DETAILED ACTION
This Office action is in response to the request for continued examination, associated fee
payment, and amendments filed on February 27th, 2026, for Application No. 17/566,624, in which claims 1-16 and 21-24 are presented for examination.
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 .
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and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114.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 02/27/2026 has been entered.
Information Disclosure Statement
The information disclosure statement submitted on 03/17/2026 is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure was considered by the examiner.
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 8-13 are rejected under 35 U.S.C. 101 because the claimed invention is directed to non-statutory subject matter.
Regarding claims 8-13, the claims do not fall within at least one of the four categories of patent eligible subject matter because the broadest reasonable interpretation of the “computer program product” encompasses software per se. The specification discloses that “[o]ne or more embodiments of the invention or elements thereof can be implemented in the form of a computer program product including a computer readable storage medium with computer usable program code for performing the method steps indicated” (Applicant’s Spec., Para. [0007]), “a computer program product can include a computer-readable storage medium with code adapted to be implemented to carry out one or more method steps described herein, including the provision of the system with the distinct software modules” (Applicant’s Spec., Para. [0090]), and “[t]he computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention” (Applicant’s Spec., Para. [0093]). These descriptions encompass both products with a physical or tangible form and products that do not have a physical or tangible form. A claim whose BRI covers both statutory and non-statutory embodiments embraces subject matter that is not eligible for patent protection and therefore is directed to non-statutory subject matter. See MPEP 2106.03(II). It is suggested that claim 1 be amended to recite structural recitations to overcome this rejection.
Additionally, the claims do not fall within at least one of the four categories of patent eligible subject matter because the broadest reasonable interpretation of the “computer readable medium” encompasses signals per se. While the specification describes a non-transitory computer readable medium (see Applicant’s Spec., Para. [0004] and Para. [0049]), it does not discuss any computer readable medium that is not specifically non-transitory. As such, the readable computer medium encompasses both transitory and non-transitory entities. A claim whose BRI covers both statutory and non-statutory embodiments embraces subject matter that is not eligible for patent protection and therefore is directed to non-statutory subject matter. See MPEP 2106.03(II). It is suggested that claim 1 be amended to recite a “non-transitory” computer readable medium to overcome this rejection.
As a result, claims 8-13 fail to recite statutory subject matter under 35 U.S.C. 101. Accordingly, the claims are rejected.
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 1-2, 8-9, 14-15, and 21-24 are rejected under 35 U.S.C. 103 as being unpatentable over Lai et al. (hereinafter Lai) (Patent Pub. No. US 2021/0011288 A1) in view of Kurtz et al. (hereinafter Kurtz) (“Inducing and Exploiting Activation Sparsity for Fast Neural Network Inference”).
Regarding Claim 1, Lai teaches a computer-implemented method, comprising (Para. [0039], “systems and methods . . . for implementing multiple neural networks across multiple devices” and Para. [0075], “Implementations of the present disclosure can be realized in a variety of apparatus including electronic devices implemented using any combination of circuitry and software”):
learning an input transformation function (Fig. 3A, where “First Device 302a” includes “Neural Network 114a”, which is within the broadest reasonable interpretation of an input transformation function, which is learned, see Para. [0045], “The input data 110 can include any type or form of data for configuring, tuning, training and/or activating a neural network 114”)
that transforms input data that is input to a downstream machine learning system that performs a task . . . (Para. [0040], “The first device can generate a reduced data set based on the input data and provide the reduced data set to the second device as an input to the neural network on the second device”; Fig. 3A; and Para. [0056], “The output data 112 can include object detection data, enhanced/translated/augmented data, a recommendation, a classification, and/or segmented data, as examples”, where “reduced” input data is within the broadest reasonable interpretation of transformed input data, which is input to a downstream machine learning system, “provide the reduced data set to the second device as an input to the neural network on the second device”, in order to perform a task, “object detection data, enhanced/translated/augmented data, a recommendation, a classification, and/or segmented data”, and where, as depicted in Fig. 3a, the “Second Device 302b” is downstream of the “First Device 302a” because the “Second Device 302b” receives the “Reduced Data Set 310” from the “First Device 302a” as downstream input, in order to produce the “Output Data 112”),
the learning using a minimization of a summation of a task loss for the task . . . [and energy consumed] for the downstream machine learning system (Fig. 1A; Fig. 3A; and Para. [0045], “Tuning or configuring can refer to or include training . . . [which can include] establishing the neural network . . . to be successful for the type of problem or objective desired”, where the learning, “Tuning or configuring can refer to or include training”, uses a minimization of a summation, “establishing the neural network . . . to be successful for the type of problem or objective desired”, of a task loss for the task, see Para. [0045], “training . . . the neural network 114 to allow the neural network to improve accuracy”, and energy consumed, see Para. [0039], “methods can be implemented to optimize neural network computations for energy savings”, for the downstream machine learning system, “Neural Network 114b”);
receiving new input data (Fig. 3B; Para. [0140], “The process 350 includes receiving input data at a first device . . . In some embodiments, the input data is received from sensors, cameras, etc., of the HMD and/or the AR system”; and Para. [0046], “The neural network(s) 114 can be deployed or used to perform data (e.g., image, audio, video) processing, object or feature recognition, recommender functions, data or image classification, data (e.g., image) analysis, etc., such as natural language processing”, where input data is received, “the input data is received”, which is new when “from sensors, cameras, etc., of the HMD and/or the AR system” and when “deployed or used”);
transforming the new input data using the learned input transformation function such that transformed new input data is generated (Fig. 3B and Para. [0141], “The process 350 includes generating a reduced data set using . . . a first neural network and the input data . . . In some embodiments, step 354 is performed by the neural network 114a using the input data received or obtained in step 352 (e.g., the input data 110)”, where the new input data is transformed using the learned input transformation function, “performed by the neural network 114a using the input data received”, such that transformed new input data is generated, “generating a reduced data”); and
inputting the transformed new input data into the downstream machine learning system such that the downstream machine learning system performs the task as an inferencing task with respect to the transformed new input data (Fig. 3B and Para. [0145] – [0146], “Process 350 includes providing the reduced data set to the second device (step 362) . . . Process 350 includes identifying one or more features of the reduced data set using . . . a second neural network and the reduced data set . . . the reduced data set is output by the first device (e.g., by the first neural network . . . ) and provided to the neural network . . . of the second device as an input”, where the transformed new input data, “the reduced data set”, is input into the downstream machine learning system, “provided to the neural network . . . of the second device as an input”, such that the downstream machine learning system performs the task as an inferencing task with respect to the transformed new input data, “identifying one or more features of the reduced data set using . . . a second neural network and the reduced data set”; see also Para. [0140], “The process 350 includes receiving input data at a first device . . . In some embodiments, the input data is received from sensors, cameras, etc., of the HMD and/or the AR system” and Para. [0046], “The neural network(s) 114 can be deployed or used to perform data (e.g., image, audio, video) processing, object or feature recognition, recommender functions, data or image classification, data (e.g., image) analysis, etc., such as natural language processing”).
Lai does not explicitly disclose . . . by using activations of zeros and non-zeros and by using a post-activation density . . . and a post-activation density loss for the post-activation density, wherein the minimization that includes the post-activation density loss reduces a number of non-zero activation values . . . .
However, Kurtz teaches . . . [a machine learning system that performs an inferencing task] by using activations of zeros and non-zeros and by using a post-activation density . . . (Pg. 1, Col. 1, Abstract, “To induce highly sparse activation maps without accuracy loss, we introduce a new regularization technique, coupled with a new threshold-based sparsification method based on a parameterized activation function called Forced Activation-Threshold Rectified Linear Unit (FATReLU). We examine the impact of our methods on popular image classification models . . . and show that it can enable significant speedups for end-to-end inference on a range of popular models”, where a machine learning system performs an inferencing task, “image classification models . . . for end-to-end inference”, by using activations of zeros and non-zeros adjusted using “a new threshold-based sparsification method based on a parameterized activation function called Forced Activation-Threshold Rectified Linear Unit (FATReLU)”, see Pg. 2, Col. 1, Para. 2, “FATReLU implements a variable threshold for the common ReLU activation function, below which all activations are set to zero”, and by using a post-activation density, “highly sparse activation maps”; see also Pg. 9, Col. 1, Para. 3, “The second noticeable trend is that Dynamic Boosting can consistently reduce the density of activations without accuracy loss” and Pg. 6, Col. 2, Para. 4, “we must compress our input data at runtime . . . we use the vcmp instruction to identify the locations of the non-zero elements in v stored in a mask register m”)
[using a minimization of a summation of a task loss for the task] and a post-activation density loss for the post-activation density (Pg. 1, Col. 1, Abstract, “we present an analysis of methods for maximizing the sparsity of the activations in a trained neural network . . . To induce highly sparse activation maps without accuracy loss, we introduce a new regularization technique, coupled with a new threshold-based sparsification method based on a parameterized activation function called Forced Activation-Threshold Rectified Linear Unit (FATReLU). We examine the impact of our methods on popular image classification models . . . and show that it can enable significant speedups for end-to-end inference on a range of popular models” and Pg. 2, Col. 1, Para. 2, “The methods we propose induce negligible (< 0.3%) accuracy loss on ImageNet-scale models, and can even result in minor accuracy increase . . . we find that sharp thresholds exist for activations, beyond which accuracy drops, and cannot be recovered. This observation simplifies the fine tuning process, since, up to this threshold, we are usually able to recover full accuracy”, where “maximizing the sparsity of the activations . . . without accuracy loss” is a minimization of a summation of a task loss for the task, “without accuracy loss” “and can even result in minor accuracy increase”, and a post-activation density loss for the post-activation density, “maximizing the sparsity of the activations”, with a constraint that “sharp thresholds exist for activations” are adhered to “recover full accuracy”, which is used for system inferencing, “inference on a range of popular models”; see also Pg. 9, Col. 1, Para. 3, “The second noticeable trend is that Dynamic Boosting can consistently reduce the density of activations without accuracy loss”),
wherein the minimization that includes the post-activation density loss reduces a number of non-zero activation values . . . (Pg. 2, Col. 1, Para. 2, “Our sparsity-boosting methods combine a regularizer following the Hoyer sparsity metric (Hoyer, 2004), together with a variant of the classic ReLU activation, which we call Forced Activation Threshold ReLU (FATReLU). Simply put, FATReLU implements a variable threshold for the common ReLU activation function, below which all activations are set to zero”, wherein the minimization that includes the post-activation density loss, “Our sparsity-boosting methods”, reduces a number of non-zero activation values, “FATReLU implements a variable threshold . . . below which all activations are set to zero”; see also Pg. 4, Col. 2, Section “3. Boosting Activation Sparsity”, Subsection “Forced-Activation Thresholds”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the learning an input transformation function that transforms input data that is input to a downstream machine learning system that performs a task, wherein the learning uses a minimization of a summation of a task loss for the task and energy consumed for the downstream machine learning system of Lai with the machine learning system that performs an inferencing task by using activations of zeros and non-zeros and by using a post-activation density through a minimization of a summation of a task loss for the task and a post-activation density loss for the post-activation density wherein the minimization that includes the post-activation density loss reduces a number of non-zero activation values of Kurtz in order to utilize Lai’s method for input data transformation (Lai, Abstract, “The method can include outputting, by the first one or more layers of the neural network implemented on the first device, a data set that is reduced in size relative to the input data while identifying one or more features of the input data for processing by a second one or more layers of the neural network”) in order to achieve significant inference speedups with maintained or even improved accuracy (Kurtz, Pg. 1, Col. 1, Abstract, “We examine the impact of our methods on popular image classification models, showing that most architectures can adapt to significantly sparser activation maps without any accuracy loss . . . these compression gains can be translated into inference speedups . . . to enable fast convolution operations over networks with sparse activations, and . . . enable significant speedups for end-to-end inference on a range of popular models on the large-scale ImageNet image classification task on modern Intel CPUs, with relatively low retraining cost”; Kurtz, Pg. 9, Col. 2, Para. 4, “Our framework is particularly well-suited for speeding-up inference on accurate, but heavy, deep and wide models”; Kurtz, Pg. 2, Col. 1, Para. 2, “The methods we propose induce negligible (< 0.3%) accuracy loss on ImageNet-scale models, and can even result in minor accuracy increase . . . we find that sharp thresholds exist for activations, beyond which accuracy drops, and cannot be recovered. This observation simplifies the fine tuning process, since, up to this threshold, we are usually able to recover full accuracy”), which can be efficiently implemented on modern hardware (Kurtz, Pg. 6, Col. 1, “the location of the non-zero elements is not known until inference time, and so we must be able to efficiently compress the activations at run time. Once compressed, we can apply Algorithm 1 to the compressed input. Importantly, both CSR compression and sparse-input convolution can be implemented efficiently on modern hardware, i.e. without the need to branch on zero elements”), and where it was well-established in the art before the effective filing date of the invention that input data can be altered to change a model’s post activation density while maintaining the model’s accuracy (Krithivasan et al., “Sparsity Turns Adversarial: Energy and Latency Attacks on Deep Neural Networks”, Pg. 4129, Col. 1, Abstract, “we propose and demonstrate adversarial sparsity attacks, which modify a DNN’s inputs so as to reduce sparsity (or the incidence of zeros) in its internal activation values . . . To prevent easy detection of the attack, we further ensure that the perturbation magnitude is within a specified constraint and that the perturbation does not affect classification accuracy”; Shumailov et al., “Sponge Examples: Energy-Latency Attacks on Neural Networks”, Pg. 215, Para. 3, “inputs that lead to less sparse activations will increase the number of operations and the number of memory accesses, and thus energy consumption”).
Regarding Claim 2, Lai in view of Kurtz teaches the computer-implemented method of claim 1, wherein the downstream machine learning system is implemented with a neural network (Lai, Para. [0039], “The second device can include a second neural network”; see also Lai, Fig. 3A, where the “Second Device 302b” is downstream of the “First Device 302a” because the “Second Device 302b” receives the “Reduced Data Set 310” from the “First Device 302a” as downstream input).
Regarding Claim 8, Lai teaches a computer program product comprising: a computer readable medium; and computer executable instructions stored on the computer readable medium to perform the method operations . . . (Para. [0150], “The present disclosure contemplates methods, systems and program products on any machine-readable media for accomplishing various operations. The embodiments of the present disclosure may be implemented using existing computer processors, or by a special purpose computer processor for an appropriate system, incorporated for this or another purpose, or by a hardwired system. Embodiments within the scope of the present disclosure include program products comprising machine-readable media for carrying or having machine-executable instructions or data structures stored thereon”).
The remaining limitations are substantially the same as the limitations of Claim 1, therefore it is rejected under the same rationale.
Regarding Claim 9, the additional elements of the dependent claim are substantially the same as the limitations of Claim 2, therefore it is rejected under the same rationale.
Regarding Claim 14, Lai teaches an apparatus, comprising: a memory; at least one processor, coupled to said memory; and program instructions stored on the memory for causing the at least one processor to perform operations . . . (Para. [0075], “Implementations of the present disclosure can be realized in a variety of apparatus including electronic devices implemented using any combination of circuitry and software”; Para. [0074], “Some implementations include electronic components, such as microprocessors, storage and memory that store computer program instructions in a computer readable storage medium. Many of the features described in this specification can be implemented as processes that are specified as a set of program instructions encoded on a computer readable storage medium. When these program instructions are executed by one or more processors, they cause the processors to perform various operation indicated in the program instructions”; and Para. [0150], “The present disclosure contemplates methods . . . implemented using . . . machine-readable media [, which] can comprise RAM, ROM, EPROM, EEPROM, or other optical disk storage . . . which can be accessed by a . . . machine with a processor”).
The remaining limitations are substantially the same as the limitations of Claim 1, therefore it is rejected under the same rationale.
Regarding Claim 15, the additional elements of the dependent claim are substantially the same as the limitations of Claim 2, therefore it is rejected under the same rationale.
Regarding Claim 21, Lai in view of Kurtz teaches the computer-implemented method of claim 1, wherein the downstream machine learning system performing the task with respect to the transformed new input data (Lai, Fig. 3B and Lai, Para. [0145] – [0146], “Process 350 includes providing the reduced data set to the second device (step 362) . . . Process 350 includes identifying one or more features of the reduced data set using . . . a second neural network and the reduced data set . . . the reduced data set is output by the first device (e.g., by the first neural network . . . ) and provided to the neural network . . . of the second device as an input”, where the transformed new input data, “the reduced data set”, is input into the downstream machine learning system, “provided to the neural network . . . of the second device as an input”, such that the downstream machine learning system performs the task as an inferencing task with respect to the transformed new input data, “identifying one or more features of the reduced data set using . . . a second neural network and the reduced data set”; see also Lai, Para. [0140], “The process 350 includes receiving input data at a first device . . . In some embodiments, the input data is received from sensors, cameras, etc., of the HMD and/or the AR system” and Lai, Para. [0046], “The neural network(s) 114 can be deployed or used to perform data (e.g., image, audio, video) processing, object or feature recognition, recommender functions, data or image classification, data (e.g., image) analysis, etc., such as natural language processing”)
includes skipping one or more instructions in response to an activation having an activation value of zero (Kurtz, Pg. 6, Col. 1, Para. 2, “the location of the non-zero elements is not known until inference time, and so we must be able to efficiently compress the activations at run time. Once compressed, we can apply Algorithm 1 to the compressed input. Importantly, both CSR compression and sparse-input convolution can be implemented efficiently on modern hardware, i.e. without the need to branch on zero elements”, where, in view of Kurtz, performing the task, “inference time”, includes skipping one or more instructions in response to an activation having an activation value of zero, “implemented efficiently on modern hardware, i.e. without the need to branch on zero elements”).
The reasons for obviousness were discussed in regard to the rejection of claim 1 above and remain applicable here.
Regarding Claim 22, Lai in view of Kurtz teaches the computer-implemented method of claim 21, wherein the downstream machine learning system includes convolution layers (Lai, Para. [0040], “The first device can generate a reduced data set based on the input data and provide the reduced data set to the second device as an input to the neural network on the second device” and Lai, Para. [0046], “A neural network 114 of the AI accelerator 108 can include any type of neural network including, for example, a convolution neural network (CNN)”, where the downstream machine learning system, “the neural network on the second device”, can be a “a convolution neural network (CNN)”, which includes convolution layers; see also Kurtz, Pg. 2, Col. 1, Para. 1, “We begin by performing an in-depth analysis of regularization and thresholding methods as a way to increase activation map sparsity in convolutional neural networks”; Kurtz, Pg. 5, Col. 1, Para. 1, “We first aim to find layer wise activation thresholds which sparsify a large fraction of the activations preserving accuracy”; and Kurtz, Pg. 5, Col. 1, Fig. 2, “the sensitivity analysis results for a single convolutional layer in ResNet18, for different threshold values”)
and is part of a hardware accelerator that performs matrix multiplication (Lai, Fig. 1A; Lai, Para. [0058], “the AI accelerator 108 can include one or more processors 124. The one or more processors 124 can include any logic, circuitry and/or processing component (e.g., a microprocessor) for pre-processing input data for any one or more of the neural network(s) 114”; Lai, Fig. 1C; and Lai, [0069], “An array or plurality of MAC units 140 (e.g., in PEs) can be arranged in a systolic array, for parallel integration, convolution, correlation, matrix multiplication, data sorting, and/or data analysis tasks”, where, as shown in Fig. 1A, the “one or more of the neural network(s) 114”, which includes the downstream machine learning system, are part of a hardware accelerator, “the AI accelerator 108 can include one or more processors 124”, and where, as shown in Fig. 1C, the “AI Accelorator(s) 108” comprise “MAC Unit(s) 140”, which are for “matrix multiplication”; see also Kurtz, Pg. 7, Col. 2, Para. 6, “We implemented our sparse-input convolution in C++, on top of an existing fully-dense baseline framework, which uses optimized direct convolution or general matrix multiply (GeMM) operations for all layers”; see generally Kurtz, Pg. 1, Col. 2, Para. 2, “A non-trivial fraction of the activations are sparse as a natural consequence of the structure of Rectified Linear Unit (ReLU) activation functions. This observation has been leveraged by hardware accelerators”).
The reasons for obviousness were discussed in regard to the rejection of claim 1 above and remain applicable here.
Regarding Claim 23, Lai in view of Kurtz teaches the computer-implemented method of claim 1, wherein the downstream machine learning system performing the task with respect to the transformed new input data (Lai, Fig. 3B and Lai, Para. [0145] – [0146], “Process 350 includes providing the reduced data set to the second device (step 362) . . . Process 350 includes identifying one or more features of the reduced data set using . . . a second neural network and the reduced data set . . . the reduced data set is output by the first device (e.g., by the first neural network . . . ) and provided to the neural network . . . of the second device as an input”, where the transformed new input data, “the reduced data set”, is input into the downstream machine learning system, “provided to the neural network . . . of the second device as an input”, such that the downstream machine learning system performs the task as an inferencing task with respect to the transformed new input data, “identifying one or more features of the reduced data set using . . . a second neural network and the reduced data set”; see also Lai, Para. [0140], “The process 350 includes receiving input data at a first device . . . In some embodiments, the input data is received from sensors, cameras, etc., of the HMD and/or the AR system” and Lai, Para. [0046], “The neural network(s) 114 can be deployed or used to perform data (e.g., image, audio, video) processing, object or feature recognition, recommender functions, data or image classification, data (e.g., image) analysis, etc., such as natural language processing”)
includes skipping one or more computations of an instruction in response to an activation having an activation value of zero (Kurtz, Pg. 6, Col. 1, Para. 2, “the location of the non-zero elements is not known until inference time, and so we must be able to efficiently compress the activations at run time. Once compressed, we can apply Algorithm 1 to the compressed input. Importantly, both CSR compression and sparse-input convolution can be implemented efficiently on modern hardware, i.e. without the need to branch on zero elements”, where, in view of Kurtz, performing the task, “inference time”, includes skipping one or more computations of an instruction in response to an activation having an activation value of zero, “implemented efficiently on modern hardware, i.e. without the need to branch on zero elements”).
The reasons for obviousness were discussed in regard to the rejection of claim 1 above and remain applicable here.
Regarding Claim 24, Lai in view of Kurtz teaches the computer-implemented method of claim 23, wherein the skipping of the one or more computations of the instruction occurs via use of an element-wise mask (Kurtz, Pg. 6, Col. 1, “the location of the non-zero elements is not known until inference time, and so we must be able to efficiently compress the activations at run time. Once compressed, we can apply Algorithm 1 to the compressed input. Importantly, both CSR compression and sparse-input convolution can be implemented efficiently on modern hardware, i.e. without the need to branch on zero elements” and Kurtz, Pg. 6-7, Col. 2-1, “Because we must compress our input data at runtime, we also require an efficient algorithm to compress a matrix M to CSR format. This can be done as follows: given a SIMD vector, v, of 16 floats, which we want to compress, we use the vcmp instruction to identify the locations of the non-zero elements in v stored in a mask register m. Then use the vcompress instruction twice: once applied to v with mask m to produce contiguous non zero elements to be written to values, and a second time applied to the vector {j,...,j + 15} with mask m (where j is the column index of the first element of v in M) to produce column indices to be written to columns. The popcnt instruction applied to m can be used to keep track of the number of non-zero elements and thereby maintain the offset for writing to values and columns, as well as to record row pointers[i]”, wherein the skipping of the one or more computations of the instruction, “implemented efficiently on modern hardware, i.e. without the need to branch on zero elements”, occurs via use of an element-wise mask, “the locations of the non-zero elements in v stored in a mask register m”).
The reasons for obviousness were discussed in regard to the rejection of claim 1 above and remain applicable here.
Claims 3-7, 10-13, and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Lai in view of Kurtz, and Park et al. (hereinafter Park) (“Big/Little Deep Neural Network for Ultra Low Power Inference”).
Regarding Claim 3, Lai in view of Kurtz teach the computer-implemented method of claim 2, wherein the learning of the input transformation function comprises balancing the task loss and the post-activation density loss of the neural network at inference time . . . (Lai, Para. [0039], “implementing multiple neural networks . . . methods can be implemented to optimize neural network computations for energy savings by splitting the computations across multiple devices . . . the accuracy and one or more corresponding threshold values to determine if the second device is required for further analysis of the input data”; where the “threshold” balances accuracy and energy consumed, which, in view of Kurtz, is balancing of the task loss and the post-activation density loss of the neural network, see Kurtz, Pg. 1, Col. 1, Abstract, “we present an analysis of methods for maximizing the sparsity of the activations in a trained neural network . . . To induce highly sparse activation maps without accuracy loss, we introduce a new regularization technique, coupled with a new threshold-based sparsification method based on a parameterized activation function called Forced Activation-Threshold Rectified Linear Unit (FATReLU). We examine the impact of our methods on popular image classification models . . . and show that it can enable significant speedups for end-to-end inference on a range of popular models” and Kurtz, Pg. 2, Col. 1, Para. 2, “The methods we propose induce negligible (< 0.3%) accuracy loss on ImageNet-scale models, and can even result in minor accuracy increase . . . we find that sharp thresholds exist for activations, beyond which accuracy drops, and cannot be recovered. This observation simplifies the fine tuning process, since, up to this threshold, we are usually able to recover full accuracy”, at inference time, Kurtz, Pg. 6, Col. 1, Para. 2, “the location of the non-zero elements is not known until inference time, and so we must be able to efficiently compress the activations at run time”).
The reasons for obviousness have been noted in the rejection of Claim 1 above and remain applicable here.
Lai in view of Kutz does not teach . . . based on a specified tradeoff factor . . . .
However, Park teaches . . . [balancing at inference time] based on a specified tradeoff factor . . . (Pg. 127, Para. 8, “The hyper-parameter [lambda] enables us to make a trade-off between energy consumption and inference accuracy”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the learning based on minimizing task-specific loss and post-activation density loss during inferencing, to develop an input transformation function for use by a neural network of Lai in view of Kurtz, with balancing based on a specified tradeoff factor of Park to allow for the development, selection, and use of a multiple machine learning systems to meet the specific needs of a user (Park, Pg. 130, Para. 2, “our current methodology . . . design is to explore all the possible combinations . . . and select the best one which gives the maximum reduction in energy consumption under the given limit of accuracy loss”).
Regarding Claim 4, Lai in view of Kurtz and Park teach the computer-implemented method of claim 1, further comprising repeating the learning of the input transformation function (Park, Pg. 126, Para. 5, “We assume that each [neural network] is trained by supervised learning”, where the neural network is the input transformation function; Park, Pg. 129, Para. 5, “We give the results of 10 runs”)
at different specified tradeoff factors to generate additional input transformation functions (Park, Fig. 7; Park, Pg. 129, Para. 3, “Figure 7 shows the relationship between energy consumption and inference accuracy which is obtained by varying the static threshold from 0 to 1”; Park, Pg. 127, Para. 8, “The hyper-parameter [lambda] enables us to make a trade-off between energy consumption and inference accuracy”; Park, Pg. 127, Equation 4, where “lambda” tradeoff factors are used to calculate “static threshold[s]”, which differentiates the input transformation function neural networks),
wherein the transforming of the new input data using the learned input transformation function (Lai, Fig. 3B and Lai, Para. [0141], “The process 350 includes generating a reduced data set using . . . a first neural network and the input data . . . In some embodiments, step 354 is performed by the neural network 114a using the input data received or obtained in step 352 (e.g., the input data 110)”, where the new input data is transformed using the learned input transformation function, “performed by the neural network 114a using the input data received”, such that transformed new input data is generated, “generating a reduced data”)
is performed using a selected one of the additional input transformation functions corresponding to a selected tradeoff factor of the different specified tradeoff factors (Park, Pg. 130, Para. 2, “our current methodology . . . design is to explore all the possible combinations . . . and select the best one”; Park, Pg. 130, Figure 7, where users can view results by threshold in order to select one of the additional input transformation functions, “explore all the possible combinations . . . and select the best one”, corresponding to a selected tradeoff factor, “The hyper-parameter [lambda]”, of the different specified tradeoff factors, “if inference accuracy is more favored than energy consumption, [lambda] needs to become larger”, see Park, Pg. 127, Col. 2, Para. 1, “The hyper-parameter [lambda] enables us to make a trade-off between energy consumption and inference accuracy . . . Thus, if inference accuracy is more favored than energy consumption, [lambda] needs to become larger”).
The reasons for obviousness have been noted in the rejection of Claim 3 above and remain applicable here.
Regarding Claim 5, Lai in view of Kurtz and Park teach the computer-implemented method of claim 4, further comprising obtaining the selection of one of the additional input transformation functions (Park, Pg. 130, Para. 2, “our current methodology . . . design is to explore all the possible combinations of given big and little DNNs and select the best one which gives the maximum reduction in energy consumption under the given limit of accuracy loss”, where selecting a possible combination, which includes the “little DNN” input transformation function, which are additional for “the possible combinations” based on considerations such as “if inference accuracy is more favored than energy consumption, [lambda] needs to become larger”, see Park, Pg. 127, Col. 2, Para. 1, “The hyper-parameter [lambda] enables us to make a trade-off between energy consumption and inference accuracy . . . Thus, if inference accuracy is more favored than energy consumption, [lambda] needs to become larger”, in a manner that “gives” the desired result indicates the selection was obtained).
The reasons for obviousness have been noted in the rejection of Claim 3 above and remain applicable here.
Regarding Claim 6, Lai in view of Kurtz and Park teach the computer-implemented method of claim 5, wherein the obtained selection is one of: the selected tradeoff factor, an energy level, an energy saving level, an identification of one of the additional input transformation functions, and an accuracy level (Park, Pg. 130, Para. 2, “As our future work, we will work on design methodologies for the selection of good little DNN (offering large energy reduction while satisfying the given limit of accuracy loss”, where the “little DNN” is one of the additional input transformation functions and “satisfying the given limit of accuracy loss” indicts the obtained selection is an accuracy level).
The reasons for obviousness have been noted in the rejection of Claim 3 above and remain applicable here.
Regarding Claim 7, Lai in view of Kurtz and Park teach the method of claim 4, wherein, in the learning, minimizing the task loss ensures satisfactory accuracy for the inferencing task (Lai, Para. [0039], “The first device may use . . . the accuracy . . . to determine whether or not it can make a sufficiently accurate determination without using the second device”, where sufficient accuracy is ensured, which is reasonably understood as minimizing accuracy loss, through use of “the second device[‘s]” neural network when accuracy is insufficient, see also Kurtz, Pg. 1, Col. 1, Abstract, “we present an analysis of methods for maximizing the sparsity of the activations in a trained neural network . . . To induce highly sparse activation maps without accuracy loss, we introduce a new regularization technique, coupled with a new threshold-based sparsification method based on a parameterized activation function called Forced Activation-Threshold Rectified Linear Unit (FATReLU). We examine the impact of our methods on popular image classification models . . . and show that it can enable significant speedups for end-to-end inference on a range of popular models” and Kurtz, Pg. 2, Col. 1, Para. 2, “The methods we propose induce negligible (< 0.3%) accuracy loss on ImageNet-scale models, and can even result in minor accuracy increase . . . we find that sharp thresholds exist for activations, beyond which accuracy drops, and cannot be recovered. This observation simplifies the fine tuning process, since, up to this threshold, we are usually able to recover full accuracy”, where minimizing the task loss ensures satisfactory accuracy, “up to this threshold, we are usually able to recover full accuracy”, for an inferencing task at inference time, see Kurtz, Pg. 6, Col. 1, Para. 2, “the location of the non-zero elements is not known until inference time, and so we must be able to efficiently compress the activations at run time”)
while minimizing the post-activation density loss ensures that a density of post-activations will be sufficiently small to attain reduced energy consumption (Lai, Para. [0039], “The systems and methods can be implemented to optimize neural network computations for energy savings by splitting the computations across multiple devices”, where energy consumed in the second device is “optimize[d]”, which is reasonably understood as minimized, which reasonably ensures it, given the inherit tradeoff with accuracy that also must be ensured; see also Kurtz, Pg. 1, Col. 1, Abstract, “we present an analysis of methods for maximizing the sparsity of the activations in a trained neural network . . . To induce highly sparse activation maps without accuracy loss, we introduce a new regularization technique, coupled with a new threshold-based sparsification method based on a parameterized activation function called Forced Activation-Threshold Rectified Linear Unit (FATReLU). We examine the impact of our methods on popular image classification models . . . and show that it can enable significant speedups for end-to-end inference on a range of popular models” and Kurtz, Pg. 6, Col. 1, Para. 2, “the location of the non-zero elements is not known until inference time, and so we must be able to efficiently compress the activations at run time. Once compressed, we can apply Algorithm 1 to the compressed input. Importantly, both CSR compression and sparse-input convolution can be implemented efficiently on modern hardware, i.e. without the need to branch on zero elements”, where a person of ordinary skill in the art would understand that minimizing the post-activation density loss, “maximizing the sparsity of the activations”, ensures that a density of post-activations will be sufficiently small to attain reduced energy consumption through, in part, “efficiently on modern hardware, i.e. without the need to branch on zero elements”; see also Kurtz, Pg. 9, Col. 1, Para. 3, “The second noticeable trend is that Dynamic Boosting can consistently reduce the density of activations without accuracy loss”),
and wherein each specified tradeoff factor serves to balance the task loss and the post-activation density loss to enable a tradeoff between an accuracy level and the energy consumption (Park Pg. 127, Para. 8, “The hyper-parameter [lambda] enables us to make a trade-off between energy consumption and inference accuracy”, wherein each specified tradeoff factor, “hyper-parameter [lambda]”, enable a tradeoff between an accuracy level and the energy consumption, “enables us to make a trade-off between energy consumption and inference accuracy”, which as discussed above, serves to balance the task loss and the post-activation density loss to, see Kurtz, Pg. 1, Col. 1, Abstract, “we present an analysis of methods for maximizing the sparsity of the activations in a trained neural network . . . To induce highly sparse activation maps without accuracy loss, we introduce a new regularization technique, coupled with a new threshold-based sparsification method based on a parameterized activation function called Forced Activation-Threshold Rectified Linear Unit (FATReLU). We examine the impact of our methods on popular image classification models . . . and show that it can enable significant speedups for end-to-end inference on a range of popular models”).
The reasons for obviousness were discussed in regard to the rejection of Claim 1 above, for the combination of Lai with Kurtz, and in regard to the rejection of Claim 3 above, for the combination of Lai and Kurtz with Park, and remain applicable here.
Regarding Claim 10, the additional elements of the dependent claim are substantially the same as the limitations of Claim 3, therefore it is rejected under the same rationale.
Regarding Claim 11, the additional elements of the dependent claim are substantially the same as the limitations of Claim 4, therefore it is rejected under the same rationale.
Regarding Claim 12, the additional elements of the dependent claim are substantially the same as the limitations of Claim 5, therefore it is rejected under the same rationale.
Regarding Claim 13, the additional elements of the dependent claim are substantially the same as the limitations of Claim 6, therefore it is rejected under the same rationale.
Regarding Claim 16, the additional elements of the dependent claim are substantially the same as the limitations of Claim 3, therefore it is rejected under the same rationale.
Response to Arguments
Applicant's arguments filed on 02/27/2026 have been fully considered. Each argument is addressed in detail below.
I. Applicant argues the rejections of the claims, under 35 U.S.C. § 101, should be withdrawn (Applicant’s Remarks, 02/27/2026, Pg. 10-14, section “REJECTION UNDER 35 U.S.C. § 101”).
Applicant’s amendments to the claims have overcome the previously communicated rejections of the claims, under 35 U.S.C. § 101, as having been directed to abstract ideas without significantly more. As a result, these rejections have been withdrawn.
However, as discussed in detail above, Applicant’s amendments to claims 8-13 render their subject matter ineligible for patenting because they no longer fall within at least one of the four categories of patent eligible subject matter. A detailed discussion of this issue is provided above.
II. Applicant argues the rejections of the claims, under 35 U.S.C. § 103, should be withdrawn (Applicant’s Remarks, 02/27/2026, Pg. 14-18, section “REJECTION UNDER 35 U.S.C. § 103”).
In response to Applicant’s amendments, the previously communicated rejections under 35 U.S.C. § 103, have been withdrawn. However, Applicants arguments are not persuasive in light of the new grounds for rejection, under 35 U.S.C. § 103, discussed in detail above. The new grounds of rejection rely on new prior art of record to teach the new combinations of elements in the amended claims, which were not presented in these arrangements in any of the previously presented claims. As a result, Applicant arguments against the previously communicated rejections under 35 U.S.C. § 103 are rendered moot.
However, for clarity of the record and in the interest of compact prosecution, arguments still relevant to the new grounds of rejection are discussed below.
Specifically, Applicant’s arguments against the teachings of Lai are insufficient because they amount to attacking a reference individually where the rejection is based on a combination of references (see MPEP 2145, “One cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references”).
As a result, the arguments are not persuasive.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MATTHEW BRYCE GOLAN whose telephone number is (571)272-5159. The examiner can normally be reached Monday through Friday, 8:00 AM to 5:00 PM ET.
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, Alexey Shmatov can be reached at (571) 270-3428. 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.
/MATTHEW BRYCE GOLAN/Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
