CTFR 18/085,555 CTFR 100537 DETAILED ACTION This office action is in response to amendments filed on 02/12/2026 . Claims 1, 3-5, 8, 10, 13-15, 17, and 19 have been amended. Claims 1-20 are pending. Notice of Pre-AIA or AIA Status 07-03-aia AIA 15-10-aia 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 Arguments Rejections Under 35 USC § 112(b): In light of applicant’s amendments to the claims (pg. 2-6) responsive to the rejections under 35 USC § 112(b), the following rejections have been withdrawn: The rejection of claim 4 due to lack of antecedent basis for the term “the image.” The rejection of claims 8 and 17 due to orientation-dependence and lack of clarity of the terms “top,” “bottom,” and “rightmost corner.” In light of applicant’s clarifying remarks (pg. 8) regarding the rejections under 35 USC § 112(b), the following rejections have been withdrawn: The rejection of claims 8 and 17 due to ambiguity of the term “variable connectivity.” However, applicant’s arguments regarding the rejection of claims 1, 13, and 14 have been fully considered but they are not persuasive. The claims are amended to define hyperuniformity as the suppression of long-range density fluctuations, consistent with the definition provided in paragraph 0034 of the specification. However, neither the specification nor the claims explain what hyperuniformity means in the context of a neural network topology. Hyperuniformity is a term in materials science referring to the density and spatial distribution of points in physical space. Nodes of a neural network, on the other hand, are not points in any continuous physical space, but rather vertices of a discreet graph, where the distances between them in a diagram are representational artifacts with no intrinsic meaning. A neural network topology does not have a spatial density, and thus it is unclear what it would mean for the topology of a neural network to be hyperuniform, be based on hyperuniformity, or suppress long-range density fluctuations, or how a hyperuniform topology could be generated by applying a sparse mask to a fully-connected neural network. Thus, the use of the term “hyperuniform” to describe the topology of a neural network remains unclear. Rejections Under 35 USC § 101: Applicant's arguments regarding the rejections under 35 USC § 101 (pg. 9-14) have been fully considered but they are not persuasive. Regarding Step 2A, Prong One, applicant argues that the independent claim is not directed to a mental process because the amended claim 1 limitation “generating a sparse topology for a feedforward neural network, where connectivity is based on a substantially hyperuniform topology, wherein the substantially hyperuniform topology that suppresses long-range density fluctuations is generated by applying a sparse mask to weight matrices of a fully connected neural network” cannot be performed mentally. This limitation includes two steps: (1) generating the substantially hyperuniform topology by applying a sparse mask to weight matrices, and (2) generating a sparse topology based on the substantially hyperuniform topology. Applicant specifically argues that the first step of applying a sparse mask to weight matrices involves large-scale high-dimensional matrix operations which cannot practically be performed in the human mind. Examiner notes that, as is explained in the rejection below, this limitation is considered an additional element amounting to insignificant extra-solution activity which is well-understood, routine, and conventional in the field of machine learning. Applicant also argues that the second step of generating the sparse topology is an operation performed on the internal data structures of a computer-implemented neural network, must be performed by a processor, and cannot be performed in the human mind. Examiner notes that while weight matrices may represent the topology of a computer-implemented neural network, they are in fact simple 2-dimensional data structures which can be written out and modified on a sheet of paper. MPEP 2106.04(a)(2)(III) notes that claims can recite a mental process even if they are claimed as being performed on a computer. Therefore, this step can be considered a mental process, and thus the claim is directed to a mental process. Regarding Step 2A, Prong Two, applicant argues that the amended independent claim integrates the abstract idea into a practical application by providing a technological improvement to machine learning systems. Applicant specifically argues that the claimed sparse neural network topology with hyperuniform connectivity is an improvement to neural network architecture because neural networks constructed in this manner are more efficient, more accurate, and more stable than conventional neural networks. Applicant references multiple drawings (9A-9D and 10) and paragraphs from the specification (0034, 0046, 0059-0059) which assert these improvements. However, the cited drawings and specification paragraphs do not explain how these improvements are achieved by the claimed method, nor do applicant’s remarks. According to MPEP 2106.05(a), “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.” The rejections under 35 USC § 101 have been updated to include the amended limitations and to clarify the reasoning given for the limitations that were not amended. Prior Art Rejections: Applicant's arguments regarding the prior art rejections (pg. 14-17) have been fully considered but they are not persuasive. Applicant argues that the cited primary reference Prabhu fails to disclose or suggest the features of the amended independent claim. Applicant specifically argues that while Prabhu teaches a sparse topology which is uniform in degree and connectivity, this is not the same as hyperuniformity, which requires the suppression of long-range density fluctuations. However, as is explained in the rejection under 35 USC § 112(b), the meaning of “hyperuniformity” and “the suppression of long-range density fluctuations” is unclear in the context of a neural network topology. Under examiner’s interpretation established in the rejection under 35 USC § 112(b), hyperuniformity refers to uniform connectivity in the neural network topology, and thus Prabhu’s uniform Cayley graph-based topology meets the claimed hyperuniformity constraint. Applicant additionally argues that Prabhu does not teach obtaining the hyperuniform topology by starting from a fully connected neural network and applying a sparse mask to its weight matrix, as recited in the amended independent claim. Examiner notes that Isakov teaches this amended limitation. Thus, the rejection of the independent claims under 35 USC § 102 in view of Prabhu has been replaced with a rejection under 35 USC § 103 in view of Prabhu and Isakov. The prior art rejections have been updated to include the amended limitations and to clarify the reasoning given for the limitations that were not amended. Claim Rejections - 35 USC § 112 07-30-02 AIA 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. 07-34-01 Claims 1-20 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. Claims 1, 13, and 14 each recite the limitation “generating a sparse topology for a feedforward neural network, where connectivity is based on a substantially hyperuniform topology, wherein the substantially hyperuniform topology that suppresses long-range density fluctuations is generated by applying a sparse mask to weight matrices of a fully connected neural network.” According to specification paragraph 0034, “as used herein, ‘hyperuniform’ refers to the suppression of long-range density fluctuations, ‘nearly hyperuniform’ refers to the near suppression of long- range density fluctuations, and ‘substantially hyperuniform’ refers to both ‘hyperuniform’ and ‘nearly hyperuniform’ density fluctuations.” In other words, hyperuniformity is a term in materials science referring to the density and spatial distribution of points in physical space. Nodes of a neural network, on the other hand, are not points in any continuous physical space, but rather vertices of a discreet graph, where the distances between them in a diagram are representational artifacts with no intrinsic meaning. A neural network topology does not have a spatial density, and thus it is unclear what it would mean for the topology of a neural network to be hyperuniform, be based on hyperuniformity, or suppress long-range density fluctuations, or how a hyperuniform topology could be generated by applying a sparse mask to a fully-connected neural network. For examination purposes, this limitation will be interpreted as referring to generating a sparse topology for a feedforward neural network, where connectivity is based on a topology having uniform connectivity. Claims 2-12 and 15-20 are additionally rejected due to their dependence on claims rejected above. Claim Rejections - 35 USC § 101 07-04-01 AIA 07-04 35 U.S.C. 101 reads as follows: Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title. Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. Claim 1 : Step 1: The claim is directed to a method , which falls within the statutory category of a process. Step 2A Prong 1: The claim is directed to an abstract idea. Specifically, the claim recites: generating a sparse topology for a feedforward neural network, where connectivity is based on a substantially hyperuniform topology; ( Abstract idea – mental process. Generating a sparse neural network topology based on a substantially hyperuniform topology can practically be performed in the human mind or with the aid of pen and paper, for example, by viewing an example of a hyperuniform topology on a display and drawing a neural network topology which mirrors its structure. The courts have recognized that claims can recite a mental process even if they are claimed as being performed on a computer. See MPEP 2106.04(a)(2)(III).) performing a processing task ( Abstract idea – mental process. Performing a processing task can practically be performed in the human mind or with the aid of pen and paper, for example, by viewing data on a display and mentally making a prediction. The courts have recognized that claims can recite a mental process even if they are claimed as being performed on a computer. See MPEP 2106.04(a)(2)(III).) Step 2A Prong 2: The additional elements recited in the claim do not integrate the abstract idea into a practical application, individually or in combination. Specifically, the claim recites the additional elements: wherein the substantially hyperuniform topology that suppresses long-range density fluctuations is generated by applying a sparse mask to weight matrices of a fully connected neural network; (Generating the substantially hyperuniform topology by applying a sparse mask to the weight matrix of a fully connected neural network amounts to adding insignificant extra-solution activity to the judicial exception – see MPEP2106.05(g).) training the feedforward neural network with the sparse topology using a set of training data; (Generic training of a feedforward neural network with a given topology using training data is standard in the field of machine learning, and thus this limitation amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).) using the trained feedforward neural network. (Performing a generic processing task using a trained model is standard in the field of machine learning, and thus this limitation amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).) Step 2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception. Specifically, the claim recites the additional elements: wherein the substantially hyperuniform topology that suppresses long-range density fluctuations is generated by applying a sparse mask to weight matrices of a fully connected neural network; (Generating the substantially hyperuniform topology by applying a sparse mask to the weight matrix of a fully connected neural network amounts to adding insignificant extra-solution activity to the judicial exception – see MPEP2106.05(g). Further, generating a sparse topology by applying a sparse mask to a dense neural network is well-understood, routine, and conventional in the field of machine learning, per Jaiswal: “Recently, a significant amount of research efforts have been focused towards developing increasingly sophisticated and efficient pruning algorithms… to identify the sparse mask of the original dense model at the initialization, and then train only the sparse subnetwork…” (Jaiswal, “Training Your Sparse Neural Network Better with Any Mask”, pg. 1, section 1) – see MPEP 2106.05(d).) training the feedforward neural network with the sparse topology using a set of training data; (Generic training of a feedforward neural network with a given topology using training data is standard in the field of machine learning, and thus this limitation amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).) using the trained feedforward neural network. (Performing a generic processing task using a trained model is standard in the field of machine learning, and thus this limitation amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).) Claims 2-20 : Claim 2 recites The method of claim 1, wherein the processing task comprises one of classification, regression, and correlations at different scales. Classification, regression, or correlation processing tasks can practically be performed in the human mind or with the aid of pen and paper (i.e. mental processes ), for example, by viewing data on a display and mentally making a classification prediction. Therefore, the claim merges with the mental process recited in claim 1, and does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 3 recites The method of claim 1, wherein the processing task comprises connecting to a medical imaging device via a network module, obtaining a medical image, processing the medical image using the trained feedforward neural network, and treating a patient based on results of the processing of the medical image. This claim merely limits the use of the trained model to the field of medical image analysis, and thus amounts to generally linking the use of a judicial exception to a particular technological environment or field of use – see MPEP 2106.05(h). Therefore, the claim does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 4 recites The method of claim 1, wherein the processing task comprises obtaining financial information via a network module, processing the financial information using the trained feedforward neural network, and detecting and mitigating financial fraud based on results of the processing of the financial information. This claim merely limits the use of the trained model to the field of financial fraud detection, and thus amounts to generally linking the use of a judicial exception to a particular technological environment or field of use – see MPEP 2106.05(h). Therefore, the claim does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 5 recites The method of claim 1, wherein the performing of the processing task comprises: feeding input elements of an input vector into a first set of hidden nodes of the trained feedforward neural network; applying a linear transformation by each of the first set of hidden nodes to a corresponding input element or corresponding input vector; applying an activation function to a result of the linear transformation; and obtaining an output value. Feeding input elements into hidden nodes, applying a linear transformation, applying an activation function, and obtaining an output value are generic functions of a feedforward neural networks which are standard in the field of machine learning, and thus this claim amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea – see MPEP 2106.05(f). Therefore, the claim does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 6 recites The method of claim 1, wherein a layer of the sparse topology comprises a set of nodes connected in one direction and an output value of a node of a given layer is an input value of one or more subsequent nodes of the given layer or another layer. Connecting nodes in one direction such that the output of one layer is the input of a subsequent layer is a generic function of feedforward neural networks which is standard in the field of machine learning, and thus this claim amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea – see MPEP 2106.05(f). Therefore, the claim does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 7 recites The method of claim 6, wherein nodes in a same layer of the trained feedforward neural network have a same activation function. Each node in a layer having the same activation function is a generic feature of a feedforward neural network which is standard in the field of machine learning, and thus this claim amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea – see MPEP 2106.05(f). Therefore, the claim does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 8 recites The method of claim 1, wherein the sparse topology comprises a network of ring configurations of nodes in two dimensions or surface configurations of nodes in greater than two dimensions, the ring configurations and surface configurations having variable connectivity between the corresponding nodes, and wherein a copy of information flows from a first node in a first direction through a first portion of a given ring configuration or a given surface configuration of the trained feedforward neural network and in a second direction through a second portion of the given ring or the given surface to meet at a second node of the given ring or the given surface. This limitation merely qualifies the sparse topology generated via the mental process recited in claim 1. A topology comprising a network of ring configurations with variable connectivity and two paths of information flow can practically be generated in the human mind or with the aid of pen and paper. Therefore, the claim merges with the mental process recited in claim 1, and does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 9 recites The method of claim 1, wherein at least one ring configuration of nodes of the trained feedforward neural network has a different number of nodes than another configuration ring of nodes of the trained feedforward neural network. This limitation merely qualifies the sparse topology generated via the mental process recited in claim 1. A topology comprising ring configurations with different numbers of nodes can practically be generated in the human mind or with the aid of pen and paper. Therefore, the claim merges with the mental process recited in claim 1, and does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 10 recites The method of claim 9, wherein bidirectional connections between two nodes of the trained feedforward neural network enable an exchange of information that allows for a mixing of information that reaches different regions of the trained feedforward neural network. This amounts to adding insignificant extra-solution activity to the judicial exception – see MPEP2106.05(g). Further, bidirectional connections between nodes of a neural network are well-understood, routine, and conventional in the field of machine learning and neural networks, per Rojas: “In 1982 the American physicist John Hopfield proposed an asynchronous neural network model which made an immediate impact in the AI community… The network is symmetric because the weight w i j for the connection between unit i and unit j is equal to the weight w j i of the connection from unit j to unit i . This can be interpreted as meaning that there is a single bidirectional connection between both units” (Rojas, “Neural Networks”, pg. 341-342, section 13.2.2). See MPEP 2106.05(d). Therefore, the claim does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 11 recites The method of claim 1, wherein a final layer of the trained feedforward neural network has one or more nodes with linear activation for a regression or logistic function for binary classification. Adding a linear activation for regression or logistic function for classification to the final layer of a neural network amounts to adding insignificant extra-solution activity to the judicial exception – see MPEP2106.05(g). Further, using a Sigmoid (i.e. logistic) function in the final layer is well-understood, routine, and conventional in the field of machine learning, per Zhu: “The Sigmoid function is commonly used in the last layer of the neural network” (Zhu et al., “OFS-NN: An Effective Phishing Websites Detection Model Based on Optimal Feature Selection and Neural Network”, pg. 73278, section IV). See MPEP 2106.05(d). Therefore, the claim does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 12 recites The method of claim 1, wherein, in the operation of generating the sparse topology for the feedforward neural network, the sparse topology comprises a network of surface configurations of nodes in greater than two dimensions. This limitation merely qualifies the sparse topology generated via the mental process recited in claim 1. A topology comprising a network of surface configurations in greater than two dimensions can practically be generated in the human mind or with the aid of pen and paper. Therefore, the claim merges with the mental process recited in claim 1, and does not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim 13 is a product claim containing substantially the same elements as method claim 1, and is rejected on the same grounds under 35 U.S.C. 101 as claim 1, mutatis mutandis. The additional components of A non-transitory computer readable medium comprising computer executable instructions which when executed by a computer cause the computer to perform the method are interpreted as a general-purpose computer and mere instructions to apply the judicial exception on the computer. The courts have recognized that claims can recite a mental process even if they are claimed as being performed on a computer. Therefore, the claims do not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claims 14-20 are apparatus claims containing substantially the same elements as method claims 1, 5, 6, 8-10, and 12, respectively, and are rejected on the same grounds under 35 U.S.C. 101 as claims 1, 5, 6, 8-10, and 12, respectively, mutatis mutandis. The additional components of An apparatus comprising: a memory; and at least one processor, coupled to said memory, and operative to perform operations are interpreted as a general-purpose computer and mere instructions to apply the judicial exception on the computer. The courts have recognized that claims can recite a mental process even if they are claimed as being performed on a computer. Therefore, the claims do not recite additional elements that are sufficient to amount to significantly more than the abstract idea. Claim Rejections - 35 USC § 103 07-20-aia AIA 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. 07-21-aia AIA Claim s 1, 2, 5-7, 11, and 13-16 are rejected under 35 U.S.C. 103 as being unpatentable over Prabhu et al. (hereinafter Prabhu), “Deep Expander Networks: Efficient Deep Networks from Graph Theory” in view of Isakov et al. (hereinafter Isakov), “NeuroFabric: Identifying Ideal Topologies for Training A Priori Sparse Networks”. Regarding Claim 1, Prabhu teaches A method comprising: generating a sparse topology for a feedforward neural network, where connectivity is based on a substantially hyperuniform topology; (Pg. 1, Abstract: “Expander graphs are used to model connections between filters in CNNs to design networks called X-Nets.” Pg. 17, Appendix A: “Explicit Expanders: Many constructions of expander graphs have been explored in the past… We will be considering Cayley graphs… For H ⊂ { 0 , 1 } n , the Cayley graph defined by H is | H | -regular.” Pg. 7, section 3.4: “Next, we show a much stronger connectivity property known as mixing for the X-Nets. The theorem essentially says that the number of edges between subsets of input and output nodes is proportional to the product of their sizes. This result implies that the connectivity properties are uniform and rich across all nodes as well as subsets of nodes of the same size. Simply put, all nodes tend to have equally rich representational power.” Pg. 2, figure 1: “…expander graph-based models produce sparse yet highly connected networks.” According to paragraph 0034 of the instant application’s specification, “as used herein… ‘substantially hyperuniform’ refers to both ‘hyperuniform’ and ‘nearly hyperuniform’ density fluctuations.” The topology of a Cayley expander graph is r-regular and uniform in connectivity, and therefore represents a hyperuniform topology. The topology of a sparse CNN (a type of feedforward neural network) is generated based on the topology of a Cayley expander graph.) training the feedforward neural network with the sparse topology using a set of training data; and (Pg. 3, section 1: “We provide memory-efficient implementations of Convolutional (X-Conv) layers using sparse matrices and propose a fast expander-specific algorithm… X-Conv layers obtain a 4% improvement in accuracy when both the techniques are applied to the MobileNet architecture trained on Imagenet… Since we enforce the sparsity before the training phase itself, our models are inherently compact and faster to train compared to pruning techniques.” Imagenet is a set of training data.) performing a processing task using the trained feedforward neural network. (Pg. 7, section4: “Our algorithms achieve speedups and save memory in the training as well as the inference phase.” An inference phase involves processing data using the trained model.) Prabhu does not appear to explicitly disclose wherein the substantially hyperuniform topology that suppresses long-range density fluctuations is generated by applying a sparse mask to weight matrices of a fully connected neural network; However, Isakov teaches wherein the substantially hyperuniform topology that suppresses long-range density fluctuations is generated by applying a sparse mask to weight matrices of a fully connected neural network; (Pg. 3-4, section IV: “Our proposed approach replaces fully-connected layers with cascades of sparse layers… We construct a sparse layer from a matrix W ∈ R m × n by multiplying it element-wise with the mask M ∈ { 0,1 } m × n with sparsity s ∈ [ 0,1 ] .” Sparse mask M is applied to the weight matrix W of a fully connected neural network to generate the sparse neural network topology (i.e. the substantially hyperuniform topology).) It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Prabhu and Isakov. Prabhu teaches enforcing neural network sparsity prior to training via a sparse but well-connected topology based on uniform and deterministic Cayley expander graphs. Isakov teaches enforcing neural network sparsity prior to training by replacing fully connected layers with sparse cascades, where the sparse layers are generated by applying a sparse mask to the fully connected layers. One of ordinary skill would have motivation to combine Prabhu and Isakov because Isakov’s method of sparse neural network initialization using sparse masks “alleviates the vanishing gradient problem by taking layer sparsity into account” (Isakov, pg. 4, section IV). The sparse cascade architecture taught by Isakov can “replace dense or convolutional layers without affecting the rest of the network architecture” (Isakov, pg. 1, section I), and “sparsifying layers can allow users to train wider networks without the quadratic growth in network size” (Isakov, pg. 2, section III). Regarding Claim 2, Prabhu and Isakov teach The method of claim 1, as shown above. Prabhu also teaches wherein the processing task comprises one of classification, regression, and correlations at different scales. (Pg. 9, section 5.1: “We compare X-Conv against grouped convolutions using MobileNet-0.5 on the ImageNet classification task.”) Regarding Claim 5, Prabhu and Isakov teach The method of claim 1, as shown above. Prabhu also teaches wherein the performing of the processing task comprises: feeding input elements of an input vector into a first set of hidden nodes of the trained feedforward neural network; (Pg. 2, figure 1: The topology labelled ‘Expander-like Approximation’ shows the architecture of an ‘expander graph-based model’, with inputs at the top and outputs at the bottom, where the top layer of nodes labeled I 1 . . . I 5 represent elements of an input vector, and downward connections to the next layer of nodes represent feeding the input elements to a first set of hidden nodes.) obtaining an output value. (Pg. 2, figure 1: The topology labelled ‘Expander-like Approximation’ shows the architecture of an ‘expander graph-based model’, where the bottom layer of nodes labeled O 1 . . . O 5 represent output values obtained by processing the input elements through the hidden layers.) Isakov teaches applying a linear transformation by each of the first set of hidden nodes to a corresponding input element or corresponding input vector; (Pg. 2, section III: “Our approach is to replace each fully-connected layer with a cascade of sparsely-connected layers, as shown in Figure 1.” Figure 1 shows the architecture of a sparse cascade network, where the nodes labeled ‘Linear’ represent hidden nodes applying a linear transformation to input elements (pg. 3).) applying an activation function to a result of the linear transformation; (Pg. 2, section III: “Our approach is to replace each fully-connected layer with a cascade of sparsely-connected layers, as shown in Figure 1.” Figure 1 shows the architecture of a sparse cascade network, where the nodes labeled ‘ReLU’ represent applying the ReLU activation function to the results of the linear transformation (pg. 3).) Regarding Claim 6, Prabhu and Isakov teach The method of claim 1, as shown above. Prabhu also teaches wherein a layer of the sparse topology comprises a set of nodes connected in one direction and an output value of a node of a given layer is an input value of one or more subsequent nodes of the given layer or another layer. (Pg. 2, figure 1: The topology labelled ‘Expander-like Approximation’ shows the architecture of an ‘expander graph-based model’, with inputs at the top and outputs at the bottom, where each horizontal row of nodes represents a layer, and downward connections from the nodes of a higher layer to the nodes of a lower layer represent the outputs of nodes of the higher layer being passed as inputs to subsequent nodes of the lower layer.) Regarding Claim 7, Prabhu and Isakov teach The method of claim 6, as shown above. Isakov also teaches wherein nodes in a same layer of the trained feedforward neural network have a same activation function. (Pg. 2, section III: “Our approach is to replace each fully-connected layer with a cascade of sparsely-connected layers, as shown in Figure 1.” Figure 1 shows the architecture of a sparse cascade network, where the nodes in the second layer of the first sparse cascade are all followed by the same ReLU activation function (pg. 3).) Regarding Claim 11, Prabhu and Isakov teach The method of claim 1, as shown above. Isakov also teaches wherein a final layer of the trained feedforward neural network has one or more nodes with linear activation for a regression or logistic function for binary classification. (Pg. 2, section III: “Our approach is to replace each fully-connected layer with a cascade of sparsely-connected layers, as shown in Figure 1.” Figure 1 shows the architecture of a sparse cascade network, where the final layer of nodes labeled ‘Sigmoid’ represents applying a sigmoid function (i.e. logistic function for binary classification) (pg. 3).) Claim 13 is a product claim containing substantially the same elements as method claim 1. Prabhu and Isakov teach the elements of claim 1, as shown above. Prabhu also teaches A non-transitory computer readable medium comprising computer executable instructions which when executed by a computer cause the computer to perform the method (Examiner notes this limitation is interpreted as implementation of the disclosed method in a computing environment. Pg. 22, Appendix C.3: “We trained all the models on a setup consisting of 10 Intel Xeon E5-2640 cores and 2 GeForce GTX 1080 Ti GPUs.”) Claims 14-16 are system claims containing substantially the same elements as method claims 1 and 5-6, respectively. Prabhu and Isakov teach the elements of claims 1 and 5-6, as shown above. Prabhu also teaches An apparatus comprising: a memory; and at least one processor, coupled to said memory, and operative to perform operations comprising: (Examiner notes this limitation is interpreted as implementation of the disclosed method in a computing environment. Pg. 22, Appendix C.3: “We trained all the models on a setup consisting of 10 Intel Xeon E5-2640 cores and 2 GeForce GTX 1080 Ti GPUs.”) 07-21-aia AIA Claim 3 i s re jected under 35 U.S.C. 103 as being unpatentable over Pr abhu in view of Isakov , and further in view of Ka thiravelu et al. (hereinafter Kathiravelu), “A DICOM Framework for Machine Learning Pipelines against Real-Time Radiology Images”. Regarding Claim 3, Prabhu and Isakov teach The method of claim 1, as shown above. Prabhu and Isakov do not appear to explicitly disclose wherein the processing task comprises connecting to a medical imaging device via a network module, obtaining a medical image, processing the medical image using the trained feedforward neural network, and treating a patient based on results of the processing of the medical image. However, Kathiravelu teaches wherein the processing task comprises connecting to a medical imaging device via a network module, obtaining a medical image, processing the medical image using the trained feedforward neural network, and treating a patient based on results of the processing of the medical image. (Pg. 1-2, section 1: “Radiology departments consist of several clinical systems such as PACS [17] and Vendor-Neutral Archives (VNAs) [26] that receive images real-time from various scanners… ML pipelines have been proposed to reliably prognosis and predict cancer… This paper presents Niffler, an ML framework that retrieves images from the PACS using DICOM network listeners, and extracts and processes metadata from the acquired images at the research clusters. It then executes ML pipelines and real-time analytics pipelines on radiology images and their textual metadata.” Scanned images (i.e. medical images) are obtained by connecting to scanners (i.e. medical imaging devices) through the PACS system and DICOM protocol (i.e. network module) and processed by ML pipelines (i.e. trained feedforward neural network) to perform prognosis (i.e. treat a patient).) It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Prabhu, Isakov, and Kathiravelu. Prabhu teaches enforcing neural network sparsity prior to training via a sparse but well-connected topology based on uniform and deterministic Cayley expander graphs. Isakov teaches enforcing neural network sparsity prior to training by replacing fully connected layers with sparse cascades, where the sparse layers are generated by applying a sparse mask to the fully connected layers. Kathiravelu teaches a system for real-time execution of machine learning models on medical imaging data. One of ordinary skill would have motivation to combine Prabhu, Isakov, and Kathiravelu because Prabhu provides efficient and accurate CNNs: “Such networks are expected to preserve accuracy (due to connectivity) while being runtime efficient (due to the sparsity). We empirically demonstrate this in the later sections” (Prabhu, pg. 4, section 3), and Kathiravelu’s Niffler framework enables efficient real-time execution of ML models—including CNNs (Kathiravelu, pg. 5, section 4.1)—on medical imaging data: “we presented Niffler, a framework that supports the seamless transfer of data from the PACS to the research clusters and enables efficient execution of ML pipelines on the images, reports, and the extracted textual metadata. Niffler facilitates the execution of ML models with a minimal tuning of infrastructure” (Kathiravelu, Pg. 1, section 1) . 07-21-aia AIA Claim 4 i s re jected under 35 U.S.C. 103 as being unpatentable over Pr abhu in view of Isakov , and further in view of Be rdouz Qrichi Aniba (hereinafter Berdouz), U.S. Patent Application Publication US 20240056471 A1. Regarding Claim 4, Prabhu and Isakov teach The method of claim 1, as shown above. Prabhu and Isakov do not appear to explicitly disclose wherein the processing task comprises obtaining financial information via a network module, processing the financial information using the trained feedforward neural network, and detecting and mitigating financial fraud based on results of the processing of the financial information. However, Berdouz teaches wherein the processing task comprises obtaining financial information via a network module, processing the financial information using the trained feedforward neural network, and detecting and mitigating financial fraud based on results of the processing of the financial information. ([0002]: “The invention belongs to the field of automated fraud prediction and detection, and finds various applications in systems wherein frauds can occur, for example financial systems…” [0008]: “To this end, the invention proposes, according to one aspect, a method of automated detection of the risk of fraud in a monitored system, from data streams generated by said monitored system…” [0036]: “The monitored system generates data streams 4 1 to 4 k , which are recorded in one or a plurality of electronic memory units 6, for example in the form of a database or any other data storage structure. For example, the unit 6 consists of a plurality of interconnected memories.” [0047]: “The data streams 4 l , stored over associated periods of time T l are accessible to an automated fraud detection system 10 according to the invention.” [0095]: “The initialization step 50 is followed by substantially concomitant steps 52 and 54, which are a step 52 of application of a first parameterized process for the estimation of the risk of fraud, thereafter called ‘process 1’…” [0118]: “For example, in one embodiment, the ‘process 1’ implements a Convolutional Neural Network (CNN)…” [0127]: “In the case where the operator identified by Id_M is presumed to be a fraudster, the step 64 is followed by a step 68 for raising an alert, e.g. by sending a notification, or any other means, to the managing authorities of the monitored system.” Data streams generated by the monitored system (i.e. financial information) are obtained from the memory units (i.e. via a network module) and processed by ‘process 1’ implementing a CNN (i.e. trained feedforward neural network) to detect fraud and raise an alert (i.e. detect and mitigate the fraud).) It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Prabhu, Isakov, and Berdouz. Prabhu teaches enforcing neural network sparsity prior to training via a sparse but well-connected topology based on uniform and deterministic Cayley expander graphs. Isakov teaches enforcing neural network sparsity prior to training by replacing fully connected layers with sparse cascades, where the sparse layers are generated by applying a sparse mask to the fully connected layers. Berdouz teaches financial fraud detection and mitigation using ML models including neural networks. One of ordinary skill would have motivation to combine Prabhu, Isakov, and Berdouz because Prabhu provides efficient and accurate CNNs: “Such networks are expected to preserve accuracy (due to connectivity) while being runtime efficient (due to the sparsity). We empirically demonstrate this in the later sections” (Prabhu, pg. 4, section 3), and Berdouz’s framework uses neural networks—including CNNs (Berdouz, 0118)—to perform financial fraud detection more reliably than other methods: “However, it has been found that most prior art fraud detection methods provide too high false-positive rates, which have a negative impact on the system in terms of performance and customer experience and do not guarantee that all false-negatives have been detected. One of the goals of the invention is to remedy such drawback by proposing a more reliable method” (Berdouz, 0006-0007) . 07-21-aia AIA Claim s 8, 9, 12, 17, 18, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Prabhu in view of Isakov , and further in view of Cohen , U.S. Patent Application Publication US 20210004344 A1. Regarding Claim 8, Prabhu and Isakov teach The method of claim 1, as shown above. Prabhu and Isakov do not appear to explicitly disclose the remaining features of claim 8. However, Cohen teaches wherein the sparse topology comprises a network of ring configurations of nodes in two dimensions or surface configurations of nodes in greater than two dimensions, the ring configurations and surface configurations having variable connectivity between the corresponding nodes, and wherein a copy of information flows from a first node in a first direction through a first portion of a given ring configuration or a given surface configuration of the trained feedforward neural network and in a second direction through a second portion of the given ring or the given surface to meet at a second node of the given ring or the given surface. ([0022]: “FIGS. 27, 28, 29, 30, 31, and 32 show varied message passing interfaces for neural networks derived from polyhedral clusters which comprise a centroid node.” In the neural network topology shown in figure 27, the nodes of layers B, C, and D form ring configurations, with copies of information flowing from each B node (i.e. first node) to both an upper C node (i.e. in a first direction through a first portion) and a lower C node (i.e. in a second direction through a second portion) before meeting at a D node (i.e. second node) at the rightmost corner of the ring. B and D nodes have degree 3, while C nodes have degree 2 (i.e. the nodes of the rings have variable connectivity).) It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Prabhu, Isakov, and Cohen. Prabhu teaches enforcing neural network sparsity prior to training via a sparse but well-connected topology based on uniform and deterministic Cayley expander graphs. Isakov teaches enforcing neural network sparsity prior to training by replacing fully connected layers with sparse cascades, where the sparse layers are generated by applying a sparse mask to the fully connected layers. Cohen teaches a scalable, high dimensional, lattice-based routing topology for neural network processing. One of ordinary skill would have motivation to combine Prabhu, Isakov, and Cohen because “The processor configurations enable function-follows-form computing. Their computing benefits include…increased number of interconnects among neighboring nodes, which offers improved neural network computing” (Cohen, 0003). Regarding Claim 9, Prabhu and Isakov teach The method of claim 1, as shown above. Prabhu and Isakov do not appear to explicitly disclose the remaining features of claim 9. However, Cohen teaches wherein at least one ring configuration of nodes of the trained feedforward neural network has a different number of nodes than another configuration ring of nodes of the trained feedforward neural network. ([0022]: “FIGS. 27, 28, 29, 30, 31, and 32 show varied message passing interfaces for neural networks derived from polyhedral clusters which comprise a centroid node.” In the neural network topology shown in figure 27, there exist different sizes of ring configurations. Information starting in a B node flows through an upper C node and a lower C node and meets back at a D node, forming a small ring configuration of 4 nodes. Information starting at the input A node flows through an upper B-C-D path and a lower B-C-D path and meets back at the output A node, forming a larger ring configuration of 8 nodes.) Regarding Claim 12, Prabhu and Isakov teach The method of claim 1, as shown above. Prabhu and Isakov do not appear to explicitly disclose the remaining features of claim 12. However, Cohen teaches wherein, in the operation of generating the sparse topology for the feedforward neural network, the sparse topology comprises a network of surface configurations of nodes in greater than two dimensions. ([0003]: “The architecture scales in a lattice model. Therefore within each cluster, the routers are capable of routing messages in hypercube topologies of at least up to six dimensions, and continue by extension to the next cluster on the scaling lattice.” Each cluster is a surface configuration in more than two dimensions, and lattice-based scaling creates a network of these clusters.) Claims 17, 18, and 20 are system claims containing substantially the same elements as method claims 8, 9, and 12, respectively. Prabhu, Isakov, and Cohen teach the elements of claims 8, 9, and 12, as shown above . 07-21-aia AIA Claim s 10 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Prabhu in view of Isakov and Cohen , and further in view of Boguñá et al. (hereinafter Boguñá), “Generalized percolation in random directed networks”. Regarding Claim 10, Prabhu, Isakov, and Cohen teach The method of claim 9, as shown above. Prabhu, Isakov, and Cohen do not appear to explicitly disclose the remaining features of claim 10. However, Boguñá teaches wherein bidirectional connections between two nodes of the trained feedforward neural network enable an exchange of information that allows for a mixing of information that reaches different regions of the trained feedforward neural network. (Pg. 1, section I: “In this paper we present a general theory for percolation in directed random networks with general two point correlations and bidirectional links.” Pg. 2, section II: “The giant connected component in undirected graphs becomes internally structured in the case of directed networks so that four different types of giant components may arise. Whether giant or not, these components are characterized as follows… The strongly connected component, SCC, the set of vertices reachable from its every vertex by a directed path.” Pg. 7, section VI: “In this case, we have also shown that bidirectional edges act as a catalyst for percolation, favoring the emergence of the GSCC [giant strongly connected component]…” Bidirectional edges between nodes of a directed network facilitate the emergence of strong connectedness (i.e. mixing of information that reaches different regions of the network).) It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Prabhu, Isakov, Cohen, and Boguñá. Prabhu teaches enforcing neural network sparsity prior to training via a sparse but well-connected topology based on uniform and deterministic Cayley expander graphs. Isakov teaches enforcing neural network sparsity prior to training by replacing fully connected layers with sparse cascades, where the sparse layers are generated by applying a sparse mask to the fully connected layers. Cohen teaches a scalable, high dimensional, lattice-based routing topology for neural network processing. Boguñá teaches the impacts of bidirectional edges on percolation in directed networks. One of ordinary skill would have motivation to combine Prabhu, Isakov, Cohen, and Boguñá because “well connectedness can preserve the expressive power of the CNNs” (Prabhu, pg. 1, abstract), and “bidirectional edges act as a catalyst for percolation, favoring the emergence of the GSCC [giant strongly connected component]” (Boguñá, pg. 7, section VI). Claim 19 is a system claim containing substantially the same elements as method claim 10. Prabhu, Isakov, Cohen, and Boguñá teach the elements of claim 10, as shown above. Conclusion 07-40 AIA Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL . See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a). A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action. Any inquiry concerning this communication or earlier communications from the examiner should be directed to BENJAMIN M ROHD whose telephone number is (571)272-6445. The examiner can normally be reached Mon-Thurs 8:00-6:00 EST. Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Viker Lamardo can be reached at (571) 270-5871. 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. /B.M.R./Examiner, Art Unit 2147 /VIKER A LAMARDO/Supervisory Patent Examiner, Art Unit 2147 Application/Control Number: 18/085,555 Page 2 Art Unit: 2147 Application/Control Number: 18/085,555 Page 3 Art Unit: 2147 Application/Control Number: 18/085,555 Page 4 Art Unit: 2147 Application/Control Number: 18/085,555 Page 5 Art Unit: 2147 Application/Control Number: 18/085,555 Page 6 Art Unit: 2147 Application/Control Number: 18/085,555 Page 7 Art Unit: 2147 Application/Control Number: 18/085,555 Page 8 Art Unit: 2147 Application/Control Number: 18/085,555 Page 9 Art Unit: 2147 Application/Control Number: 18/085,555 Page 10 Art Unit: 2147 Application/Control Number: 18/085,555 Page 11 Art Unit: 2147 Application/Control Number: 18/085,555 Page 12 Art Unit: 2147 Application/Control Number: 18/085,555 Page 13 Art Unit: 2147 Application/Control Number: 18/085,555 Page 14 Art Unit: 2147 Application/Control Number: 18/085,555 Page 15 Art Unit: 2147 Application/Control Number: 18/085,555 Page 16 Art Unit: 2147 Application/Control Number: 18/085,555 Page 17 Art Unit: 2147 Application/Control Number: 18/085,555 Page 18 Art Unit: 2147 Application/Control Number: 18/085,555 Page 19 Art Unit: 2147 Application/Control Number: 18/085,555 Page 20 Art Unit: 2147 Application/Control Number: 18/085,555 Page 21 Art Unit: 2147 Application/Control Number: 18/085,555 Page 22 Art Unit: 2147 Application/Control Number: 18/085,555 Page 23 Art Unit: 2147 Application/Control Number: 18/085,555 Page 24 Art Unit: 2147 Application/Control Number: 18/085,555 Page 25 Art Unit: 2147 Application/Control Number: 18/085,555 Page 26 Art Unit: 2147 Application/Control Number: 18/085,555 Page 27 Art Unit: 2147 Application/Control Number: 18/085,555 Page 28 Art Unit: 2147 Application/Control Number: 18/085,555 Page 29 Art Unit: 2147 Application/Control Number: 18/085,555 Page 30 Art Unit: 2147 Application/Control Number: 18/085,555 Page 31 Art Unit: 2147
Prosecution Insights