DETAILED ACTION
This action is in response to the original filing on March 3, 2023 and the Remarks and Amendments filed on August 11, 2026. Claims 1-13 are pending and have been considered below. Claims 1 and 9 are independent claims. Claims 1-2, 6, 9, and 12 are amended.
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 Arguments
Applicant's arguments and amendments filed August 11, 2026, regarding the abstract idea rejections from the previous office action made under 35 U.S.C. 101 have been fully considered but are unpersuasive.
On page 6 of the Remarks, Applicant argues that “The claimed subject matter is not directed to a mere abstract idea, but rather implements a concrete technical process…” Applicant further asserts on pages 6-7 that the “sequence of technical operations–activation measurement, threshold-based weight removal, and transfer-learning retraining–cannot be practically performed in the human mind or with pen and paper, particularly for large-scale neural network models comprising millions of weights, and therefore integrates any alleged abstract idea into a practical application under Step 2A, Prong Two and amounts to significantly more under Step 2B.” Examiner respectfully disagrees. While paragraphs 41 and 42 in the specification describes reducing a “large model” with “a large number of parameters” into “a small model,” the features upon which applicant relies (i.e., a large-scale neural network model comprising millions of weights) is neither recited in the rejected claim nor the specification. Given its broadest reasonable interpretation, the “reconstructed first model” of claim 1 may be a model with a large number of parameters while still allowing a human to reasonably perform the “sequence of technical operations” within the human mind or with the aid of a pen and paper. Even though the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Vans Geuns, 988 F .2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993). Additionally, Applicant’s assertion that the claim integrates the abstract idea into a practical application and amounts to significantly more is conclusory, and identifies no additional elements beyond the abstract idea itself. As explained in the rejection below, the only additional elements listed in the “technical process” on page 6 of the remarks include “inputting the second dataset to the reconstructed first model” and “generating the second model by training the first model from which the weights are removed, based on the second dataset,” which amount to insignificant extra-solution activity of data gathering and outputting (MPEP 2106.05(g) “all uses of the recited judicial exception require such data gathering or data output”).
On page 7 of the Remarks, Applicant argues that “Generating the second model using the pruning mask, as recited in the claims, results in an improved machine learning model, similar to the claims at issue in Ex Parte Desjardins, and therefore integrates any alleged abstract idea into a practical application under Step 2A, Prong Two and amounts to significantly more under Step 2B.” Examiner respectfully disagrees. As stated above, Applicant identifies no additional elements in the claim that would integrate the abstract idea into a practical application or amount to significantly more. Furthermore, Ex Parte Desjardins cites paragraph 21 of the specification for support that “the claimed subject matter provides technical improvements over conventional systems.” Applicant’s assertion that the claim integrates the abstract idea into a practical application and amounts to significantly more is conclusory, as no paragraph from the specification is cited as a technical improvement. Paragraphs 2-3 and 53 in the specification come the closest to disclosing a technical improvement, as they describe how “Pruning… enables a great improvement in the amount of computation and the number of parameters compared to an existing model,” “transfer learning has proven more effective when data is insufficient,” and “by efficiently converging a model compression technique and a transfer learning technique, which have been developed as independent techniques, it is possible to gain independent advantages of the two simultaneously,” respectively. However, the amended claim 1 fails to reflect the simultaneously disclosed “improvement in the amount of computation” and effective training “when data is insufficient” as recited in the specification, while the specification itself only sets forth the improvements in a conclusory manner (MPEP 2106.04(d)(1) “if the specification explicitly sets forth an improvement but in a conclusory manner (i.e., a bare assertion of an improvement without the detail necessary to be apparent to a person of ordinary skill in the art), the examiner should not determine the claim improves technology”).
In consideration of this conclusion, the amended independent claims 1 and 9, and their associated dependent claims 2-8, and 10-13, respectively, are subject-matter ineligible and, thus, are rejected under 35 U.S.C. 101.
Applicant's arguments and amendments filed August 11, 2026, regarding the rejections from the previous office action made under 35 U.S.C. 102(a)(1) and 103 have been fully considered but are unpersuasive.
On page 8 of the Remarks under “Claim Rejections - 35 USC § 102,” Applicant argues that “Liu merely prunes a fixed number of the lowest-activated neurons based on the target data, and does not disclose deriving a pruning mask by minimizing a loss function that takes, as a variable, the loss difference before and after weight removal. Applicant further argues on pages 8-9 under “Claim Rejections - 35 USC § 103,” that neither Liu nor Xu discloses or suggests the amended feature of claim 1, and that “In Xu, a reinforcement-learning agent assigns a sparsity ratio to each layer and maximizes a reward function… through trial and error; the reward received by the agent is the absolute accuracy of the compressed model… not a loss difference between the model before and after weight removal. Xu therefore neither derives, nor minimizes, a loss function taking the loss difference before and after weight removal as a variable.” Examiner respectfully disagrees. As explained in the previous office action, while Liu fails to teach the “pruning mask,” Xu teaches a “sparsity ratio” applied to each layer, which functions in substantially the same way as a pruning mask. Similarly, while Xu fails to explicitly teach minimizing a “loss function,” the disclosed mechanism of maximizing the reward function, which can be the accuracy of a compressed model, functions in substantially the same way as minimizing inaccurate predictions when compared to correctly labeled data or calculating a loss using “cross-entropy” in paragraphs 39-40 of the specification (MPEP 2183 “The prior art element performs the identical function specified in the claim in substantially the same way, and produces substantially the same results as the corresponding element disclosed in the specification”). Furthermore, Xu explicitly teaches considering an accuracy difference between a baseline and pruned sparse model (see ¶98 below; “an accuracy within 1%, 2%, 5%, etc., of an accuracy of the baseline”) while Figures 6-8 depict a column with “DIFF FROM BASELINE TOP-1 ACCURACY.” Although Xu fails to explicitly consider a “loss difference,” its consideration of an “accuracy” difference between a model before and after weight pruning functions in substantially the same manner, as explained above. Additionally, using “weight removal as a variable” is not recited in the currently amended claim 1.
On page 9 of the Remarks, Applicant argues that neither Liu nor Xu discloses or suggests the amended feature of claim 6. Applicant asserts that “Xu’s ‘target cycle reduction’ or ‘number of cycles’ is a single, absolute evaluation metric of a pre-set constraint used as a binary threshold condition… and is not a quantitative difference between a target hardware resource and the hardware resource of the compressed model that is summed into a loss function and minimized.” Examiner respectfully disagrees. Xu, in paragraphs 91 to 98, explicitly describes continuously updating a per-layer pruning sparsity in order to maximize a reward function consisting of “log(number_of_cycles)” (see Xu’s Equation (1) in the rejection below) and to reach a desired cycle reduction on the target hardware by comparing the quantitative difference between a number of cycles for the target hardware and the achieved cycle reduction of the compressed model (see ¶96 below; Xu “aggressively prunes all the remaining layers to reach the target cycle reduction”). Furthermore, while Xu fails to teach deriving a pruning mask by summing a loss difference and a hardware-resource difference into a reward function or “loss function,” the “quantitative difference between a target hardware resource and the hardware resource of the compressed model that is summed into a loss function and minimized” and “summing the loss difference and the hardware-resource difference within a single loss function and minimizing that loss function” as describe on page 9 of the Remarks is not recited in the currently amended claim 6.
On page 9 of the Remarks, Applicant argues that “Because the feature of claim 1 is derived by minimizing a loss function that considers the loss difference before and after weight removal, whereas Liu performs simple threshold-based pruning and Xu performs macroscopic, reward-maximizing search by reinforcement learning, the two references are directed to fundamentally different technical approaches. A person of ordinary skill in the art, even combining Liu and Xu, would not arrive at the claimed loss-function-minimization-based derivation of a pruning mask.” Examiner respectfully disagrees. As explained below, both Liu and Xu are both in the same field of endeavor of machine learning, and are both directed to neural network pruning while maintaining the accuracy of the original model. Liu’s own disclosure states that “In the future, it would be interesting to apply other compression strategies” (Liu, Page 2251, Col. 1, Section 5, ¶1). For example, one of ordinary skill in the art may be motivated to apply the compression strategies of Xu “that are optimized and/or otherwise increases compared to conventional network compression techniques while maintaining state-of-the-art accuracy” (Xu, ¶63) which also takes target hardware into consideration when pruning (Xu, ¶38). Neither Liu nor Xu disparages or teaches away from the combination of the two references, while their combination discloses or suggests all elements of the amended claim elements (MPEP 2145(III) “The test for obviousness is not whether the features of a secondary reference may be bodily incorporated into the structure of the primary reference.... Rather, the test is what the combined teachings of those references would have suggested to those of ordinary skill in the art”).
In consideration of this conclusion, the claims 1-4 and 7 are now rejected under 35 U.S.C. 103 while the previous rejections under 103 still stand.
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 1-13 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1:
Step 1 – The claim is directed to a method: A method… of generating a transfer learning model, the method comprising….
Step 2A, Prong 1 – A judicial exception is recited in this claim as it recites mathematical concepts (see MPEP 2106.04(a)(2)(I)):
using a pruning mask that minimizes a loss function that considers a loss difference before and after removing at least some weights from the reconstructed first model, to remove the at least some weights from the reconstructed first model based on a second dataset that is different from the first dataset… using a pruning mask that minimizes a loss function is a mathematical calculation.
Step 2A, Prong 2 – The following limitations are additional elements that fail to integrate the judicial exception into a practical application:
performed by a computer device comprising at least one processor… a computer used as a mere tool to apply an exception is a generic element for performing or applying the abstract idea using a generic computing environment (see MPEP 2106.05(f)).
generating a transfer learning model, the method comprising: reconstructing… a first model that is pre-trained based on a first dataset… generating a model based on model that was trained on a training dataset is data gathering and outputting (see MPEP 2106.05(g)).
generating… a second model by using a pruning mask… generating a second model is data outputting (see MPEP 2106.05(g)).
and generating the second model that is trained with transfer learning by using the second dataset, from the first model from which the at least some weights are removed… generating a second model by inputting a training dataset is data gathering and outputting (see MPEP 2106.05(g)).
reconstructing/generating, by the at least one processor… a processor used as a mere tool to apply an exception is a generic element for performing or applying the abstract idea using a generic computing environment (see MPEP 2106.05(f)).
Step 2B – These elements are recited at such a high level of generality that they fail to integrate the abstract idea into a practical application, since they provide nothing more than mere instructions to implement an abstract idea on a generic computer (MPEP 2106.05(f)) or only amount to data gathering or outputting (MPEP 2106.05(g)) without significantly more. These limitations, taken either alone or in combination, fail to provide an inventive concept. Thus, the claim is not patent eligible.
Claims 2-7 recite limitations which further narrow the abstract ideas of claim 1 by specifying more details of the mathematical concept that occurs:
Regarding claim 2, this claim further limits the abstract idea of claim 1 to be based on a mental process (MPEP 2106.04(a)(2)(III)): determining weights to be removed from the reconstructed first model using a result of the inputting the second dataset and removing the determined weights from the reconstructed first model can be reasonably performed within the human mind or with the aid of a pen and paper, which is a mental process. Furthermore, inputting the second dataset to the reconstructed first model and generating the second model by training the first model from which the weights are removed, based on the second dataset, wherein the second dataset is configured for a target task of the transfer learning different from a task of the first model and includes data different from data used to pre-train the first model is still data gathering or outputting (MPEP 2106.05(g)).
Regarding claim 3, this claim further limits the abstract idea of claim 1 to be based on a mental process: determining, from among the weights of the reconstructed first model, the at least some weights to be removed from the first model, based on degrees to which the weights of the reconstructed first model are activated can be reasonably performed within the human mind or with the aid of a pen and paper, which is a mental process. Furthermore, wherein the generating of the second model comprises determining… in response to the second dataset being input into the reconstructed first model is still data gathering or outputting (MPEP 2106.05(g)).
Regarding claim 4, this claim further limits the abstract idea of claim 3 to be based on a mental process: determining… weights of which the degrees to which the weights are activated are less than or equal to a threshold value can be reasonably performed within the human mind or with the aid of a pen and paper, which is a mental process. Furthermore, wherein the generating of the second model further comprises… is still data outputting (MPEP 2106.05(g)). Furthermore, determining, as the at least some weights to be removed from the first model… would be mere instructions to apply an abstract idea using a generic computing environment (see MPEP 2106.05(f)).
Regarding claim 5, this claim further limits the abstract idea of claim 4 to be based on a mental process: wherein the threshold value is determined based on at least one of performance of target hardware, an amount of computation allowed for the second model, and a number of parameters allowed for the second model… Given a simple enough performance of target hardware, amount of computation allowed, and a number of parameters, a human can reasonably perform determining a threshold value within the human mind or with the aid of a pen and paper model, which is a mental process.
Regarding claim 6, this claim further limits the abstract idea of claim 1 to be based on a math concept: using a pruning mask that minimizes a loss function that considers a hardware difference between a target hardware resource and a hardware resource of the first model from which the at least some weights are removed calculates a hardware difference between two resources to minimize a loss function, which is a math concept. Furthermore, the generating of the second model comprises generating the second model by using a pruning mask is still data outputting (MPEP 2106.05(g)).
Regarding claim 7, specifying wherein the reconstructing of the first model comprises reconstructing the first model according to a target task to be transferred in this manner does not overcome the rejection of claim 1 as modifying the reconstructing does not make the abstract idea of claim 1 to not be a math concept.
Regarding claim 8, a computer-readable recording medium having recorded thereon a program for executing, on a computer device, the method of claim 1 would be mere instructions to apply the abstract ideas of 1-7 using a generic computing environment (see MPEP 2106.05(f)).
Claims 9-13 recite a device that parallels the method claims of 1-3 and 6-7, respectively. Therefore, the analysis discussed above with respect to claims 1-3 and 6-7 applies to claims 9-13, respectively. Accordingly, claims 9-13 are rejected based on substantially the same rationale as set forth above with respect to claims 1-3 and 6-7, respectively.
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 1-13 are rejected under 35 U.S.C. 103 as being unpatentable over Liu et al. (“Sparse Deep Transfer Learning for Convolutional Neural Network,” 2017, hereinafter Liu) in view of Xu et al. (US 20200401891 A1, hereinafter Xu).
Regarding claim 1:
Liu teaches a method, performed by a computer device comprising at least one processor, of generating a transfer learning model (Page 2245, Col. 2, ¶1 “we perform our sparse deep transfer learning approach on a number of challenging transfer learning tasks for the famous AlexNet and VGGNet,” wherein to perform transfer learning on known neural networks in the art implicitly requires a computer device comprising at least one processor, for example a CPU or GPU), the method comprising: reconstructing, by the at least one processor, a first model that is pre-trained based on a first dataset (Page 2248, Col. 2, ¶1 “we choose AlexNet (trained by ImageNet)… we perform the iterative pruning strategy… to reduce model redundancy in the source domain… we obtain our Sparse-SourceNet,” Page 2245, Col. 2, ¶1 “Second, we propose a Hybrid-TransferNet in which an extra branch is incorporated into the modified Sparse-SourceNet,” Page 2246, Fig. 1 depicts a “Hybrid-TransferNet,” or a first model, that is reconstructed from “pruning” and adding additional branches to a pre-trained “Reference SourceNet”).
Regarding the limitation generating, by the at least one processor, a second model by using a pruning mask that minimizes a loss function that considers a loss difference before and after removing at least some weights from the reconstructed first model, to remove the at least some weights from the reconstructed first model based on a second dataset that is different from the first dataset, Liu teaches generating, by the at least one processor, a second model by removing at least some weights from the reconstructed first model (Page 2245, Col. 2, ¶1 “we introduce a Sparse-SourceNet by pruning CNN in the source domain… we propose a Hybrid-TransferNet in which an extra branch is incorporated into the modified Sparse-SourceNet”) based on a second dataset that is different from the first dataset (Page 2246, Fig. 1 depicts a model with a “Sparse-TargetNet” branch, or a second model, Caption: “deep transfer learning approach for AlexNet (8-layer CNN), from ImageNet (source, 1000 object classes) to MIT Indoor67 (target, 67 scene classes),” Page 2248, Col. 1, Section 3.3, ¶1 “the top Nm layers in the main branch of our Hybrid-TransferNet are recovered to be dense… After the domain is transferred to target, these layers may be redundant. Additionally, compared to the large source data set, the target data set is limited. In this case, many connections and even neurons may not be useful after transferring from source to target,” ¶2 “Since our main branch of Hybrid-TransferNet is used for prediction in the target domain, we propose to prune the main branch of our Hybrid-TransferNet,” Page 2249, Col. 1, ¶2 “we propose to prune the main branch of our Hybrid-TransferNet in the target domain and denote the resulting net as our Sparse-TargetNet,” Page 2250, Col. 1, Table 2, Caption: “when we prune neurons, the weights are automatically removed”). However, Liu fails to teach by using a pruning mask that minimizes a loss function that considers a loss difference before and after removing at least some weights from the reconstructed first model, to remove the at least some weights from the reconstructed first model…
Xu teaches generating a pruned model by using a pruning mask that minimizes a loss function that considers a loss difference before and after removing at least some weights from a baseline model, to remove the at least some weights from the baseline model (Fig. 1 – 104A-E, 124, ¶61 “the model training controller 104A-E assigns sparsity ratios to respective layers of the machine learning model(s) 124… the third model training controller 104C can generate a first action including assigning a first sparsity ratio of 70% to a first level, a second action including assigning a second sparsity ratio of 65% to a second level,” ¶62 “responsive to the sparsity ratio assignments… the third model training controller 104C can execute the first one of the machine learning model(s) 124 to generate an output (e.g., a model output), such as a reward. In such examples, the reward can be an accuracy of the first one of the machine learning model(s) 124. In some such examples, the third model training controller 104C can generate a new set of one or more actions to adjust the sparsity ratios for respective layers of the first one of the machine learning model(s) 124 based on the reward,” Fig. 3 – 304- 308, 312-314, ¶96 “the model training controller 104A-E learns to prune the weights in each layer to find the best combination of cycle reduction for each layer. If… the model training controller 104A-E determines that the action (e.g., the action 306) taken cannot reach the target percentage, then the model training controller 104A-E aggressively prunes all the remaining layers to reach the target cycle reduction… the model training controller 104A-E can reach the target percentage reduction while achieving the best reward by using the reward function:
PNG
media_image1.png
59
564
media_image1.png
Greyscale
¶97 “in response to assigning a sparsity ratio to every layer of the baseline machine learning model 304… the environment 308 can perform an evaluation by invoking the baseline machine learning model 304 having the assigned sparsity ratios to process a subset of training data… the environment 308 can generate an example reward 312 responsive to the processing of the subset of the training data. The reward 312 is an evaluation result that is fed back to the model training controller 104A-E to update the next prediction (e.g., the next action for each layer),” ¶98 “the best reward sparse model 314 is a neural network trained by the model training controller 104A-E to have high performance on the target hardware platform with substantially similar accuracy as the baseline machine learning model 304 (e.g., an accuracy within 1%, 2%, 5%, etc., of an accuracy of the baseline machine learning model 304),” wherein a “sparsity ratio” that “prunes the weights in each layer” functions in substantially the same way as a pruning mask; to generate “sparse model” by pruning a “baseline machine learning model” so that its “accuracy” is similar to that of a “baseline” functions in substantially the same way as a pruning mask that minimizes a loss function that considers a loss difference before and after removing the at least some weights).
Liu further teaches and generating the second model that is trained with transfer learning by using the second dataset, from the first model from which the at least some weights are removed (Page 2247, Col. 1, Section 3.2, ¶1 “to extract the implicit knowledge from the source domain and incorporate it into transfer learning, we design the following Hybrid-TransferNet with the main branch, the implicit-knowledge-extraction branch and the extra branch,” Page 2246, Fig. 1 depicts the “Hybrid Transfer-Net,” or the reconstructed first model with a “Main Branch” and an “Implicit-Knowledge Extraction Branch” with a final “Softened Softmax” layer, Page 2247, Col. 2, ¶2-3 “Hybrid-TransferNet Training is performed by using the following total loss,
PNG
media_image2.png
38
464
media_image2.png
Greyscale
Where λ is a trade-off weight. The main loss Lmain is the cross entropy between the softmax output of our main branch pmain and the target label ytarget
PNG
media_image3.png
32
537
media_image3.png
Greyscale
Note that our main branch is sparse… the sparsity of our main branch can alleviate overfitting for transfer learning via Lmain,” Page 2248, Col. 2, ¶2 “For the implicit-knowledge-extraction branch, we feed target images (MIT Indoor67 or Flower102) into Reference SourceNet, and output the implicit knowledge by the softened softmax,” Page 2249, Col. 1, ¶2-3 “we… prune the main branch of our Hybrid-TransferNet in the target domain and denote the resulting net as our Sparse-TargetNet… we fine-tune the pruned model with our total loss (Eq. (3)) in the retraining phase, to incorporate implicit knowledge for prediction… the implicit knowledge provides our Sparse-TargetNet with important source-domain supervision to improve accuracy in the target domain,” Col. 2, ¶2 “we randomly choose 50 / 10 images per class from the training set of MIT Indoor67 / Flower102, and feed these images into our Sparse-TargetNet to compute the mean of activations (values after ReLU) for each neuron of FC6 and FC7. In each pruning iteration, we prune 512 lowest activated neurons and perform retraining with our total loss in Eq. (3),” wherein “fine-tuning” and “retraining” the first model from which at least some weights are removed using “implicit knowledge” and the loss from Equation (3) which are computed based on “target images” and “target labels,” respectively, encompass generating the second model that is trained with transfer learning by using the second dataset).
Liu and Xu are analogous to the claimed invention as both are from the same field of endeavor of 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 combine the pruning mask that minimizes a loss function of Xu with the methodology of Liu. The motivation to do so is “to improve a speed at which to train a neural network and an efficiency of utilizing hardware resources to train the neural network” (Xu, ¶40).
Regarding claim 2, Liu in view of Xu teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Liu further teaches wherein the generating of the second model comprises: inputting the second dataset to the reconstructed first model (Page 2246, Fig. 1 as explained above with respect to claim 1, Page 2247, Col. 1, 3.2 ¶1 “to extract the implicit knowledge from the source domain and incorporate it into transfer learning, we design the following Hybrid-TransferNet,” Page 2248, Col. 2, ¶2 “For the implicit-knowledge-extraction branch, we feed target images (MIT Indoor67 or Flower102) into Reference SourceNet, and output the implicit knowledge by the softened softmax,” Col. 2, Equation (3) and ¶3 “The main loss Lmain is the cross entropy between the softmax output of our main branch pmain and the target label ytarget,” wherein to perform “Hybrid-TransferNet Training” or training the reconstructed first model using a loss calculated from “the target label ytarget” implies inputting, the second dataset).
Liu further teaches determining weights to be removed from the reconstructed first model using a result of the inputting the second dataset (Page 2246, Fig. 1, Page 2248, Col. 1, Section 3.3, ¶2 “Since our main branch of Hybrid-TransferNet is used for prediction in the target domain, we propose to prune the main branch of our Hybrid-TransferNet,” ¶3 “we propose to prune neurons of the inner-product layers in our main branch, due to the fact that many high level features (neurons) of the transferred model may not be quite useful in the target domain, especially when the size of the target data is limited… for an inner-product layer to be pruned, we compute the mean of the activation vectors over a data subset (randomly chosen from the training set), and proportionally delete the low-activated neurons. Then, we perform re-training with our total loss in Eq. (3). The resulting pruned main branch is denoted as Sparse TargetNet,” Page 2249, Col. 2, ¶2 “we prune neurons… we randomly choose 50 / 10 images per class from the training set of MIT Indoor67 / Flower102,” Page 2250, Col. 1, Table 2, Caption: “when we prune neurons, the weights are automatically removed”).
Liu further teaches removing the determined weights from the reconstructed first model (Page 2249, Col. 1, ¶2 “we propose to prune the main branch of our Hybrid-TransferNet in the target domain and denote the resulting net as our Sparse-TargetNet,” ¶3 “Firstly, we prune connections,” Col. 2, ¶2 “Secondly, we prune neurons”).
Liu further teaches and generating the second model by training the first model from which the weights are removed, based on the second dataset (Page 2247, Col. 1, Section 3.2, ¶1 “Suppose that source is object (ImageNet) and target is scene (MIT Indoor67),” Col. 2, Equation (3) and ¶3 “The main loss Lmain is the cross entropy between the softmax output of our main branch pmain and the target label ytarget,” Page 2248, Col. 2, ¶2 “For the implicit-knowledge-extraction branch, we feed target images (MIT Indoor67 or Flower102) into Reference SourceNet, and output the implicit knowledge by the softened softmax,” Page 2249, Col. 1, ¶3 “we prune connections… we fine-tune the pruned model… in the retraining phase, to incorporate implicit knowledge for prediction,” Col. 2, ¶2 “we prune 512 lowest activated neurons and perform retraining with our total loss in Eq. (3)”).
Liu further teaches wherein the second dataset is configured for a target task of the transfer learning different from a task of the first model and includes data different from data used to pre-train the first model (Page 2248, Col. 1, Section 4.1, ¶1 “For AlexNet, we evaluate our approach on two popular transfer learning tasks… where the source domain of both tasks is object recognition with ImageNet ILSVRC-2012… The target domains are respectively scene recognition with MIT Indoor67… and fine-grained flower recognition with Flower102”).
Regarding claim 3, Liu in view of Xu teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Liu further teaches wherein the generating of the second model comprises determining, from among the weights of the reconstructed first model, the at least some weights to be removed from the first model, based on degrees to which the weights of the reconstructed first model are activated in response to the second dataset being input into the reconstructed first model (Page 2248, Col. 1, Section 3.3, ¶3 “we propose to prune neurons of the inner-product layers in our main branch, due to the fact that many high level features (neurons) of the transferred model may not be quite useful in the target domain, especially when the size of the target data is limited… for an inner-product layer to be pruned, we compute the mean of the activation vectors over a data subset (randomly chosen from the training set), and proportionally delete the low-activated neurons… The resulting pruned main branch is denoted as Sparse TargetNet,” Page 2249, Col. 2, ¶2 “we randomly choose 50 / 10 images per class from the training set of MIT Indoor67 / Flower102, and feed these images into our Sparse-TargetNet to compute the mean of activations (values after ReLU) for each neuron of FC6 and FC7. In each pruning iteration, we prune 512 lowest activated neurons,” Page 2250, Col. 1, Table 2, Caption: “when we prune neurons, the weights are automatically removed”).
Regarding claim 4, Liu in view of Xu teaches the method of claim 3 (and thus the rejection of claim 3 is incorporated).
Liu further teaches wherein the generating of the second model further comprises determining, as the at least some weights to be removed from the first model, weights of which the degrees to which the weights are activated are less than or equal to a threshold value (Page 2248, Col. 2, ¶3 “The first base line is Hybrid-TransferNet without both model sparsity and implicit knowledge, where there is only the main branch in Hybrid-TransferNet and this branch is initialized from the dense Reference-SourceNet… we denote it as Standard-TransferNet,” Page 2249, Col. 2, ¶2 “we… compute the mean of activations (values after ReLU) for each neuron of FC6 and FC7. In each pruning iteration, we prune 512 lowest activated neurons… This iterative procedure is stopped when the accuracy is the level of Standard-TransferNet,” wherein to “prune 512 lowest activated neurons” until an “accuracy” is reached implies that the at least some weights to be removed from the first model are based on weights of which the degrees to which the weights are activated are less than or equal to a threshold value).
Regarding claim 5, Liu in view of Xu teaches the method of claim 4 (and thus the rejection of claim 4 is incorporated).
Liu fails to teach wherein the threshold value is determined based on at least one of performance of target hardware, an amount of computation allowed for the second model, and a number of parameters allowed for the second model. However, Xu, in the same field of endeavor, teaches this limitation (¶33 “training is performed until the level of error… satisfies a threshold… training is performed until a number or quantity of cycles (e.g., clock cycles, instruction cycles, processor cycles, etc.) to execute a trained machine learning model… satisfies a threshold (e.g., a cycle threshold, a clock cycle threshold, an instruction cycle threshold, a processor cycle threshold, a training threshold, etc.),” ¶39 “an example model training controller applies hardware-aware sparsity to a neural network based on an architecture (e.g., a hardware, software, and/or firmware architecture) of a target hardware platform or portion(s)… the model training controller identifies the sparsity ratios based on the architecture of the target hardware platform,” Fig. 2 – 104A-E, 260, 274, ¶81 “the model training controller 104A-E includes the deployment controller 260 to deploy a trained version of the machine learning model 274 to a target hardware resource and/or, more generally, a target hardware platform… in response to at least one of (a) an accuracy of the machine learning model 274 satisfies an accuracy threshold or (b) a number of reduction in cycles satisfies a cycle reduction threshold”).
Liu and Xu are analogous to the claimed invention as both are from the same field of endeavor of 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 combine the threshold based on performance of target hardware of Xu with the methodology of Liu. The motivation to do so is “to improve a speed at which to train a neural network and an efficiency of utilizing hardware resources to train the neural network” (Xu, ¶40).
Regarding claim 6, Liu in view of Xu teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Regarding the limitation wherein the generating of the second model comprises generating the second model by using a pruning mask that minimizes a loss function that considers a hardware difference between a target hardware resource and a hardware resource of the first model from which the at least some weights are removed, Liu teaches wherein the generating of the second model comprises generating the second model (Page 2245, Col. 2, ¶1 “we develop a Sparse-TargetNet, by further pruning our Hybrid-TransferNet in the target domain”) and the first model from which the at least some weights are removed (Page 2245, Col. 2, ¶1, Page 2246, Fig. 1). However, Liu fails to teach wherein the generating of the second model comprises generating the second model by using a pruning mask that minimizes a loss function that considers a hardware difference between a target hardware resource and a hardware resource of the first model from which the at least some weights are removed.
Xu teaches generating a pruned model by using a pruning mask that minimizes a loss function that considers a hardware difference between a target hardware resource and a hardware resource of a baseline (Fig. 1 – 104A-E, 124, ¶¶61-62 Fig. 3 – 304- 308, 312-314, ¶¶96-98 and Equation (1) as explained above with respect to claim 1, ¶96 “the action space for the baseline machine learning model 304 is limited by setting a target percentage reduction for the number of cycles on the target hardware platform… the model training controller 104A-E aggressively prunes all the remaining layers to reach the target cycle reduction,” Fig. 5 – 502, 506, 514, ¶114 “the model training controller 104A-E invokes and/or otherwise causes the environment 514 to output the best reward sparse model 506 based on the trained version of the baseline machine learning model 504 satisfying… a percentage of the reduction in a number of cycles to be executed by the target hardware platform satisfying a cycle threshold,” Fig. 1 – 108, Fig. 2 – 276, Fig. 12 – 1208, ¶151 “the example model training controller 104A-E determines whether a quantity of clock cycles to execute the machine learning model satisfies a threshold… whether a quantity of clock cycles of the first acceleration resource 108 to process the training data 276… using the assigned sparsity ratios satisfies a threshold quantity of clock cycles by being less than the threshold quantity of clock cycles”).
Liu and Xu are analogous to the claimed invention as both are from the same field of endeavor of 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 have combined the pruning mask that minimizes a loss function and the difference between a target hardware resource and a hardware resource of Xu with the methodology of Liu. The motivation to do so is “to improve a speed at which to train a neural network and an efficiency of utilizing hardware resources to train the neural network” (Xu, ¶40).
Regarding claim 7, Liu in view of Xu teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Liu further teaches wherein the reconstructing of the first model comprises reconstructing the first model according to a target task to be transferred (Page 2247, Col. 1, Section 3.2, ¶1-2 “to extract the implicit knowledge from the source domain and incorporate it into transfer learning, we design the following Hybrid-TransferNet with the main branch, the implicit-knowledge-extraction branch and the extra branch… Main Branch is used to make prediction for target. It is our Sparse-SourceNet but with the target-domain-related modifications… we change the output layer of Sparse-SourceNet to the target classes, in order to perform transfer learning from source to target. Furthermore, we propose to recover the top Nm layers of Sparse-SourceNet to be dense and re-initialize these layers randomly to increase transferability”).
Regarding claim 8, Liu in view of Xu teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Xu teaches a computer-readable recording medium having recorded thereon a program for executing, on a computer device (Fig. 12-13, ¶145 “the example processes of FIGS. 12-13 may be implemented using executable instructions (e.g., computer and/or machine readable instructions) stored on a non-transitory computer and/or machine readable medium”).
Liu and Xu are analogous to the claimed invention as both are from the same field of endeavor of 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 combine the computer-readable recording medium of Xu with the methodology of Liu. The motivation to do so is “to improve a speed at which to train a neural network and an efficiency of utilizing hardware resources to train the neural network” (Xu, ¶40).
Regarding claim 9, Liu in view of Xu teaches a computer device comprising at least one processor configured to execute instructions readable by the computer device, wherein the at least one processor is further configured to… (Xu, Fig. 12-13, ¶145 “the example processes of FIGS. 12-13 may be implemented using executable instructions (e.g., computer and/or machine readable instructions) stored on a non-transitory computer and/or machine readable medium,” Fig. 14, ¶166 “an example processor platform 1400 structured to execute the instructions of FIGS. 12-13”).
Claims 9-13 recite a device that parallels the method claims of 1-3 and 6-7, respectively. Therefore, the analysis discussed above with respect to claims 1-3 and 6-7 applies to claims 9-13, respectively. Accordingly, claims 9-13 are rejected based on substantially the same rationale as set forth above with respect to claims 1-3 and 6-7, respectively.
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 WILLIAM M LEE whose telephone number is (571)272-4761. The examiner can normally be reached Mon-Fri. 8am-5pm.
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, Cesar Paula can be reached at (571)272-4128. 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.
/WILLIAM M LEE/
Examiner, Art Unit 2145
/CESAR B PAULA/Supervisory Patent Examiner, Art Unit 2145
