to adjust the values of both the auxiliary prediction neural network parameters and the action selection neural network parameters to perform a demonstrated task of training neural network system [col. 2, lines 20-65]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated the concept of using supervised learning to adjust the values of neural network parameters as suggested by Hester into the system of selecting candidate architectures based on candidate parameters as suggested by Ravindran because both of these systems addressing the process of training neural networks based on the performances of the candidate architectures. Doing so would be desirable because the system of Ravindran may need an additional step of adjusting parameter values to enhance the step of selecting candidate parameters for the candidate architectures.
However, the combination of Ravindran and Hester does not particularly disclose using a performance prediction neural network.
In the same field of endeavor, Bowen teaches using a performance prediction neural network (modeling learning curves use features based on time-series (TS) validation accuracies, architecture parameters (AP) and hyperparameters (HP). Image classification datasets are trained on deep CNNs (ResNet or Cuda-Convnet) and with LSTMs. By performing final neural network performance, 100 randomly sampled neural network configurations were trained to obtain the best performing method using random hyperparameter search over 3-fold cross-validation. Later, the neural network performance prediction of SRMs is compared with three existing learning curve prediction methods: Bayesian Neural Network, the learning curve extrapolation method and the last seen value heuristic [pages 2-5, 2. Related Work & 3. Neural Network Performance Prediction]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated the concept of performing optimization on neural network hyperparameters and meta-modeling as suggested by Bowen into the combination of Ravindran and Hester because all of these systems addressing the process of making predictions and training neural networks based on the performance of input data. Doing so would be desirable because the combination of Ravindran and Hester would be more efficient by predicting the final performance of partially trained model configurations using features based on network architectures, hyperparameters, and time-series validation performance data (Bowen, [page 1, Abstract]).
With respect to claim 3, the combination of Ravindran, Hester, and Bowen teaches wherein the particular machine learning task comprises image processing (Ravindran, the main task is the image processing which builds and trains ensembles of CNNs [par. 0010]).
With respect to claim 4, the combination of Ravindran, Hester, and Bowen teaches wherein the particular machine learning task comprises image or video recognition (Ravindran, besides images, videos are also used as input to train the CNNs on the still and moving objects [par. 0001, 0013]).
With respect to claim 5, the combination of Ravindran, Hester, and Bowen teaches wherein the particular machine learning task comprises speech recognition (Hester, another input can be used to train the CNNs includes speech which is mentioned as another interaction method with users [col. 14, lines 15-20]).
With respect to claim 6, the combination of Ravindran, Hester, and Bowen teaches the method further comprising:
training a task neural network having the determined architecture (Ravindran, an architecture is determined when training a task of CNN, for instance, a predetermined number of flagged intermediate CNNs is used where each flagged CNN is built for the selected candidate parameters [par. 0010-0012]); and
using the trained task neural network having the determined architecture to perform the particular machine learning task on received network inputs (Ravindran, a set of digital images may be used to train an intermediate CNN. The CNN builder may select a number of architectures and parameters to train this intermediate CNN [par. 0019]).
With respect to claim 7, the combination of Ravindran, Hester, and Bowen teaches the method further comprising: for each new candidate architecture:
generating a new set of candidate architectures by selecting one or more of the new candidate architectures based on the performance predictions for the new candidate architectures (Ravindran, the iterative process selects the candidate architecture from a plurality of candidate architectures and also validates a set of candidate parameters for the selected candidate architectures. The selected candidate parameters may be determined by the ensemble predictions of the intermediate CNNs [par. 0010, 0011]).
With respect to claim 8, the combination of Ravindran, Hester, and Bowen teaches the method further comprising: selecting one of the new candidate architectures in the new set as the architecture for the task neural network (Ravindran, one of the task neural networks provided is an advanced deep learning architecture which exhibits superior classification accuracy to assess property damage [par. 0012]).
With respect to claim 9, the combination of Ravindran, Hester, and Bowen teaches wherein the selecting comprises: for each new candidate architecture in the new set:
generating an instance of the task neural network having the new candidate architecture (Ravindran, a CNN generated may have fewer parameters and may be a more efficient architecture to use to assess the property damage. A CNN may also be used for classifying the extent of damage to a property [par. 0012, 0013]);
training the instance to perform the particular machine learning task (Ravindran, the damage may be an injury or a harm caused to the appearance of the property image. The digital images may be input to the image processing system and be classified based on the amount of damage, and then the system may use the architecture to detect the extent of damage to the property [par. 0012, 0013]); and
evaluating a performance of the trained instance on the particular machine learning task to determine an actual performance of the trained instance (Ravindran, the intermediate CNN may be evaluated based on the performance during the training to determine whether the CNN meets a certain threshold [par. 0011, 0019, 0025]); and
selecting a new candidate architecture corresponding to the trained instance having the best actual performance as the architecture for the task neural network (Ravindran, there are 5 CNNs with the most accuracy will be selected as an ensemble, and the ensemble will be combined with the aggregate predictions to achieve a higher accuracy of almost 90% [par. 0041-0043]).
With respect to claim 10, the combination of Ravindran, Hester, and Bowen teaches wherein the architecture for the task neural network comprises a plurality of convolutional cells that each share one or more hyperparameters (Ravindran, convolution cells are basically defined as convolution neural networks. The architecture, which may be later determined as one of candidate architectures, has a multitude of parameters associated with it [par. 0010]), each of the plurality of convolutional cells comprising one or more operation blocks that each receive one or more respective input hidden states and generate a respective output hidden state (Ravindran, the candidate parameters of candidate architectures include input layers, hidden layers and a number of output layers [par. 0010]), and wherein each candidate architecture and each new candidate architecture defines values for the hyperparameters that are shared by each convolutional cell (Ravindran, each candidate architecture has different values for the selected candidate parameters after each training of an intermediate CNN in each round of iterative process [par. 0010, 0011]).
With respect to claim 11, the combination of Ravindran, Hester, and Bowen teaches wherein each candidate architecture defines an architecture for a convolutional cell having a first number of operation blocks (Ravindran, the iterative process is repeated until a predetermined number of CNNs has met a predefined threshold, such that a candidate architecture may be selected after the iterative process finishes [par. 0010, 0011]).
With respect to claim 14, it is a system claim that corresponding to the method of claim 2. Therefore, it is rejected for the same reason as claimed in claim 2 above.
With respect to claim 15, it is a system claim that corresponding to the method of claim 6. Therefore, it is rejected for the same reason as claimed in claim 6 above.
With respect to claim 16, it is a system claim that corresponding to the method of claim 7. Therefore, it is rejected for the same reason as claimed in claim 7 above.
With respect to claim 17, it is a system claim that corresponding to the method of claim 8. Therefore, it is rejected for the same reason as claimed in claim 8 above.
With respect to claim 18, it is a system claim that corresponding to the method of claim 9. Therefore, it is rejected for the same reason as claimed in claim 9 above.
With respect to claim 19, it is a system claim that corresponding to the method of claim 10. Therefore, it is rejected for the same reason as claimed in claim 10 above.
With respect to claim 20, it is a system claim that corresponding to the method of claim 11. Therefore, it is rejected for the same reason as claimed in claim 11 above.
With respect to claim 23, it is a non-transitory computer storage media claim that corresponding to the method of claim 2. Therefore, it is rejected for the same reason as claimed in claim 2 above.
Claims 12, 13, 21 and 22 are rejected under 35 U.S.C. 103 as being unpatentable over Ravindran et al (US 20160259994 A1) hereafter Ravindran, in view of Hester et al (US 11604941 B1) hereafter Hester, further in view of Bowen et al (“Accelerating Neural Architecture Search Using Performance Prediction”) hereafter Bowen, as claimed in claims 2 and 14 above, and further in view of Audhkhasi et al (US 20170270100 A1) hereafter Audhkhasi.
Bowen was filed on 05/30/2017.
With respect to claim 12, the combination of Ravindran, Hester, and Bowen does not particularly disclose wherein the input specifying the candidate architecture is a sequence of embeddings that define the candidate architecture, and wherein the performance prediction neural network is a recurrent neural network.
In the same field of endeavor, Audhkhasi teaches wherein the input specifying the candidate architecture is a sequence of embeddings that define the candidate architecture (Word embeddings are part of natural language processing. One example is the neural network language model architectures that incorporate external word embeddings as input to produce semantic word embeddings [par. 0020]), and wherein the performance prediction neural network is a recurrent neural network (Continuous space language model configured to use word embeddings to make predictions. Neural network uses this concept to avoid extremely large amount of unique words, and the architecture may be recurrent neural network model [par. 0047]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated the introduction of word embeddings in neural network as suggested by Audhkhasi into the combination of Ravindran, Hester and Bowen because all of these systems addressing the process of training neural networks architecture by performance predictions gathered in the process. Doing so would be desirable because the combination of Ravindran, Hester and Bowen would be more efficient in training the neural network by using word embedding neural network models to expand the knowledge of making predictions based on the amount of input data in a wide variety of applications (Audhkhasi, [par. 0004-0008]).
With respect to claim 13, the combination of Ravindran, Hester, Bowen and Audhkhasi teaches wherein the performance prediction is an output of the recurrent neural network after processing a last embedding in the sequence of embeddings (Audhkhasi, neural network language model is mostly used in processing a sequence of embeddings as a recurrent neural network. This technique predicts and processes a current word based on a predicted current word before ending the operation [par. 0047, 0052, 0074-0078 and FIG. 10]).
With respect to claim 21, it is a system claim that corresponding to the method of claim 12. Therefore, it is rejected for the same reason as claimed in claim 12 above.
With respect to claim 22, it is a system claim that corresponding to the method of claim 13. Therefore, it is rejected for the same reason as claimed in claim 13 above.
Response to Arguments
The examiner respectfully acknowledges the applicant’s amendments to claims 2, 14 and 23.
Applicant’s amendments filed on 12/19/2025 regarding the rejections to claims 2, 14 and 23 under 35 U.S.C. 112(b) have been considered and consequently withdrawn.
Applicant’s arguments filed on 12/19/2025 regarding the rejections to claims 2-23 under 35 U.S.C. 101 have been fully considered but are not persuasive.
Applicant argued that “The claims recite a specific, non-generic technical process: (i) generating performance predictions using a performance prediction neural network before instantiating or training any task neural network having the candidate architecture; (ii) selecting a subset of candidate architectures based on those predictions; (iii) training only that subset; (iv) updating the performance prediction neural network using the actual performance of the trained subset via supervised learning; and (v) recursively generating and evaluating new architectures using the updated performance prediction neural network. As a whole, this ordered combination improves the computational efficiency of computer systems performing neural architecture search by reducing training cost while iteratively refining the prediction model based on selectively trained architectures. The Office has not provided any evidence that it is routine or conventional for a neural architecture search system to dynamically train, update, and recursively interact with a performance prediction neural network in the manner required by claim 2 to determine an architecture for a task neural network configured to perform a particular machine learning task.”
Examiner respectfully disagrees.
Throughout the amendment, Applicant keeps repeating the steps of claim 2: generating performance prediction neural network, selecting a subset of candidate architectures based on predictions, or updating the performance prediction neural network using actual performance; without actually providing a technical solution to a technical problem that can be integrated into a practical application. The techniques reduce training time by using a performance prediction neural network, but this is just a functional result. “Saving time” and “improving efficiency” are often considered a result, not a means.
Even though Applicant uses paragraph 0006 from the Specifications, the paragraph just summarizes the results of using performance prediction neural network. While the Specification describes a reduction in training instances, this is a result that flows from the abstract idea of performance prediction. The claim does not recite the specific technical features, such as the repeated output cell that provides the improvement. The claim is too broad that it covers any method of predicting performance. As such, the claim merely recites a mathematical concept or mental process without an inventive concept that transforms it into a practical application.
Hence, independent claim 2 and its corresponding claims are not patent-eligible for at least the reasons above. Dependent claims 3-13 and 14-22, those are either depended on independent claims 2 and 14, are not patent-eligible for the same reasons.
Applicant’s arguments filed on 12/19/2025 regarding the rejections to claims 2-23 under 35 U.S.C. 103 have been considered and moot in view of new ground of rejection (see rejection above).
Conclusion
Applicant’s amendment necessitated the new grounds of rejection presented in this Office Action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP 706.07(a). Applicant is remined 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 filled 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 extension fee 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 date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Quoc Phung whose telephone number is (703) 756 1330. The examiner can normally be reached on Monday through Friday from 9am to 5pm PT.
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, Jennifer Welch can be reached on 571-272-7212. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Q.L.P./Examiner, Art Unit 2143
/JENNIFER N WELCH/Supervisory Patent Examiner, Art Unit 2143
Prosecution Insights