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 .
Response to Amendments
Claims 1,7-9, and 15 have been amended.
Claims 5-6 and 13-14 have been canceled.
Claims 1-4, 7-12, and 15 remain pending in the application.
The amendment filed 03/03/2026 is sufficient to overcome the 35 U.S.C. 101
rejection of claim 8. The previous rejection has been withdrawn.
Response to Arguments
Argument 1, regarding the 101 rejection, applicant argues that the 101 rejection of claim 8 should be withdrawn in view of amendments made to claim 8. Examiner agrees and the rejection has been withdrawn.
Argument 2, regarding the prior art rejections, applicant argues that none of the cited art teaches “grouping the plurality of neural network architectures as a learning team”. Applicant argues that the cited portion of Alabbasi is directed towards reinforcement learning processes involving selection or evaluation of subsets of nodes or model components and not towards grouping neural network architectures as a learning team. Applicant argues that the claims require multiple neural network architectures within a super network organized together as a team. Examiner notes that the broadest reasonable interpretation of “the plurality of neural network architectures” includes multiple subsets of nodes within a neural network because a subset of nodes may be interpreted as a subnetwork within a neural network. Furthermore, Alabbasi is combined with Benyahia which more explicitly recites multiple neural network architectures. Alabbasi teaches grouping the plurality of neural network architectures as a learning team (Reinforcement learning may include multiple subsets of nodes with different combinations to determine an optimal subset for accuracy and efficiency. Different subsets of nodes are interpreted here as different neural network architectures, P0060, P0080, P0090).
Applicant also argues that none of the cited art teaches “obtaining node classification results corresponding to the plurality of neural networks by classifying the nodes of the subgraph through each neural network architecture in the plurality of neural network architectures”. Applicant argues that the cited portion of Alabbasi (P0145) discusses classification within model framework but does not describe using multiple neural network architectures to classify the same nodes of a subgraph and generating separate node-level classification outputs corresponding to each architecture for comparative purposes. Examiner respectfully disagrees because Alabbasi teaches the model may be trained by minimizing the loss function to encourage the classification model to select a subset of nodes so as to optimize accuracy, (see Alabasi, P0145). Furthermore, a subset of nodes within a neural network is considered to be a “neural network architecture” under the broadest reasonable interpretation of the claim language, and Alabbasi is combined with Benyahia which more explicitly recites multiple neural network architectures.
Applicant also argues that none of the cited art teaches “performing accuracy statistics on the node classification results, and selecting a neural network architecture corresponding to a node classification result with a highest accuracy as the optimal architecture”. Applicant argues that the limitation requires selecting a neural network architecture from multiple neural network architectures based on which neural network architecture has the highest scoring accuracy statistics, and that the cited portion of Alabbasi does not teach this comparative analysis. Examiner respectfully disagrees because P0051 and P0145 recite selecting a subset of nodes for optimizing accuracy, meaning different subsets of nodes are compared based on accuracy and the subset with optimal accuracy is selected. Furthermore, a subset of nodes within a neural network is considered to be a “neural network architecture” under the broadest reasonable interpretation of the claim language, and Alabbasi is combined with Benyahia which more explicitly recites multiple neural network architectures.
Applicant also argues that none of the cited art teaches “obtaining a classification difficulty value of each of the nodes in the subgraph by evaluating classification difficulty of each of the nodes based on the node classification result corresponding to the optimal architecture”. Applicant argues that P0049-P0051 of Alabbasi was cited as teaching this limitation, and argues that these paragraphs are instead directed towards general training objectives, prediction processes, or performance evaluation within a model and does not include an architecture-dependent, node-specific difficulty determination mechanism. Examiner notes that P0145 of Alabbasi was cited as teaching selecting an optimal architecture from the learning team, and obtaining a classification difficulty value of each of nodes in the subgraph by evaluating classification difficulty of each of the nodes through the optimal architecture (Model may be trained by minimizing the loss function to encourage the classification model to select a subset of nodes so as to optimize accuracy, P0145). P0145 of Alabbasi recites selecting a subset of nodes for a classification model based on optimal accuracy statistics, with P0049-P0051 teaching collecting pass/fail rates for separate subsets of nodes.
Applicant also argues that none of the cited art teaches “setting a weight for a loss value of each of the nodes based on the classification difficulty value of each of the nodes”. Applicant argues the cited portion of Benyahia is directed towards reinforcement learning reward signals and architecture evaluation mechanisms, not a node-specific loss weighting scheme based on a classification difficulty value of each node. Examiner respectfully disagrees because P0089-P0093 of Benyahia recites “The cross-entropy loss function may be defined so that minimizing the cross-entropy likelihood provides a maximum likelihood for a classification problem” (see Benyahia P0090) and “the modified loss function may take into account the importance of that weighting (and its associated node) in an earlier-trained model. This may be based on, e.g., an indication of corresponding information (e.g., Fisher information) for that weighting (and node)” (see Benyahia P0089). Thus, a cross-entropy loss function may be updated based on weights associated with specific nodes and likelihood of a classification problem.
Applicant also argues that none of the cited art teaches “obtaining the plurality of trained neural network architectures by adjusting parameters of the plurality of neural network architectures based on the loss value”. Applicant argues that P0090, P0093 of Benyahia is directed towards updating controller parameters or adjusting architecture selection strategies based on reward feedback instead of adjusting parameters of multiple neural network architectures based on a loss value. Examiner respectfully disagrees because P0090 and P0093 of Benyahia recite a loss function being defined for each neural network architecture, with adjusting weights of architectures based on their associated cross-entropy loss function.
Regarding the remaining arguments on the 103 rejection, the examiner finds them not persuasive for the reasons above. The examiner did not rely on hindsight reasoning, and the examiner finds that the combination does not destroy the functionality of the original art or change the principle of operation. The examiner further does not find the arguments regarding the Omeprazole case to be persuasive. Further this case is relevant only to rejections under KSR rationale A (Combining Prior Art Elements According to Known Methods to Yield Predictable Results) and is not relevant to this case.
The full prior art rejections are outlined below.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-4, 7-12, and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Benyahia et al (Pub. No.: US 20240354579 A1), hereafter Benyahia in view of Xu et al (Pub. No.: US 20220108546 A1), hereafter Xu and Alabbasi et al (Pub. No.: US 20240172283 A1), hereafter Alabbasi.
Regarding claims 1, 8, and 9, Benyahia teaches An architecture search method for a large-scale graph (“Systems and/or methods are provided for neural architecture search”, P0004), comprising a computer-readable storage medium storing a computer program, electronic device, comprising a memory, a processor, and a computer program, wherein the computer program is stored on the memory and executable by the processor, wherein when the computer program is executed by the processor, the computer program allows the processor to execute operations of (“As utilized herein, for example, a particular processor and memory (e.g., a volatile or non-volatile memory device, a general computer-readable medium, etc.) may comprise a first “circuit” when executing a first one or more lines of code and may comprise a second “circuit” when executing a second one or more lines of code”, P0013): subgraph sampling for obtaining a subgraph of the large-scale graph by performing local sampling on the large-scale graph (“a controller (e.g., a reinforcement learning controller) may be used to evaluate trained subgraphs and to suggest further subgraphs for sampling based on the results of the evaluation of the trained subgraphs”, P0059); architecture sampling for sampling a plurality of neural network architectures in a pre-constructed super network according to a pre-customized importance sampling strategy (Stochastic policy and reward function outline how architectures are sampled from an existing set of architectures, P0112-P0113); architecture training for obtaining a plurality of trained neural network architectures by training, …, the plurality of neural network architectures with the subgraph (“Providing the neural network may comprise providing a neural network having an architecture corresponding to the subgraph of the best-performing suggested candidate. The weightings for the nodes in the subgraph may be re-trained. For example, re-training may start from the weightings found during the search process or it may start from a set of random weightings”, P0043); … setting a weight for a loss value of each of the nodes based on the classification difficulty value of each of the nodes (The cross-entropy loss function may be defined so that minimizing the cross-entropy likelihood provides a maximum likelihood for a classification problem. The loss may be calculated for each weighting and its associated node, P0090-P0093); obtaining the plurality of trained neural network architectures by adjusting parameters of the plurality of neural network architectures based on the loss value (Weightings for architectures may be updated based on their associated loss function, P0090, P0093); obtaining a trained super network by iteratively executing the subgraph sampling, the architecture sampling, and the architecture training (“Identifying the preferred model may be an iterative process in which a plurality of candidate models are obtained which are then sequentially trained and then evaluated”, P0040); and obtaining an optimal architecture corresponding to the large-scale graph by performing architecture search on the trained super network (After the last iteration, the architecture corresponding to the subgraph of the best-performing candidate is determined to be the optimal model, P0043).
Benyahia does not appear to explicitly teach “according to a peer learning method”.
Xu teaches according to a peer learning method (neural network uses a peer-to-peer learning method on node feature information and structure information for a graph learning task, P0118).
Accordingly, it would have been obvious to a person having ordinary skill in the
art before the effective filing date of the claimed invention, having the teachings of
Benyahia and Xu before them, to include Xu’s specific teaching of a neural network using a peer-to-peer learning method on node feature information for a graph learning task in Benyahia’s system of Neural Network Architecture Search. One would have been motivated to make such a combination of peer-to-peer learning method on node feature information for a graph learning task (see Xu P0118) and reinforcement learning being used to select specific nodes and edges for the purpose of evaluating trained subgraphs and suggest further subgraphs for sampling based on the evaluation (see Benyahia P0059) to improve object detection (see Xu P0005).
Benyahia in view of Xu does not appear to explicitly teach “wherein the architecture comprises: grouping the plurality of neural network architectures as a learning team; selecting an optimal architecture from the learning team; obtaining node classification results corresponding to the plurality of neural networks by classifying the nodes of the subgraph through each neural network architecture in the plurality of neural network architectures; performing accuracy statistics on the node classification results, and selecting a neural network architecture corresponding to a node classification result with a highest accuracy as the optimal architecture; obtaining a classification difficulty value of each of the nodes in the subgraph by evaluating classification difficulty of each of the nodes based on the node classification result corresponding to the optimal architecture”.
Alabbasi teaches wherein the architecture comprises: grouping the plurality of neural network architectures as a learning team (Reinforcement learning may include multiple subsets of nodes with different combinations to determine an optimal subset for accuracy and efficiency. Different subsets of nodes are interpreted here as different neural network architectures, P0060, P0080, P0090); selecting an optimal architecture from the learning team (Model may be trained by minimizing the loss function to encourage the classification model to select a subset of nodes so as to optimize accuracy, P0145); obtaining node classification results corresponding to the plurality of neural networks by classifying the nodes of the subgraph through each neural network architecture in the plurality of neural network architectures (Model may be trained by minimizing the loss function to encourage the classification model to select a subset of nodes so as to optimize accuracy, P0145); performing accuracy statistics on the node classification results (Accuracy of nodes may be measured by success and fail rates, P0051), and selecting a neural network architecture corresponding to a node classification result with a highest accuracy as the optimal architecture (Reinforcement learning is used to encourage the classification model to select nodes to optimize accuracy, P0145); obtaining a classification difficulty value of each of the nodes in the subgraph by evaluating classification difficulty of each of the nodes based on the node classification result corresponding to the optimal architecture (Reinforcement learning agent evaluates the success and fail rate of each node when determining optimal parameters of the classification model, P0049-P0051. Model may be trained by minimizing the loss function to encourage the classification model to select a subset of nodes so as to optimize accuracy, P0145).
Accordingly, it would have been obvious to a person having ordinary skill in the
art before the effective filing date of the claimed invention, having the teachings of
Benyahia, Xu, and Alabbasi before them, to include Alabbasi’s specific teachings of reinforcement learning being used to select a specific subset of nodes and minimizing a loss function to encourage a classification model to select a subset of nodes to optimize accuracy in Benyahia’s system of Neural Network Architecture Search. One would have been motivated to make such a combination of minimizing a loss function to encourage a classification model to select a subset of nodes through reinforcement learning to optimize accuracy (see Alabbasi P0060, P0080, P0090 and P0145) and reinforcement learning being used to select specific nodes and edges for the purpose of evaluating trained subgraphs and suggest further subgraphs for sampling based on the evaluation (see Benyahia P0059) to improve accuracy of a classification network (see Alabbasi P0032).
Regarding claims 2 and 10, Benyahia in view of Xu teaches the limitations of claims 1 and 9 as outlined below. Benyahia further teaches determining a sampling area in the large-scale graph (Input data corresponds to specific nodes and/or edges of a computational graph, P0023); and obtaining the subgraph by sampling nodes and edges in the sampling area (Different subgraphs are defined in the computational graph based on a selected plurality of nodes, edges, and weights that may be unique to their subgraph, P0023).
Regarding claims 3 and 11, Benyahia in view of Xu teaches the limitations of claims 1 and 9 as outlined below. Benyahia further teaches agent decision-making for making, by an agent, a decision to determine the plurality of neural network architectures in the pre-constructed super network in the architecture sampling, and sampling the plurality of neural network architectures (Controller evaluates different model architectures for the neural network, P0043. A controller selects subgraphs, P0107); graph data processing for obtaining a graph data processing result by processing graph data through the plurality of neural network architectures (At step 440, subgraphs are evaluated using validation data sets. stochastic policy is used to evaluate all possible architectures based on validation perplexity, P0109, P0113); reward value returning for returning a reward value to the agent based on an accuracy of the graph data processing result (Controller may use reinforcement learning where the controller receives a reward based on the actions it takes, such as selecting an architecture or subgraph, P0110-P0113); strategy adjusting for adjusting, based on the reward value, a strategy for a next sampling by the agent (Stochastic policy and reward function outline how architectures are sampled from an existing set of architectures, P0112-P0113); and obtaining a trained agent by iteratively executing the agent decision-making, the graph data processing, the reward value returning, and the strategy adjusting, wherein an architecture sampling strategy executed by the agent is the importance sampling strategy (Steps 420-470 may be performed and looped, with each iteration the policy for the controller to follow based on evaluations of trained models, P0121).
Regarding claims 4 and 12, Benyahia in view of Xu teaches the limitations of claims 1 and 9 as outlined below. Benyahia further teaches wherein the super network comprises all graph network layers, each comprising all information transport methods (Each layer of neural network contains multiple edges used for transporting information throughout the network, P0066, P0079, FIG 1, FIG 2, P0161).
Regarding claims 7 and 15, Benyahia in view of Xu and further in view of Alabbasi teaches the limitations of claims 1 and 9 as outlined above. Alabbasi further teaches setting a low weight for a loss value of a node in response to determining that the classification difficulty value of the node is high (Loss function may encourage the classification model to avoid nodes associated with generating false data, P0145, P0148); and setting a high weight for the loss value of the node in response to determining that the classification difficulty value of the node is low (Loss function may encourage the classification model to select nodes that may optimize accuracy, P0145, P0086).
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ISHAN MOUNDI whose telephone number is (703)756-1547. The examiner can normally be reached 8:30 A.M. - 5 P.M..
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, Matthew Ell can be reached at (571) 270-3264. 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.
/I.M./Examiner, Art Unit 2141
/MATTHEW ELL/Supervisory Patent Examiner, Art Unit 2141
Prosecution Insights