DETAILED ACTION
Remarks
Claims 1-20 have been examined and rejected. This Office Action is responsive to the amendment filed on 04/17/2026, 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 1-3, 5-7, 9-11, 13-15 and 17-19 are presented for examination.
Response to Amendment
Applicant’s amendment filed on 04/17/2026 has been entered. Claims 1, 5, 9, 13 and 17 are amended. Claims 4, 8, 12, 16 and 20 are cancelled. Claims 1-3, 5-7, 9-11, 13-15 and 17-19 are pending in the application.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-3, 5-7, 9-11, 13-15 and 17-19 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
With respect to claim 1,
It is unclear how the limitation “the error being a difference between a tensor before reduction and a tensor after reduction of elements of the layer the threshold being an upper limit of a reduction ratio of the elements” [line 6] is structured. The phrase “elements of the layer the threshold being an upper limit” is confusing. The phrase may need a proper punctuation mark to separate the sentences. For the purposes of examination, Examiner will interpret the limitation as “the error being a difference between a tensor before reduction and a tensor after reduction of elements of the layer; with the threshold being an upper limit of a reduction ratio of the elements”
It is unclear what “a plurality of the thresholds calculated for each of the layers” [line 12] refers to. Claim 1 first recites “a threshold of an error in a tensor” and the error is a difference of elements of the layer. The phrase “a plurality of the thresholds calculated for each of the layers” actually refers to multiple thresholds calculated for a single layer, which contradicts with what was previously recited. For the purposes of examination, Examiner will interpret the limitation as “each threshold of a plurality of the thresholds calculated for each of the layers.”
With respect to claims 9 and 17, they are corresponding to claim 1 and are rejected for the same reasons as explained above.
With respect to claims 2, 3, 5-7, 10, 11, 13-15, 18 and 19, they are rejected based on their virtual dependency of claims 1, 9 and 17.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-3, 5-7, 9-11, 13-15 and 17-19 are rejected under 35 U.S.C. 103 as being unpatentable over Pravendra Singh et al (“Play and Prune: Adaptive Filter Pruning for Deep Model Compression”) hereafter Singh, further in view of Cho et al (US 20210081798 A1) hereafter Cho, and further in view of Ji et al (US 10832135 B2) hereafter Ji.
Singh was cited in the IDS filed on 03/22/2023.
With respect to claim 1, Singh teaches a non-transitory computer-readable recording medium having stored therein a machine learning program for causing a computer to execute a process (a Play & Prune (PP) framework configured to prune and to fine-tunes CNN model parameters with an adaptive pruning rate while maintaining the model’s predictive performance [page 1, Abstract]) comprising:
calculating for each of a plurality of layers included in a trained model of a neural network, a threshold of an error in a tensor, the error being a difference between a tensor before reduction and a tensor after reduction of elements of the layer (L1 regularization constant is employed in the cost function. An adaptive weight threshold is chosen for each layer Li, such that results in negligible accuracy drop after removing. Adaptive Filter Pruning (AFP) is configured to minimize the number of filters in the model, wherein the weight thresholds W is calculated initially. The pruning rate is changed using the pruning rate controller (PRC). Equation (6) is used to calculate the adaptive thresholds WA [page 3, 3.3. Weight Threshold Initialization – page 4, 3.5. Pruning Rate Controller (PRC)]);
selecting, for each of the plurality of layers, a reduction ratio candidate to be applied to the layer based on a plurality of the thresholds calculated for each of the layers and the errors in tensors in cases where the elements are reduced by each of a plurality of reduction ratio candidates in the layers (the pruning module P needs to identify the candidates to be pruned. U and I are the set of unimportant and important filters respectively. Equation (2) and (3) indicate the approach of calculating filter importance uses the L1 norm. U is treated as a candidate set of filters to be pruned which is a subset that will be pruned eventually. Parameter α is treated as the reduction ratio in the Equation (3) [page 3, 3.2. Convolutional Filter Partitioning – 3.3. Weight Threshold Initialization]); and
determining, for each of the plurality of layers, a reduction ratio to be applied to the layer based on retrained model of a reduced model, the reduced model being obtained by reducing the elements of the plurality of layers in the trained model according to the reduction ratio candidate to be applied (the AFP minimizes the number of filters in the network, and the PRC optimizes the accuracy given that number of filters. Figure 1 shows how AFP minimizes while PRC maximizes the accuracy during training. Equation (7) is given where C(#w) is the accuracy with #w remaining filters, ε is the accuracy of the unpruned network, and C(#w)-( ε- ϵ) indicates the gap of accuracy and tolerance error level [page 2, 2. Related Work – page 4, 3. Proposed Approach and FIG. 1]).
However, Singh does not explicitly teach determining, for each of the plurality of layers, a reduction ratio to be applied to the layer based on inference accuracy of the trained model and inference accuracy of a retrained model of a reduced model; and the calculating includes scaling the thresholds such that an L2 norm of the thresholds of the plurality of layers becomes equal to or smaller than a threshold upper limit; the process further comprises repeating execution of the calculating, the selecting, and the determining until execution times or the reduction ratios satisfy a predetermined condition; and outputting the reduction ratios determined when the predetermined condition is satisfied, and the predetermined condition is satisfied when a size of a memory of the computer required to store the reduced model reaches a saturated size.
In the same field of endeavor, Cho teaches determining, for each of the plurality of layers, a reduction ratio to be applied to the layer based on inference accuracy of the trained model and inference accuracy of a retrained model of a reduced model (the change in inference accuracy may include calculating sensitivity for each of the plural layers based on the difference between an inference accuracy before pruning on each layer is performed and an inference accuracy after pruning on each of the plural layers is performed. A compression method is performed to reduce the size of a neural network, reduce system costs, and reduce the amount of computations in the implementation of a neural network. The pruning process of a neural network may comprise the compression or removal of the connectivity between nodes [par. 0010-0014, 0066-0071]).
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 pruning a neural network with setting a weight threshold value based on a weight distribution of layers included in a neural network as suggested by Cho into the concept of using Play and Prune framework to prune and to fine-tune the CNN parameters as suggested by Singh because both of these systems addressing the process of pruning neurons/parameters in the neural networks. Doing so would be desirable because the concept of Singh would be more efficient by predicting a change in inference accuracy by calculating a sensitivity each of the plurality of layers based on the difference between an inference accuracy before pruning and an inference accuracy after pruning, while adjusting the weight threshold value (Cho, [par. 0010-0015]).
However, the combination of Singh and Cho does not particularly disclose the calculating includes scaling the thresholds such that an L2 norm of the thresholds of the plurality of layers becomes equal to or smaller than a threshold upper limit; the process further comprises repeating execution of the calculating, the selecting, and the determining until execution times or the reduction ratios satisfy a predetermined condition; and outputting the reduction ratios determined when the predetermined condition is satisfied, and the predetermined condition is satisfied when a size of a memory of the computer required to store the reduced model reaches a saturated size.
In the same field of endeavor, Ji teaches the calculating includes scaling the thresholds such that an L2 norm of the thresholds of the plurality of layers becomes equal to or smaller than a threshold upper limit (a method includes a layer of a neural network (NN) having multiple layers using a threshold, and repeating pruning a plurality of layers of NN using automatically determined thresholds. L2 norm of the thresholds becomes equal to or smaller than a threshold upper limit is just a comparison of a certain threshold to a predetermined threshold. The threshold may be initialized to a range of values, such as 0 or 1. The threshold may be determined using empirical rules and may be set to a value that was automatically determined for another layer. the threshold may be the same for all layers. The NN is then pruned using the threshold. The pruning error is a function of the weights before and after pruning. [col. 1, lines 25-45; col. 3, line 15 – col. 5, line 35]);
the process further comprises
repeating execution of the calculating, the selecting, and the determining until execution times or the reduction ratios satisfy a predetermined condition (the method repeats the pruning of the layer of the NN using different threshold until a pruning error of the pruned layer reaches a pruning error allowance. Number of iterations is compared with a threshold. If the number is less than a threshold, the pruning and retraining are repeated. After the retraining, when the pruning is performed, some non-zero weights may have fallen below the pruning threshold for the associated layer [col. 1, lines 25-45; col. 7, line 35 – col. 8, line 5]); and
outputting the reduction ratios determined when the predetermined condition is satisfied (if the number of iterations has reached the threshold in a set of layers of the NN that have a type different from those being pruned, the fixed layers are not retrained during a subsequent retraining. The pruning and retraining may be repeated until a desired number of iterations are completed [col. 7, lines 35-60]), and
the predetermined condition is satisfied when a size of a memory of the computer required to store the reduced model reaches a saturated size (the NNs may be created with reduced parameter size. A level of performance for image recognition task may be maintained while reducing the load on NN hardware. The NNs may be pruned to reduce the parameter size. The GoogLeNet when pruned has the smallest size of weight parameters for a NN capable of delivering higher accuracy. Both the pruned training and inference NNs are able to achieve over 89% top-5 accuracy with the smallest size of weight parameters [col. 3, lines 15-35; col. 8, lines 40-55]).
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 pruning a layer of a NN having multiple layers using a threshold as suggested by Ji into the combination of Singh and Cho because all of these systems addressing the process of pruning neurons/parameters in the neural networks. Doing so would be desirable because the combination of Singh and Cho would be more efficient by repeating pruning a plurality of layers of a NN using automatically determined thresholds (predetermined thresholds) and retraining the NN using only weights remaining after pruning, such that the process repeats the pruning of the layer of the NN using a different threshold until a pruning error of the pruned layer reaches a pruning error allowance (Ji, [col. 1, lines 25-45]).
With respect to claim 2, the combination of Singh, Cho and Ji teaches wherein the calculating the thresholds includes calculating the thresholds based on values of loss functions of the trained model at a time of reducing elements of each of the plurality of layers and weight gradients of each of the plurality of layers (Singh, an objective function is used as the original cost function along with the L1 regularization constant. W indicates the initial weight thresholds for the K layers, and the adaptive thresholds are calculated in Equation (6). The AFP is used as an assistance to minimize the number of filters in the model. The weight thresholds is updated dynamically by the PRC module [page 3, 3.3. Weight Threshold Initialization – page 4, 3.5. Pruning Rate Controller]).
With respect to claim 3, the combination of Singh, Cho and Ji teaches wherein the determining the reduction ratios includes:
discarding a plurality of the selected reduction ratio candidates when a sum of the inference accuracy of the reduced model after machine learning and a margin is lower than the inference accuracy of the trained model (Cho, operation 704 measures inference accuracy of a neural network with respect to the pruning data set. Operation 706 is performed when the measured inference accuracy is lower than the threshold accuracy. The processor updates the weight threshold value by increasing the weight threshold value. When the measured inference accuracy is less than the threshold accuracy, the processor terminates the pruning on the current layer [par. 0119-0128]); and
determining to adopt a plurality of the selected reduction ratio candidates as the reduction ratios to be applied one to each of the plurality of layers when the sum of the inference accuracy of the reduced model after machine learning and the margin is equal to or higher than the inference accuracy of the trained model (Cho, operation 704 measures inference accuracy of a neural network with respect to the pruning data set. Operation 707 is performed when the measured inference accuracy is greater than the threshold accuracy. The processor determines whether the pruning is completed regarding all layers of the neural network [par. 0119-0128]).
With respect to claim 5, the combination of Singh, Cho and Ji teaches wherein the calculating the thresholds includes:
decreasing the threshold upper limit when the sum of the inference accuracy of the reduced model after machine learning and the margin is lower than the inference accuracy of the trained model (Cho, the processor may adjust the weight threshold value when a decrease in the inference accuracy is less than or equal to a certain level. Pruning may be performed on the current layer. The processor updates the weight threshold value until the inference accuracy of a neural network with respect to the pruning data set is decreased to the threshold accuracy [par. 0119-0128]); and
increasing the threshold upper limit when the sum of the inference accuracy of the reduced model after machine learning and the margin is equal to or higher than the inference accuracy of the trained model (Cho, the processor may update the weight threshold value by increasing the weight threshold value by δ, wherein δ may be a value that is arbitrarily set based on various factors such as the weight distribution of a neural network, a pruning rate to the current layer and the like [par. 0119-0128]).
With respect to claim 6, the combination of Singh, Cho and Ji teaches wherein the calculating the thresholds includes updating the threshold upper limit such that combinations of reduction ratio candidates of the plurality of layers differ in each execution of selecting the reduction ratio candidates (Cho, operations 702 to 707 to update the weight threshold value are repeatedly performed from the current layer to the next layer based on the result of operation 707, whether the operation 707 determines whether the pruning is completed. The retraining of a neural network is repeatedly performed to reduce a decrease in the accuracy of pruning. When a pruning of a current layer is completed, the processor repeatedly performs pruning on another layer of the neural network [par. 0119-0130, 0137]).
With respect to claim 7, the combination of Singh, Cho and Ji teaches wherein the calculating the thresholds includes setting an initial value of the threshold upper limit so as to calculate thresholds that causes, among the plurality of layers, an element of a layer in which the threshold is maximum to be reduced and that causes an element of a layer other than the layer in which the threshold is maximum not to be reduced (Singh, the initial weight threshold is calculated for each layer with an initial regularization constant in Figure 2, which creates 2 clusters of filters. Left cluster uses the binary search to find the maximum threshold for a layer such that the accuracy drop is nearly zero [page 3, 3.3. Weight Threshold Initialization]).
With respect to claim 9, it is a computer-implemented method claim that is corresponding to the non-transitory computer-readable recording medium of claim 1. Therefore, it is rejected for the same reason as claimed in claim 1 above.
With respect to claim 10, it is a computer-implemented method claim that is corresponding to the non-transitory computer-readable recording medium of claim 2. Therefore, it is rejected for the same reason as claimed in claim 2 above.
With respect to claim 11, it is a computer-implemented method claim that is corresponding to the non-transitory computer-readable recording medium of claim 3. Therefore, it is rejected for the same reason as claimed in claim 3 above.
With respect to claim 13, it is a computer-implemented method claim that is corresponding to the non-transitory computer-readable recording medium of claim 5. Therefore, it is rejected for the same reason as claimed in claim 5 above.
With respect to claim 14, it is a computer-implemented method claim that is corresponding to the non-transitory computer-readable recording medium of claim 6. Therefore, it is rejected for the same reason as claimed in claim 6 above.
With respect to claim 15, it is a computer-implemented method claim that is corresponding to the non-transitory computer-readable recording medium of claim 7. Therefore, it is rejected for the same reason as claimed in claim 7 above.
With respect to claim 17, it is an information processing apparatus claim that is corresponding to the non-transitory computer-readable recording medium of claim 1. Therefore, it is rejected for the same reason as claimed in claim 1 above.
With respect to claim 18, it is an information processing apparatus claim that is corresponding to the non-transitory computer-readable recording medium of claim 2. Therefore, it is rejected for the same reason as claimed in claim 2 above.
With respect to claim 19, it is an information processing apparatus claim that is corresponding to the non-transitory computer-readable recording medium of claim 3. Therefore, it is rejected for the same reason as claimed in claim 3 above.
Response to Arguments
The examiner respectfully acknowledges the applicant’s amendments to claims 1, 5, 9, 13 and 17.
Applicant’s amendments filed on 04/17/2026 regarding the claims rejections to claims 1-20 under 35 USC 112(b) have been considered and are partly withdrawn as other amended limitations are not persuasive.
Applicant’s arguments filed on 04/17/2026 regarding the claims rejections to claims 1-20 under 35 USC 101 have been fully considered and are consequently withdrawn.
Applicant’s arguments filed on 04/17/2026 regarding the claims rejections to claims 1-20 under 35 USC 103 have been considered and moot in view of new ground of rejections (see rejections 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