DETAILED ACTION
The action is in response to the original filing on February 1, 2023 and the Remarks and Amendments filed on March 2, 2026. Claims 1-20 are pending and have been considered below. Claims 1 and 10 are independent claims. Claims 1-6, 10-14, 16, and 20 are amended.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. The analysis of the claims will follow the 2025 “Reminders on evaluating subject matter eligibility of claims under 35 U.S.C. 101,” (“SME Memo”).
Regarding claim 1:
Step 2A, Prong 1 – A judicial exception is recited in this claim as it recites mental processes (see MPEP 2106.04(a)(2)(III)):
identifying a plurality of compressible target blocks among a plurality of blocks included in the trained model, wherein the identifying comprises excluding predefined non-compressible blocks from the plurality of blocks, wherein the predefined non-compressible blocks are defined based on structural characteristics of the blocks in the trained model and a compression method… Given a small enough “plurality of blocks included in the trained model” and a simple enough “structural characteristics of the blocks” and “compression method,” “identifying a plurality of compressible target blocks” can reasonably be performed within the human mind or with the aid of a pen and paper model. Hence, “identifying a plurality of compressible target blocks among a plurality of blocks included in the trained model, wherein the identifying comprises excluding predefined non-compressible blocks from the plurality of blocks, wherein the predefined non-compressible blocks are defined based on structural characteristics of the blocks in the trained model and a compression method” is a mental process.
determining a set of compression parameters including block compression configuring values to be applied to a respective one of the plurality of target blocks, based on compression configuring information and latency information received from the device farm… Given a simple enough and small enough amount of “compression configuring information” and “latency information,” determining a small enough “set of compression parameters including block compression configuring values” can reasonably be performed within the human mind or with the aid of a pen and paper model. Hence, “determining a set of compression parameters including block compression configuring values to be applied to a respective one of the plurality of target blocks, based on compression configuring information and latency information received from the device farm” is a mental process.
Step 2A, Prong 2 – The following limitations are additional elements without significantly more than the abstract idea:
providing a neural network model… Providing a model amounts to no more than mere data gathering and outputting which is insignificant extra-solution activity that does not amount to an inventive concept (see MPEP 2106.05(g)).
that is performed by a computing device, comprising… A “computing device” used as a mere tool to apply an exception is a generic element for performing or applying the abstract idea using a generic computing environment (see MPEP 2106.05(f)).
receiving, at a processor of the computing device, a trained model that has been trained based on a data set and a target device identified in a device farm using information about the target device that has been inputted by a user… “receiving… a trained model that has been trained based on… information about the target device that has been inputted by a user” amounts to no more than mere data gathering and outputting which is insignificant extra-solution activity that does not amount to an inventive concept (see MPEP 2106.05(g)).
compressing the plurality of target blocks based on the set of compression parameters to generate a compressed trained model… Compressing a “plurality of target blocks” to “generate a compressed trained model” amounts to no more than mere data gathering and outputting which is insignificant extra-solution activity that does not amount to an inventive concept (see MPEP 2106.05(g)).
providing download data corresponding to the compressed trained model so that the compressed trained model is deployed on the target device… providing “download data… so that the compressed trained model is deployed” amounts to no more than mere data gathering and outputting which is insignificant extra-solution activity that does not amount to an inventive concept (see MPEP 2106.05(g)).
Step 2B – These additional elements are recited at such a high level of generality that they fail to integrate the abstract idea into a practical application, since they provide nothing more than mere instructions to implement an abstract idea on a generic computer (see MPEP 2106.05(f)) or only amount to data gathering or outputting without significantly more (see MPEP 2106.05(g)). These limitations, either taken alone or in combination, fail to provide an inventive concept. Thus, the claim is not patent eligible.
Claims 2-9 and 20 recite limitations which further narrow the abstract ideas of claim 1 by specifying more details of the mental concepts that occur:
Regarding claim 2, this claim further limits the abstract ideas of claim 1 to be based on a mathematical concept (see MPEP 2106.04(a)(2)(I)): and wherein, when the first compression mode is configured, the determining the set of compression parameters further comprises: deriving the block compression configuring values based on the model compression configuring value and a predefined algorithm… Determining “the set of compression parameters” which comprises “deriving the block compression configuring values based on… a predefined algorithm” requires mathematical calculations by referring to an algorithm, which is a mathematical concept. Furthermore, specifying wherein the compression configuring information includes a first compression mode indicating that the trained model is compressed based on a model compression configuring value that is configured by the user… is still insignificant extra-solution activity of mere data gathering (see MPEP 2106.05(g)).
Regarding claim 3, specifying providing the block compression configuring values to the user and when the block compression configuring values are modified by the user, updating the set of compression parameters to include a second set of compression parameters including the modified block compression configuring values is still insignificant extra-solution activity of mere data gathering (see MPEP 2106.05(g)).
Regarding claim 4, specifying wherein the compression configuring information includes a second compression mode indicating that information on a block included in the trained model is provided and the trained model is compressed based on block compression configuring values configured by the user…, and wherein, when the second compression mode is configured, the determining the set of compression parameters further comprises: providing information on the plurality of target blocks to the user…, and receiving the block compression configuring values configured by the user is still insignificant extra-solution activity of mere data gathering (see MPEP 2106.05(g)).
Regarding claim 5, specifying wherein the information on the block included in the trained model includes at least one of identification information of the block, a latency corresponding to the block, or a quantity of channels included in the block in this manner does not overcome the rejection of claim 1 as modifying the information on the block does not make the abstract ideas of claim 1 to not be mental processes.
Regarding claim 6, this claim further limits the abstract ideas of claim 1 to be based on a mathematical concept: wherein each latency data of the plurality of latency data is derived… to “derive” latency data is to calculate the delay between a request and its response, which is a mathematical concept. Furthermore, … derived by executing an associated block of the plurality of blocks by the target device would be mere instructions to apply an abstract idea using a generic computing environment (see MPEP 2106.05(f)). Furthermore, specifying receiving a plurality of latency data from the target device… and wherein each latency data of the plurality of latency data is associated with a respective one of the plurality of blocks… is insignificant extra-solution activity of mere data gathering (see MPEP 2106.05(g)).
Regarding claim 7, specifying wherein the compression configuring information includes at least one of compression methods, compression configuring values, or reference information for determining a compression target among a plurality of channels included in the trained model in this manner does not overcome the rejection of claim 1 as modifying the compression configuring information does not make the abstract ideas of claim 1 to not be mental processes.
Regarding claim 8, specifying receiving, at the processor, a user command for retraining the compressed trained model…, generating a retrained model based on the compressed trained model…, and providing download data corresponding to the retrained model… is still insignificant extra-solution activity of mere data gathering (see MPEP 2106.05(g)).
Regarding claim 9, this claim further limits the abstract ideas of claim 1 to be based on a mathematical concept: performing… at least one quantization or calibration operation on the compressed trained model based on the information about the target device… performing quantization or calibration operations are mathematical calculations, which is a mathematical concept. Furthermore, performing, at the processor would be mere instructions to apply an abstract idea using a generic computing environment (see MPEP 2106.05(f)).
Regarding claim 20, a non-transitory computer-readable recording medium storing a program that causes a computing device to execute the method of claim 1 would be mere instructions to apply an abstract idea using a generic computing environment (see MPEP 2106.05(f)).
Regarding claim 10:
Step 2A, Prong 1 – A judicial exception is recited in this claim as it recites mental processes (see MPEP 2106.04(a)(2)(III)):
identify a plurality of compressible target blocks among a plurality of blocks included in the trained model, wherein the identifying comprises excluding predefined non- compressible blocks from the plurality of blocks, wherein the predefined non-compressible blocks are defined based on structural characteristics of the blocks in the trained model and a compression method… Given a small enough “plurality of blocks included in the trained model” and a simple enough “structural characteristics of the blocks” and “compression method,” to “identify a plurality of compressible target blocks” can reasonably be performed within the human mind or with the aid of a pen and paper model. Hence, “identify a plurality of compressible target blocks among a plurality of blocks included in the trained model, wherein the identifying comprises excluding predefined non- compressible blocks from the plurality of blocks, wherein the predefined non-compressible blocks are defined based on structural characteristics of the blocks in the trained model and a compression method” is a mental process.
determine a set of compression parameters including block compression configuring values to be applied to a respective one of the plurality of target blocks, based on compression configuring information and latency information received from the device farm… Given a simple enough and small enough amount of “compression configuring information” and “latency information,” to determine a small enough “set of compression parameters including block compression configuring values” can reasonably be performed within the human mind or with the aid of a pen and paper model. Hence, “determine a set of compression parameters including block compression configuring values to be applied to a respective one of the plurality of target blocks, based on compression configuring information and latency information received from the device farm” is a mental process.
Step 2A, Prong 2 – The following limitations are additional elements without significantly more than the abstract idea:
providing a neural network model… Providing a model amounts to no more than mere data gathering and outputting which is insignificant extra-solution activity that does not amount to an inventive concept (see MPEP 2106.05(g)).
a communication interface, configured to send and receive data via a data network, including at least one communication circuit… A “communication interface… including at least one communication circuit” used as a mere tool to apply an exception is a generic element for performing or applying the abstract idea using a generic computing environment (see MPEP 2106.05(f)).
a memory configured to store at least one operation instruction… A “memory” used as a mere tool to apply an exception is a generic element for performing or applying the abstract idea using a generic computing environment (see MPEP 2106.05(f)).
a processor, wherein execution of the at least one operation causes the processor to… A “processor” used as a mere tool to apply an exception is a generic element for performing or applying the abstract idea using a generic computing environment (see MPEP 2106.05(f)).
receive a trained model that has been trained based on a data set and a target device identified in a device farm using information about the target device that has been inputted by a user… to “receive… a trained model that has been trained based on… information about the target device that has been inputted by a user” amounts to no more than mere data gathering and outputting which is insignificant extra-solution activity that does not amount to an inventive concept (see MPEP 2106.05(g)).
compress the plurality of target blocks based on the set of compression parameters to generate a compressed trained model… To compress a “plurality of target blocks” to “generate a compressed trained model” amounts to no more than mere data gathering and outputting which is insignificant extra-solution activity that does not amount to an inventive concept (see MPEP 2106.05(g)).
provide download data corresponding to the compressed trained model so that the compressed trained model is deployed on the target device… To “provide download data… so that the compressed trained model is deployed” amounts to no more than mere data gathering and outputting which is insignificant extra-solution activity that does not amount to an inventive concept (see MPEP 2106.05(g)).
Step 2B – These additional elements are recited at such a high level of generality that they fail to integrate the abstract idea into a practical application, since they provide nothing more than mere instructions to implement an abstract idea on a generic computer (see MPEP 2106.05(f)) or only amount to data gathering or outputting without significantly more (see MPEP 2106.05(g)). These limitations, either taken alone or in combination, fail to provide an inventive concept. Thus, the claim is not patent eligible.
Claim 19 recites limitations which further narrow the abstract ideas of claim 1 by specifying more details of the mental concepts that occur:
Regarding claim 19, specifying wherein the processor is further configured to determine a compression configuring value of the trained model based on the latency information in this manner does not overcome the rejection of claim 10 as modifying the processor does not make the abstract ideas of claim 10 to not be mental processes.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-5, 7, 9-14, 16, and 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over Chiu et al. (“C2S2: Cost-aware Channel Sparse Selection for Progressive Network Pruning,” 2019, hereinafter Chiu) in view of Welsh et al. (US 20220172119 A1, hereinafter Welsh) and further in view of Yashima (US 20220318563 A1, hereinafter Yashima).
Regarding claim 1:
Regarding the limitation a method for providing a neural network model that is performed by a computing device, comprising: receiving, at a processor of the computing device, a trained model that has been trained based on a data set and a target device identified in a device farm using information about the target device that has been inputted by a user, Chiu teaches a method for providing a neural network model that is performed by a computing device (Page 1, Col. 1, Section 1, ¶2 “network pruning… to alleviate the storage requirement and the computational burden,” wherein “storage” and “computation” imply performed by a computing device, Page 8, Col. 1, Section 6, ¶1 “When the pruning process is completed, we… obtain a pruned model, which is ready for deployment”), comprising: receiving, at a processor of the computing device, a trained model that has been trained based on a data set (Abstract: “a given pre-trained model for compression,” Page 9, Algorithm 1: Require: Dataset… and ConvNet M,” Page 5, Col. 1, Section 5, ¶1 “We evaluate our method on several popular ConvNets,” Col. 2, ¶3 “We thus perform a pilot experiment on ResNet-20 pre-trained on CIFAR-10. The result is reported in Figure 4,” Page 6, Fig. 4 depicts calculating “Error rate” for various pruning ratios; one of ordinary skill in the art would recognize that a processor of the computing device is implicit, for example, a CPU or GPU, in order to evaluate neural network performance after receiving… a trained model). However, Chiu fails to teach a trained model that has been trained based on a data set and a target device identified in a device farm using information about the target device that has been inputted by a user.
Welsh, in the same field of endeavor, teaches optimizing a model based on optimization data and a target device identified in a device farm (¶27 “the facility optimizes a subject model by using the optimization result data as a starting point for optimizing the subject model,” ¶28 “the facility manages a “device farm” in which target devices of a variety of types are used to test the execution of candidate implementations of the subject model… the facility chooses one or more of these target devices to optimize the subject model based on comparing the device's hardware and the hardware target of the subject model”) using information about the target device that has been inputted by a user (Fig. 9 – 901, ¶50 “FIG. 9 is a model hardware target selection screen… The hardware target list 901 includes one or more hardware targets which a user can select to identify which hardware targets the facility should use when optimizing the model”).
Chiu further teaches identifying a plurality of compressible target blocks among a plurality of blocks included in the trained model, wherein the identifying comprises excluding predefined non-compressible blocks from the plurality of blocks (Page 3, Col. 1, Section 3, ¶1 “At the core of our method is a new type of generic network layer called pruning layer that plays the crucial role of selecting and masking channels during the pruning process… After inserting the pruning layers to a neural network, the function of each pruning layer is to select informative channels from its inputs, i.e., the outputs from its preceding layer. Hence the pruned channels can be either feature maps or simply neurons, depending on whether a pruning layer is positioned right after a convolutional layer or a fully-connected layer. However, in this study we restrict the use of pruning layers to expunging redundant feature maps from convolutional layers”), wherein the predefined non-compressible blocks are defined based on structural characteristics of the blocks in the trained model (Page 3, Col. 1, Section 3, ¶2 “We augment the model with pruning layers, each of which follows right next to a convolutional layer. After performing channel sparse selection on the augmented network, we unplug the pruning layers and remove the masked channels to obtain a compressed network… that contains a reduced number of channels for each convolutional layer,” Page 4, Fig. 2a depicts how “pruning layers” are only added after “convolutional layers” as opposed to “fully connected layers” to reduce the “number of channels for each convolutional layer,” hence the identifying comprises excluding predefined non-compressible blocks from the plurality of blocks, wherein the predefined non-compressible blocks are defined based on structural characteristics of the blocks) and a compression method (Abstract: “network pruning”).
Regarding the limitation determining a set of compression parameters including block compression configuring values to be applied to a respective one of the plurality of target blocks, based on compression configuring information and latency information received from the device farm, Chiu further teaches determining a set of compression parameters including block compression configuring values (Page 3, Col. 1, Section 3, ¶3 “λ = {λ1, λ2} are the weight parameters in the objective function (2) for the layer-wise pruning,” Page 5, Col. 2, ¶3 “λ1 and λ2 are fixed at 0.002 in all experiments,” wherein a “weight parameters” set “λ” with values “{λ1 , λ2}” encompasses a set of compression parameters including block compression configuring values) to be applied to a respective one of the plurality of target blocks (Page 3, Col. 2, Section 3.1.1, ¶1 “the objective function for layer-wise pruning can be written as
PNG
media_image1.png
39
539
media_image1.png
Greyscale
”), based on compression configuring information (Page 4, Col. 2, Section 4, ¶3 to Page 5, Col. 1, ¶1 “At the beginning of pruning a specific convolutional layer, we trigger the state transition to move from Start to Pruning… the detection of excessive pruning… causes the process to make the state transition to Restoring,” ¶2 “Once the process is at Restoring state, our method automatically restores the pruned channels by changing the sign of λ1… then moves to the End state to conclude the channel selection for the convolutional layer,” wherein “changing the sign of λ1” encompasses determining a set of compression parameters including block compression configuring values… based on compression configuring information or “the state transition”). However, Chiu fails to teach determining a set of compression parameters… based on compression configuring information and latency information received from the device farm.
Welsh teaches the device farm (Fig. 3 - 300, ¶41 “The device farm 300”). However, Welsh fails to teach based on compression configuring information and latency information received from the device farm.
Yashima, in the same field of endeavor, teaches compression conditions and latency information received from a setting (Fig. 24 – 431, ¶214 “The child screen 431 is a screen for setting a permissible range (compression condition) for each of indices, namely, latency, the memory, the intermediate buffer…”).
Chiu further teaches compressing the plurality of target blocks based on the set of compression parameters to generate a compressed trained model (Page 3, Col. 1, Section 3, ¶1 “we restrict the use of pruning layers to… convolutional layers,” ¶3 “λ = {λ1, λ2} are the weight parameters,” Page 9, Col. 1, Equation (1) and ¶1 “Equation (1) is the objective function for learning pruning weights,” Col. 2, Algorithm 1 depicts the “Pruning” state, which updates each “pruning layer” with Equation (1) which uses λ1 and λ2 before updating “all convolution layers” to generate “the compressed model”).
Regarding the limitation providing download data corresponding to the compressed trained model so that the compressed trained model is deployed on the target device, Chiu further teaches the compressed trained model (Abstract: “a given pre-trained model for compression”) and the compressed trained model is deployed (Page 8, Col. 1, Section 6, ¶1 “When the pruning process is completed, we remove the augmented pruning layers and the identified insignificant channels to obtain a pruned model, which is ready for deployment”). However, Chiu fails to teach providing download data corresponding to the compressed trained model so that the compressed trained model is deployed on the target device.
Welsh teaches providing download data corresponding to a model so that the model is downloaded on the target device (Fig. 10 – 1002, ¶57 “when a user activates the download button 1002, the facility presents a model download screen,” Fig. 12 – 1203, ¶59 “When a user activates the download button 1203, the facility packages the optimized machine learning model for each of the selected hardware targets and packaging options, and allows the user to download the packaged machine learning model”).
Chiu, Welsh, and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the target device, device farm, and download data of Welsh and the latency of Yashima with the methodology of Chiu. The motivation to do so is to reduce “the dynamic display area, processing, storage, and/or data transmission resources needed to perform a certain task, thereby enabling the task to be permitted by less capable, capacious, and/or expensive hardware devices, and/or be performed with lesser latency, and/or preserving more of the conserved resources for use in performing other tasks” (Welsh, ¶33) and to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Regarding claim 2, Chiu in view of Welsh and further in view of Yashima teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Regarding the limitation wherein the compression configuring information includes a first compression mode indicating that the trained model is compressed based on a model compression configuring value that is configured by the user, Chiu teaches wherein the compression configuring information includes a first compression mode (Page 4, Col. 2, Section 4, ¶3 “At the beginning of pruning a specific convolutional layer, we trigger the state transition to move from Start to Pruning”) indicating that the model is compressed based on a model compression configuring value (Page 4, Col. 2, Section 4, ¶2 “we run all the training examples with the given pre-trained ConvNet and obtain the average error rate Ebase. Then, with the augmented network, we maintain an incrementally updated error rate, denoted as Eema, which is an exponential moving average of the training error rate,” ¶3 to Page 5, Col. 1, ¶1 “At the beginning of pruning a specific convolutional layer, we trigger the state transition to move from Start to Pruning. Meanwhile, we set a scaling coefficient cp to form a tolerable upper bound… the algorithm continues to prune redundant channels until Eema exceeds cp · Ebase. It signals the detection of excessive pruning and causes the process to make the state transition to Restoring,” ¶2 “Once the process is at Restoring state, our method automatically restores the pruned channels by changing the sign of λ1. Particularly, we set a proper coefficient cr such that the algorithm continues to restore the pruned channels until Eema is smaller than cr · Ebase and then moves to the End state to conclude the channel selection for the convolutional layer,” wherein the pruning being determined by the product of “average error rate Ebase” of the model and “scaling coefficient cp to form a tolerable upper bound” and “proper coefficient cr” to “restore the pruned channels” encompasses indicating that the model is compressed based on a model compression configuring value). However, Chiu fails to teach wherein the compression configuring information includes a first compression mode indicating that the trained model is compressed based on a model compression configuring value that is configured by the user.
Yashima teaches a compression rate that is configured by the user (Fig. 23 – 411, ¶211 “as for compression using the compression technique selected in the dropdown list 411, an accuracy deterioration tolerance value which is an index of the extent to which accuracy deterioration is tolerated and a target compression rate may be set by the user”).
Chiu further teaches and wherein, when the first compression mode is configured (Page 4, Col. 2, Section 4, ¶3 “we trigger the state transition to move from Start to Pruning,” Page 5, Col. 1, ¶1 “the detection of excessive pruning and causes the process to make the state transition to Restoring”), the determining the set of compression parameters (Page 3, Col. 1, Section 3, ¶3 “λ = {λ1, λ2} are the weight parameters… for layer-wise pruning”) further comprises: deriving the block compression configuring values based on the model compression configuring value (Page 5, Col. 1, ¶2 “Once the process is at Restoring state, our method automatically restores the pruned channels by changing the sign of λ1. Particularly, we set a proper coefficient cr such that the algorithm continues to restore the pruned channels until Eema is smaller than cr · Ebase and then moves to the End state”) and a predefined algorithm (Page 9, Algorithm 1: “if Eema > cp X Ebase then… λ1 [Wingdings font/0xDF] -λ1”).
Chiu and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the user configuration of Yashima with the methodology of Chiu. The motivation to do so is to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Regarding claim 3, Chiu in view of Welsh and further in view of Yashima teaches the method of claim 2 (and thus the rejection of claim 2 is incorporated).
Regarding the limitation further comprising: providing the block compression configuring values to the user, Chiu teaches the block compression configuring values (Page 3, Col. 1, Section 3, ¶3 “{λ1, λ2}”). However, Chiu fails to teach further comprising: providing the block compression configuring values to the user.
Yashima teaches further comprising: providing compression results to the user (Fig. 25-26 – 421, 441-442, ¶218 “A compression result for each computation layer included in the base model 421 is indicated on the right of the base model 421 subject to compression… A compression rate for the index selected in the index selection section 411 is indicated as a compression result for each computation layer”).
Regarding the limitation and when the block compression configuring values are modified by the user, updating the set of compression parameters to include a second set of compression parameters including the modified block compression configuring values, Chiu teaches updating the set of compression parameters (Page 3, Col. 1, Section 3, ¶3 “λ={λ1, λ2},” Page 9, Algorithm 1: “if Eema > cp X Ebase then… λ1 [Wingdings font/0xDF] -λ1”). However, Chiu fails to teach and when the block compression configuring values are modified by the user, updating the set of compression parameters to include a second set of compression parameters including the modified block compression configuring values.
Yashima further teaches and when the block compression configuring values are modified by the user (Fig. 24 – 421, 431, ¶214 “from among the computation layers included in the base model 421, an “Affine_3” layer is selected, and a child screen 431 is displayed. The child screen 431 is a screen for setting a permissible range (compression condition) for each of indices… for the selected computation layer,” ¶215 “A radio button for enabling a setting of a permissible range for each of the indices and text boxes for inputting a minimum value and a maximum value of the permissible range are provided in the child screen 431. A compression condition associated with the selected computation layer is set by enabling the setting of the permissible range and inputting the minimum and maximum values of the permissible range,” wherein “a permissible range (compression condition) for each of indices” encompasses the block compression configuring values), updating a compression setting to include a second set of compression parameters including the modified block compression configuring values (Fig. 3 – 211, Fig. 21 – S55-S56, ¶191 “it is determined in step S55 whether or not to change the compression setting in response to a user action,” ¶192 “In a case where it is determined in step S55 that the compression setting will be changed, the process proceeds to step S56, and the acceptance section 211 accepts selection of a computation layer,” wherein updating a “compression setting” to include a second set of compression parameters including the modified block compression configuring values, or the setting of “a permissible range (compression condition) for each of indices,” is implied to happen after “selection of a computation layer,” for example selecting an “Affine_3” layer, when “the compression setting will be changed”).
Chiu and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the providing of results to a user and the modifying of values by a user of Yashima with the methodology of Chiu. The motivation to do so is to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Regarding claim 4, Chiu in view of Welsh and further in view of Yashima teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Regarding the limitation wherein the compression configuring information includes a second compression mode indicating that information on a block included in the trained model is provided and the trained model is compressed based on block compression configuring values configured by the user, Chiu teaches the compression configuring information (Page 4, Col. 2, Section 4, ¶3 “we trigger the state transition to move from Start to Pruning,” Page 5, Col. 1, ¶2 “Once the process is at Restoring state, our method automatically restores the pruned channels… and then moves to the End state”) and the trained model (Abstract: “a given pre-trained model”) is compressed based on block compression configuring values (Page 3, Col. 1, Section 3, ¶3 “{λ1,λ2} are the weight parameters… for the layer-wise pruning”). However, Chiu fails to teach wherein the compression configuring information includes a second compression mode indicating that information on a block included in the trained model is provided and the trained model is compressed based on block compression configuring values configured by the user.
Yashima teaches wherein compression includes a second compression mode (Fig. 23 – 411, ¶207 “Three compression techniques, namely, “Pruning,” “Quantization,” and “Distillation” are displayed in the dropdown list 411”) indicating that information on a block included in a model is provided (Fig. 23 – 421, ¶209 “A calculation amount for each computation layer included in the base model 421 is indicated on the right of the base model 421. The calculation amount for each computation layer is indicated as a ratio of memory usage by each computation layer,” ¶211 “as for compression using the compression technique selected in the dropdown list 411, an accuracy deterioration tolerance value which is an index of the extent to which accuracy deterioration is tolerated and a target compression rate may be set by the user”) and compression is based on values configured by the user (Fig. 24 – 431, ¶215 “A radio button for enabling a setting of a permissible range for each of the indices and text boxes for inputting a minimum value and a maximum value of the permissible range are provided in the child screen 431. A compression condition associated with the selected computation layer is set by enabling the setting of the permissible range and inputting the minimum and maximum values of the permissible range”).
Regarding the limitation and wherein, when the second compression mode is configured, the determining the set of compression parameters further comprises: providing information on the plurality of target blocks to the user, Chiu teaches the determining the set of compression parameters (Page 3, Col. 1, Section 3, ¶3 “λ={λ1, λ2} are the weight parameters,” Page 5, Col. 1, ¶2 “Once the process is at Restoring state, our method automatically restores the pruned channels by changing the sign of λ1,” Col. 2, ¶4 “λ1 and λ2 are fixed at 0.002 in all experiments”). However, Chiu fails to teach and wherein, when the second compression mode is configured, the determining the set of compression parameters further comprises: providing information on the plurality of target blocks to the user.
Yashima teaches and wherein, when the second compression mode is configured (Fig. 23 – 411, ¶207 “Three compression techniques… are displayed in the dropdown list 411, and the user can select any one of the three compression techniques”), compression further comprises: providing information on the plurality of target blocks to the user (Fig. 23 – 412, 421, ¶208 “The button 412 is a GUI part for performing compression by the compression technique selected in the dropdown list 411,” ¶209 “a base model 421 subject to compression is displayed. A calculation amount for each computation layer included in the base model 421 is indicated on the right of the base model 421. The calculation amount for each computation layer is indicated as a ratio of memory usage by each computation layer”).
Yashima further teaches and receiving the block compression configuring values configured by the user (Fig. 24 – 431, ¶214 “The child screen 431 is a screen for setting a permissible range (compression condition) for each of indices… for the selected computation layer,” ¶215 “A radio button for enabling a setting of a permissible range for each of the indices and text boxes for inputting a minimum value and a maximum value of the permissible range are provided in the child screen 431”).
Chiu and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the second compression mode and the providing and configuring of values by a user of Yashima with the methodology of Chiu. The motivation to do so is to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Regarding claim 5, Chiu in view of Welsh and further in view of Yashima teaches the method of claim 4 (and thus the rejection of claim 4 is incorporated).
Yashima teaches wherein the information on the block included in the trained model includes at least one of identification information of the block, a latency corresponding to the block, or a quantity of channels included in the block (Fig. 24 – 421, 431, ¶214 “from among the computation layers included in the base model 421, an “Affine_3” layer is selected… The child screen 431 is a screen for setting a permissible range (compression condition) for each of indices, namely, latency”).
Chiu and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the information on the block of Yashima with the methodology of Chiu. The motivation to do so is to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Regarding claim 7, Chiu in view of Welsh and further in view of Yashima teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Yashima teaches wherein the compression configuring information includes at least one of compression methods, compression configuring values, or reference information for determining a compression target among a plurality of channels included in the trained model (Fig. 23 – 411-412, ¶207 “the user can select any one of three compression techniques,” Fig. 24 – 421, 431, ¶214 “an “Affine_3” layer is selected… The child screen 431 is a screen for setting a permissible range (compression condition) for each of indices,” ¶215 “A radio button for enabling a setting of a permissible range for each of the indices and text boxes for inputting a minimum value and a maximum value of the permissible range are provided in the child screen 431”).
Chiu and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the compression configuring information of Yashima with the methodology of Chiu. The motivation to do so is to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Regarding claim 9, Chiu in view of Welsh and further in view of Yashima teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Regarding the limitation further comprising: performing, at the processor, at least one quantization or calibration operation on the compressed trained model based on the information about the target device, Chiu teaches the compressed trained model (Abstract: “a given pre-trained model for compression”). However, Chiu fails to teach further comprising: performing, at the processor, at least one quantization or calibration operation on the compressed trained model based on the information about the target device.
Welsh teaches optimizing a model based on the information about the target device (¶28 “the facility chooses one or more of these target devices to optimize the subject model based on comparing the device's hardware and the hardware target of the subject model”). However, Welsh fails to teach further comprising: performing, at the processor, at least one quantization or calibration operation on the compressed trained model.
Yashima teaches further comprising: performing, at the processor (Fig. 3 – 110, 215, ¶52 “The control section 110 includes processors,” ¶58 “The control section 110 in FIG. 3 includes… a display control section 215,” ¶63 “The display control section 215 controls the display… of the GUI associated with the designing of a neural network and various pieces of information”), at least one quantization or calibration operation (Fig. 23 – 412, ¶207 “Three compression techniques, namely… Quantization”).
Chiu, Welsh, and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the processor and target device of Welsh and quantization of Yashima with the methodology of Chiu. The motivation to do so is to reduce “the dynamic display area, processing, storage, and/or data transmission resources needed to perform a certain task, thereby enabling the task to be permitted by less capable, capacious, and/or expensive hardware devices, and/or be performed with lesser latency, and/or preserving more of the conserved resources for use in performing other tasks” (Welsh, ¶33) and to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Regarding claim 20, Chiu in view of Welsh and further in view of Yashima teaches a non-transitory computer-readable recording medium storing a program that causes a computing device to execute (Yashima, ¶224 “The above series of processes can be performed by hardware or software. In a case where the series of processes are performed by software, the program included in the software is installed to a computer incorporated in dedicated hardware, a general-purpose personal computer, or the like from a program recording medium”) the method of claim 1 (Chiu in view of Welsh and further in view of Yashima teaches the method of claim 1, and thus the rejection of claim 1 is incorporated).
Chiu, Welsh, and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the non-transitory computer-readable medium storing a program of Yashima with the methodologies of Chiu and Welsh. The motivation to do so is to reduce “the dynamic display area, processing, storage, and/or data transmission resources needed to perform a certain task, thereby enabling the task to be permitted by less capable, capacious, and/or expensive hardware devices, and/or be performed with lesser latency, and/or preserving more of the conserved resources for use in performing other tasks” (Welsh, ¶33) and to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Regarding claim 10:
Regarding the limitation an electronic apparatus for providing a neural network model, comprising: a communication interface, configured to send and receive data via a data network, including at least one communication circuit, Chiu teaches for providing a neural network model (Page 8, Col. 1, Section 6, ¶1 “a pruned model, which is ready for deployment”). However, Chiu fails to teach an electronic apparatus for providing a neural network model, comprising: a communication interface, configured to send and receive data via a data network, including at least one communication circuit.
Yashima teaches an electronic apparatus for designing a model (Abstract: “an information processing apparatus, and a program that allow a neural network tailored to a desired task to be designed with ease”), comprising: a communication interface, configured to send and receive data via a data network (Fig. 1 – 10-30, ¶46 “The information processing terminal 10 and the information processing server 30 are connected via a network 20 in such a manner as to be able to communicate with each other”), including at least one communication circuit (Fig. 1 – 30, Fig. 2 – 100, Fig. 27 – 1000, 1009, ¶223 “it is sufficient if each of the processes performed by the above information processing apparatus 100 is performed by… the information processing server 30,” ¶226 “The above information processing apparatus 100 is realized by a computer 1000,” Fig. 27 – 1009, ¶228 “a communication section 1009 that includes a network interface”).
Yashima further teaches a memory configured to store at least one operation instruction (Fig. 27 –1008 ¶229 “the above series of processes are performed, for example, as a result of loading and execution of a program stored in the storage section 1008”).
Yashima further teaches and a processor (Fig. 27 – 1001 “the CPU 1001”).
Regarding the limitation wherein execution of the at least one operation instruction causes the processor to: receive a trained model that has been trained based on a data set and a target device identified in a device farm using information about the target device that has been inputted by a user, Chiu teaches to: receive a trained model that has been trained based on a data set (Abstract, Page 9, Algorithm 1, Page 5, Col. 1, Section 5, ¶1 and Col. 2, ¶3, Page 6, Fig. 4 all as explained above with respect to claim 1). However, Chiu fails to teach wherein execution of the at least one operation instruction causes the processor to: receive a trained model that has been trained based on a data set and a target device identified in a device farm using information about the target device that has been inputted by a user.
Yashima teaches wherein execution of the at least one operation instruction causes the processor to perform operations (Fig. 27 – 1000, 1001, 1008, ¶230 “The program executed by the CPU 1001… is installed to the storage section 1008,” ¶231 “the program executed by the computer 1000 may be a program that performs the processes… described in the present specification”). However, Yashima fails to teach … based on a data set and a target device identified in a device farm using information about the target device that has been inputted by a user.
Welsh teaches optimizing a model based on optimization data and a target device identified in a device farm (¶27-28) using information about the target device that has been inputted by a user (Fig. 9 – 901, ¶50).
Chiu further teaches identify a plurality of compressible target blocks among a plurality of blocks included in the trained model, wherein the identifying comprises excluding predefined non-compressible blocks from the plurality of blocks (Page 3, Col. 1, Section 3, ¶1), wherein the predefined non-compressible blocks are defined based on structural characteristics of the blocks in the trained model (Page 3, Col. 1, Section 3, ¶2, Page 4, Fig. 2a, all as explained above with respect to claim 1) and a compression method (Abstract: “network pruning”).
Regarding the limitation determine a set of compression parameters including block compression configuring values to be applied to a respective one of the plurality of target blocks, based on compression configuring information and latency information received from the device farm, Chiu further teaches determine a set of compression parameters including block compression configuring values (Page 3, Col. 1, Section 3, ¶3, Page 5, Col. 2, ¶3 all as explained above with respect to claim 1) to be applied to a respective one of the plurality of target blocks (Page 3, Col. 2, Section 3.1.1, ¶1 and Equation (2)), based on compression configuring information (Page 4, Col. 2, Section 4, ¶3 to Page 5, Col. 1, ¶1, ¶2 all as explained above with respect to claim 1). However, Chiu fails to teach determine a set of compression parameters… based on compression configuring information and latency information received from the device farm.
Welsh teaches the device farm (Fig. 3 - 300). However, Welsh fails to teach based on compression configuring information and latency information received from the device farm.
Yashima, in the same field of endeavor, teaches compression conditions and latency information received from a setting (Fig. 24 – 431, ¶214).
Chiu further teaches compress the plurality of target blocks based on the set of compression parameters to generate a compressed trained model (Page 3, Col. 1, Section 3, ¶1, ¶3, Page 9, Col. 1, Equation (1) and ¶1, Col. 2, Algorithm 1 all as explained above with respect to claim 1).
Regarding the limitation provide download data corresponding to the compressed trained model so that the compressed trained model is deployed on the target device, Chiu further teaches the compressed trained model (Abstract: “a given pre-trained model for compression”) and the compressed trained model is deployed (Page 8, Col. 1, Section 6, ¶1). However, Chiu fails to teach provide download data corresponding to the compressed trained model so that the compressed trained model is deployed on the target device.
Welsh teaches provide download data corresponding to a model so that the model is downloaded on the target device (Fig. 10 – 1002, ¶57, Fig. 12 – 1203, ¶59).
Chiu, Welsh, and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the target device, device farm, and download data of Welsh and the communication interface, memory, processor, and latency of Yashima with the apparatus of Chiu. The motivation to do so is to reduce “the dynamic display area, processing, storage, and/or data transmission resources needed to perform a certain task, thereby enabling the task to be permitted by less capable, capacious, and/or expensive hardware devices, and/or be performed with lesser latency, and/or preserving more of the conserved resources for use in performing other tasks” (Welsh, ¶33) and to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Regarding claim 19, Chiu in view of Welsh and further in view of Yashima teaches the electronic apparatus of claim 10 (and thus the rejection of claim 10 is incorporated).
Regarding the limitation wherein the processor is further configured to determine a compression configuring value of the trained model based on the latency information, Chiu teaches to determine a compression configuring value of the trained model (Page 3, Col. 1, Section 3, ¶3 “λ={λ1, λ2} are the weight parameters,” Page 5, Col. 1, ¶2 “Once the process is at Restoring state, our method automatically restores the pruned channels by changing the sign of λ1,” Col. 2, ¶4 “λ1 and λ2 are fixed at 0.002 in all experiments”). However, Chiu fails to teach wherein the processor is further configured to determine a compression configuring value of the trained model based on the latency information.
Yashima teaches wherein the processor is further configured compress a model based on the latency information (Fig. 3 – 110, 215, ¶52 “The control section 110 includes processors,” ¶58 “The control section 110 in FIG. 3 includes… a display control section 215,” ¶63 “The display control section 215 controls the display… of the GUI associated with the designing of a neural network and various pieces of information,” Fig. 24 – 431, ¶214 “The child screen 431 is a screen for setting a permissible range (compression condition) for each of indices, namely, latency… for the selected computation layer”).
Chiu and Yashima are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the processor and latency of Yashima with the apparatus of Chiu. The motivation to do so is to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”).
Claims 11-14, 16, and 18 recite an apparatus that parallels the method claims of 2-5, 7, and 9 respectively. Therefore, the analysis discussed above with respect to claims 2-5, 7, and 9 also applies to claims 11-14, 16, and 18, respectively. Accordingly, claims 11-14, 16, and 18 are rejected based on substantially the same rationale as set forth above with respect to claims 2-5, 7, and 9, respectively.
Claims 6 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Chiu in view of Welsh and further in view of Yashima, and further in view of Zhou et al. (US 12307365 B2, hereinafter Zhou).
Regarding claim 6, Chiu in view of Welsh and further in view of Yashima teaches the method of claim 5 (and thus the rejection of claim 5 is incorporated).
Regarding the limitation further comprising: receiving a plurality of latency data from the target device, Welsh teaches the target device (Fig. 3 – 305, ¶41 “the target devices”). However, Welsh fails to teach further comprising: receiving a plurality of latency data from the target device.
Yashima teaches further comprising: receiving a plurality of latency data from a setting (Fig. 24 – 431, ¶214 “from among the computation layers included in the base model… The child screen 431 is a screen for setting a permissible range (compression condition) for each of indices, namely, latency,” ¶215 “A radio button for enabling a setting of a permissible range for each of the indices and text boxes for inputting a minimum value and a maximum value of the permissible range are provided in the child screen 431”).
Yashima teaches wherein each latency data of the plurality of latency data is associated with a respective one of the plurality of blocks (Fig. 24 – 431, ¶214 “an “Affine_3” layer is selected… The child screen 431 is a screen for setting… latency”).
Regarding the limitation and wherein each latency data of the plurality of latency data is derived by executing an associated block of the plurality of blocks by the target device, Welsh teaches the target device (Fig. 3 – 305, ¶41). However, Welsh fails to teach and wherein each latency data of the plurality of latency data is derived by executing an associated block of the plurality of blocks by the target device.
Zhou, in the same field of endeavor, teaches and wherein each latency data of the plurality of latency data is derived by executing an associated block of the plurality of blocks by devices (Fig. 1 – S102, Col. 8 Lines 3-16 “The multiple preset network models are trained on multiple different types of devices, which may determine overall invoke latency of the preset network models on the devices. The first latency data refers to invoke latency data of the preset network model. Latency of the network layer is a part of the first latency data, and the first latency data may be utilized for obtaining the second latency data corresponding to the multiple types of network layers. After the second latency data is obtained, the preset network layer latency information is generated based on the corresponding relationship among the second latency data, the network layer type of the each network layer of the multiple network layers and the device types”).
Chiu, Welsh, Yashima, and Zhou are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the target device of Welsh, the latency of Yashima, and the deriving of latency by executing blocks of a neural network of Zhou with the methodology of Chiu. The motivation to do so is to reduce “the dynamic display area, processing, storage, and/or data transmission resources needed to perform a certain task, thereby enabling the task to be permitted by less capable, capacious, and/or expensive hardware devices, and/or be performed with lesser latency, and/or preserving more of the conserved resources for use in performing other tasks” (Welsh, ¶33), to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”), and to reduce latency when running a model on a device (Zhou, Col. 4 Lines 44-47 “the target model finally trained has a minimum latency when running on the device corresponding to the other device type. The problem that the latency is high in the related technologies is solved”).
Claim 15 recites an apparatus that parallels the method claim of 6, respectively. Therefore, the analysis discussed above with respect to claim 6 also applies to claim 15, respectively. Accordingly, claim 15 is rejected based on substantially the same rationale as set forth above with respect to claim 6, respectively.
Claims 8 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Chiu in view of Welsh and further in view of Yashima, and further in view of Choudhury et al. (WO 2020260991 A1, hereinafter Choudhury).
Regarding claim 8, Chiu in view of Welsh and further in view of Yashima teaches the method of claim 1 (and thus the rejection of claim 1 is incorporated).
Regarding the limitation receiving, at the processor, a user command for retraining the compressed trained model, Chiu teaches the compressed trained model (Abstract: “a given pre-trained model for compression”). However, Chiu fails to teach receiving, at the processor, a user command for retraining the compressed trained model.
Yashima teaches receiving, at the processor, a user command (Fig. 3 – 110, 211, 215, ¶52 “The control section 110 includes processors,” ¶58 “The control section 110 in FIG. 3 includes an acceptance section 211… a display control section 215,” ¶59 “the acceptance section 211 accepts a user input associated with the designing of a neural network,” ¶63 “The display control section 215 controls the display… of the GUI associated with the designing of a neural network and various pieces of information”). However, Yashima fails to teach receiving, at the processor, a user command for retraining the compressed trained model.
Choudhury, in the same field of endeavor, teaches a module for retraining (Fig. 1 – 108-110, ¶27 “The retraining module 110 retrains the compressed model output by the weighted decomposition module 108 resulting in a trained compressed model”).
Choudhury further teaches generating a retrained model based on the compressed trained model (Fig. 1 – 110, 112, ¶27 “The retraining module 110 retrains the compressed model… resulting in a trained compressed model 112”).
Regarding the limitation providing download data corresponding to the retrained model, Welsh teaches providing download data corresponding to an optimized model (Fig. 10 – 1002, ¶57 “when a user activates the download button 1002, the facility presents a model download screen,” Fig. 12 – 1203, ¶59 “When a user activates the download button 1203, the facility packages the optimized machine learning model… and allows the user to download the packaged machine learning model”). However, Welsh fails to teach providing download data corresponding to the retrained model.
Choudhury teaches outputting the retrained model (Fig. 5 – 510, ¶32 “Step 510 includes outputting the compressed trained model,” wherein it is implicit that “the compressed trained model” may be “a trained compressed model” produced by “the retraining module”).
Chiu, Yashima, Welsh, and Choudhury are analogous art to the claimed invention as all are from the same field of endeavor of machine learning. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the user command of Yashima and the retraining and download data of Choudhury with the methodology of Chiu. The motivation to do so is to reduce “the dynamic display area, processing, storage, and/or data transmission resources needed to perform a certain task, thereby enabling the task to be permitted by less capable, capacious, and/or expensive hardware devices, and/or be performed with lesser latency, and/or preserving more of the conserved resources for use in performing other tasks” (Welsh, ¶33), to better optimize a neural network for a specific task (Yashima, ¶141 “This makes it possible to design a neural network tailored to a desired task with ease and, by extension, makes it possible to optimize the structure of a neural network tailored to a wide range of tasks”), and to reduce the “size and/or flops of neural networks compared to existing compression techniques” (Choudhury, ¶79).
Claim 17 recites an apparatus that parallels the method claim of 8, respectively. Therefore, the analysis discussed above with respect to claim 8 also applies to claim 17, respectively. Accordingly, claim 17 is rejected based on substantially the same rationale as set forth above with respect to claim 8, respectively.
Response to Amendments
In review of Applicant’s amendments, filed March 2, 2026, the objections to the claims, specification, and drawings made in the previous office action have been withdrawn.
The rejection of claim 14 under 35 U.S.C. 112(b) set forth in the previous office action is withdrawn in view of the amendment to the claim.
Response to Arguments
In review of Applicant’s amendments, filed March 2, 2026, the non-statutory subject matter rejections from the previous office action made under 35 U.S.C. 101 have been withdrawn.
Applicant’s arguments and amendments, filed March 2, 2026, regarding the abstract idea rejections from the previous office action made under 35 U.S.C. 101 have been fully considered but are not persuasive.
On page 12 of the Remarks, Applicant argues that “even though the claimed limitation “compressing the plurality of target blocks based on the set of compression parameters” involves a broad array of techniques and/or activities that may involve or rely upon mathematical concepts, the limitation does not set forth or describe any mathematical relationships.” Applicant cites as reasoning an excerpt from “SME Memo” pertaining to distinguishing “claims that recite a judicial exception… from claims that merely involve a judicial exception.” Applicant submits that the “compressing” of claim 1 “merely involves a mathematical concept, which is patent eligible.” Examiner agrees with the distinction. However, the “compressing” and generating of claim 1, when evaluated under Step 2A, Prong Two, is merely inputting and outputting data (See MPEP 2106.05(g)). This aligns with the inputting and outputting of data in claim 2 of example 47. Per the claim, the trained model is received and used in the “compressing” step to “generate” a compressed model that is later deployed to a target device, which is merely inputting and outputting data recited at a high level of generality and a post-solution step of transmitting data output that does not meaningfully limit the claim, and thus is insignificant extra-solution activity and not subject matter eligible.
On page 13 of the Remarks, Applicant argues that the amended “identifying” and “excluding predefined non-compressible blocks” of claim 1 do not recite a mental process. Applicant cites as reasoning an excerpt from “SME Memo” pertaining to “claim limitations that encompass AI in a way that cannot be practically be performed within the human mind.” Applicant submits that claim 1 does not recite a mental process that can be practically performed in the human mind. Applicant further cites as reasoning passages from the specification discussing “reliance on non-human processing.” Examiner respectfully disagrees with the conclusion that the “identifying” and “excluding” cannot be performed in the human mind. The amended claim limitation “identifying a plurality of compressible target blocks… comprises excluding predefined non-compressible blocks” does not “encompass AI in a way that cannot be practically performed in the human mind,” since given a simple pen and paper neural network model with “a plurality of blocks,” a human can reasonably perform “identifying” and “excluding” blocks in the model with the aid of a pen and paper (See MPEP 2106.04(a) “With the exception of generic computer-implemented steps, there is nothing in the claims themselves that foreclose them from being performed by a human, mentally or with pen and paper”). Furthermore, using the “processor” of paragraph [0239] and “a computer” of [0313] to perform the mental process would be mere instructions to apply the exception using a generic computer (See MPEP 2106.05(f) “Use of a computer or other machinery in its ordinary capacity for economic or other tasks… or simply adding a general purpose computer or computer components after the fact to an abstract idea… does not integrate a judicial exception into a practical application or provide significantly more”).
On page 14 of the Remarks, Applicant argues that “even if claim 1 is directed to a judicial exception, additionally recited elements in the claim integrate the judicial exception into a practical application.” Applicant cites as reasoning the “2019 Revised Patent Subject Matter Eligibility Guidance” pertaining to the recited elements in the claim “integrating the judicial exception into a practical application,” an excerpt from “SME Memo” pertaining to the specification describing “the invention such that the improvement would be apparent to one of ordinary skill in the art,” and reasoning excerpts from the MPEP pertaining to whether the claim amounts to “an improvement to a computer or technical field” (See MPEP 2106.05(d)(1) and 2106.05(a)). Applicant submits that claim 1 is patent eligible because it directly improves the technology by improving user convenience and satisfaction by allowing the user to acquire a “lightweight model” optimized for “a target device,” and that the amended features in claim 1 reflect an “improvement in the technical field of user acquisition of a neural network model optimized for a target device.” Applicant further cites as reasoning passages from the specification outlining the alleged improvements. Examiner respectfully disagrees with the conclusion that the claim amounts to an improvement in technology. Although Applicant asserts that paragraphs [0243], [0272], and [0276] in the specification “details the improvements to the technical field of user acquisition of a neural network model optimized for a target device using two compression modes,” the specification only sets forth the improvement in a conclusory manner, while the claim itself fails to reflect the disclosed improvements of “user convenience” or “user satisfaction” after obtaining a lightweight model optimized for a target device using two compression modes (See MPEP 2106.04(d)(1) “if the specification explicitly sets forth an improvement but in a conclusory manner (i.e., a bare assertion of an improvement without the detail necessary to be apparent to a person of ordinary skill in the art), the examiner should not determine the claim improves technology”).
In consideration of this conclusion, independent claims 1 and 10, and their associated dependent claims 2-9, 20, and 11-19, respectively, are subject-matter ineligible and, thus, the rejections under 35 U.S.C. 101 stand.
Applicant’s arguments, filed March 2, 2026 regarding the rejections from the previous office action made under 35 U.S.C. 103 have been fully considered but are moot as they do not apply to the references Chiu and Welsh being used in the current rejections of claims 1 and 10 and their associated dependent claims 2-9, 20 and 11-19, respectively, to teach the amended claim limitation directed to identifying a plurality of compressible target blocks by excluding predefined non-compressible blocks based on structural characteristics of the blocks in the trained model and a compression method and the original claim limitations of identifying a device in a device farm and providing download data corresponding to the trained model.
Specifically, Chiu teaches the amended claim limitation wherein identifying compressible target blocks comprises excluding predefined non-compressible blocks or layers, and wherein the predefined non-compressible blocks are defined based on their structural characteristics, such as compression being limited to reducing channels in “convolutional layers” only, and a compression method, or “pruning.”
Additionally, in the absence of the previous references Choudhury and Yang, the reference Welsh teaches the original claim limitations of identifying a device in a device farm using information inputted by a user and specifying providing download data corresponding to the compressed trained model to deploy the compressed trained model to a target device.
With the addition of the reference Chiu teaching the subject matter introduced in the amendments and the addition of reference Welsh teaching the subject matter in the original claim limitations, the rejections under 35 U.S.C. 103 stand.
Conclusion
Applicant’s amendments necessitated the new ground(s) of rejection presented in this office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to WILLIAM M LEE whose telephone number is (571)272-4761. The examiner can normally be reached Mon-Fri. 8am-5pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Cesar Paula can be reached at (571)272-4128. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/WILLIAM MICHAEL LEE/
Examiner, Art Unit 2145
/CESAR B PAULA/ Supervisory Patent Examiner, Art Unit 2145
Prosecution Insights