DETAILED ACTION
This action is responsive to the application filed on 03/03/2026. Claims 1,3-10, and 18 are pending and have been examined. This action is Non-final.
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 .
Priority
Applicant’s claim for the benefit of a prior-filed application under 35 U.S.C. 119(e) or under 35 U.S.C.
120, 121, 365(c), or 386(c) is acknowledged.
Continued Examination Under 37 CFR 1.114
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and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114.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 03/03/2026 has been entered.
Response to Arguments
Argument 1: Applicant mainly argues that the amended claims, especially independent claims 1 and 18, now require a more specific layer-based pruning and scaling arrangement than the art teaches. In particular, they argue that claim 1 now makes clear that the first layer includes the up-scaled weights, the second layer includes the down-scaled weights, and the third layer includes the remaining data after the first data is pruned, and the new amendment further clarifies that all weights of the second and third layers are not up-scaled and all weights of the first and third layers are not down-scaled. They contend that Luo may generally discuss successive CNN layers and scaling channel weights to minimize reconstruction error, but Luo’s Equations 5 and 6 only describe a minimization condition and do not teach the claimed layer-specific placement of up-scaled and down-scaled weights. They also argue that the examiner’s prior reliance on Luo’s removed channel/filter is improper because Luo removes a less important channel to reduce reconstruction error, while the claim ties the pruned first data to a first plurality of weights that includes a “major value” weight having greater influence on the convolution result. Thus, applicant says Luo is different from, or even opposite to, the claimed combination, and modifying Luo to meet the claim would change Luo’s principle of operation or make it unsatisfactory for its intended purpose. They further assert that Mishra, Hu, and Dally do not cure these alleged deficiencies, so claim 1, dependent claims 3-10, and similarly amended independent claim 18 should be allowed.
Examiner Response to Argument 1: The examiner has considered applicant's arguments regarding the 103 rejection of claims 1, 18, and their dependents, and finds them unpersuasive. Applicant first argues that the amended claims require a specific layer-based arrangement in which the first layer includes the up-scaled weights, the second layer includes the down-scaled weights, and the third layer includes the remaining data after the first data is pruned, and that Luo's Equations 5 and 6 describe only a reconstruction-error minimization condition rather than the claimed layer-specific placement, however this does not distinguish over the rejection of record, because, as set forth in the Final Rejection and maintained above, Luo operates on a convolutional neural network with successive layers i, i+1, and i+2 in which each layer performs its own convolution that multiplies weights by input data and sums the results and the forward pass applies these layers one after another so that the convolution in each layer is performed at a different time than the convolutions in other layers, and the claims as amended do not require that a given layer contains only up-scaled weights, or only down-scaled weights, or that a layer performs only pruning, such that Luo's channel-wise scaling factors that increase some channel weights and decrease others, together with Luo's pruning steps that remove channels and their filters, are consistent with a first layer that includes up-scaled weights and data to be convolved, a second layer that includes intermediate data and down-scaled weights, and a third layer that carries forward the remaining data after pruning, with the same Luo teachings that supported the three-layer mapping for former claim 5 continuing to support the three-layer limitations now recited in independent claims 1 and 18. Applicant further argues that the examiner's reliance on Luo's removed channel is improper because Luo removes a less important channel to reduce reconstruction error whereas the claim ties the pruned first data to a first plurality of weights that includes a major-value weight having greater influence on the convolution result, such that Luo is said to be different from, or even opposite to, the claimed combination; this argument conflates two separately mapped limitations, because the major-value weight limitation is not mapped to Luo's removed channel but rather to Hu, which expressly teaches per-channel weighting in which the activations act as channel weights adapted to the input-specific descriptor so that some channels are weighted more heavily than others when forming the convolution output, while Luo supplies the selection, down-scaling, and removal of channel data and the associated weights, so there is no inconsistency between Luo's removal of a channel and the claimed major-value weight because the two are directed to different limitations and applicant's "opposite to" argument is directed to a mapping the examiner has not made. With respect to the newly added limitation reciting that all weights of the second and third layers are not up-scaled and all weights of the first and third layers are not down-scaled, the examiner agrees that Luo, Mishra, and Hu do not by themselves teach this direction-segregated and layer-exclusive arrangement, and this limitation is therefore newly mapped to Shen, which expressly teaches that only the input, the first layer, and the output of the first layer have the second, lower precision while the second layer, the third layer, and their outputs have the first, higher precision, so that the precision-reducing operation is confined to the first layer and the corresponding precision-restoring operation is confined to the single transition immediately following the first layer, which directly teaches that the second and third layers are not up-scaled and the first and third layers are not down-scaled, with Shen relied upon solely for this newly added limitation while every other limitation of amended claims 1 and 18 continues to be taught by Luo, Mishra, and Hu exactly as set forth in the Final Rejection. Applicant additionally argues that modifying Luo would change its principle of operation or render it unsatisfactory for its intended purpose, but this is not persuasive because Luo's principle of operation is the structured pruning of channels and the scaling of channel weights to balance compression against reconstruction accuracy, and the combination does not alter that principle, since Mishra supplies only the type of input data and does not change how Luo prunes or scales, Hu supplies per-channel weighting that operates in the same channel-scaling framework Luo already employs, and Shen supplies a layer-segregated precision scheme that, like Luo, scales values associated with network layers to balance computational cost against accuracy, such that incorporating Shen's confinement of precision reduction to a first layer leaves Luo's pruning and reconstruction-error minimization intact and does not render Luo unsatisfactory for its intended purpose. Finally, because the only limitation that Luo does not teach for independent claims 1 and 18 as amended is the newly added negative limitation now taught by Shen, while Mishra and Hu continue to supply the multimedia-data and major-value-weight limitations respectively, and because applicant has not separately argued the dependent claims apart from their dependence on claims 1 and 18, the rejections of claims 1, 18, and their dependent claims under 35 U.S.C. 103 are maintained.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1,3-8, 10, and 18 are rejected under 35 U.S.C. 103 as being unpatentable over NPL reference “ThiNet: Pruning CNN Filters for a Thinner Net” by Luo (referred to herein as Luo) in view of US 10573313 B2 by Mishra (referred to herein as Mishra) in view of NPL reference “Squeeze-and-Excitation Networks” by Hu (referred to herein as Hu) and further in view of US 20220129736 A1 by Shen (referred to herein as Shen).
Regarding claim 1, Luo teaches:
An operation method of a computing device, the method comprising: ([Luo, page 2525, 1 Introduction] “With our pruned network, some important transfer learning tasks such as object detection or fine-grained recognition can run much faster (in both training and inference), especially in small devices.”, wherein the examiner interprets “small devices” to be the same as a computing device and “run much faster” in the context of executing neural-network-based transfer learning tasks to be the same as an operation method of the computing device because they are both directed to a computing device performing the claimed method steps for pruning and inference.)
selecting, by the computing device implementing a neural network and among a plurality of data, first data on which a first pruning is to be performed; ([Luo, page 2528, sec. 3.2.3-3.2.4] “After obtaining the subset S, we can safely remove the n th channel in each filter of layer i + 1 if n ∈ S. The corresponding filters in the previous layer i can be pruned, too. We further minimize the reconstruction error (cf. Eq. (5)) by scaling the channel weight ... only those channels which have minimal reconstruction loss after rescaling are preserved.”, wherein the examiner interprets “remove the n th channel in each filter of layer i + 1 if n ∉ S” to be the same as selecting, among a plurality of channel data, first data on which a first pruning is to be performed because in Luo each output channel corresponds to feature map data produced by the neural network, so evaluating channels based on reconstruction loss and designating particular channels for removal is equivalent to evaluating a plurality of data and selecting specific data items as the targets of pruning and removal.)
down-scaling, by the computing device, a first plurality of weights included in a first output channel associated with the first data ([Luo, page 2528, sec. 3.2.4] “We further minimize the reconstruction error (cf. Eq. (5)) by scaling the channel weights, which is formulated as: [Eq. 6] only those channels which have minimal reconstruction loss after rescaling are preserved.”, wherein the examiner interprets “scaling the channel weights” and “after rescaling” to be the same as down-scaling a first plurality of weights in a channel associated with data because they are both directed to altering the magnitudes of channel weights, including reducing some weights to lessen their contribution before deciding which channels to prune.)
up-scaling, by the computing device, a second plurality of weights used to generate second data to be multiplied by the one weight of the first plurality of weights ([Luo, page 2528, sec. 3.2.4] “If we can assign a scaling factor w_j to each filter channel, Eq. (5) is rewritten as: ... Of course, these scaling factors are not globally optimal. Hence, we need to perform another least squares after selection as we have done in Eq. (7). Algorithm 1 summarizes our final pruning strategy.”, wherein the examiner interprets “assign a scaling factor w_j to each filter channel” to be the same as up-scaling a second plurality of weights used to generate second data because they are both directed to multiplying channel weights by scaling factors, including factors greater than one that increase some weights to compensate for others being reduced or removed.)
calculating, by the computing device, the second data based on the up-scaled second plurality of weights; ([Luo, page 2528, sec. 3.2.4] “Of course, these scaling factors are not globally optimal. Hence, we need to perform another least squares after selection as we have done in Eq. (7). Algorithm 1 summarizes our final pruning strategy.”, wherein the examiner interprets “perform another least squares after selection” to be the same as calculating the second data based on the up-scaled second plurality of weights because they are both directed to recomputing network outputs or activations using the newly scaled weights after the selection step.)
performing, by the computing device, the first pruning on the first data in the neural network, to remove the first data from the plurality of data and all the down-scaled first plurality of weights included in the first output channel associated with the first data; ([Luo, page 2528, sec. 3.2.3-3.2.4] “After obtaining the subset S, we can safely remove the n th channel in each filter of layer i + 1 if n ∈ S. The corresponding filters in the previous layer i can be pruned, too. We further minimize the reconstruction error (c.f. Eq. (5)) by scaling the channel weights, which is formulated as: [Eq. 6], only those channels which have minimal reconstruction loss after rescaling are preserved.”, and, “Note that, if a filter in W is removed, its corresponding channel in I and W_l are also discarded.” wherein the examiner interprets “remove the n th channel in each filter of layer i + 1” and “its corresponding channel in I and W_l are also discarded” to be the same as removing the first data from the plurality of data and all the down-scaled first plurality of weights included in the first output channel associated with the first data because they are both directed to deleting an entire channel’s activation data and all filter weights associated with that channel.)
wherein the computing device includes a first layer, a second layer, and a third layer in which convolution operations are performed at a time different from each other, ([Luo, page 2527, sec. 3.2.1-3.2.2] “We use a triplet to denote the convolution process in layer i, where I is the input tensor, which has C channels, H rows and W columns ... The convolution operation is computed with a corresponding bias b as follows: ...”, wherein the examiner interprets “the convolution process in layer i” together with Luo’s teaching that the convolution operation is defined and applied for each layer i of the network to be the same as convolution operations in a first layer, a second layer, and a third layer being performed at a time different from each other because they are both directed to a multi-layer (layer i, performing a convolution process is interpreted that multiple layers, including layers, one, two and three) convolutional neural network in which each layer performs its convolution on the output of a preceding layer as a separate step in the forward pass, so the convolution in each successive layer necessarily occurs at a different stage in time.)
wherein the first layer includes the up-scaled second plurality of weights and a plurality of third data to be convolved with the up-scaled second plurality of weights based on multiplication and summation of the up-scaled second plurality of weights and the plurality of third data ([Luo, page 2527, sec. 3.2.2 and Fig. 3 caption] “The convolution operation is computed with a corresponding bias b as follows: ... We first randomly sample an element y from the activation tensor of layer i + 1 with random spatial location and random channel index. According to the spatial location and channel index of y, the corresponding filter W and sliding window x can also be determined.” AND ([Luo, page 2528, sec 3.2.3-3.2.4]] “After obtaining the subset S, we can safely remove the n th channel in each filter of layer i + / if n & S. The corresponding filters in the previous layer i can be pruned, too. We further minimize the reconstruction error (c.f. Eq. (5)) by scaling the channel weights, “, wherein the examiner interprets the convolution between “filter W” and input “x” as a sum of products and “We further minimize the reconstruction error by scaling the channel weights [weights are scaled up and down according to Equation (6) above]” to be the same as a first layer(s) including the up-scaled second plurality of weights and a plurality of third data to be convolved based on multiplication and summation because they are both directed to computing output activations as weighted sums of input data by channel-specific filter weights.)
wherein the second layer includes the second data and the down-scaled first plurality of weights included in the first output channel, ([Luo, page 2528, sec. 3.2.3-3.2.4] “We further minimize the reconstruction error (cf. Eq. (5)) by scaling the channel weight ... only those channels which have minimal reconstruction loss after rescaling are preserved.” AND [Luo, page 2527, sec. 3.2.2 and Fig. 3 caption] “We first randomly sample an element y from the activation tensor of layer i + 1 with random spatial location and random channel index.”, wherein the examiner interprets channels with “minimal reconstruction loss after rescaling” and their scaled weights, and Layer i+1 to be the same as the second layer including the second data and the down-scaled first plurality of weights included in the first output channel because they are both directed to holding activation data and their associated scaled weights in a layer (or multiple layers proceeding the first one) after rescaling but before final pruning.)
wherein after removing the first data by the first pruning, the third layer includes data other than the first data among the plurality of data. ([Luo, page 2529, sec. 3.2.4] “With the preserved channel index set S and scaling vector w, we can safely remove those weak filters in layer i. As for layer i + 1, we first discard corresponding channels to reduce model size, and then rescale the filter weights using w. After that, the pruned model can be fine-tuned.”, wherein the examiner interprets the “preserved channel index set S” and “discard corresponding channels to reduce model size”, as well as layer i+1 to be the same as, after removing the first data by the first pruning, the third layer including data other than the first data among the plurality of data because they are both directed to retaining only the remaining channels (the preserved subset) and carrying those channels’ data forward into subsequent layers (i.e., in a multi-layer neural network, the next layer also has a +1 layer, resulting in i+2).)
Luo does not teach wherein the plurality of data is related to user data about one or more of an image, a video, an audio, a voice, and a text; ... wherein the first plurality of weights includes one weight having a major value equal to or greater than a predetermined critical value, such that the one weight has a greater influence on a result of a convolution operation to obtain the first data compared to other weights of the first plurality of weights;.
Mishra teaches wherein the plurality of data is related to user data about one or more of an image, a video, an audio, a voice, and a text; ([Mishra, p.1] “Video data is obtained, on a first computing device, wherein the video data includes images of one or more people. Audio data is obtained, on a second computing device, which corresponds to the video data. A face within the video data is identified. A first voice, from the audio data, is associated with the face within the video data.”, wherein the examiner interprets “video data ... includes images of one or more people”, “audio data ... corresponds to the video data”, and “a first voice ... is associated with the face within the video data” to be the same as user data about an image, a video, an audio, and a voice because they are all directed to multimedia user content comprising images or video and corresponding audio or voice streams.)
Luo and Mishra do not teach wherein the first plurality of weights includes one weight having a major value equal to or greater than a predetermined critical value, such that the one weight has a greater influence on a result of a convolution operation to obtain the first data compared to other weights of the first plurality of weights;.
Hu teaches wherein the first plurality of weights includes one weight having a major value equal to or greater than a predetermined critical value, such that the one weight has a greater influence on a result of a convolution operation to obtain the first data compared to other weights of the first plurality of weights; ([Hu, page 4] “The final output of the block is obtained by rescaling transformation output U with the activations:The activations act as channel weights adapted to the input-specific descriptor z.”, wherein the examiner interprets “The activations act as channel weights adapted to the input-specific descriptor z” and “s_c · u_c” to be the same as one weight having a major value equal to or greater than a predetermined critical value such that the one weight has a greater influence on the result of a convolution operation compared to other weights because they are both directed to per-channel scaling factors that weight some channels more heavily than others when forming the convolution output.)
Luo, Mishra, and Hu do not teach wherein all of weights of the second and third layers are not up-scaled and all of weights of the first and third layers are not down-scaled.
Shen teaches wherein all of weights of the second and third layers are not up-scaled and all of weights of the first and third layers are not down-scaled. ([Shen, 022] “The second layer L2, the output X3’ of the second layer L2, the third layer L3, and the output X4’ of the third layer L3 have the first precision. That is, in FIG. 4, only the input X11 of the first layer L1’, the first layer L1’, and the output X12 of the first layer L1’ have the second precision.”, wherein the examiner interprets “only the input X11 of the first layer L1’, the first layer L1’, and the output X12 of the first layer L1’ have the second precision” and “The second layer L2 ... the third layer L3 ... have the first precision” to be the same as all of weights of the second and third layers not being up-scaled and all of weights of the first and third layers not being down-scaled because they are both directed to confining a down-scaling operation exclusively to a first layer and confining a corresponding up-scaling operation exclusively to the transition immediately following the first layer, such that the second layer and the third layer are not subject to either operation.)
Luo, Mishra, Hu, Shen, and the instant application are analogous art because they are all directed to operation methods of computing devices implementing CNNs that process user multimedia data (images, video, audio, and voice) and apply pruning and channel-weighting.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the scaled pruning of weights disclosed by Luo to include the video and audio data presented by Mishra and the channel weight adaptation technique disclosed by Hu. One would be motivated to do so to efficiently compress and accelerate convolutional neural networks that operate on user multimedia inputs while preserving or improving discriminability by emphasizing more informative channels, as suggested by Mishra ([Mishra, Abstract] “Video data is obtained ... Audio data ... corresponds to the video data.”) and Hu ([Hu, page 4] “The activations act as channel weights adapted to the input-specific descriptor z ... helping to boost feature discriminability.”).
It would have also been obvious to a person of ordinary skill in the art before the effective filing date of the invention to further to include the layer-segregated quantization and de-quantization scheme disclosed by Shen, wherein a first layer is quantized to a lower precision and the output of that first layer is de-quantized back to a higher precision before being provided to a second layer, while the second and third layers remain at the higher, unscaled precision. One would be motivated to do so to efficiently balance computational cost and prediction precision across the layers of the pruned network, as suggested by Shen ([Shen, 0026] “the layer with a larger quantization loss is quantized with the third precision which has higher precision of quantization ... The layer with a smaller quantization loss is quantized with the second precision which has the lower precision of quantization”)
Regarding claim 3, Luo, Mishra, Hu, and Shen teaches The operation method of claim 1, (see rejection of claim 1).
Luo further teaches wherein the downscaling of the first plurality of weights included in the first output channel associated with the first data includes: down-scaling the first plurality of weights included in the first output channel with a plurality of predetermined scaling values, respectively, and wherein the up-scaling of the value from among the downscaled first plurality of weights included in the first output channel includes: up-scaling the second plurality of weights used to generate the second data with the plurality of predetermined scaling values, respectively. ([Luo, page 2528, sec 3.2.3-3.2.4] “After obtaining the subset S, we can safely remove the n th channel in each filter of layer i+ 1ifn €S. The corresponding filters in the previous layer i can be pruned, too. We further minimize the reconstruction error (c.f. Eq. (5)) by scaling the channel weights, which is formulated as:
PNG
media_image1.png
47
284
media_image1.png
Greyscale
... only those channels which have minimal reconstruction loss after rescaling are preserved”, and “Note that, if a filter in W; is removed, its corresponding channel in [+ and Wi, are also discarded.”, AND [Luo, page 2528, sec 3.2.4] “Let us revisit the optimization goal. If we can assign a scaling factor w?j to each filter channel, Eq. (5) is rewritten as:
PNG
media_image2.png
91
418
media_image2.png
Greyscale
... Of course, these scaling factors are not globally optimal. Hence, we need to perform another least squares after selection as we have done in Eq. (7). Algorithm 1 summarizes our final pruning strategy” wherein the examiner interprets “layer i can be pruned, too. We further minimize the reconstruction error by scaling the channel weights [weights are scaled up and down according to Equation (6)]” and “filter in Wi is removed” to be the same as down scaling the first plurality of weights in the first output channel because in a CNN, removing filters also includes removing associated weights and its corresponding channel. The scaling operation could involve reducing some weights (down-scaling) while increasing others (up-scaling) to minimize reconstruction error, and “remove the nth channel” and “prune corresponding filters” to correspond to “performing the first pruning”, as shown in Eq. (6)). The examiner also interprets “these scaling factors are not globally optimal. Hence, we need to perform another least squares after selection...” to be the same as generating the second data based on the up-scaled second plurality of weights.)
Luo, Mishra, Hu, Shen, and the instant application are analogous art, because they are all directed to the operation method stated in claim 1, and the down-scaling and up-scaling of weight values from an output channel associated with predetermined values for scaling.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the operation method of claim 1 as disclosed by Luo, Mishra, Hu, and Shen to include the scaling of first plurality of weights and “scaling the channel weights... With the preserved channel index set S and scaling vector w” disclosed by Luo. One would be motivated to do so to efficiently use a scaling vector w determined from first data to generate second data from fine tuning, as suggested by Luo (Luo, page 2528, sec 3.2.3] in Algorithm 1 and “Each element in w can be interpreted as a scaling factor of its corresponding filter channel ... this scaling operation provides a better initialization for fine-tuning).
Regarding claim 4, Luo, Mishra, Hu, and Shen teaches The operation method of claim 1, (see rejection of claim1).
Luo further teaches wherein the downscaling of the first plurality of weights included in the first output channel associated with the first data includes: performing a convolution operation on the up-scaled second plurality of weights and a plurality of third data to obtain the second data. ([Luo, Page. 2527, sec. 3.2.2] “The convolution operation is computed with a corresponding bias b as follows:
PNG
media_image3.png
46
326
media_image3.png
Greyscale
AND [Luo, page 2528, sec 3.2.3-3.2.4]] “We further minimize the reconstruction error (c.f. Eq. (5)) by scaling the channel weights ... only those channels which have minimal reconstruction loss after rescaling are preserved.”, wherein the examiner interprets “The convolution operation” to minimize error to be the same as performing convolution on upscaled weights to generate next data.)
Luo, Mishra, Hu, Shen, and the instant application are analogous art, because they are all directed to the operation method stated in claim 1, and multiple layers to which convolution operations is performed, and up-scaled group of weights.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the operation method of claim 1 as disclosed by Luo, Mishra, Hu, and Shen to include “The convolution operation is computed with a corresponding bias b as follows:... scaling the channel weights” as disclosed by Luo. One would be motivated to do so to perform a convolution operation to either reduce reconstruction error as disclosed by Luo or to use a convolution operation on current or new data that is based on a group of up-scaled weights as disclosed by the instant application (see Luo, [page 2528, sec 3.2.3-3.2.4] quote above.)
Regarding claim 5, Luo, Mishra, Hu, and Shen teaches The operation method of claim 1, (see rejection of claim 1).
Luo further teaches wherein convolution operations in the first layer, the second layer, and the third layer are sequentially performed; ([Luo, page 2526-2527, Fig. 2 and sec. 3.2.1] “Starting from a pre-trained model, we prune it layer by layer with a predefined compression rate. We use a triplet ⟨Iᵢ, Wᵢ, φᵢ⟩ to denote the convolution process in layer i, where Iᵢ is the input tensor … and Wi...is a set of filters with K x K kernel size, whichoutputs a new tensor with D channels.,” wherein the examiner interprets “prune it layer by layer” together with “the convolution process in layer i” to be the same as convolution operations in the first layer, the second layer, and the third layer are sequentially performed because they are both directed to a multi-layer convolutional neural network in which each layer performs its own convolution step in order, one after another, on the activations produced by the previous layer.)
Luo, Mishra, Hu, Shen, and the instant applicant application are analogous art, because they are all directed to sequentially performing convolution on layers.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the operation method of claim 1 as disclosed by Luo, Mishra, Hu, and Shen to include the “convolution process of layer i” disclosed by Luo. One would be motivated to do so to efficiently sequentially, layer-by-layer, perform convolution as suggested by Luo ([Luo, page 2526-2527, Fig. 2 and sec. 3.2.1] “Starting from a pre-trained model, we prune it layer by layer with a predefined compression rate. We use a triplet ⟨Iᵢ, Wᵢ, φᵢ⟩ to denote the convolution process in layer i, where Iᵢ is the input tensor…”). Claim 14 is analogous to claim 5, and thus would face the same rejection set forth above.
Regarding claim 6, Luo, Mishra, Hu, and Shen teaches The operation method of claim 5, (see rejection of claim 5).
Luo further teaches wherein the plurality of data of the third layer includes fourth data, wherein the second layer further includes a second output channel associated with the fourth data, and wherein the operation method further comprises: calculating the fourth data based on the second data and a weight included in the second output channel. ([Luo, Page 2527, sec 3.2.1] “we use layer i + / to guide the pruning in layer i. The key idea is: if we can use a subset of channels in layer (i + 1)’s input to approximate the output in layer i + /, the other channels can be safely removed from the input of layer i + 1. Note that one channel in layer (i + 1)’s input is produced by one filter in layer i, hence we can safely prune the corresponding filter in layer i: ... our data-driven channel selection method, which determines the channels (and their associated filters) that are to be pruned away.”, wherein the examiner interprets “our data-driven channel selection method, which determines the channels” and “we use layer i+ to guide pruning in layer i” to be the same as using data from multiple layers (in this case layer i=3) and data associated with channels in a network.)
Luo, Mishra, Hu, Shen, and the instant application are analogous art, because they are all directed to the operation method of both claim 1 and claim 5, as well as layers of a neural network containing output channels, as well as calculating data based on weight values in a channel.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the operation method of claim 5 disclosed by Luo, Mishra, Hu, and Shen to include the “data-driven channel selection method, which determines the channels (and their associated filters [including weights] ) that are to be pruned....we use layer i+1 to guide pruning in layer i~” as disclosed by Luo. One would be motivated to do so to efficiently select channels and determine which corresponding layer(s) that contains either multiple layers or weights to perform the pruning, as suggested by Luo (Luo, [page 2527, sec. 3.2.1] “we can safely prune the corresponding filter in layer i.)
Regarding claim 7, Luo, Mishra, Hu, and Shen teaches The operation method of claim 6 (see rejection of claim 6).
Luo further teaches wherein the plurality of data of the second layer includes a third output channel associated with the fifth data, and wherein the operation method further comprises: calculating the fifth data based on the second data and a weight included in the third output channel. ([Luo, Page 2527, sec 3.2.1] “we use layer i + 1 to guide the pruning in layer i. The key idea is: if we can use a subset of channels in layer (i + 1)’s input to approximate the output in layer i + 1, the other channels can be safely removed from the input of layer i + 1. Note that one channel in layer (i + 1)’s input is produced by one filter in layer i, hence we can safely prune the corresponding filter in layer i: ... our data-driven channel selection method, which determines the channels (and their associated filters) that are to be pruned away.”, wherein the examiner interprets “our data-driven channel selection method, which determines the channels” and “we use layer i+1 to guide pruning in layer i” to be the same as using data from multiple layers (in this case layer i=3) and data associated with channels in a network.)
Luo, Mishra, Hu, Shen, and the instant application are analogous art, because they are all directed to the operation method of both claim 1 and claim 5, as well as layers of a neural network containing output channels, as well as calculating data based on weight values in a channel.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the operation method of claim 6 disclosed by Luo, Mishra, and Shen to include the “data-driven channel selection method, which determines the channels (and their associated filters [including weights] ) that are to be pruned....we use layer i+1 to guide pruning in layer i” as disclosed by Luo. One would be motivated to do so to efficiently select channels and determine which corresponding layer(s) that contains either multiple layers or weights to perform the pruning, as suggested by Luo (Luo, [page 2527, sec. 3.2.1] “we can safely prune the corresponding filter in layer i.)
Regarding claim 8, Luo, Mishra, Hu, and Shen teaches The operation method of claim 1, (see rejection of claim 1).
Luo further teaches:
by the computing device, ([Luo, Page 2525, 1.Introduction] “With our pruned network, some important transfer learning tasks such as object detection or fine- grained recognition can run much faster (in both training and inference), especially in small devices.”, wherein the examiner interprets “small devices” to be equivalent to “computing device”, and “run much faster” in the context of transfer learning tasks to indicate an “operation method”. Mobile devices mentioned throughout the paper will also be considered a computing device that will perform pruning-related tasks.)
selecting… third data on which a second pruning is to be performed; calculating, … at least one value based on the third data and at least one weight to be convolved with the third data; and performing, by the computing device the second pruning. ([Luo, page 2528, sec 3.2.3-3.2.4] “After obtaining the subset S, we can safely remove the nth channel in each filter of layer i + / if n € S. The corresponding filters in the previous layer i can be pruned, too. We further minimize the reconstruction error (c.f. Eq. (5)) by scaling the channel weight ... only those channels which have minimal reconstruction loss after rescaling are preserved”, wherein the examiner interprets “obtaining the subset S” to correspond to “selecting first [second or third] data”, and “remove the nth channel” and “prune corresponding filters” to correspond to performing the any instance of pruning).
performing … a convolution operation, based on the multiplication and summation of the least one weight and the third data, to obtain one fourth data; applying, … the at least one value to the at least one fourth data corresponding to a convolution result of the third data for purpose of compensation ([Luo, page 2528, sec 3.2.2] “Eq. (13.1) can be simplified as:
PNG
media_image4.png
85
333
media_image4.png
Greyscale
In which y*=y-b [b is bias value]. It is worthwhile to keep in mind that x* and y” are random variables whose instantiations require fixed spatial locations....we can find a subset $ = {1, 2, .. C} and the equality Eq. (13.4)
PNG
media_image5.png
81
337
media_image5.png
Greyscale
Eq. (13.4) cannot always be true for all instances of the random variables x” and y”. However, we can manually extract instances of them to find a subset S such that Eq. (13.4) is approximately correct.”, wherein the examiner interprets “manually extract instances …the approximated bias value of y” “ in Eq. 13.4 to be the same as adding/applying a value, bias value being “a [convolved] value” (a.k.a. approximation) that makes output approximately correct for the next data.)
Luo, Mishra, Hu, Shen, and the instant application are analogous art, because they are all directed to the operating method of claim 1 as well as calculating, applying, and performing pruning on data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the operating method of claim 1 as disclosed by Luo, Mishra, Hu, and Shen to include the operation to convolve the values for the purpose of compensation and obtaining data as disclosed by Luo. One would be motivated to do so to efficiently to perform a summated operation on convolved values onto data as suggested by Luo (see [Luo, page 2528, sec 3.2.3-3.2.4, and sec 3.2.2] quote above).
Regarding claim 10, Luo, Mishra, Hu, and Shen teaches The operation method of claim 8 (see rejection of claim 8).
Luo further teaches wherein the applying of the at least one value to the at least one fourth data corresponding to the convolution result of the third data for purpose of compensation includes: adding a corresponding value of the at least one value to a bias value of each of the at least one fourth data. ([Luo, page 2528, sec 3.2.2] “Eq. (13.1) can be simplified as:
PNG
media_image4.png
85
333
media_image4.png
Greyscale
In which y*=y-b [b is bias value]. It is worthwhile to keep in mind that x* and y” are random variables whose instantiations require fixed spatial locations....we can find a subset $ = {1, 2, .. C} and the equality Eq. (13.4)
PNG
media_image5.png
81
337
media_image5.png
Greyscale
Eq. (13.4) cannot always be true for all instances of the random variables x” and y”. However, we can manually extract instances of them to find a subset S such that Eq. (13.4) is approximately correct.”, wherein the examiner interprets “manually extract instances …”manually extract instances …the approximated bias value of y” in Eq. 13.4 to be the same as adding/applying a value, bias value being “a [convolved] value” (a.k.a. approximation) that makes output approximately correct for the next data.)
Luo, Mishra, Hu, Shen, and the instant application are analogous art, because they are all directed to the operation method that is stated in claim 8 and 1, as well as determining applying a certain average value on data that is derived from a computed result and applying it to data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the operation method of claim 1 as disclosed by Luo, Mishra, Hu, and Shen to include the “manually extract instances …the approximated bias value of y” as disclosed by Luo .One would be motivated to do so to efficiently maintain a high accuracy while reducing the number of parameters needed, a suggested by Luo (see [Luo, page 2528, sec 3.2.2] quote above).
Regarding claim 18, Luo teaches:
A computing device implementing a neural network, comprising: a buffer memory configured to store a first layer, a second layer and a third layer, and a memory interface configured to communicate with an external memory device and the buffer memory; a processor configured to communicate with the buffer memory, and a non-transitory computer-readable recording medium storing instructions, when executed by the processor, to cause the processor to:; ([Luo, page 2534, sec. 4.2] “We run these 5 models on 3 different devices: GPU, CPU and mobile phone.”, wherein the examiner interprets the GPU, CPU, and mobile phone devices that run the convolutional neural network models as including processors, memory subsystems (buffer memories) that store layer data, and memory interfaces that communicate with external memory and internal buffers, to be the same as a computing device implementing a neural network with a buffer memory, a memory interface, a processor, and a non-transitory computer-readable recording medium storing instructions because all of these elements describe standard computing hardware used to store and execute neural network layers and pruning instructions.)
select first data, among a first plurality of data, on which a first pruning is to be performed and second data, among a second plurality of data, on which second pruning is to be performed; ([Luo, page 2528, sec. 3.2.3-3.2.4] “After obtaining the subset S, we can safely remove the n th channel in each filter of layer i + 1 if n ∉ S. The corresponding filters in the previous layer i can be pruned, too. We further minimize the reconstruction error (cf. Eq. (5)) by scaling the channel weights … only those channels which have minimal reconstruction loss after rescaling are preserved.”, wherein the examiner interprets “remove the n th channel in each filter of layer i + 1 if n ∉ S” together with the repeated pruning procedure summarized in Algorithm 1 to be the same as selecting first data, among a first plurality of data, on which a first pruning is to be performed and second data, among a second plurality of data, on which second pruning is to be performed because in Luo each channel corresponds to feature-map data, and evaluating all channels to determine which channel indices (those not in S) will be removed in successive pruning steps is equivalent to selecting specific data items from first and second pluralities of data as the targets of first and second pruning operations.)
down-scale a first plurality of weights included in a first output channel associated with the first data, [wherein the first plurality of weights includes one weight having a major value equal to or greater than a predetermined critical value, such that the one weight has a greater influence on a result of a convolution operation to obtain the first data compared to other weights of the first plurality of weights]; ([Luo, page 2528, sec. 3.2.3-3.2.4] “We further minimize the reconstruction error (cf. Eq. (5)) by scaling the channel weights … only those channels which have minimal reconstruction loss after rescaling are preserved.”, wherein the examiner interprets scaling the channel weights to reduce reconstruction error, and preserving only channels with minimal reconstruction loss after rescaling, to be the same as down-scaling a first plurality of weights in an output channel associated with first data while effectively treating some weights as more influential than others because Luo adjusts the magnitudes of weights for different channels so that channels contributing more to accurate reconstruction are preserved and those contributing less are reduced and removed, which corresponds to having a weight of major value relative to a threshold or critical value that dominates the convolution result compared to other weights.)
up-scale a second plurality of weights used to generate third data to be multiplied by the one weight of the first plurality of weights; ([Luo, page 2528, sec. 3.2.4] “If we can assign a scaling factor wj to each filter channel, Eq. (5) is rewritten as:
PNG
media_image6.png
69
318
media_image6.png
Greyscale
Of course, these scaling factors are not globally optimal. Hence, we need to perform another least squares after selection as we have done in Eq. (7).”, wherein the examiner interprets assigning scaling factors wj that can increase the magnitude of some filter channels to be the same as up-scaling a second plurality of weights used to generate third data to be multiplied by the one weight of the first plurality of weights because both describe increasing the effective contribution of particular channels via scaling factors so that the resulting third data reflects enhanced influence from the selected weights.)
calculate at least one value based on the second data and at least one weight to be convolved with the second data based on multiplication and summation of the at least one weight and the second data, and apply the at least one value to at least one fourth data corresponding to a convolution result of the second data for purpose of compensation; ([Luo, page 2528, sec. 3.2.2] “Eq. (13.1) can be simplified as: <see previous Eq.> In which y* = y - b [b is bias value]. It is worthwhile to keep in mind that x* and y′ are random variables whose instantiations require fixed spatial locations … we can find a subset S = {1, 2, … , C} and the equality Eq. (13.4), <see previous Eq.>, Eq. (13.4) cannot always be true for all instances of the random variables x′ and y′. However, we can manually extract instances of them to find a subset S such that Eq. (13.4) is approximately correct.”, wherein the examiner interprets the computation of y from the sum over channel-wise products of weights and activations together with the bias b (y* = y - b) and subsequent approximated equality in Eq. (13.4) to be the same as calculating at least one value based on second data and at least one weight to be convolved with the second data and applying this value to fourth data corresponding to a convolution result of the second data for compensation because Luo derives numerical values (including bias and channel-wise contributions) from the convolution outputs and then uses these values to adjust or approximate the activations so that the pruned network compensates for removed channels.)
and calculate the third data based on the upscaled second plurality of weights; ([Luo, page 2528, sec. 3.2.4] “If we can assign a scaling factor wj to each filter channel, Eq. (5) is rewritten as: <see previous Eq.>, Of course, these scaling factors are not globally optimal. Hence, we need to perform another least squares after selection as we have done in Eq. (7). Algorithm 1 summarizes our final pruning strategy.”, wherein the examiner interprets computing the network outputs after scaling the channel weights by factors wj and then refining via least squares as calculating third data based on the upscaled second plurality of weights because both involve recalculating the network’s output using the updated, upscaled weights.)
wherein the non-transitory computer-readable recording medium stores the instructions, when executed by the processor, to cause the processor to perform first pruning on the first data to remove the first data from the first plurality of data and all the weights included in the first output channel associated with the first data and perform second pruning on the second data to remove the second data from the plurality of second data and all weights included in a second output channel associated with the second data in the neural network to increase operation speed and reduce power consumption of the neural network; ([Luo, page 2528, sec. 3.2.3-3.2.4] “After obtaining the subset S, we can safely remove the n th channel in each filter of layer i + 1 if n ∉ S. The corresponding filters in the previous layer i can be pruned, too.” AND [Luo, page 2525, sec. 1 Introduction] “With our pruned network, some important transfer learning tasks such as object detection or fine-grained recognition can run much faster (in both training and inference), especially in small devices.”, wherein the examiner interprets removing channels and their corresponding filters in adjacent layers as performing first and second pruning on selected data to remove both the data (channel activations) and all weights in the associated output channels, and interprets “run much faster … especially in small devices” as being the same as increasing operation speed and reducing power consumption of the neural network because pruning channels and filters reduces computation and memory access, thereby improving speed and efficiency on resource-constrained devices.)
wherein the first layer includes the up-scaled second plurality of weights and a plurality of fifth data to be convolved with the up-scaled second plurality of weights based on multiplication and summation of the up-scaled second plurality of weights and the plurality of fifth data; ([Luo, page 2527, sec. 3.2.1-3.2.2] “We use a triplet to denote the convolution process in layer i, where I is the input tensor, which has C channels, H rows and W columns … The convolution operation is computed with a corresponding bias b as follows:”, wherein the examiner interprets the filters W_i in layer i that are scaled by factors wj (as described in sec. 3.2.4) and applied to the input tensor I via convolution (multiplication and summation) to be the same as a first layer that includes an up-scaled second plurality of weights and a plurality of fifth data to be convolved with the up-scaled weights based on multiplication and summation because Luo’s description of layer i specifies weights and input data that are multiplied and summed to produce activations, and the channel-wise scaling factors up-scale those weights prior to convolution.)
wherein the second layer includes the third data, the down-scaled first plurality of weights included in the first output channel, the second data, and the at least one weight to be convolved with the second data; ([Luo, page 2528, sec. 3.2.3-3.2.4] “We further minimize the reconstruction error (cf. Eq. (5)) by scaling the channel weights … only those channels which have minimal reconstruction loss after rescaling are preserved.” AND [Luo, page 2527, sec. 3.2.2] “We first randomly sample an element y from the activation tensor of layer i + 1 with random spatial location and random channel index. According to the spatial location and channel index of y, the corresponding filter Wc and sliding window x can also be determined.”, wherein the examiner interprets the activation tensor at the next layer (layer i+1) holding third data, together with scaled (including down-scaled) channel weights and additional input data and filters corresponding to those channels, to be the same as a second layer including third data, a down-scaled first plurality of weights in a first output channel, second data, and at least one weight to be convolved with second data because Luo describes a layer whose activations depend on both scaled weights and input data, and where some weights are effectively reduced in magnitude while activations and filters define the next convolution step.)
and wherein after removing the first data by the first pruning, the third layer includes data other than the first data among the first plurality of data including the at least one fourth data; ([Luo, page 2528, sec. 3.2.3-3.2.4] “Iff a filter in W_i is removed, the corresponding channel in I_{i+1} and in W_{i+1} is also discarded, while the output tensor I_{i+2} remains the same size.”, wherein the examiner interprets discarding the channel corresponding to a removed filter while keeping the output tensor I_{i+2} the same size to be the same as a third layer including data other than the first data among the first plurality of data, including the compensated or adjusted activations (fourth data), after removing the first data via pruning because Luo’s pruning scheme removes particular channels and their associated weights while the subsequent layer’s output retains only the remaining channels’ data, which corresponds to data other than the pruned first data plus the compensated outputs.)
Luo does not teach the first plurality of data and the second plurality of data are related to user data about one or more of an image, a video, an audio, a voice, and a text … wherein the second plurality of weights includes one weight having a major value equal to or greater than a predetermined critical value, such that the one weight has a greater influence on a result of a convolution operation to obtain the third data compared to other weights of the second plurality of weights
Mishra teaches the first plurality of data and the second plurality of data are related to user data about one or more of an image, a video, an audio, a voice, and a text ([Mishra, Abstract] “video data is obtained … Audio data … A face within the data is identified. A first voice, from the audio data, is associated with the face within the video data.”, wherein the examiner interprets “video data … Audio data … [and] a first voice, from the audio data, … associated with the face within the video data” to be the same as user data about an image, a video, an audio, and a voice because Mishra explicitly processes user-related video and audio streams and associates face and voice content, which falls within the recited categories of user multimedia data.)
Mishra does not teach wherein the second plurality of weights includes one weight having a major value equal to or greater than a predetermined critical value, such that the one weight has a greater influence on a result of a convolution operation to obtain the third data compared to other weights of the second plurality of weights
Hu teaches wherein the second plurality of weights includes one weight having a major value equal to or greater than a predetermined critical value, such that the one weight has a greater influence on a result of a convolution operation to obtain the third data compared to other weights of the second plurality of weights ([Hu, page 4] “The final output of the block is obtained by rescaling transformation output U with the activations:
PNG
media_image7.png
37
302
media_image7.png
Greyscale
The activations act as channel weights adapted to the input-specific descriptor z.”, wherein the examiner interprets “The activations act as channel weights adapted to the input-specific descriptor z” and “s_c · u_c” to be the same as one weight having a major value equal to or greater than a predetermined critical value such that the one weight has a greater influence on the result of a convolution operation compared to other weights because they are both directed to per-channel scaling factors that weight some channels more heavily than others when forming the convolution output.)
Luo, Mishra, and Hu do not teach wherein all of weights of the second and third layers are not up-scaled and all of weights of the first and third layers are not down-scaled.
Shen teaches wherein all of weights of the second and third layers are not up-scaled and all of weights of the first and third layers are not down-scaled. ([Shen, 0022] “The second layer L2, the output X3’ of the second layer L2, the third layer L3, and the output X4’ of the third layer L3 have the first precision. That is, in FIG. 4, only the input X11 of the first layer L1’, the first layer L1’, and the output X12 of the first layer L1’ have the second precision.”, wherein the examiner interprets “only the input X11 of the first layer L1’, the first layer L1’, and the output X12 of the first layer L1’ have the second precision” and “The second layer L2 ... the third layer L3 ... have the first precision” to be the same as all of weights of the second and third layers not being up-scaled and all of weights of the first and third layers not being down-scaled because they are both directed to confining a down-scaling operation exclusively to a first layer and confining a corresponding up-scaling operation exclusively to the transition immediately following the first layer, such that the second layer and the third layer are not subject to either operation.)
Luo, Mishra, Hu, Shen, and the instant application are analogous art, because they are all directed to operation methods of computing devices implementing neural networks that process user multimedia data, including images, video, audio, and voice, apply pruning and channel weighting to reduce computational load, and apply scaling or precision adjustments to weights or values associated with multiple layers of the neural network.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the computing device and pruning method disclosed by Luo to include the video and audio data presented by Mishra and the channel weight adaptation technique disclosed by Hu. One would be motivated to do so to efficiently compress and accelerate convolutional neural networks that operate on user multimedia inputs while preserving or improving discriminability by emphasizing more informative channels, as suggested by Mishra ([Mishra, Abstract] "Video data is obtained ... Audio data ... corresponds to the video data.") and Hu ([Hu, page 4] "The activations act as channel weights adapted to the input-specific descriptor z ... helping to boost feature discriminability.").
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to further include the layer-segregated quantization and de-quantization scheme disclosed by Shen, wherein a first layer is quantized to a lower precision and the output of that first layer is de-quantized back to a higher precision before being provided to a second layer, while the second and third layers remain at the higher, unscaled precision. One would be motivated to do so to efficiently balance computational cost and prediction precision across the layers of the pruned network, as suggested by Shen ([Shen, 0026] "the layer with a larger quantization loss is quantized with the third precision which has higher precision of quantization ... The layer with a smaller quantization loss is quantized with the second precision which has the lower precision of quantization")
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Luo in view of Mishra in view of Hu in view of Shen in further view of NPL reference “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding” by Dally et. al (referred herein as Dally).
Regarding claim 9, Luo, Mishra, Hu, and Shen teaches The operation method of claim 8, (see rejection of claim 8).
Luo further teaches removing all weights included in a second output channel associated with the third data. ([Luo, Page 2527, sec 3.2.1] “Our goal is to remove some unimportant filters in Wi. Note that, if a filter in Wi is removed, its corresponding channel in fi+1 and Wi+1 are also discarded. However, since the number of filters in layeri + / is not changed, the size of its output tensor, i.e., [i+2, will remain exactly the same. Inspired by this observation, we believe that if we remove filters that have little influence on /i+2 (which is also the output of layeri + 1), it will have little influence on the overall performance, too. In other words, minimizing the reconstruction error of fi+2 is closely related to the network’s classification performance.”, wherein the examiner interprets “filter in Wi is removed, its corresponding channel in /i+1 and Wi+1 are also discarded” to be the same as the removal of all weights included in a given i th output channel associated with i+/ th data.)
Luo, Mishra, Hu, and Shen do not teach wherein the performing of the second pruning includes:.
Dally teaches wherein the performing of the second pruning includes: ([Dally, Page 1, abstract ] “Our method first prunes the network by learning only the important connections”, wherein the examiner interprets “first prunes ... by learning only the important connections” to be the same as selecting data to be pruned.)
Luo, Mishra, Hu, Shen, Dally, and the instant application are analogous art, because they are all directed to the operation method of claim 8 (which is also dependent on the operation method stated in claim 1), as well as removing weights that are in channels associated with data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the operation method of claim 8 as disclosed by Luo, Mishra, Hu, and Shen to include the “second” pruning will occur and a method to “remove some unimportant filters in Wi. Note that, if a filter in Wi is removed, its corresponding channel in /i+1, Wi+17” and “first prunes ... by learning only the important connections” as disclosed by Dally. One would be motivated to do so to effectively remove unimportant filters and their corresponding channels as well as minimize reconstruction error, as suggested by Luo (Luo [Page 2527, sec 3.2.1], “minimizing reconstruction error ... is closely related to … performance.”).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DEVAN KAPOOR whose telephone number is (703)756-1434. The examiner can normally be reached Monday - Friday: 9:00AM - 5:00 PM EST (times may vary).
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, David Yi can be reached at (571) 270-7519. 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.
/DEVAN KAPOOR/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights