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 . A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 2/9/2026 has been entered.
Remarks
This Office Action is responsive to Applicants' Amendment filed on February 9, 2026, in which claims 1, 6, 12, and 17 are currently amended. Claims 8 and 19 are canceled. Claims 1-7, 9-18, and 20 are currently pending.
Response to Arguments
Applicant’s arguments with respect to rejection of claims 1-7, 9-18, and 20 under 35 U.S.C. 102/103 based on amendment have been considered.
With respect to Applicant's arguments on p. 9 of the Remarks submitted 2/9/2026 that "Guo explicitly distinguishes its approach from pruning, and in fact teaches away from pruning", Examiner respectfully disagrees. Guo cites other pruning methods explicitly in the "Related Work" section on p. 2. and shows how the described method is an improvement over existing pruning methods, rather than teaching away from pruning. Neural network pruning is defined as "the task of reducing the size of a network by removing parameters" ("What Is The State Of Neural Network Pruning", 2020). Guo objectively and explicitly reduces/contracts the size of the expanded neural network model by removing parameters ([p. 8 §5.1] "our ExpandNets can be contracted”).
The remaining arguments are moot in view of a new ground of rejection set forth below.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis 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.
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.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1, 3, 11, 12, and 14 are rejected under U.S.C. §103 as being unpatentable over the combination of Guo (“ExpandNets: Linear Over-parameterization to Train Compact Convolutional Networks”, 2020) and Wang (“Pruning from Scratch”, 2019).
Regarding claim 1, Guo teaches An automatic lightweight method performed by a computer device comprising at least one processor, the method comprising: receiving, by the at least one processor, a first model;([p. 2] "training a given compact, nonlinear convolutional network" See FIG. 1 compact network interpreted as first model)
generating, by the at least one processor, a second model by injecting a trainable bottleneck layer including trainable bottleneck parameters into the first model,([p. 2 §1] "to expand a compact network: (i) replacing a kxk convolution by three convolutional layers with kernel size 1x1, kxk and 1x1,respectively;" [p. 4] "kxk kernels with k > 3 can be equivalently represented with a series of l 3x3 convolutions" Inserting A 1x1 convolution followed by kxk and 1x1 convolutions is interpreted as injecting a bottleneck block/layer.)
wherein the trainable bottleneck layer is configured to control an information flow of the second model during a forward pass([p. 2] "we focus on practical, nonlinear, compact convolutional neural networks, and demonstrate the use of linear expansion as a means to introduce over-parameterization and facilitate training, so that a given compact network achieves better performance." [p. 15] "We used standard stochastic gradient descent" Guo explicitly uses the inserted layers during training to overparameterize the network and explicitly uses SGD for training which requires both forward and backward pass)
training, by the at least one processor, the bottleneck parameters of the second model using training data ([p. 2] "the use of linear expansion as a means to introduce over-parameterization and facilitate training […] training a given compact, nonlinear convolutional network" [p. 8 §5.2] "utilize the open-source implementation of Zhang et al. [62] to generate three CIFAR-10 and CIFAR-100 training sets, containing 20%, 50% and 80% of random labels, respectively, while the test set remains clean")
determining, by the at least one processor, an optimal threshold for the trained bottleneck parameters;([p. 4] "we rely on the notion of expansion rate. Specifically, for an original layer with m input channels and n output ones, given an expansion rate r, we define the number of output channels of the first 1 1 layer as p = rm and the number of output channels of the intermediate k layer as q = rn [...] For an expansion rate r, we set the number of output channels of the first 3 3 layer to p1 = rm and that of the subsequent layers to pi = rn. The same matrix-based strategy as before can be used to algebraically contract back the expanded unit" Guo explicitly varies expansion rate ([p. 15] "varying the expansion rate r E{2, 4, 8}) which controls the bottleneck layer parameters p and q by p=r*m and 1=r*n where p,q is explicitly greater than bottleneck parameters n, m ([p. 4]) such that r*m, r*n > n, m such that expanded channel r*m, r*n is interpreted as optimal threshold and the trained bottleneck weight parameters correspond to the channels of size p,q or n,m where the channel weights are trained.)
and determining, by the at least one processor, at least one filter to be removed from the second model associated with a trained bottleneck parameter lower than the optimal threshold ([p. 4] "we rely on the notion of expansion rate. Specifically, for an original layer with m input channels and n output ones, given an expansion rate r, we define the number of output channels of the first 1 1 layer as p = rm and the number of output channels of the intermediate k layer as q = rn [...] For an expansion rate r, we set the number of output channels of the first 3 3 layer to p1 = rm and that of the subsequent layers to pi = rn. The same matrix-based strategy as before can be used to algebraically contract back the expanded unit" Guo explicitly varies expansion rate ([p. 15] "varying the expansion rate r E{2, 4, 8}) which controls the bottleneck layer parameters p and q by p=r*m and 1=r*n where p,q is explicitly greater than bottleneck parameters n, m ([p. 4]) such that r*m, r*n > n, m such that expanded channel r*m, r*n is interpreted as optimal threshold and the trained bottleneck weight parameters correspond to the channels of size p,q or n,m where the channel weights are trained.)
removing, by the at least one processor, the injected trainable bottleneck layer from the second model([p. 3] "a convolution operation can be expressed in matrix form. Specifically, let Xbmwh be the input tensor to a convolutional layer, with batch size b, m input channels, height h and width w, and Fnmkk be the tensor encoding the convolutional filters" [p. 8 §5.1] "since our ExpandNets can be contracted back to the original network, at test time, they have exactly the same number of parameters, MACs and inference time as the original networks, but our networks achieve better performance." Contracting network to eliminate added bottleneck layer interpreted as pruning by removing the injected trainable bottleneck layer in light of the instant specification ([¶0015] "pruning the second model by removing a filter with a trained bottleneck parameter" [¶0098] "prune the second model by removing a filter with a trained bottleneck parameter"))
pruning, by the at least one processor, the second model in which the injected trainable bottleneck layer is removed by discarding the determined filter to reduce a total number of parameters of the second model compared to the first model.([p. 2] "we focus on practical, nonlinear, compact convolutional neural networks, and demonstrate the use of linear expansion as a means to introduce over-parameterization [...] An expanded network can then be contracted back to the compact one algebraically" [p. 3] "a convolution operation can be expressed in matrix form. Specifically, let Xb*m*w*h be the input tensor to a convolutional layer, with batch size b, m input channels, height h and width w, and Fn*m*k*k be the tensor encoding the convolutional filters" [p. 8 §5.1] "since our ExpandNets can be contracted back to the original network, at test time, they have exactly the same number of parameters, MACs and inference time as the original networks, but our networks achieve better performance." Contracting network to eliminate added bottleneck layer interpreted as pruning the injected trainable bottleneck layer. Guo explicitly states that the purpose of expansion is to over-parameterize the model and explicitly states that the contracted model has the same number of parameters as the original model, but fewer parameters than the widened/expanded model).
However, Guo does not explicitly teach to quantify an importance of each filter in the second model;
wherein the determining the optimal threshold comprises calculating the optimal threshold based on the trained bottleneck parameters to satisfy a target floating point operations per second (FLOPs) of the second model.
Wang, in the same field of endeavor, teaches to quantify an importance of each filter in the second model;([p. 2] "we propose a novel network pruning pipeline that a pruned network structure can be directly learned from the randomly initialized weights (as shown in Figure 1(c)). Specifically, we utilize a similar technique in Network Slimming [19] to learn the channel importance by associating scalar gate values with each layer. The channel importance is optimized to improve the model performance under the sparsity regularization" [p. 5] "When learning channel importance for the models on CIFAR10 dataset, we use Adam optimizer with an initial learning rate of 0.01 [...] When learning channel importance for the models on ImageNet dataset, we use Adam optimizer with an initial learning rate of 0.01 and a batch-size of 100")
wherein the determining the optimal threshold comprises calculating the optimal threshold based on the trained bottleneck parameters to satisfy a target floating point operations per second (FLOPs) of the second model; ([p. 2] "we utilize a similar technique in Network Slimming [19] to learn the channel importance by associating scalar gate values with each layer [...] After finishing the learning of channel importance, we utilize a simple binary search strategy to determine the channel number configurations of the pruned model given resource constraints (e.g., FLOPS)" [p. 2 §3] "Then with the learned channel importance values, a global threshold is set to determine which channels are preserved given a predefined resource constraint" [p. 8] "ResNet50" ResNet50 uses bottleneck blocks.).
Guo as well as Wang are directed towards neural network pruning to improve efficiency. Guo achieves the efficiency by only expanding the network during training, and Wang achieves this by pruning filters based on learned importance to satisfy a FLOPs constraint. It would have been obvious before the effective filing date of the claimed invention to apply both pruning methods together. Wang provides as additional motivation for combination ([p. 6] "When compressing the models, our method outperforms both uniform expansion models and other complicated pruning strategies across all three architectures"). This motivation for combination also applies to the remaining claims which depend on this combination.
Regarding claim 3, the combination of Guo and Wang teaches The method of claim 1, wherein the training the bottleneck parameters comprise updating the trainable bottleneck parameters based on a loss of the second model.(Guo [p. 21] "in Figure S4, we compare the original model with an over-parameterized one relying on a product L2 regularizer as in [5], and with an over-parameterized network with normal L2 regularization, corresponding to our FC expansion strategy. Even though the overall loss of Arora et al. [5]’s over-parameterized model decreases faster than that of the baseline, the cross-entropy loss term, the training error and the test error do not show the same trend. The test errors of the original model, Arora et al. [5]’s over-parameterized model with product L2 norm and our ExpandNet-FC with normal L2 norm are 0.9%, 1.1% and 0.8%, respectively").
Regarding claim 11, the combination of Guo and Wang teaches A non-transitory computer-readable recording medium storing a program to perform the method of claim 1 on a computer device.(Guo [p. 20] "Here, we compare the complexity of our expanded networks and the original networks in terms of number of parameters, number of MACs and GPU speed").
Regarding claims 12 and 14, claims 12 and 14 are directed towards a device for performing the methods of claims 1 and 3, respectively. Therefore, the rejections applied to claims 1 and 3 also apply to claims 12 and 14. Claims 1 and 3 also recite additional elements at least one processor configured to execute a computer-readable instruction, wherein the at least one processor is configured to (Guo [p. 20] "Here, we compare the complexity of our expanded networks and the original networks in terms of number of parameters, number of MACs and GPU speed").
Claims 2 and 13 are rejected under U.S.C. §103 as being unpatentable over the combination of Guo, Wang, and Tamada (US20220253708A1).
Regarding claim 2, the combination of Guo and Wang teaches The method of claim 1.
However, the combination of Guo and Wang doesn't explicitly teach, further comprising: finetuning, by the at least one processor, the pruned second model using the training data.
Tamada, in the same field of endeavor, teaches finetuning, by the at least one processor, the pruned second model using the training data. ([¶0060] ", the optimization module 128 can retrain and fine-tune (e.g., using fine tuning component 130) the compressed NN model 120 using the training dataset 102 to generate optimized compressed NN model 132").
The combination of Guo and Wang as well as Tamada are directed towards convolutional neural networks. Therefore, the combination of Guo and Wang as well as Tamada are reasonably pertinent analogous art. It would have been obvious before the effective filing date of the claimed invention to the teachings of the combination of Guo and Wang with the teachings of Tamada by performing additional pruning/finetuning to that in the combination of Guo and Wang using the pruning method described in Tamada. Tamada provides as additional motivation for combination ([¶0086] "the disclosed GS compression technique provides a systematic and simple method for filter pruning which can be used to achieve high compression ratios while preserving model performance"). This motivation for combination also applies to the remaining claims which depend on this combination.
Regarding claim 13, claim 13 is directed towards a device for performing the method of claim 2. Therefore, the rejection applied to claim 2 also applies to claim 13.
Claims 4, 9, 15, and 20 are rejected under U.S.C. §103 as being unpatentable over the combination of Guo, Wang, and Teig (US12112254B1).
Regarding claim 4, the combination of Guo and Wang teaches The method of claim 3, wherein the loss includes a cross-entropy loss, (Guo [p. 21] "in Figure S4, we compare the original model with an over-parameterized one relying on a product L2 regularizer as in [5], and with an over-parameterized network with normal L2 regularization, corresponding to our FC expansion strategy. Even though the overall loss of Arora et al. [5]’s over-parameterized model decreases faster than that of the baseline, the cross-entropy loss term, the training error and the test error do not show the same trend. The test errors of the original model, Arora et al. [5]’s over-parameterized model with product L2 norm and our ExpandNet-FC with normal L2 norm are 0.9%, 1.1% and 0.8%, respectively")
a first loss designed to satisfy constraints such that all the modules that belong to the same convolution block are pruned, (Guo [p. 3] "a convolution operation can be expressed in matrix form. Specifically, let Xbmwh be the input tensor to a convolutional layer, with batch size b, m input channels, height h and width w, and Fnmkk be the tensor encoding the convolutional filters" [p. 8 §5.1] "since our ExpandNets can be contracted back to the original network, at test time, they have exactly the same number of parameters, MACs and inference time as the original networks, but our networks achieve better performance." Contracting network to eliminate added bottleneck layer interpreted as pruning by removing the injected trainable bottleneck layer in light of the instant specification ([¶0015] "pruning the second model by removing a filter with a trained bottleneck parameter" [¶0098] "prune the second model by removing a filter with a trained bottleneck parameter"). Guo explicitly contracts all of the expansion layers after training.).
However, the combination of Guo and Wang doesn't explicitly teach and a second loss designed to force a bottleneck parameter to converge toward a binary solution indicating presence or absence of a filter.
Teig, in the same field of endeavor, teaches and a second loss designed to force a bottleneck parameter to converge toward a binary solution indicating presence or absence of a filter. ([Col. 16 l. 14-44] "The first term in the loss function, I(X; {circumflex over (X)}) measures the mutual information between the input data and the bottleneck variable. [...] During optimization, the expected behavior for these terms is that I(X; {circumflex over (X)}) will start at a large value and decrease over time, while βI({circumflex over (X)}; Y) will start as a small value and increase. This corresponds to the network learning how to compress unnecessary bits out of the input data and how to decode the correct category from the remaining bits." [Col. 22 l. 5-15] "The result of this gradient is that the VIB loss function pushes harder to increase noise on a channel with small noise variance than on a channel with a large noise variance (e.g., a channel that is on the threshold of being pruned)." See also Eqn. 5. IB loss interpreted as second loss. Compressing unnecessary bits out of the input data is interpreted as synonymous with forcing a bottleneck parameter (bit) to converge toward a binary solution indicating presence or absence of a filter.).
The combination of Guo and Wang as well as Teig are directed towards optimizing convolutional neural networks. Therefore, the combination of Guo and Wang as well as Teig are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of the combination of Guo and Wang with the teachings of Teig by substituting the pruning loss function in Tamada with the loss function in Teig, specifically a composite loss function. Teig provides as additional motivation for combination ([Col. 18 l. 23-35] "The goal is to make sure {circumflex over (X)}i includes the necessary information for decoding Yi, without dictating a specific decoding technique. One approach is to use a second loss function term (e.g., cross-entropy loss"). This motivation for combination also applies to the remaining claims which depend on this combination.
Regarding claim 9, the combination of Guo and Wang teaches The method of claim 1.
However, the combination of Guo and Wang doesn't explicitly teach wherein the injecting comprises injecting the trainable bottleneck layer including the trainable bottleneck parameters and noise into the first model.
Teig, in the same field of endeavor, teaches wherein the injecting comprises injecting the trainable bottleneck layer including the trainable bottleneck parameters and noise into the first model. ([Col. 21 l. 1-22] "VIB builds on the information bottleneck concept by introducing a variational bound of the information bottleneck loss function. In some embodiments, VIB moves layer-by-layer to identify portions of the network (e.g., nodes, edges, or even entire filters) that are not passing important information. To accomplish this, in some embodiments, VIB introduces probabilistic (e.g., Gaussian) noise into the output values of a set of computation nodes of the network (e.g., the nodes of one or more layers of the network). That is, the outputs of such nodes (which are passed to nodes in the next layer) are made to vary probabilistically around the actual computed output value during training. This noise enables the training system to identify nodes that are less important to the eventual output of the network (e.g., the classification decision, etc.) and remove these nodes").
The combination of Guo and Wang as well as Teig are directed towards optimizing convolutional neural networks. Therefore, the combination of Guo and Wang as well as Teig are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of the combination of Guo and Wang with the teachings of Teig by substituting the pruning loss function in Tamada with the loss function in Teig, specifically a composite loss function. Teig provides as additional motivation for combination ([Col. 18 l. 23-35] "The goal is to make sure {circumflex over (X)}i includes the necessary information for decoding Yi, without dictating a specific decoding technique. One approach is to use a second loss function term (e.g., cross-entropy loss"). This motivation for combination also applies to the remaining claims which depend on this combination.
Regarding claims 15 and 20, claims 15 and 20 are directed towards a device for performing the method of claims 4 and 9, respectively. Therefore, the rejections applied to claims 4 and 9 also apply to claims 15 and 20.
Claims 5 and 16 are rejected under U.S.C. §103 as being unpatentable over the combination of Guo and Wang and He (“Asymptotic Soft Filter Pruning for Deep Convolutional Neural Networks”, 2020).
Regarding claim 5, the combination of Guo and Wang teaches The method of claim 1.
However, the combination of Guo and Wang doesn't explicitly teach wherein the determining the optimal threshold comprises estimating floating point operations per second (FLOPs) of the pruned second model without actual pruning by pseudo-pruning the second model based on a threshold.
He, in the same field of endeavor, teaches the determining the optimal threshold comprises estimating floating point operations per second (FLOPs) of the pruned second model without actual pruning by pseudo-pruning the second model based on a threshold. ([p. 3600 Col. 2] "When the pruning rate is small, we find the performance of SFP [50] and ASFP is competitive. But ASFP outperforms SFP when a large portion of FLOPs are pruned" Asymptomatic soft filter pruning (ASFP) interpreted as synonymous with pseudo-pruning without actual pruning.).
The combination of Guo and Wang as well as He are directed towards optimizing convolutional neural networks including through pruning. Therefore, the combination of Guo and Wang as well as He are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of the combination of Guo and Wang with the teachings of He by soft-pruning with respect to FLOPS in addition to or in place of the pruning in Tamada. He provides as additional motivation for combination ([p. 7 Col. 2] "Our ASFP could achieve a better performance than other state-of-the-art HFP methods. For example, PFEC [23] accelerates ResNet-110 by 38.6% speedup ratio with 0.61% accuracy drop when pruning the scratch models. In contrast, our ASFP can accelerate ResNet-110 to 52.3% speedup with only 0.48% accuracy drop [...] When pruning the scratch ResNet-56, we can achieve more acceleration ratio than CP [24] (52.6% versus 50.0%) with less accuracy drop (1.15% versus 1.90%) Notably, we can even improve 0.24% accuracy when pruning 28.2% FLOPs of scratch ResNet-56").
Regarding claim 16, claim 16 is directed towards a device for performing the method of claim 5. Therefore, the rejection applied to claim 5 also applies to claim 16.
Claims 6, 7, 17, and 18 are rejected under U.S.C. §103 as being unpatentable over the combination of Guo and He2 (“AMC: AutoML for Model Compression and Acceleration on Mobile Devices”, 2018).
Regarding claim 6, the combination of Guo and Wang teaches The method of claim 1.
However, the combination of Guo and Wang doesn't explicitly teach, further comprising updating the optimal threshold to reduce a distance between current FLOPs of a pseudo-pruned second model and the target FLOPs through the dichotomy algorithm
when a difference between the current FLOPs and the target FLOPs is greater than or equal to a preset FLOPs error.
He2, in the same field of endeavor, teaches updating the optimal threshold to reduce a distance between current FLOPs of a pseudo-pruned second model and the target FLOPs through the dichotomy algorithm ([p. 5 §3.2] "The State Space For each layer t, we have 11 features that characterize the state st: (t, n, c, h, w, stride, k, F LOP s[t], reduced, rest, at−1) (1) where t is the layer index, the dimension of the kernel is n×c×k×k, and the input is c × h × w. F LOP s[t] is the FLOPs of layer Lt. Reduced is the total number of reduced FLOPs in previous layers. Rest is the number of remaining FLOPs in the following layers" REST is the difference between the target FLPS and the current FLOPS (Reduced).)
when a difference between the current FLOPs and the target FLOPs is greater than or equal to a preset FLOPs error.([p. 2] "Reward= -Error*log(FLOP)" [p. 8] "RErr corresponds to FLOPs-constrained compression with channel pruning" See also FIG. 1).
The combination of Guo and Wang as well as He2 are directed towards optimizing convolutional neural networks including through pruning. Therefore, the combination of Guo and Wang as well as He2 are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of the combination of Guo and Wang with the teachings of He2 by using the reward function in He2 to perform soft-pruning in addition or replacement of the pruning in Tamada. He2 provides as additional motivation for combination ([p. 12] "Even for the current state-of-the-art efficient model design MobileNet-V2, AMC can still improve its accuracy by 1.0% at the same computation (Table 3). The pareto curve of MobileNet-V1 is presented in Figure 5a").
Regarding claim 7, the combination of Guo, Wang, and He2 teaches The method of claim 6, wherein the updating the optimal threshold is iteratively performed while the difference between the current FLOPs and the target FLOPs is greater than or equal to the preset FLOPs error.(He2 [p. 10] "The pruning policy (sparsity ratio) given by our reinforcement learning agent for ResNet-50. With 4 stages of iterative pruning" See also FIG. 1).
Regarding claims 17 and 18, claims 17 and 18 are directed towards a device for performing the methods of claims 6 and 7, respectively. Therefore, the rejections applied to claims 6 and 7 also apply to claims 17 and 18.
Claim 10 is rejected under U.S.C. §103 as being unpatentable over the combination of Guo and Sandler (“MobileNetV2: Inverted Residuals and Linear Bottlenecks”, 2018).
Regarding claim 10, the combination of Guo and Wang teaches The method of claim 1.
However, the combination of Guo and Wang doesn't explicitly teach\ wherein the injecting comprises restricting a trainable parameter layer-wisely by injecting a bottleneck parameter into each convolution block of the first model.
Sandler, in the same field of endeavor, teaches the injecting comprises restricting a trainable parameter layer-wisely by injecting a bottleneck parameter into each convolution block of the first model. ([p. 3] "assuming the manifold of interest is low-dimensional we can capture this by inserting linear bottleneck layers into the convolutional blocks").
The combination of Guo and Wang as well as Sandler are directed towards convolutional neural network layer expansion and optimization. Therefore, the combination of Guo and Wang as well as Sandler are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of the combination of Guo and Wang with the teachings of Sandler by injecting a bottleneck layer in every convolutional block. While it would be obvious to one of ordinary skill in the art that every layer in SmallNet is eligible for expansion using ExpandNet. As both the combination of Guo and Wang and Sandler are directed towards CNN expansion, it would have been obvious to apply the expansion and contraction of the combination of Guo and Wang to each convolutional block in a network as done in Sandler. Sandler provides as additional motivation for combination ([p. 3 §3.2] “we can capture this by inserting linear bottleneck layers into the convolutional blocks. Experimental evidence suggests that using linear layers is crucial as it prevents non linearities from destroying too much information”).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
Liu (“SNF: Filter Pruning via Searching the Proper Number of Filters”, 2021) is directed towards filter pruning based on calculation of filter-wise importance values.
Dai (“NeST: A Neural Network Synthesis Tool Based on a Grow-and-Prune Paradigm”, 2018) is directed towards a system to grow neural networks and then explicitly prune them to reach a desired number of FLOPs.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SIDNEY VINCENT BOSTWICK whose telephone number is (571)272-4720. The examiner can normally be reached M-F 7:30am-5:00pm EST.
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, Miranda Huang can be reached on (571)270-7092. 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.
/SIDNEY VINCENT BOSTWICK/Examiner, Art Unit 2124
Prosecution Insights