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 Amendment
The office action is responsive to the amendment filed on 10/24/2025. As directed by the amendments claims 1-3, 10 and 18 are amended. Claim 16 is canceled. Claim 26 is new. Claims 1-15 and 17-26 are pending for examination.
Response to Arguments
Regarding the 35 U.S.C § 112 Rejection:
In view of amendments to claims 2 and 3, see Remarks pg. 2 filled on 10/24/2025, the rejection of claims 2 and 3 under 35 U.S.C § 112 has been withdrawn.
Regarding the 35 U.S.C § 101 Rejection:
Applicant’s arguments, see pg. 6-9, filed 10/24/2025 with respect to rejection of claims 1-25 under 35 U.S.C § 101 have been fully considered and are persuasive. The rejection of claims 1-25 under 35 U.S.C § 101 has been withdrawn.
Regarding the 35 U.S.C § 103 for claims 1, 10, and 18:
Applicant's further arguments see pg. 10-13 filed 10/24/2025 have been fully considered but they are not persuasive.
APPLICANT ARGUMENT:
Applicant argues the following: Regarding independent claims 1, 10, and 18, the cited references do not teach or suggest, "expand the pruned and permuted machine learning model to un-prune zero values within partially-zero-valued pruned packing shapes".
EXAMINER RESPONSE: Examiner respectfully disagree, applicant argument is not persuasive. Independent claim 1 and similar independent claims 10 and 18 were amended to include the limitation drawn from cancelled claim 16, however, the claim as presented does not teach or suggest what constitutes “expand the pruned and permuted machine learning model to un-prune zero values within partially-zero-valued pruned packing shapes” (emphasis added), in particular, partially uprunning does not suggest that the model is not pruned completely, therefore under the broasted reasonable interpretation, “expanding the pruned and permuted machine learning model to un-prune zero values within partially-zero-valued pruned packing shapes” can be view as uprunning the machine learning model in order to restore accuracy. For which Romaszkan pg. 7, para. 2, teaches “Before pruning is fixed, the pruned packs can still be unpruned if sum of another pack drops lower than the pruned pack. The network is trained for a few more epochs to determine which packs to prune. Finally, the packs to be pruned are fixed, and the model is fine-tuned to further improve” thus suggesting the model being pruned and unpruned during multiple epochs. In addition, the cited reference, Wang, that was used to disclose the limitation from cancelled claim 16 also teaches restoring (unpruned) the model, when the accuracy of the model is being degraded (see Fig. 2 element 208 & [0031]).
Therefore, for the reasons stated above, the combination of the references Cai, Romaszkan and in addition with Wang teach the limitations of amended independent claims 1, 10 and 18. Accordingly, claims 1,10, and 18 are not patent eligible under 35 U.S.C § 103. Claims 2-15 and 17-26 are dependent on claims 1,10, and 18, thus are not allowable for the same reasons as claims 1,10 and 18 discussed above.
Regarding the 35 U.S.C § 103 for claim 2:
Applicant’s arguments with respect to claim 2 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Regarding the 35 U.S.C § 103 for claim 15:
Applicant's further arguments see pg. 11-12 filed 10/24/2025 have been fully considered but they are not persuasive.
APPLICANT ARGUMENT:
Applicant argues the following: Regarding dependent claim 15, the cited references do not teach or suggest, "alternating between permuting rows and columns of weight matrices
corresponding to weights between layers of the machine learning model until a
convergence is detected".
EXAMINER RESPONSE: Examiner respectfully disagree, applicant argument is not persuasive. Chen pg. 4, Fig. 3 teaches an example of weight permutation that comprising alternating between the row and the column of a matrix (see Fig. 3 below). Further, Chen teaches the use of an “annealing algorithm for optimization” such as “simulated annealing (SA)... to find a close-to optimal solution efficiently” thus to find a “convergence” during the permutation ( see Chen pg. 3, right colm, last para.; pg. 4 left colm. para. 1; pg. 5, sec: C. Simulated Annealing Based Permutation, para. 1). Accordingly, claim 15 is not patent eligible under 35 U.S.C § 103.
PNG
media_image1.png
300
527
media_image1.png
Greyscale
Regarding the 35 U.S.C § 103 for claim 26:
Applicant’s arguments with respect to claim 26 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Information Disclosure Statement
The information disclosure statement filed 08/01/2025 fails to comply with the provisions of 37 CFR 1.97, 1.98 and MPEP § 609 because a copy of cited reference 1, was not provide. Examiner notes that applicant provided copy for reference 1 with a different date as the one cited in the IDS. It has been placed in the application file, but the information referred to therein has not been considered as to the merits. Applicant is advised that the date of any re-submission of any item of information contained in this information disclosure statement or the submission of any missing element(s) will be the date of submission for purposes of determining compliance with the requirements based on the time of filing the statement, including all certification requirements for statements under 37 CFR 1.97(e). See MPEP § 609.05(a).
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 1-2, 4, 8-12, 14, 18-21, and 25 are rejected under 35 U.S.C. 103 as being unpatentable over Cai et al. Hunter: HE-Friendly Structured Pruning for Efficient Privacy-Preserving Deep Learning (hereinafter Cai) as disclosed in the Information Disclosure Statement (IDS) dated 07/05/2022 in view of Romaszkan et al. 3PXNet: Pruned-Permuted-Packed XNOR Networks for Edge Machine Learning (hereinafter Romaszkan)as disclosed in IDS dated 09/20/2023 in further view of Wang et al. US 2019/0080238 A1 (hereinafter Wang).
Regarding claim 1:
Cai teaches A system, comprising a processor to: ( Cai pg. 939, right col., sec: 4 Evaluation, para. 1, teaches “We implement Hunter with Pytorch on a machine with Nvidia RTX5000 16G GPU, Intel Xeon Silver 4214 12-core CPU and 32G RAM”. As would be familiar to one skilled in the art, a machine is a type of system, and both the CPU and RAM will require a processor to run).
prune a machine learning model based on an importance of neurons or weights; ( Cai pg. 934, right col., para. 2., teaches a proposed framework named “Hunter” that prunes the model parameter (weights) and pg. 932, left col., sec: Preliminary observation, para. 1 and 3 teaches model pruning is done based on the importance of the parameters such the parameter below a threshold).
...reduce an amount of ciphertext computation under a selected constraint; and (Cai pg. 934, left col., para. 3, teaches “Hunter” the proposed framework “features with a significant reduction for HE operations including Perm [permutation] and thus also contributes to the reduction for PerDecomp and HstPerm operations... Hunter aims to minimize the computation complexity of Perm and thus correspondingly reduces the HE-based computation in privacy preserving MLaaS” this suggest, Hunter ability to reduce cyphertext computation by enabling “HE friendly pruning strategies” to reduce computation complexity).
While Cai teaches the propose method Hunter can be used to reduce an amount of ciphertext computation under selected constrains, Cai does not explicitly teach permute and pack remaining neurons or weights of the pruned machine learning model to reduce an amount of ciphertext computation under a selected constraint; and expand the pruned and permuted machine learning model to un-prune zero values within partially-zero-valued pruned packing shapes.
Nevertheless, Romaszkan teaches the following:
permute and pack remaining neurons or weights of the pruned machine learning model to reduce an amount of (Romaszkan, abstract, pg. 7 teaches a pruned, permuted and packed binarized network referred to as 3PXNet. Furthermore, Romaszkan pg. 4, Sec: 3.1 XNOR Networks, teaches packing weights and pg. 6, para. 2., teaches performing weights permutation on a fully connected or convolutional layer and Fig. 2 (b) teaches packing constrains of 4 bits being performed. In addition, pg. 2, sec: 1.1 A Case for Sparse XNOR Networks, para. 3, teaches the proposed model (3PXNet) aims “to combine binarization and pruning in a way that is computationally efficient and does not significantly degrade accuracy”).
...the pruned and permuted machine learning model... ( Romaszkan pg. 2, sec: 1.1 A Case for Sparse XNOR Networks, para. 3 teaches 3PXNet, a pruned, permuted and packed XNOR neural network).
Romaszkan is also in the same field of endeavor as Cai (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of permute and pack a model, as being disclosed and taught by Romaszkan, in the system taught by Cai to yield the predictable results of efficiently implementing the prosed Pruned-Permuted-Packed XNOR Networks (3PXNet) on even the smallest of embedded microcontrollers and to provide a software implementation that is “complete with training methodology and model generating scripts, making it easy and fast to deploy” (Romaszkan Abstract).
While Romaszkan pg. 7, para. 2, teaches uprunning, specifically “for packs that are not pruned, the weights inside are forced into {-1, +1}, even if they were originally 0. Before pruning is fixed, the pruned packs can still be unpruned if sum of another pack drops lower than the pruned pack”. Neither Cai or Romaszkan suggest expand the pruned and permuted machine learning model to un-prune zero values within partially-zero-valued pruned packing shapes.
Nevertheless, Wang teaches the following:
expand the pruned (Wang teaches Fig. 1 where a machine learning model is being pruned element 104 and [0021] teaches the pruning includes "clamping the weights of connections that are close to zero to be zero". Further, Fig. 2 element 208 and [0031] teaches restore (unpruned) the model, when the accuracy of the model is being degraded. This suggest the “restoring” would unpruned the model, to return it to its original state rather than utilizing the latest version of the model (prune model)).
Wang is also in the same field of endeavor as Cai and Romaszkan (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of restore (unpruned) the model to its original state as being disclosed and taught by Wang, in the system taught by Cai and Romaszkan to yield the predictable results of proving “an improved approach for pruning a CNN in order to increase the number of zero coefficients while still achieving high levels of accuracy” ( Wang [0004]).
Regarding claim 2:
Cai, Romaszkan and Wang teaches The system of claim 1. Cai Abstract specially teaches the proposed framework Hunter, output a pruned model without any loss in accuracy that achieves a significant reduction in Homomorphic Encryption (HE) and Romaszkan teaches wherein the processor is to prune and pack based on a combination of pruning, packing, and permutation determined using an objective function ( Romaszkan pg. 2, sec: 1.1 A Case for Sparse XNOR Networks, para. 3 teaches 3PXNet, a pruned, permuted and packed XNOR neural network and teaches “the network being trained for few more epochs” (see pg. 7, para. 2). In addition Romaszkan, pg. 7, para. 2, teaches an objective function (i.e., sum of another pack drops lower than the pruned pack) is being used to ensure the network is not losing accuracy and to ensure the accuracy of the network is maintained).
Regarding claim 4:
Cai, Romaszkan and Wang teaches The system of claim 1. Cai specifically teaches wherein the importance is based on values of the weights ( Cai pg. 933, right col., sec: 2.2 Model Pruning, lines 5-8, teaches pruning the model parameter (weights) that are smaller than a threshold, this implies weights are pruned based on their values such as threshold values).
Regarding claim 8:
Cai, Romaszkan and Wang teaches The system of claim 1. Cai specifically teaches wherein pruning the machine learning model comprises eliminating an operation from the machine learning model (Cai pg. 936, left col., sec: Hunter’s diagonal pruning structure, para. 2, teaches reducing (eliminating) the Perm operations...in one fully-connected layer of AlexNet (machine learning model)).
Regarding claim 9:
Cai, Romaszkan and Wang teaches The system of claim 1. Cai specifically teaches wherein the ciphertext computation comprises an execution of a homomorphically encrypted inference of the pruned, permuted, and packed machine learning model ( Cai pg. 933, right col., sec: 2.4 Packed Homomorphic Encryption, para. 1, teaches “Homomorphic Encryption (HE) is a kind of encryption that allows linear computation over ciphertext without decryption” and pg. 933, left col., para. 1, teaches the client encrypts its private data to obtain ciphertext, which is then used to perform model inference and teaches how the proposed framework “Hunter” considers packed nature of Homomorphic Encryption (HE) to train, prune, permute and pack a model (pg. 934, left col., para. 2)). Regarding claim 10: is similar to claim 1, therefore the rejection of claim 1 applies. Claim 10 only recites the additional element of A computer-implemented method, comprising... via processor for which Cai Abstract teaches Hunter a computer implemented method and pg. 939, right col., sec: 4 Evaluation, para. 1, teaches a machine, that inheritably require a processor to run.
Regarding claim 11: is similar to claim 9, therefore the rejection of claim 9 applies.
Regarding claim 12:
Cai, Romaszkan and Wang teaches The computer-implemented method of claim 10. Cai specifically teaches wherein pruning the machine learning model comprises pruning a weight of the machine learning model by setting weights with values that do not exceed a threshold to zero ( Cai pg. 932, left col., last para. and right col., para. 1, teaches pruning the parameters (weights) of the model (machine learning model) by setting the weights values below a threshold to zero).
Regarding claim 14:
Cai, Romaszkan and Wang teaches The computer-implemented method of claim 10. Romaszkan specifically teaches wherein permuting the machine learning model comprises using a balanced clustering ( Romaszkan pg. 6, para. 2, teaches a permuted neural network (machine learning model) where “Weight permutations are performed on input channels (NI) of a fully connected or convolution layer”. Furthermore, pg. 6, para. 3, teaches “Permutation tries to group (cluster) similar input channels into packs of 32 in a method resembling Prim’s algorithm” which implies the cluster will be the same size, that is a balanced clustering).
Regarding claim 18: is similar to claim 1, therefore the rejection of claim 1 applies. Claim 18 only recites the additional element of A computer program product for pruning and packing machine learning models, the computer program product comprising a computer-readable storage medium having program code embodied therewith, the program code executable by a processor to cause the processor to... for which Cai pg. 939, right col., sec: 4 Evaluation, para. 1, teaches “We implement Hunter with Pytorch on a machine with Nvidia RTX5000 16G GPU, Intel Xeon Silver 4214 12-core CPU and 32G RAM”, which inheritably will required a computer program product comprising a computer-readable storage medium having program code.
Regarding claim 19: is similar to claim 12, therefore the rejection of claim 12 applies.
Regarding claim 20: is similar to claim 14, therefore the rejection of claim 14 applies.
Regarding claim 21: is similar to claim 14, therefore the rejection of claim 14 applies.
Regarding claim 25: is similar to claim 9, therefore the rejection of claim 9 applies.
Claims 3, 13, and 23 are rejected under 35 U.S.C. 103 as being unpatentable over Cai, Romaszkan, Wang in further view of Shen et al. US 2022/0292360 A1 (hereinafter Shen).
Regarding claim 3:
Cai, Romaszkan and Wang teaches The system of claim 1. However, neither Cai, Romaszkan and Wang disclose wherein the importance is based on the criticality of the neurons.
Nevertheless, Shen teaches the following:
wherein the importance is based on the criticality of the neurons ( Shen [0091] teaches “a neural network pruning 114 prunes a neural network 102 such that only top-k neurons of neural network 102 remain... top-k neurons correspond to one or more neurons (e.g., nodes) that are most likely to affect performance of a neural network ”, this implies that the pruning is being performed based on the criticality of the neurons).
Shen is also in the same field of endeavor as Cai, Romaszkan and Wang (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of pruning based on the importance of the neurons, as being disclosed and taught by Shen, in the system taught by Cai, Romaszkan and Wang to yield the predictable results of “determine an optimal time to prune one or more neurons from a neural network; an ability to reduce complexity of a neural network during one or more training processes... [and] ...reduce training time of a pruned neural network” ( Shen [0061]).
Regarding claim 13:
Cai, Romaszkan and Wang teaches The computer-implemented method of claim 10. Neither Cai, Romaszkan or Wang suggest wherein pruning the machine learning model comprises pruning a neuron of the machine learning models.
However, Shen teaches the following:
wherein pruning the machine learning model comprises pruning a neuron of the machine learning models ( Shen [0058] teaches pruning refers to remove (pruning) a neuron of the neural network (machine learning model)).
Regarding claim 23:
Cai, Romaszkan and Wang teaches The computer program product of claim 18. Cai specifically teaches further comprising program code executable by the processor to retrain the pruned machine learning model and execute the pruned machine learning model to obtain an accuracy score for the pruned machine learning model ( Cai pg. 939, right col., sec: 4 Evaluation, para. 1, teaches implementing the proposed framework “Hunter with Pytorch on a machine with Nvidia RTX5000 16G GPU, Intel Xeon Silver 4214 12-core CPU and 32G RAM”, which inheritably will required a computer program product comprising a computer-readable storage medium having program code). Furthermore, Cai pg. 932, left col., para. 1, & pg. 934, Figure 2, teaches retraining or finetuning the pruned model to recover accuracy” and pg. 940, Table 2, teaches obtaining accuracy percentages (accuracy score) associated with Hunter (pruned model)).
While Cai teaches retraining or finetuning the pruned model to recover accuracy, and teaches obtain accuracy scores. Neither Cai, Romaszkan or Wang explicitly disclose the pruned machine learning model is particular associated with a pruning threshold.
However, Shen teaches the following:
...the pruned machine learning model associated with a particular pruning threshold ( Shen [0065] teaches a pruning ratio indicating a ratio of a number of neurons of a neural network (e.g. pruned machine learning model) that are to be removed to a total number of neurons of said neural network, and is implemented using a data type such as an “integer, floating-point number, character, string, and/or variations thereof. For example, a prune ratio with a value of 0.3 indicates that 30% of neurons of a neural network are to be pruned, resulting in 70% of said neurons of said neural network remaining after one or more pruning processes”).
Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over Cai, Romaszkan, Wang in further view of Desappan et al. US 2019/0012559 A1 (hereinafter Desappan).
Regarding claim 5:
Cai, Romaszkan and Wang teaches The system of claim 1. Neither Cai, Romaszkan or Wang suggest wherein the selected constraint comprises an inference accuracy constraint.
Nevertheless, Desappan teaches the following:
wherein the selected constraint comprises an inference accuracy constraint (Desappan [0004] teaches the constrain comprises “inference accuracy”).
Desappan is also in the same field of endeavor as Cai, Romaszkan and Wang (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of inference accuracy constraints, as being disclosed and taught by Desappan, in the system taught by Cai, Romaszkan and Wang to yield the predictable results of “reduce the inference cost of neural networks for widespread application of DNNs to mobile and embedded systems” (Desappan [0003]).
Claims 6-7 are rejected under 35 U.S.C. 103 as being unpatentable over Cai, Romaszkan, Wang in further view of Laganakos et al. US 2023/0177116 A1 (hereinafter Laganakos).
Regarding claim 6:
Cai, Romaszkan and Wang teaches The system of claim 1. Neither, Cai, Romaszkan and Wang suggest wherein the selected constraint comprises a memory constraint.
Nevertheless, Laganakos teaches the following:
wherein the selected constraint comprises a memory constraint ( Laganakos [0021] teaches the constraint comprises memory constraints).
Laganakos is also in the same field of endeavor as Cai, Romaszkan and Wang (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of memory related constraints, as being disclosed and taught by Laganakos, in the system taught by Cai, Romaszkan and Wang to yield the predictable results of “synthesize a neural network for performing said task subject to a given set of constraints” (Laganakos Abstract).
Regarding claim 7:
Cai, Romaszkan and Wang teaches The system of claim 1. Neither, Cai, Romaszkan and Wang suggest wherein the selected constraint comprises a latency constraint.
Nevertheless, Laganakos teaches the following:
wherein the selected constraint comprises a latency constraint ( Laganakos [0021] teaches the constraint comprises latency constraints).
Claims 15, 22 and 26 are rejected under 35 U.S.C. 103 as being unpatentable over Cai, Romaszkan, Wang in further view of Chen et al. Tight Compression: Compressing CNN Through Fine-Grained Pruning and Weight Permutation for Efficient Implementation (hereinafter Chen).
Regarding claim 15:
Cai, Romaszkan and Wang teaches The computer-implemented method of claim 10.
Cai, Romaszkan and Wang do not explicitly teach wherein permuting the machine learning model comprises alternating between permuting rows and columns of weight matrices corresponding to weights between layers of the machine learning model until a convergence is detected.
However, Chen teaches the following:
wherein permuting the machine learning model comprises alternating between permuting rows and columns of weight matrices corresponding to weights between layers of the machine learning model until a convergence is detected ( Chen pg. 4, Algorithm 1, teaches weight matrix compression of a neural network model
(
M
)
and teaches after training, the compression goes to a second stage line 10, “where the sparse weight matrix
(
m
l
)
of each layer is further compressed to a small dense format through weight permutation” such that an optimal permutation result can be obtained by applying a “simulated annealing algorithm for optimization” (pg. 3, right col., last para. & pg. 4, left col., para. 1). In addition, Chen pg. 4, Fig. 3 teaches an example of weight permutation that comprises alternating between the row and the column of the matrix).
Chen is also in the same field of endeavor as Cai, Romaszkan and Wang (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of weight permutation that comprises alternating between the row and the column, as being disclosed and taught by Chen, in the system taught by Cai, Romaszkan and Wang to yield the predictable results of improved the compression performance for the model ( pg. 1, right col., 2nd bullet point).
Regarding claim 22: is similar to claim 15, therefore the rejection of claim 15 applies.
Regarding claim 26:
Cai, Romaszkan, Wang and Chen The computer-implemented method of claim 15. Chen specifically teaches wherein the convergence is detected in response to a row and column permutation not resulting in any additional zero- tiles ( Chen Fig. 3 and pg. 4, right colm. para. 1, teaches after multiple permutations the “the number of zero weight is reduced from 17 to 1, and the size of the weight matrix is reduced by 50 %” thus suggesting an optimal permutation result (i.e., convergence) is detected in response of a row and column permutation not resulting in any additional zero-tile. This can also be seen in Fig. 4, where multiple permutations are being performed in order to obtained a compressed matrix with less zero-tiles).
Claims 17 and 24 are rejected under 35 U.S.C. 103 as being unpatentable over Cai, Romaszkan, Wang in further view of Aharoni et al. Tile Tensors: A versatile data structure with descriptive shapes for homomorphic encryption (hereinafter Aharoni) as disclosed in the IDS dated 07/05/2022 in further view of Dong et al. DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures (hereinafter Dong).
Regarding claim 17:
Cai, Romaszkan and Wang teaches The computer-implemented method of claim 10 and teaches the pruned, permuted, and packed machine learning model ( Romaszkan pg. 2, sec: 1.1 A Case for Sparse XNOR Networks, para. 3 teaches 3PXNet, a pruned, permuted and packed XNOR neural network).
Cai, Romaszkan and Wang do not teach further comprising simulating the pruned and packed machine learning model to obtain a latency score and memory score associated with a plurality of packing shapes and pruning thresholds, wherein the pruned, permuted, and packed machine learning model comprises a pruning threshold and a packing shape that minimizes an objective function based on the selected constraint.
Nevertheless, Aharoni teaches the following:
further comprising simulating the ( Aharoni version (2020) pg. 11, TABLE III, teaches performing inference (simulation) using different tile shape (packing shapes) and it obtains the latency score and the memory scores).
Aharoni is also in the same field of endeavor as Cai, Romaszkan and Wang (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of simulating the machine learning model to obtain latency and memory scores related to tile shapes, as being disclosed and taught by Aharoni, in the system taught by Cai, Romaszkan and Wang to yield the predictable results of “improved latency for neural network inference under encryption using a new technique for matrix multiplication, and some automated optimizations” ( Aharoni, pg. 13, right col., sec: Conclusion).
Aharoni does not teach a latency score and memory score associated with a plurality of pruning threshold , and wherein the pruned, permuted, and packed machine learning model comprises a pruning threshold and a packing shape that minimizes an objective function based on the selected constraint.
Dong teaches the following:
...a latency score and memory score associated with a plurality of ...pruning threshold (Dong pg. 10, sec: 6.2 Results on CIFAR-10, para. 2, teaches the latency score (latency time) is associated with the device-agnostic objectives of the model (configurations such as the pruning threshold and packing shapes ) and pg. 11, Table 2, teaches how the device agnostic objective and the device aware metrics such as latency scores and memory score are associated).
wherein the ( Dong pg. 2, para. 3 and Fig. 1 teaches the given device related and device agnostic objectives (pruning threshold and a packing shape) can be used to optimize for multiple objective. In addition, pg. 8, sec: 5.2 Pareto Optimality teaches using “Pareto Optimality” to optimizing the problem, as would be familiar to one ordinary in the skilled art, “Pareto Optimality” can aid minimizing objective functions by identifying a solution where no solution can be improved in one objective without worsening another ).
Dong is also in the same field of endeavor as Cai, Romaszkan, Wang and Aharoni (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of device agnostic objectives associated with latency and memory score as well the device agnostic objectives minimizing the objective function with the aid of Pareto Optimality, as being disclosed and taught by Dong, in the system taught by Cai, Romaszkan, Wang and Aharoni to yield the predictable results of achieving better performances, higher accuracy & shorter inference time on various devices ( Dong Abstract).
Regarding claim 24: is similar to claim 17, therefore the rejection of claim 17 applies.
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 GISEL G FACCENDA whose telephone number is (703)756-1919. The examiner can normally be reached Monday - Friday 8:00 am - 4:00 pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Abdullah Al Kawsar can be reached at (571) 270-3169. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/G.G.F./Examiner, Art Unit 2127
/ABDULLAH AL KAWSAR/Supervisory Patent Examiner, Art Unit 2127
Prosecution Insights