DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
The amendment filed 03/23/2026 has been entered. As directed, claims 1, 10 and 19 have
been amended, no claim have been canceled or added. Thus claims 1-19 and 21 remain pending in the application. The applicant's amendments to the claims fail to overcome the rejection under 35 U.S.C. § 112(b) set forth in the Non-final Office Action mailed 12/23/2025.
Response to Arguments
With respect to the Applicant’s argued rejection under 35 U.S.C. § 112(b) in “Applicant Arguments/Remarks Made in an Amendment”:
Applicant argues:
…
Applicant submits that the computational graph model is a neural network comprised in the executing network node or cloud. (See Paragraph [0042] of the Specification). The computational graph model acts as a parent model (See Paragraph [0030] of the Specification). The intermediate network node either obtains the imitation model from the computational graph model or builds the imitation model based on the input parameters from the requesting node. (See Paragraphs [0029] and [0063] of the Specification).
Thus, based on the above cited text, the imitation model can be understood as a limited form of a Computational Graph Model-specifically, a Convolutional Neural Network (CNN) (see Para. [0026] of the Specification), operating either in the cloud or at an executing network node. The imitation model is deployed within an intermediate network and is either constructed by that network using input parameters received from a requesting node or derived from a parent model, such as a Computational Graph Model (CNN) running in the cloud or at an executing network node.
(see Response filed 03/23/2026 [page 8]).
Applicant’s arguments have been fully considered but are not persuasive. Although, based on the instant specification, the imitation model may be understood as a limited form of the computational graph model, such as a CNN, the claim does not clearly recite that relationship. For example, claim 1 recites “storing an imitation model that is executed by the at least one processor to execute a Convolutional Neural Network (CNN) for object detection, wherein the object detection through CNN for object detection … wherein the imitation model is a limited version of the computational graph model” which does not clearly specify whether the imitation model itself is the recited CNN, whether the computational graph model corresponds to the recited CNN, or whether the imitation model is merely used to execute or approximate the CNN. Thus, the claim fails to particularly point out and distinctly claim the relationship among the imitation model, the CNN, and the computational graph model. Claims 10 and 19 recite similar limitations regarding an imitation model, a computational graph model, and a CNN, and are rejected under 35 U.S.C. § 112(b) for the same reason. Therefore, the rejection under 35 U.S.C. § 112(b) is maintained.
With respect to the Applicant’s argued rejection under 35 U.S.C 101 in “Applicant Arguments/Remarks Made in an Amendment”:
Applicant argues:
…
Amended independent Claim 1 recites steps performed by an intermediate network node for object detection in a communication network such as obtaining and storing an imitation model, execution of the stored imitation model to execute a CNN for object detection involving feature learning by creating one or more activation maps from an input image and classification of an input image by converting pooled activation maps of preceding layers into a N dimension vector, which are not mathematical calculations, but steps performed by a processor for object detection.
In particular, amended independent Claim 1 recites steps to provide: (a) an imitation model closer to the edge of the communication network, (b) provide low latency and real-time imitation models, (c) cost-efficient wireless communication network. Thus, Applicant submits that the claimed disclosure showcases the ability to improve overall efficiency of communication systems and user experience.
Further, Applicant respectfully submits that the structure of intermediate network node in amended independent Claim 1 in connection with core network equipped with computational graph model, to provide a method for real-time handling of data of communication network in a technically efficient and reliable manner.
Applicant further submits that, in accordance with the MPEP § 2106.04(a)(2), subsection I, the above subject matter does not constitute mathematical concept rather it indicates steps performed by an intermediate network node while executing an imitation model involving execution of CNN or object detection in an image. (The enumerated groupings of abstract ideas are defined as: 1) Mathematical concepts - mathematical relationships, mathematical formulas or equations, mathematical calculations (see MPEP § 2106.04(a)(2), subsection I))
Referring to MPEP § 2106.04(a)(2) Abstract Idea Groupings Subsection I, states that "When determining whether a claim recites a mathematical concept (i.e., mathematical relationships, mathematical formulas or equations, and mathematical calculations), examiners should consider whether the claim recites a mathematical concept or merely limitations that are based on or involve a mathematical concept. A claim does not recite a mathematical concept (i.e., the claim limitations do not fall within the mathematical concept grouping), if it is only based on or involves a mathematical concept. See, e.g., Thales Visionix, Inc. v. United States, 850 F.3d 1343, 1348-49, 121 USPQ2d 1898, 1902-03 (Fed. Cir. 2017). For example, a limitation that is merely based on or involves a mathematical concept described in the specification may not be sufficient to fall into this grouping, provided the mathematical concept itself is not recited in the claim.
Further, Referring to MPEP § 2111.01(II) Plain Meaning Subsection II, states that "Though understanding the claim language may be aided by explanations contained in the written description, it is important not to import into a claim limitations that are not part of the claim. For example, a particular embodiment appearing in the written description may not be read into a claim when the claim language is broader than the embodiment." Superguide Corp. v. DirecTV Enterprises, Inc., 358 F.3d 870, 875, 69 USPQ2d 1865, 1868 (Fed. Cir. 2004)... In Zletz,supra, the examiner and the Board had interpreted claims reading "normally solid polypropylene" and "normally solid polypropylene having a crystalline polypropylene content" as being limited to "normally solid linear high homopolymers of propylene which have a crystalline polypropylene content." The court ruled that limitations, not present in the claims, were improperly imported from the specification. See also In re Marosi, 710 F.2d 799, 802, 218 USPQ 289, 292 (Fed. Cir. 1983) ("'[C]laims are not to be read in a vacuum, and limitations therein are to be interpreted in light of the specification in giving them their 'broadest reasonable interpretation."' (quoting In re Okuzawa, 537 F.2d 545, 548, 190 USPQ 464, 466 (CCPA 1976)). The court looked to the specification to construe "essentially free of alkali metal" as including unavoidable levels of impurities but no more.)"
Thus, from the above cited guidelines (from MPEP § 2106.04(a)(2) and MPEP § 2111.01(11)) Applicant submits that, the limitations cited by examiner (from Pages 5, 8, and 11 of the specification of the pending application) does not constitute the current application to be falling into the mathematical concepts grouping as the cited text is obtained from specification of the patent and is not reflected directly in the claims. Accordingly, requests the examiner to withdraw the rejections under 35 U.S.C. 101.
(see Response filed 03/23/2026 [pages 9-12]).
Applicant’s arguments have been fully considered but are not persuasive. Applicant argues that the rejection improperly relies on mathematical descriptions from the specification rather than the claim language. However, the rejection does not import unclaimed limitations from the specification. Claim 1 expressly recites CNN object detection including feature learning, creation of activation maps, conversion of pooled activation maps of preceding layers into an N-dimensional vector, and values of the vector representing probabilities. The specification is used only to interpret these expressly recited CNN terms under their broadest reasonable interpretation. This interpretation is consistent with Example 47 discussed in the “2024 Guidance Update on Patent Subject Matter Eligibility, Including on Artificial Intelligence,” where the claim recites training an ANN using back propagation and gradient descent, and the analysis determines that the recited training limitation includes mathematical concepts because the specification/background explains that backpropagation and gradient descent are mathematical calculations. Similarly, the instant specification explains that the recited CNN feature learning and classification operations include numerical matrices, multiplication/dot-product operations, pooling/down-sampling, vector transformation and probability values. Therefore, the claim recites operations of object detection through CNN are not merely “based on” or “involve” a mathematical concept. Rather, under BRI in light of the specification, the claim itself recites CNN operation can be considered as represent a mathematical concept including mathematical calculations and mathematical relationships.
Further, Applicant’s argument that the limitation are “… steps performed by a processor for object detection” is not persuasive. Mathematical calculations do not cease to be mathematical concepts because they are performed by a processor or used for object detection. The processor executes the recited CNN object detection operations. Therefore, the recited operations are reasonably interpreted as mathematical concepts under MPEP §2106.04(a)(2). The processor execute the CNN for object detection, which is perform the mathematical concepts.
Additionally, the alleged improvement or benefit (e.g., low latency, real-time operation, cost efficiency, and improvement to communication-system efficiency) only reflect to the abstract idea itself. As explained in MPEP 2106.05(a), II.: "it is important to keep in mind that an improvement in the abstract idea itself (e.g. a recited fundamental economic concept) is not an improvement in technology."
Therefore, the claim limitation is a “mathematical concept”, similar to the comparison steps in MPEP 2106.04(a)(2)(I), and rejection under 35 U.S.C. 101, Step 2A, Prong One is maintained.
With respect to the Applicant’s argued rejection under 35 U.S.C 101 in “Applicant Arguments/Remarks Made in an Amendment”:
Applicant argues:
…
Further, regarding prong two of Step 2A of the 2019 Revised Patent Subject Matter Eligibility Guidance, even if one were to arrive at a conclusion satisfying the Prong One of such
analysis, assuming arguendo, to which Applicant does not concede, Applicant submits the alleged abstract idea is integrated into practical implementation. For example, amended independent Claim 1 recites "[a]n intermediate network node ...comprising: at least one processor; at least one memory connected to the at least one processor and storing an imitation model ... to execute a Convolutional Neural Network (CNN) for object detection, wherein the object detection through CNN comprises performing feature learning and classification of an input image, wherein the feature learning comprises creation of one or more activation maps and the classification process comprises conversion of a pooled activation maps of preceding layers into a N dimension vector, ..."
According to Applicant's disclosure, for example, "[tlo understand environment such as images, sounds etc. one may use different ways to detect certain event, objects or similar. A way of learning is using machine learning (ML) algorithms to improve accuracy. Computational graph models such as ML models are currently used in different applications... Training of these computational graph models is typically an offline process...and takes several minutes to hours and days, depending on the underlying technology, the capabilities of the infrastructure used for training and the complexity of the computational graph model,...execution of these computational graph models is done anywhere from an edge of the communication network ... In many cases, execution time is critical, i.e. time to obtain an output from the computational graph model, in particular for real-time applications, requiring low-latency feedback loops. A typical scenario is mobile devices in a wireless communication network requesting a decision from a centralized ML model execution environment, e.g. small devices that make use of object detection or augmented reality but have limited resources, e.g. glasses or other wearables, small robots like drones, etc. There are some solutions already proposed including placing computing nodes executing ML models closer to the network edge, i.e. closer to the nodes requesting ML executions...However, placing these computing nodes close to the edge is an expensive task... in order for the detector to identify an object (e.g. a particular radio unit model) in different types of lighting conditions and against different backgrounds, not only a larger training set needs to be used, but also more layers to increase precision. On the other hand, a smaller model (in terms of classes and layers) on the network edge (e.g. close to the base station), would be sufficient to identify the radio unit in question. Thus, the imitation model such as a detection model may be run closer to the edge and have respectable accuracy and real-time performance with cheap commercial off the self (COTS) hardware, shortening control loops, whereas larger object detection models cannot." See Specification as published at Paragraphs [0002]-[0005], [0085], and [0086].
More particularly, Applicant's disclosure points to applications in, for example, mobile phones, drones, Glasses with augmented reality, where quick object detection is critical.
Accordingly, Applicant has shown teaching in the Specification that describes a practical implementation and how the technology is improved and has thus established a clear nexus between the claim language and the practical implementation and improvements to the technology.
Applicant submits that the claimed subject matter results in a solution for object detection by smaller models implemented on the network edge in a communication system, while providing accuracy and real-time performance with cheap commercial off the shelf hardware. By implementing smaller models at the edge of the communication network for object recognition, a reduction of processing time and operational costs, is enabled, while providing accuracy in communication systems.
These features not only provide highly valuable solutions for object detection in drones and mobile phones as well as in augmented reality related applications. (Id. paragraph [0004].)
Further, Applicant submits that the disclosure to which the claim language is directed to is more than a desired result or a network node merely practicing a known process. The claimed subject matter achieves accurate object detection useful for multiple applications including drones, wearables and mobile phones by reciting a small computational graph model with lesser computational resources.
Based on the above, it is submitted that the disclosure provides improved accuracy of object detection in communication systems, and therefore reflects a practical implementation of the alleged judicial exception. Thus, the subject matter of amended independent Claim 1 of the present application recites additional features which integrate the alleged abstract idea into a practical application.
Accordingly, the features of amended independent Claim 1 do not describe an abstract concept, or a concept similar to those found by the Courts to be abstract and recites additional features which integrate the alleged abstract idea into a practical application. At least, for these reasons, Applicant respectfully submits that amended independent Claim 1 meets the standard for patent eligibility under 35 U.S.C. 101.
Further, Applicant respectfully submits that the patent subject matter eligibility guidelines clearly indicate that a patent subject matter is eligible if the Claims do not recite an Abstract Idea and if the claims recite additional elements that integrate the judicial exception into a practical application. See MPEP § 2106.04 for more information on Step 2A.
Referring to the guidelines from MPEP § 2106.04, "... If the additional elements in the claim integrate the recited exception into a practical application of the exception, then the claim is not directed to the judicial exception (Step 2A: NO) and thus is eligible at Pathway B. This concludes the eligibility analysis.") and Pathway B: Claims taken as a whole that fall within a statutory category (Step 1: YES) and are not directed to a judicial exception (Step 2A: NO) are eligible at Pathway B. These claims do not need to go to Step 2B. See MPEP § 2106.04 for more information about this pathway and Step 2A.
Thus, Applicant believes it is not necessary to argue the merits of the invention regarding the rejections under Step 2B of 35 U.S.C. 101 separately.
Accordingly, Applicant further submits that amended independent Claims 10 and 19 recite, inter alia, features similar to those recited in amended independent Claim 1. Accordingly, amended independent Claims 10 and 19 also meet the standard for patent eligibility under 35 U.S.C. 101, allowance of which is respectfully requested.
Applicant respectfully submits that taking all the elements of Claims 1-19 and 21 individually, and in combination, as a whole amount to an eligible subject matter under 35 U.S.C. 101. Therefore, Applicant respectfully requests the rejection of Claims 1-19 and 21 under 35 U.S.C. 101 be withdrawn.
(see Response filed 03/23/2025 [pages 12-16]).
In response to applicant's argument, the examiner disagrees that “the subject matter of amended independent Claim 1 of the present application recites additional features which integrate the alleged abstract idea into a practical application.”
In order to determine if additional element is integrating the abstract idea into a practical application, See MPEP 2106.04(d)(1), “first the specification should be evaluated to determine if the disclosure provides sufficient details such that one of ordinary skill in the art would recognize the claimed invention as providing an improvement. The specification need not explicitly set forth the improvement, but it must describe the invention such that the improvement would be apparent to one of ordinary skill in the art. Conversely, 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. Second, if the specification sets forth an improvement in technology, the claim must be evaluated to ensure that the claim itself reflects the disclosed improvement. That is, the claim includes the components or steps of the invention that provide the improvement described in the specification. The claim itself does not need to explicitly recite the improvement described in the specification (e.g., "thereby increasing the bandwidth of the channel").” In other words, the specification should describe the claimed improvement over the background invention or existing technology, and the claimed improvement should be reflected at least in the additional elements (emphasis added) by specifying how the claimed improvement perform the additional element different from existing technology, functioning of a computer or existing technical field.
However, the additional elements – “… the intermediate network node comprising: at least one processor; at least one memory connected to the at least one processor and storing an imitation model that is executed by the at least one processor to execute a Convolutional Neural Network (CNN)…” and “A non-transitory computer readable storage medium including instructions, which, when executed on at least one processor of an intermediate network node cause the intermediate network node to execute a Convolutional Neural Network (CNN) for object detection operations comprising:” which are mere instructions to implement an abstract idea on a computer, or merely uses a computer as tool to perform an abstract idea (see MPEP § 2106.05(f)) with the broad reasonable interpretation in light of specification, which does not integrate a judicial exception into practical application. Further, the additional elements “An intermediate network node in a communication network, the communication network comprises a requesting node and an executing network node, the executing network node comprising a computational graph model,” which is merely adding the words "apply it" (or an equivalent) with the judicial exception, or instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. (See MPEP 2106.05(f)). This additional limitation merely identifies that the recited CNN based mathematical concepts are executed within a communication network environment involving an intermediate node, requesting node, and executing node, but do not recite any particular manner in which the network nodes improve the execution of the computation graph model or otherwise improve computer functionality or network technology. Rather, the limitations are recited at a high level of generality and merely use generic network components as tools to perform the recited abstract idea. Therefore, this additional limitation does not impose any meaningful limitation on practicing the abstract idea. Alternatively, the recited communication network environment merely links the use of the judicial exception to a particular technological environment or field of use. Limiting the recited CNN based mathematical operations to an intermediate network node in a communication network comprising a requesting node and an executing network node does not add significantly more, similar to how limiting the abstract idea in Flook to petrochemical and oil-refining industries was insufficient. (See MPEP § 2106.05(h)).
Applicant’s arguments regarding reduced latency, real-time performance, reduced computational resources, shortened control loops, and lower cost are not persuasive because the alleged benefits arise from the recited mathematical concepts themselves, including the reduced imitation/computational graph model having fewer nodes/edges and requiring fewer computational resources to perform the recited CNN based feature learning and classification operations. Thus, the alleged improvement reflects an improvement in the abstract idea itself, rather than an improvement to computer functionality or network technology. As explained in MPEP 2106.05(a), II.: "it is important to keep in mind that an improvement in the abstract idea itself (e.g. a recited fundamental economic concept) is not an improvement in technology." (emphasis added). Further, the additional elements, including the processor, memory, intermediate network node, requesting node, executing node, and communication network environment, merely provide generic computer/network components for performing the recited mathematical concepts. The additional limitations recite at high level of generality, and do not recite/reflect any particular improvement to the operation of the computer components or communication network. Therefore, even when considered as a whole, the claim merely applies the judicial exception in a generic network environment and does not integrate the judicial exception into a practical application under Step 2A, Pong Two, and nor does it amount to significantly more than the recited judicial exception under Step 2B. Therefore, the rejection under 35 U.S.C. 101 for independent claims 1, 10 and 19 is maintained.
Applicant's arguments filed “Applicant Arguments/Remarks Made in an Amendment,” on
03/23/2026, pages 16-17, have been fully considered but they are not persuasive.
The reference Krizhevsky (ImageNet Classification with Deep Convolutional Neural Networks, published in 2012) teaches the newly amended limitation “wherein each value of the vector represents a probability of an object exists in the input image.” Specifically, Krizhevsky teaches that “The output of the last fully-connected layer is fed to a 1000-way softmax which produces a distribution over the 1000 class labels.” Krizhevsky further explains that the top-5 error rate concerns whether the correct label is among “the five labels considered most probable by the model,” and Figure 4 states that “the probability assigned to the correct label is also shown with a red bar.” (see pages 85 and 89). Thus, Krizhevsky teaches that CNN output includes values representing probabilities associated with respective class/object labels for the input image. Therefore, Sundström US 20200401944 A1 in view of Krizhevsky (ImageNet Classification with Deep Convolutional Neural Networks, published in 2012) and Han (“DEEP COMPRESSION: COMPRESSING DEEP NEURAL NETWORKS WITH PRUNING, TRAINED QUANTIZATION AND HUFFMAN CODING,” published in 2016) combined together teach or suggest claims 1, 10 and 19. Therefore, the rejection of claims 1, 10 and 19, and the claims dependent thereon, under 35 U.S.C. 103 is maintained.
Claim Objections
Claims 1, 10 and 19 are objected to because of the following informalities:
Claim 1 recites “conversion of a pooled activation maps of preceding layers into a N dimension vector,” which should read “conversion of pooled activation maps of preceding layers into an N-dimensional vector”.
Claims 10 and 19 also recite the same limitation are objected to for the same reason.
Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-19 and 21 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second
paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claim 1 recites the limitation " … wherein the object detection through CNN … " which renders the claim indefinite because the relationship among the recited “imitation model,” “computational graph model,” and “Convolutional Neural Network (CNN)” is unclear.
Specifically, Claim 1 recites “ … storing an imitation model that is executed by the at least one processor to execute a Convolutional Neural Network (CNN) for object detection, wherein the object detection through CNN comprises performing feature learning and classification of an input image … wherein the imitation model is a limited version of the computational graph model, …”. However, the claim does not clearly specify whether: 1) The imitation model itself is the recited CNN, 2) the computational graph model corresponds to the recited CNN, or 3) the imitation model is merely used to execute or approximate the CNN. Further, the claim separately recites that the imitation model is a “limited version” of the computational graph model, while also reciting execution of the CNN for object detection, but does not clearly define how these entities are related to one another. As a result, it is unclear what structure or model actually performs the recited object detection operations including the feature learning and classification operations. For the purpose of examination, the limitation is interpreted as reciting that the stored imitation model is executed by the processor to perform object detection through a Convolutional Neural Network (CNN) corresponding to computational graph model, and the limitation model is a limited version of the computational graph model.
Claim 10 recites the limitation “… obtaining an imitation model capable of performing the object detection through CNN …” which renders the claim indefinite because the relationship among the recited “imitation model,” “computational graph model,” and “Convolutional Neural Network (CNN)” is unclear.
Specifically, Claim 10 recites “… the executing network node comprising a computational graph model to execute a Convolutional Neural Network (CNN) for object detection, the method comprising: obtaining an imitation model capable of performing the object detection through CNN … wherein the imitation model is a limited version of the computational graph model …”. However, the claim does not clearly specify whether: 1) The imitation model itself is the recited CNN, 2) the computational graph model corresponds to the recited CNN, or 3) the imitation model is merely used to execute or approximate the CNN. Further, the claim separately recites that the imitation model is a “limited version” of the computational graph model, while also reciting through CNN for object detection, but does not clearly define how these entities are related to one another. As a result, it is unclear what structure or model actually performs the recited object detection operations including the feature learning and classification operations. For the purpose of examination, the limitation is interpreted as reciting that the obtained imitation model performs object detection through a Convolutional Neural Network (CNN) corresponding to computational graph model, and the limitation model is a limited version of the computational graph model.
Claim 19 recites the limitation “… obtaining an imitation model capable of performing the object detection through CNN …” which renders the claim indefinite because the relationship among the recited “imitation model,” computational graph model,” and “Convolutional Neural Network (CNN)” is unclear.
Specifically, Claim 19 recites “cause a computational graph model of the intermediate network node to execute a Convolutional Neural Network (CNN) for object detection operations comprising:
obtaining an imitation model capable of performing the object detection through CNN … wherein the imitation model is a limited version of the computational graph model …”. However, the claim does not clearly specify whether: 1) The imitation model itself is the recited CNN, 2) the computational graph model corresponds to the recited CNN, or 3) the imitation model is merely used to execute or approximate the CNN. Further, the claim separately recites that the imitation model is a “limited version” of the computational graph model, while also reciting through CNN for object detection, but does not clearly define how these entities are related to one another. As a result, it is unclear what structure or model actually performs the recited object detection operations including the feature learning and classification operations. For the purpose of examination, the limitation is interpreted as reciting that the obtained imitation model performs object detection through a Convolutional Neural Network (CNN) corresponding to computational graph model, and the limitation model is a limited version of the computational graph model.
The remaining claims are dependent upon one of the claims listed above and rejected for the same reason.
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.
The claim(s) 1-19 and 21 are rejected under 35 USC § 101 because the claimed invention is directed
to judicial exception an abstract idea, it has not been integrated into practical application and the claims further do not recite significantly more than the judicial exception. Examiner has evaluated
the claims under the framework provided in the 2019 Revised Patent Subject Matter Eligibility Guidance
published in the Federal Register 01/07/2019, as well as subsequent USPTO eligibility guidance updates,
and has provided such analysis below.
Step 1: Are the claims to a process, machine, manufacture or composition of matter?"
Yes, claims 1-9 are directed to intermediate network node and fall within the statutory category of machine;
Yes, claims 10-18 are directed to method and fall within the statutory category of process;
Yes, claims 19 and 21 are directed to non-transitory computer readable storage medium and fall within the statutory category of article of manufacture.
In order to evaluate the Step 2A inquiry "Is the claim directed to a law of nature, a natural phenomenon, or an abstract idea?" we must determine, at Step 2A Prong 1, whether the claim recites a law of nature, a natural phenomenon, or an abstract idea and further whether the claim recites additional elements that integrate the judicial exception into a practical application.
Step 2A Prong 1:
MPEP 2106.4(a)(2)(I): “The mathematical concepts grouping is defined as mathematical relationships, mathematical formulas or equations, and mathematical calculations”.
MPEP 2106.04(a)(2)(I)(A), “A mathematical relationship is a relationship between variables or numbers. A mathematical relationship may be expressed in words or using mathematical symbols.”
Further, MPEP recites: “For example, a step of "determining" a variable or number using mathematical methods or "performing" a mathematical operation may also be considered mathematical calculations when the broadest reasonable interpretation of the claim in light of the specification encompasses a mathematical calculation.
Additionally, the limitation of claim 1, “… an imitation model that is executed by the at least one processor to execute a Convolutional Neural Network (CNN) for object detection, wherein the object detection through CNN comprises performing feature learning and classification of an input image, wherein the feature learning comprises creation of one or more activation maps and the classification process comprises conversion of a pooled activation maps of preceding layers into a N dimension vector, wherein each value of the vector represents a probability of an object exists in the input image, wherein the imitation model is a limited version of the computational graph model, wherein the imitation model is a model requiring less computational resources to converge when compared to the computational graph model, and wherein the imitation model comprises at least one less internal vertex or node and at least one less edge of the computational graph model compared to the computational graph model,” as drafted, under its broadest reasonable interpretation (BRI) in light of specification, can be reasonably considered to represent mathematical concept, as recites in the specification: Page.13, lines 30-35, “an input image can be represented as a matrix of values … The input image may be RGB image and may have 32 pixels height and 32 pixels width, therefore the input image is represented a 32×32×3 matrix. Page.14, lines 4-10, “Filters are multidimensional matrices which have the same depth as the image and are essentially what was described above as neurons. What happens in each step is the process of convolving the receptive field of the image with the values of the filter and writing the result to a new matrix known as activation map. This process involves a multiplication operation of the receptive field data with that of the filter and the application of a weight to this operation (dot product).” Page.14, lines 20-23, “Pooling is another function that down-samples (i.e. “densifies”) the activation map without losing valuable information. The reason for pooling is to reduce computational complexity in subsequent operations. Page.14, lines 32-35, “This layer takes as input the pooled feature maps of the preceding layer and transforms them into a vector of N dimensions (N being the number of classes—or different objects to be detected). Each value of the vector represents a probability that an object exists in the image.” Page.15, lines 1-4, “This would mean that we would convert all activation maps to a two-dimensional vector. For example, if result is [0, 0.8] and first dimension is radio unit and second an antenna, that means that there is 80% probability that an antenna exists in the image. SoftMax is a function used typically for this purpose.” Page.15, lines 16-19, “There are several factors affecting CNN execution, including the number of convolutions, but the matrix multiplications (“convolutions”) when calculating the activation matrices are the ones that are the most computationally expensive and the ones taking the most time.” Page 11, “ … Graphs are data structures that can be ingested by various algorithms, … Algorithms can "embed" each node of a graph into a real vector. The result will be vector representation of each node …” Page.8, “… A subgraph function produces a smaller version of the original computational graph model. A possible way to implement this would be by using an adjacency matrix to represent a neural network such as:
PNG
media_image1.png
126
187
media_image1.png
Greyscale
A subgraph function i.e. the imitation model may be the following where we
intentionally omit the last row and the last column.
PNG
media_image2.png
105
156
media_image2.png
Greyscale
Examiner note: The specification explains that recited CNN feature learning and classification operations as mathematical calculations performed on numerical matrices. For example, the specification states that activation maps are produced by “multiplication” and dot-product” operations between receptive field matrix and filter matrices, and that pooling “down-samples” the activation map, which constitutes a mathematical dimensional reduction operation, and that the classification stage “transforms” pooled activation maps into an “N-dimensional vector,” where each value represents a probability that an object exists in the image. Additionally, the specification explains that the imitation model may be produced as a reduced subgraph of the full computational graph model by removing rows and columns from an adjacency matrix, thereby reducing nodes and edges of the graph. Therefore, these limitations are reasonably interpreted as mathematical concepts, including mathematical relationships and calculations, under MPEP 2106.04(a)(2).
Claims 10 and 19 recite the similar elements as claim 1, and is rejected for the same reasons under 35 U.S.C. 101.
Therefore, claims 1, 10 and 19 recite judicial exceptions. The claims have been identified to recite judicial exceptions, Step 2A Prong 2 will evaluate whether the claims as a whole integrates the exception into a practical application of that exception.
Step 2A Prong 2: Claims 1, 10 and 19: The judicial exception is not integrated into a practical application.
In particular, the claims recite the following additional elements – “… the intermediate network node comprising: at least one processor; at least one memory connected to the at least one processor and storing an imitation model that is executed by the at least one processor to execute …” and “A non-transitory computer readable storage medium including instructions, which, when executed on at least one processor of an intermediate network node cause the intermediate network node to execute a Convolutional Neural Network (CNN) for object detection operations comprising:” which are mere instructions to implement an abstract idea on a computer, or merely uses a computer as tool to perform an abstract idea (See MPEP § 2106.05(f)) with the broad reasonable interpretation, which does not integrate a judicial exception into practical application.
The following additional elements “An intermediate network node in a communication network, the communication network comprises a requesting node and an executing network node, the executing network node comprising a computational graph model,” which is merely adding the words "apply it" (or an equivalent) with the judicial exception, or instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. (See MPEP 2106.05(f)). This additional limitation merely identify that the recited CNN based mathematical concepts are executed within a communication network environment involving an intermediate node, requesting node, and executing node, but do not recite any particular manner in which the network nodes improve the execution of the computation graph model or otherwise improve computer functionality or network technology. Rather, the limitations are recited at a high level of generality and merely use generic network components as tools to perform the recited abstract idea. Therefore, this additional limitation does not impose any meaningful limitation on practicing the abstract idea. Alternatively, the recited communication network environment merely links the use of the judicial exception to a particular technological environment or field of use. Limiting the recited CNN-based mathematical operations to an intermediate network node in a communication network comprising a requesting node and an executing network node does not add significantly more, similar to how limiting the abstract idea in Flook to petrochemical and oil-refining industries was insufficient. (See MPEP § 2106.05(h)).
Therefore, "Do the claims recite additional elements that integrate the judicial exception into a practical application? No, these additional elements do not integrate the abstract idea into a practical application and they do not impose any meaningful limits on practicing the abstract idea. The claims are directed to an abstract idea.
After having evaluated the inquires set forth in Steps 2A Prong 1 and 2, it has been concluded that claims 1, 10 and 19 not only recite a judicial exception but that the claims are directed to the judicial exception as the judicial exception has not been integrated into practical application.
Step 2B: Claims 1, 10 and 19: The claims do not include additional elements, alone or in combination, that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements amount to no more than generic computing components which do not amount to significantly more than the abstract idea. Limitations that the courts have found not to be enough to qualify as "significantly more" when recited in a claim with a judicial exception include:
i. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, e.g., a limitation indicating that a particular function such as creating and maintaining electronic records is performed by a computer, as discussed in Alice Corp., 573 U.S. at 225-26, 110 USPQ2d at 1984 (see MPEP § 2106.05(f));
ii. Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception, e.g., a claim to an abstract idea requiring no more than a generic computer to perform generic computer functions that are well-understood, routine and conventional activities previously known to the industry, as discussed in Alice Corp., 573 U.S. at 225, 110 USPQ2d at 1984 (see MPEP § 2106.05(d));
iii. Adding insignificant extra-solution activity to the judicial exception, e.g., mere data gathering in conjunction with a law of nature or abstract idea such as a step of obtaining information about credit card transactions so that the information can be analyzed by an abstract mental process, as discussed in CyberSource v. Retail Decisions, Inc., 654 F.3d 1366, 1375, 99 USPQ2d 1690, 1694 (Fed. Cir. 2011) (see MPEP § 2106.05(g)); or
iv. Generally linking the use of the judicial exception to a particular technological environment or field of use, e.g., a claim describing how the abstract idea of hedging could be used in the commodities and energy markets, as discussed in Bilski v. Kappos, 561 U.S. 593, 595, 95 USPQ2d 1001, 1010 (2010) or a claim limiting the use of a mathematical formula to the petrochemical and oil-refining fields, as discussed in Parker v. Flook, 437 U.S. 584, 588-90, 198 USPQ 193, 197-98 (1978) (MPEP § 2106.05(h)).
MPEP 2106.05(b)(II) recites “… additional elements that invoke computers or other machinery merely as a tool to perform an existing process will generally not amount to significantly more than a judicial exception. See, e.g., Versata Development Group v. SAP America, 793 F.3d 1306, 1335, 115 USPQ2d 1681, 1702 (Fed. Cir. 2015) (explaining that in order for a machine to add significantly more, it must "play a significant part in permitting the claimed method to be performed, rather than function solely as an obvious mechanism for permitting a solution to be achieved more quickly").
MPEP 2016.05(f) recites “Use of a computer or other machinery in its ordinary capacity for economic or other tasks (e.g., to receive, store, or transmit data) or simply adding a general purpose computer or computer components after the fact to an abstract idea (e.g., a fundamental economic practice or mathematical equation) does not integrate a judicial exception into a practical application or provide significantly more. See Affinity Labs v. DirecTV, 838 F.3d 1253, 1262, 120 USPQ2d 1201, 1207 (Fed. Cir. 2016) (cellular telephone); TLI Communications LLC v. AV Auto, LLC, 823 F.3d 607, 613, 118 USPQ2d 1744, 1748 (Fed. Cir. 2016) (computer server and telephone unit). Similarly, "claiming the improved speed or efficiency inherent with applying the abstract idea on a computer" does not integrate a judicial exception into a practical application or provide an inventive concept. Intellectual Ventures I LLC v. Capital One Bank (USA), 792 F.3d 1363, 1367, 115 USPQ2d 1636, 1639 (Fed. Cir. 2015).
Further, as explained in MPEP 2106.05(h), “ The courts often cite to Parker v. Flook as providing a classic example of a field of use limitation. See, e.g., Bilski v. Kappos, 561 U.S. 593, 612, 95 USPQ2d 1001, 1010 (2010) ("Flook established that limiting an abstract idea to one field of use or adding token postsolution components did not make the concept patentable") (citing Parker v. Flook, 437 U.S. 584, 198 USPQ 193 (1978)) … Although the additional elements did limit the use of the abstract idea, the court explained that this type of limitation merely confines the use of the abstract idea to a particular technological environment (cellular telephones) and thus fails to add an inventive concept to the claims. 838 F.3d at 1259, 120 USPQ2d at 1204.”
Therefore, "Do the claims recite additional elements that amount to significantly more than the judicial exception? No, these additional elements, alone or in combination, do not amount to significantly more than the judicial exception. Having concluded analysis within the provided framework, claims 1, 10 and 19 do not recite patent eligible subject matter under 35 U.S.C. § 101.
Dependent claims 2-9, 11-18 and 21 are also similar rejected under same rationale as cited above wherein these claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. These claims are merely further elaborate the mental process itself (and/or mathematical operations) or providing additional definition of process which does not impose any meaningful limits on practicing the abstract idea. Claims 2-9, 11-18 and 21 are also rejected for incorporating the deficiency of their independent claims 1, 10 and 19.
Claim 2 recites “the imitation model comprises one or more of following compared to the computational graph model: at least one input parameter less than the computational graph model; at least one output parameter less than the computational graph model; and/or one or more functions of less computational complexity.”
This merely further defines imitation model of claim 1 by specifying that the imitation model includes fewer input parameters, fewer output parameters, and/or reduced computational complexity relative to the computational graph model, which are characteristics of mathematical models and algorithms represent as mathematical relationships. Therefore, the claim 2 does not recite patent eligible subject matter under 35 U.S.C. § 101.
Claim 3 recites “the intermediate network node builds the imitation model based on received one or more input parameters from the requesting node. ”
This merely specifies the imitation model can be trained based on received input parameters from the requesting node; therefore, it merely adding insignificant extra-solution activities as data gathering (i.e., received inputs) to the judicial exception (see MPEP § 2106.05(g)) with broadest reasonable interpretation, which does not integrate a judicial exception into practical application. Therefore, the claim 3 does not recite patent eligible subject matter under 35 U.S.C. § 101.
Claim 4 recites “the imitation model is built by removing one or more parts of the imitation model that have not been used within a set interval.”
This merely further defines imitation model of claim 1 by specifying that the imitation model is refined by removing one or more parts that have not been used within a set interval; therefore, it merely adding the words "apply it" (or an equivalent) with the judicial exception, or instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea, and applying a computer component to perform a generic computer function to adjust or refine a mathematical model based on input parameters and usage criteria at high level of generality is simply the to apply a computer to the judicial exception does not integrate a judicial exception into a practical application or provide significantly more - see MPEP 2106.05(f). Therefore, the claim 4 does not recite patent eligible subject matter under 35 U.S.C. § 101.
Claim 5 recites “the intermediate network node obtains the imitation model from the computational graph model.”
This merely specifies the imitation model is trained by computational graph model; therefore, it merely adding insignificant extra-solution activities as data gathering (i.e., received inputs) to the judicial exception (see MPEP § 2106.05(g)) with broadest reasonable interpretation, which does not integrate a judicial exception into practical application. Therefore, the claim 5 does not recite patent eligible subject matter under 35 U.S.C. § 101.
Claim 6 recites “the at least one memory connected to the at least one processor stores program code that is executed by the at least one processor to receive a request from the requesting node, wherein the request comprises one or more input parameters; and determine whether to respond to the request or to forward the one or more parameters towards the executing network node by comparing the one or more input parameters to one or more needed input parameters of the imitation model and/or based on one or more output parameters of the imitation model.”
This merely specifies request is received and determination from imitation model to respond or forward parameters; therefore, it merely adding insignificant extra-solution activities as data gathering (i.e., receive request/parameter) and output data (i.e., forward request/parameter) in conjunction with a law of nature, and merely the words "apply it" (or an equivalent) with the judicial exception, or instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea, and applying a computer component to perform a generic computer function (e.g., data reception, data comparison, and data routing) to respond or forward parameter by comparing parameters and/or based on output parameters at high level of generality is simply the to apply a computer to the judicial exception does not integrate a judicial exception into a practical application or provide significantly more - see MPEP 2106.05(f). Therefore, the claim 6 does not recite patent eligible subject matter under 35 U.S.C. § 101.
Claim 7 recites “the one or more parameters is forwarded to a second intermediate node comprising a second imitation model being a version of the computational graph model requiring less computational resources to converge when compared to the computational graph model but more computational resources than the imitation model.”
This merely specifies data is received by second intermediate node, and a second imitation model requiring less computational resources to converge, but more capable than first imitation model; therefore, it merely a mathematical concept (similar to imitation model) and insignificant extra-solution activity (i.e., receive parameter as data gathering). Therefore, the claim 7 does not recite patent eligible subject matter under 35 U.S.C. § 101.
Claim 8 recites “computational graph model is a neural network and/or a decision tree.”
This merely further defines computational graph model is a neural network and/or a decision tree; therefore, it merely a description of computation graph model as a particular type of mathematical model does not add any additional technical detail regarding how the model is implemented, trained or executed. Therefore, the claim 8 does not recite patent eligible subject matter under 35 U.S.C. § 101.
Claim 9 recites “the intermediate network node operates between the requesting node and the executing network node.”
This merely specifies the intermediate network node is operated between requesting node and the executing network node within the communication network; therefore, it merely adding the words "apply it" (or an equivalent) with the judicial exception, or instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f). The limitation merely specifies the relative placement of the intermediate network node within the communication network, but do not recite any particular technical arrangement, data flow protocol, or network architecture that constrains how the mathematical operations are performed. Therefore, the claim 9 does not recite patent eligible subject matter under 35 U.S.C. § 101.
Claims 11-18 and 21 recite substantially the same elements as claims 2-9, and are rejected for the same reasons under 35 U.S.C. 101.
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.
Claim(s) 1-3, 5-12, 14-19 and 21 are rejected under 35 U.S.C. 103 as being unpatentable over
Sundström US 20200401944 A1 in view of Krizhevsky (ImageNet Classification with Deep Convolutional Neural Networks, published in 2012) and Han (“DEEP COMPRESSION: COMPRESSING DEEP NEURAL NETWORKS WITH PRUNING, TRAINED QUANTIZATION AND HUFFMAN CODING,” published in 2016).
Claim 1, Sundström teaches (Currently Amended) An intermediate network node (fig.1, 110 and 120; [0006], “…The higher node may be an intermediate network node 110, 120…”) in a communication network (fig.1, 100, 110, 120 and 130), the communication network comprises a requesting node (fig.1, 100; [0006], “…a compute task may be provided in an edge device 100, and data may be provided for the task to be carried out, such as sensor data from a connected or built-in sensor.”) and an executing network node (fig.1, 130; [0006], “…a compute node 130 executed in a cloud server.”), the executing network node comprising a computational graph model ([0056], “the estimation model could be a specific design of a Deep Neural Network (DNN) … a device 300 acting as a node 100 can escalate its sensor data vertically to a more capable node 110, 120, 130 with a more complex estimation model…”), the intermediate network node comprising:
at least one processor;
at least one memory connected to the at least one processor and storing an imitation model that is executed by the at least one processor to execute a Convolutional Neural Network (CNN) for object detection ([0054], “control circuitry 303 may include a processing device 304 and a data memory 305 holding computer program code representing a local estimation model (Note: i.e., imitation model).” [0056], “the estimation model could be a specific design of a Deep Neural Network (DNN) acting as an “object detector”.”), ([0005], “The computational power of these edge devices is constrained by limitations of resources such as memory, CPU and energy. In practice, the limitations mean that these devices need to make use of simplified computational models, e.g. simplified Deep Neural Networks.”), and wherein the imitation model is a model (fig,1, node 110 or 120 with estimation model) requiring less computational resources ([0005], “… limitations of resources …”) (fig.1, 130 with a more complex estimation model, i.e., DNN), and wherein the imitation model comprises
However, Sundström fails to teach the object detection through CNN comprises performing feature learning and classification of an input image, wherein the feature learning comprises creation of one or more activation maps and the classification process comprises conversion of a pooled activation maps of preceding layers into a N dimension vector, wherein each value of the vector represents a probability of an object exists in the input image.
Krizhevsky teaches the object detection through CNN comprises performing feature learning and classification of an input image, wherein the feature learning comprises creation of one or more activation maps and the classification process comprises conversion of a pooled activation maps of preceding layers into a N dimension vector, wherein each value of the vector represents a probability of an object exists in the input image (Abstract, “We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes.” Pages. 87 , 4.5 Overall Architecture “The output of the last fully-connected layer is fed to a 1000-way softmax which produces a distribution over the 1000 class labels … The first convolutional layer filters the 224 x 224 x 3 input image with 96 kernels of size 11 x 11 x 3 with a stride of 4 pixels … The second convolutional layer takes as input the (response-normalized and pooled) output of the first convolutional layer and filters it with 256 kernels of size 5 x 5 x 48 … The fully-connected layers have 4096 neurons each.” Page.86, 4.4 Overlapping Pooling “Pooling layers in CNNs summarize the outputs of neighboring groups of neurons in the same kernel map.” Figure.2: … The network’s input is 150,528-dimensional, and the number of neurons in the network’s remaining layers is given by 253,440–186,624–64,896–64,896–43,264–4096–4096–1000. Page.89, right column, Paragraph two, “Another way to probe the network’s visual knowledge is to consider the feature activations induced by an image at the last, 4096-dimensional hidden layer.” Page.85, The DATASET, “… On ImageNet, it is customary to report two error rates: top-1 and top-5, where the top-5 error rate is the fraction of test images for which the correct label is not among the five labels considered most probable by the model.” Fig.4, … The correct label is written under each image, and the probability assigned to the correct label is also shown with a red bar (if it happens to be in the top 5) … Examiner note: the reference teaches a convolutional neural network in which convolutional layers apply learned kernels to an input image to produce intermediate feature maps at each layer (feature maps corresponding to neuron activations). The feature map are subsequently processed by polling layer that summarize neighboring neuron responses, and the pooled outputs are provided to fully connected layers having a fixed length numerical of neurons. The reference further teaches that the output of the last fully-connected layer is followed by a 1000-way softmax that produces a distribution over 1000 class labels, that top-5 labels are the labels “considered most probable by the model, “ and Figure 4 shows “the probability assigned to the correct label.” A POSITA would understand that the output of a fully connected layer having a defined number of neurons constitutes a fixed length numerical vector with dimensionality equal to the number of neurons (i.e., pooled outputs of preceding layers are represented as an N-dimensional vector at the fully connected layer output). Further, the 1000-way softmax distribution over 1000 class labels teaches that the output vector includes values corresponding to probabilities for respective class/object labels in the input image).
It would have been obvious for a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Sundström to incorporate the teachings of Krizhevsky, and apply a convolutional neural network that performs image feature learning using convolutional and pooling layers and performs classification using fully connected layers and a softmax output layer that produces a probability distribution over class labels in order to apply an established and effective neural network approach for image recognition and to provide classification outputs indicating the probability that respective objects/classes are present in an input image. The combination of teachings would predictably improve image-recognition performance.
However, Sundström and Krizhevsky fail to teach the imitation model is a model requiring less computational resources to converge when compared to the computational graph model, and wherein the imitation model comprises at least one less internal vertex or node and at least one less edge of the computational graph model compared to the computational graph model.
Han teaches the imitation model is a model requiring less computational resources to converge when compared to the computational graph model, and wherein the imitation model comprises at least one less internal vertex or node and at least one less edge of the computational graph model compared to the computational graph model (Abstract, “Our method first prunes the network by learning only the important connections.” Page.2, 2 NETWORK PRUNING “Network pruning has been widely studied to compress CNN models. In early work, network pruning proved to be a valid way to reduce the network complexity and over-fitting … we prune the small-weight connections: all connections with weights below a threshold are removed from the network. Finally, we retrain the network to learn the final weights for the remaining sparse connections. Pruning reduced the number of parameters by 9x and 13x for AlexNet and VGG-16 model.” Fig. 9: Compared with the original network, pruned network layer achieved 3x speedup on CPU, 3.5x on GPU and 4.2x on mobile GPU on average. Figure 10: Compared with the original network, pruned network layer takes 7x less energy on CPU, 3.3x less on GPU and 4.2x less on mobile GPU on average. Examiner note: The reference teaches pruning a trained neural network by removing connections having weights below a threshold and retraining the reduced network to obtain final weights. A POSITA would understand that a neural network is represented as a computational graph in which internal vertices or nodes correspond to computational units and edges correspond to weighted connections. By removing the connections would produces a reduced computational graph having fewer edges and a reduced effective set of internal vertices, thereby produce a limited version of the original computational graph model. Further, because the pruned network has reduced parameters and fewer connections than the original network, and training is performed on the reduced computational structure, A POSITA would understand that the pruned network required fewer computational operations, memory accesses, and energy to reach its trained state (i.e., to converge) when compared to the original dese computational graph model as evidenced by demonstrated speedup and energy reduction results in Figs 9-10).
It would have been obvious for a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Sundström and Krizhevsky to incorporate the teachings of Han, and apply network pruning and compression techniques in order to reduce the computational complexity and resource requirement of the computation graph model, thereby producing a model that contains fewer parameters and connections while maintain comparable predictive performance.
Claim 2, Sundström further teaches The intermediate network node according to claim 1, wherein the imitation model comprises one or more of following compared to the computational graph model:
at least one input parameter less than the computational graph model;
at least one output parameter less than the computational graph model; and/or
one or more functions of less computational complexity ([0056], “…one example is when a deployed model in a device 300 acting as a node 100 can escalate its sensor data vertically to a more capable node 110, 120, 130 with a more complex estimation model…” examiner note: A POSITA would understand that estimation model of edge device node 100 has functions less computational complexity than estimation model of intermediate node 110, the estimation model of intermediate node 110 has functions less computational complexity than estimation model of another intermediate node 120…etc.).
Claim 3, Sundström further teaches The intermediate network node according to claim 1, wherein the intermediate network node builds the imitation model (i.e., estimation model) based on received one or more input parameters from the requesting node ([0006], “The reason may be that the sensor device cannot host a sufficiently complex estimation model given its limited resources, hence for this specific input it decides to transfer the image data to a higher end node 110, which may escalate further to higher nodes 120, 130, and request a more qualitative decision to this estimation task. Transmission in the uplink 160 from the edge device compute node 100 may thus include sensor data and a particular task associated with the data. An improved result, such as e.g. data representing the number of people detected in the image, may thereafter be received 170 in the downlink. [0056], “a deployed model in a device 300 acting as a node 100 can escalate its sensor data vertically to a more capable node 110, 120, 130 with a more complex estimation model, which can provide a “ground truth” estimation and at the same time use the escalated sensor data to re-train the edge device model in the device 300 with some of its recently collected inputs, thereby adjusting the less capable device's 300 estimation model to its actual input.” Examiner Note: A POSITA would understand that the intermediate node 110 or 120 with simplified computational models (i.e., DNN) is trained by received input from edge device node 100 and escalate to cloud node 130 via uplink 160, and re-trained (i.e., built) node 110 and 120 by escalated date via downlink 170).
Claim 5, Sundström further teaches The intermediate network node according to claim 1, wherein the intermediate network node obtains the imitation model (i.e., estimation model) from the computational graph model ([0006], “The reason may be that the sensor device cannot host a sufficiently complex estimation model given its limited resources, hence for this specific input it decides to transfer the image data to a higher end node 110, which may escalate further to higher nodes 120, 130, and request a more qualitative decision to this estimation task. Transmission in the uplink 160 from the edge device compute node 100 may thus include sensor data and a particular task associated with the data. An improved result, such as e.g. data representing the number of people detected in the image, may thereafter be received 170 in the downlink. [0056], “a deployed model in a device 300 acting as a node 100 can escalate its sensor data vertically to a more capable node 110, 120, 130 with a more complex estimation model, which can provide a “ground truth” estimation and at the same time use the escalated sensor data to re-train the edge device model in the device 300 with some of its recently collected inputs, thereby adjusting the less capable device's 300 estimation model to its actual input.” Examiner Note: A POSITA would understand that the intermediate nodes 110 or 120 with simplified computational models (i.e., DNN) is trained by received input from edge device node 100 and escalate to cloud node 130 via uplink 160, and re-trained (i.e., obtain or built) node 110 and 120 by escalated date from cloud node 130 (i.e., computational graph model) via downlink 170).
Claim 6, Sundström teaches The intermediate network node according claim 1, wherein the at least one memory connected to the at least one processor stores program code that is executed by the at least one processor ([0054], “control circuitry 303 may include a processing device 304 and a data memory 305 holding computer program code representing a local estimation model (Note: i.e., imitation model).”) to comprising:
receive a request from the requesting node (fig.1, edge device compute node 100), wherein the request comprises one or more input parameters ([0006], “… a compute task may be provided in an edge device 100, and data may be provided for the task to be carried out, such as sensor data from a connected or built-in sensor. Dependent on the compute deployment, the task may be carried out in the edge device node 100, or the task and the data may be escalated 160 from the edge device node 100 to a higher (more capable) compute node 110, 120.”); and
determine whether to respond to the request or to forward the one or more parameters towards the executing network node by comparing the one or more input parameters to one or more needed input parameters of the imitation model and/or based on one or more output parameters of the imitation model (fig.2, S210, S220, S230 and S260 illustrates receiving request form lower layer, make classification decision and determination, then respond the request from lower layer by use classification (i.e., output parameters) of intermediate nodes.
Claim 7, Sundström further teaches The intermediate network node according to claim 6, wherein the one or more parameters (i.e., sensor data) is forwarded to a second intermediate node (fig.1, intermediate node 120) comprising a second imitation model (i.e., estimation model of 120) being a version of the computational graph model requiring less computational resources ([0056], “the estimation model could be a specific design of a Deep Neural Network (DNN) … a device 300 acting as a node 100 can escalate its sensor data vertically to a more capable node 110, 120, 130 with a more complex estimation model…”; [0005], “The computational power of these edge devices is constrained by limitations of resources such as memory, CPU and energy. In practice, the limitations mean that these devices need to make use of simplified computational models, e.g. simplified Deep Neural Networks.” Examiner note: A POSITA would understand that the estimation model of intermediate node 110 has less computational resource than estimation model of another intermediate node 120 , the estimation model of another intermediate node 120 has less computational resource than estimation model of cloud node 130).
However, Sundström and Krizhevsky fail to teach a version of the computational graph model requiring less computational resources to converge.
Han teaches a version of the computational graph model requiring less computational resources to converge (Abstract, “Our method first prunes the network by learning only the important connections.” Page.2, 2 NETWORK PRUNING “Network pruning has been widely studied to compress CNN models. In early work, network pruning proved to be a valid way to reduce the network complexity and over-fitting … we prune the small-weight connections: all connections with weights below a threshold are removed from the network. Finally, we retrain the network to learn the final weights for the remaining sparse connections. Pruning reduced the number of parameters by 9x and 13x for AlexNet and VGG-16 model.” Fig. 9: Compared with the original network, pruned network layer achieved 3x speedup on CPU, 3.5x on GPU and 4.2x on mobile GPU on average. Figure 10: Compared with the original network, pruned network layer takes 7x less energy on CPU, 3.3x less on GPU and 4.2x less on mobile GPU on average. Examiner note: The reference teaches pruning a trained neural network by removing connections having weights below a threshold and retraining the reduced network to obtain final weights. A POSITA would understand that a neural network is represented as a computational graph in which internal vertices or nodes correspond to computational units and edges correspond to weighted connections. By removing the connections would produces a reduced computational graph having fewer edges and a reduced effective set of internal vertices, thereby produce a limited version of the original computational graph model. Further, because the pruned network has reduced parameters and fewer connections than the original network, and training is performed on the reduced computational structure, A POSITA would understand that the pruned network required fewer computational operations, memory accesses, and energy to reach its trained state (i.e., to converge) when compared to the original dese computational graph model as evidenced by demonstrated speedup and energy reduction results in Figs 9-10).
It would have been obvious for a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Sundström and Krizhevsky to incorporate the teachings of Han, and apply network pruning and compression techniques in order to reduce the computational complexity and resource requirement of the computation graph model, thereby producing a model that contains fewer parameters and connections while maintain comparable predictive performance.
Claim 8, Sundström further teaches The intermediate network node according to claim 1, wherein the computational graph model is a neural network and/or a decision tree ([0056], “the estimation model could be a specific design of a Deep Neural Network (DNN) … a device 300 acting as a node 100 can escalate its sensor data vertically to a more capable node 110, 120, 130 with a more complex estimation model…”).
Claim 9 Sundström further teaches The intermediate network node according to claim 1, wherein the intermediate network node (fig.1, 110) operates between the requesting node (fig.1, 100) and the executing network node (fig.1, 130).
The elements of claims 10-12, 14-19 and 21 are substantially the same as those of claims 1-3 and 5-9. Therefore, the elements of claims 10-12, 14-19 and 21 are rejected due to the same reasons as outlined above for claims 1-3 and 5-9.
Claim(s) 4 and 13 are rejected under 35 U.S.C. 103 as being unpatentable over Sundström and
Krizhevsky and Han as applied to claims 3 and 12 above, and further in view of Durdanovic US20170337472A1.
Claim 4, Sundström and Krizhevsky and Han teach The intermediate network node according to claim 3, wherein the imitation model (see Sundström, i.e., estimation model) is built by removing one or more parts of the imitation model (see Han, page.2, NETWORK PRUNING; fig.9-10)
However, Sundström and Krizhevsky and Han fail to teach model is built by removing one or more parts of the model that have not been used within a set interval.
Durdanovic teaches the model is built by removing one or more parts of the model that have not been used within a set interval ([0014] … active pruning of filters in convolutional neural networks (CNNs). During training (note: i.e., a set interval), the present embodiments reduce the size of all weights between each iteration, driving the weight values toward zero. Once a set of weights falls below a threshold, the weights are removed from the CNN along with associated kernels (note: i.e., remove parts),…” [0005], “A block of weights is pruned from the layer if the block of weights in the layer has a contribution to an output of the layer that is below a threshold.” [0018] “…a column 106 has fallen below the threshold, representing weights which do not contribute (note: i.e., interpreted as not been used) to the accuracy of the output. This column 106 is pruned from the first array of weights 104.”).
It would have been obvious for a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Sundström and Krizhevsky and Han to incorporate the teachings of Durdanovic, and apply systems and methods are provided for active pruning of filters in convolutional neural networks (CNNs) and a block of weights is pruned from the layer if the block of weights in the layer has a contribution to an output of the layer that is below a threshold during training in order to reducing the computational cost of using the pruned CNN without increasing the sparsity of the CNN ([0014]). in this case, removing a block of weights in the layer of CNN or DNN would improve training speed when dealing with limited training data.
The elements of claim 13 is substantially the same as those of claim 4. Therefore, the elements of claim 13 is rejected due to the same reasons as outlined above for claim 4.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Simonyan (VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION, published in 2015), discloses a thorough evaluation of networks of increasing depth using an architecture with very small (3×3) convolution filters, which shows that a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16–19 weight layers.
Sharma US20170032285A1, discloses Layer 5 consists of convolution, ReLU, max-pooling and normalization operations to obtain a feature map … The next two layers (layers 6, 7) 507 may be fully connected which outputs a 4096 dimensional vector. The final layer is C-way softmax function 508 that outputs the probabilities across C classes.([0054]). After fc7, the softmax function takes the 4096 vector as input and outputs the scores/probabilities for each class.([0064]).
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to whose telephone number is (571)270-1303. The examiner can normally be reached Monday - Friday.
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, Emerson Puente can be reached at (571)272-3652. 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.
/YI . HAO/
Examiner, Art Unit 2187
/JOHN E JOHANSEN/Examiner, Art Unit 2187
Prosecution Insights