DETAILED ACTION
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
This office action is in response to the amendment filed on 11/21/2025. Claims remain pending in the application. Claims 1, 17, and 19 are independent.
Applicant's amendment to the specification corrects previous drawing and specification objections; therefore, the previous drawing and specification objections are withdrawn.
Applicant's amendment to claims corrects previous 112 rejections; therefore, the previous 112 rejections are withdrawn.
Response to Arguments
Applicant’s arguments regarding 101 rejections, see Pages 8-9 of the Remarks, filed on 11/21/2025, with respect to Claim have been fully considered and are persuasive. The 101 rejections of Claims 1-3 and 5-20 have been withdrawn.
Applicant's arguments regarding 102/103 rejections filed on 11/21/2025 have been fully considered but they are not persuasive.
Applicant argues on Pages 10-12 of the Remarks that Zhou, either alone or in combination with Hewage, fails to disclose, inter alia, ''wherein training the machine-learning model comprises clustering the training dataset into a plurality of clusters, wherein the non-smoothness penalization function associates a weight with each of the plurality of clusters, and adaptively learning the weights for the plurality of clusters".
In response, examiner respectfully disagrees. Zhou discloses in ¶ [0012] that consider a general ranking function that depends on incoming paths of various lengths weighted by some chosen damping function that decreases with distance; i.e., links from pages that are a greater distance from the subject web page are weighted by weight that is damped less than links from closer web pages. Zhou further discloses in ¶¶ [0029]-[0032] that (1) training a classifier is performed by generating a smooth classification function based on the probabilities 120 and 122 over the detected graph 104; (2) the closeness between the classification function and the known values can be measured in a variety of different ways; (3) the classification function is not only close to known values at known nodes, but it is relatively smooth in that it changes relatively slowly on densely connected subgraphs; i.e., the nodes that reside close to one another on the subgraph may likely have values which are relatively close to one another; (4) if they are known to be one spam node and one normal node, respectively, then the classification function changes by a large amount between those nodes, but this lack of smoothness is penalized in the chosen cost function that is optimized; (5) because pages that are relatively close to one another on the directed graph are assumed to be the same type (pages close to a known spam page are likely to be spam pages, while pages close to a known normal page are likely to be normal pages) by making the function smooth and relatively slow changing pages in the directed subgraph that are close to a known normal content page will have classification function values that more likely indicate it to be a normal content page; and (6) similarly, those pages in the directed subgraph that are close to a spam page will have classification function values that are likely to indicate that it will be a spam page. In other words, Zhou teaches different closeness between nodes will have different weights to penalize non-smoothness. Zhou also discloses in ¶¶ [0039]-[0041] that (1) a graph G is weighted if it is associated with a function w; (2) Let G=(V, E, w) denote a weighted directed graph; (3) a path is a tuple of vertices with the property that (vi, vi+1) [Symbol font/0xCE] E for 1 [Symbol font/0xA3] i [Symbol font/0xA3] p-1; and (4) for a strongly connected graph, there is an integer k [Symbol font/0xB3] 1 and a unique partition V=V0 U V1 ... U Vk-1 such that for all 0 [Symbol font/0xB3] r [Symbol font/0xB3] k-1 each edge (u, v) [Symbol font/0xCE] E with u [Symbol font/0xCE] V, has v[Symbol font/0xCE]Vr+1, where Vk= V0; and k is maximal, that is, there is no other such partition with k > k. It is also well-known in the art that partitioning, segmenting, grouping, or clustering nodes/data are based on distance/closeness between nodes/data (i.e., similarity measures). Therefore, Zhou teaches "wherein the non-smoothness penalization function associates a weight with each of the plurality of clusters, and adaptively learning the weights for the plurality of clusters". Zhou only fails to explicitly disclose "clustering the training dataset into a plurality of clusters". Hewage teaches this this deficiency of Zhou in ¶¶ [0022]. [0027]-[0029], [0033]-[0034], and [0123] with FIG. 2a that (1) regularising the style vector z further comprises training an encoder network of the autoencoder with input training data to enforce a selected probability distribution on at least a portion of the style vector z; (2) regularising the style vector z during training causes the label vectors y associated with input data to form multiple or two or more clusters of label vectors y, wherein each contains a subgroup of label vectors y that are substantially the same or similar, and the set of label vectors y are substantially time invariant; (3) each cluster is defined by a region or boundary and the subgroup of label vectors y for each cluster are contained within the defined region or boundary, and label vectors y are substantially the same or similar when they are contained within the region or boundary of the same cluster, wherein the cluster relates to a true state or class label; (4) clustering the set of label vectors y to form multiple clusters of label vectors y in which each cluster contains a subgroup of label vectors y that are substantially the same or similar; and mapping each of the clusters of label vectors y to a class or state label from a set of class or state labels S associated with the input data for use by the trained autoencoder in classifying input data; (5) outputting the generator loss function value associated with label vector y for use by the autoencoder in training an encoder network to enforce the categorical distribution on the label vector y; (6) the autoencoder further comprising a decoding network coupled to the latent representation layer, wherein the training set of input data comprises a training set of neurological sample vector sequences in which Lk is the length of the k-th neurological sample vector sequence and T is the number of training neurological sample vector sequences, for each k-th neurological sample vector sequence corresponding to a k-th neural activity that is passed through the autoencoder; (7) generating a loss or cost function based on the output of the one or more regularising networks and/or the adversarial network, an estimate of k-th neurological sample vector sequence output from the decoding network, the original k-th neurological sample vector sequence; and updating the weights of the hidden layer(s) of the encoding network and/or decoding network based on the generated loss of cost function; and (8) the training set may be generated from a collected set of unlabelled neurological sample vector sequences using autoencoder 200 as a classifier that outputs, from encoder network 202a, the label vector y 206 (e.g. this may be a soft vector) for each of the input neurological sample vector sequences 201a. In other words, the clustering is performed within the encoder network 202a, and the label vector y 206 (for each of the input neurological sample vector sequences 201a) outputted from the encoder network 202a has already been clustered according to similarity measures. Therefore, Hewage DOES NOT perform clusters after the encoder network 202a as described in Page 10 of the Remarks. Therefore, the combination of Zhou and Hewage teaches "wherein training the machine-learning model comprises clustering the training dataset into a plurality of clusters, wherein the non-smoothness penalization function associates a weight with each of the plurality of clusters, and adaptively learning the weights for the plurality of clusters". Also, this limitation "the non-smoothness penalization function associates a weight with each of the plurality of clusters" is quite common in the art (see a list of prior art included in the Conclusion section).
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-3, 5-7, and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Zhou et al. (US 2008/0275833 A1, pub. date: 11/06/2008), hereinafter Zhou in view of HEWAGE et al. (US 2020/0387798 A1, pub. date: 12/10/2020), hereinafter HEWAGE.
Independent Claims 1 and 17
Zhou discloses a method of manifold regularization while training a machine-learning model (Zhou, ¶¶ [0029], [0045], and [0064]-[0065] with FIG. 1: training a classifier is performed by generating a smooth classification function based on the probabilities 120 and 122 over the detected graph 104; a number of discrete operators on directed graphs are discrete analogs of the corresponding differential operators on Riemannian manifolds; the classification algorithm for directed graphs is derived from the discrete regularization; for classifying those unclassified vertices in Sc, define a discrete regularization in Eq. 17, where C>0 is the regularization parameter), the method comprising using at least one hardware processor (Zhou, ¶¶ [0082] and [0084] with 320 in FIG. 6: multiprocessor; processing unit 320) to:
acquire a training dataset that comprises a plurality of feature sets, wherein each of the plurality of feature sets comprises a feature value for each of a plurality of features and is labeled with a target value for each of one or more targets (Zhou, ¶ [0013]: features must be selected that are useful in detecting spam, and each web page is then represented as a vector having each element described by one type of spam feature; the features can be the number of inlinks, the number of outlinks, scores under any of the above-mentioned algorithms, etc.; ¶¶ [0023]-[0024], [0027]-[0028], and [0034]-[0036] with 114, 102, 116, 118, 104, 120, and 122 in FIG. 1, 152-154 in FIG.2, and FIG. 3: the collection of web pages 114 is considered a directed graph, in that the web pages themselves in collection 114 are nodes in the graph while hyperlinks between those web pages are directed edges in the directed graph; for applying at the domain/host level, the domains/hosts are nodes in the graph and hyperlinks among the web pages in the domains/hosts are the directed edges; the web page collection 114 is also provided to trusted entity 102, such as a human expert, in identifying link spam; trusted entity 102 then identifies some examples of spam web pages in web page collection 114 as spam training examples 116. Trusted entity 102 also identifies good web pages (or normal web pages) in web page collection 114 as normal training examples 118; random walk component 106 then receives a definition of a random walk on directed graph 104 (or each strongly connected component in directed graph 104), the random walk being defined by transition probabilities; based on the defined random walk, component 106 obtains stationary probabilities associated with each node in directed graph 104; the “transition probabilities' are the probabilities of transitioning from any given node on graph 104 to another node; the stationary probability distributions are the probabilities of being in any given node on directed graph 104);
generate an optimization problem comprising an objective function and a non-smoothness penalization function, wherein the objective function calculates an estimated error between the target values and corresponding output values that the machine-learning model outputs for the plurality of feature sets, and wherein the non-smoothness penalization function is configured to increase the estimated error in the optimization problem as a smoothness of the machine-learning model decreases (Zhou, ¶¶ [0029]-[0032] and [0039]-[0080] in generating the classification function, the values of the classification function are forced to be close to known values for examples 116 and 118; in other words, the values of the classification function are forced to be close to the values that indicate spam and normal pages at the nodes in the graph that are actually known to be spam and normal pages as identified by the trusted entity 102; the closeness between the classification function and the known values can be measured in a variety of different ways; e.g. the closeness can be measured using least square loss, hinge loss, precision/recall measure, the F1-score, the ROC score, or the AUC score; the classification function is not only close to known values at known nodes, but it is relatively smooth in that it changes relatively slowly on densely connected subgraphs; in other words, the nodes that reside close to one another on the subgraph may likely have values which are relatively close to one another; however, if they are known to be one spam node and one normal node, respectively, then the classification function changes by a large amount between those nodes, but this lack of smoothness is penalized in the chosen cost function that is optimized; i.e., the classification function value can change abruptly, if necessary; again, however, this is penalized; given a directed graph G=(V, E, w), and a discrete label set L={-1, 1}, the vertices in a subset S [Symbol font/0xCD] V have labels in L; the task is to predict the labels of those unclassified vertices in Sc, the complement of S; define a function y with y(v)=1 or -1 if v[Symbol font/0xCE]S, and 0 if v[Symbol font/0xCE]Sc; for classifying those unclassified vertices in Sc, define a discrete regularization in Eq. 17, where C>0 is the regularization parameter; when choosing the basis, Eq. 17 can be written as Eq. 18; in the objective function, the first term forces the classification function to be relatively smooth, and perhaps as smooth as possible and the second term forces the classification function to fit the given labels as well as possible; the first term makes the function relatively smooth over all nodes while the second term forces the function to fit the labeled nodes to a desired closeness; a random walk over a given directed graph can be defined in many different ways; three exemplary types of random walk used in spam detection are: (a) following outlinks uniformly at random; (b) following links uniformly at random regardless of directionality; (c) following inlinks uniformly at random; assigning values to the nodes in directed graph 104 basically requires selection of a random walk definition (transition probabilities) and solving Eq. 18 above for each of the nodes; to solve the optimization problem in Eq. 18, differentiate the objective function with respect to s and then obtain Eq. 23; the classification is based on the sign of the function value on each vertex, which is equivalent to setting the classification threshold to 0; Claims 2-3, 12-13, and 15: optimizing a cost function that assigns a penalty based on a size of a difference in classification function values between nodes in the directed graph; optimizing the cost function that assigns a penalty based on a difference between classification function values assigned to nodes representing the training examples and the target function values for those nodes); and
train the machine-learning model by adjusting the machine-learning model to minimize the estimated error, produced by the training dataset, in the optimization problem (Zhou, ¶ [0013]: a classifier is chosen, such as a neural network, a decision tree, a support vector machine (SVM), etc., and it is trained with a set of examples of normal and spam web pages which have been judged by human experts; ¶¶[0029]-[0030] with 108 in FIG. 1 and 156 in FIG. 2: once examples 116 and 118 and probabilities 120 and 122 are obtained, classifier training component 108 trains a classifier that can be used in link spam detection; training a classifier is performed by generating a smooth classification function based on the probabilities 120 and 122 over the detected graph 104; in generating the classification function, the values of the classification function are forced to be close to known values for examples 116 and 118; in other words, the values of the classification function are forced to be close to the values that indicate spam and normal pages at the nodes in the graph that are actually known to be spam and normal pages as identified by the trusted entity 102; the closeness between the classification function and the known values can be measured in a variety of different ways; e.g. the closeness can be measured using least square loss, hinge loss, precision/recall measure, the F1-score, the ROC score, or the AUC score; Claim 9: a classifier training component configured to receive a first set of training pages labeled as normal pages and a second set of training pages labeled as spam pages and to train a web page classifier based on both the first set of training pages and the second set of training pages), wherein training the machine-learning model comprises (Zhou, ¶ [0012]: consider a general ranking function that depends on incoming paths of various lengths weighted by some chosen damping function that decreases with distance; i.e., links from pages that are a greater distance from the subject web page are weighted by weight that is damped less than links from closer web pages; ¶¶ [0029]-[0033]: training a classifier is performed by generating a smooth classification function based on the probabilities 120 and 122 over the detected graph 104; the closeness between the classification function and the known values can be measured in a variety of different ways; e.g., the closeness can be measured using least square loss, hinge loss, precision/recall measure, the F1-score, the ROC score, or the AUC score, as examples; the classification function is not only close to known values at known nodes, but it is relatively smooth in that it changes relatively slowly on densely connected subgraphs; i.e., the nodes that reside close to one another on the subgraph may likely have values which are relatively close to one another; if they are known to be one spam node and one normal node, respectively, then the classification function changes by a large amount between those nodes, but this lack of smoothness is penalized in the chosen cost function that is optimized; because pages that are relatively close to one another on the directed graph are assumed to be the same type (pages close to a known spam page are likely to be spam pages, while pages close to a known normal page are likely to be normal pages) by making the function smooth and relatively slow changing pages in the directed subgraph that are close to a known normal content page will have classification function values that more likely indicate it to be a normal content page; similarly, those pages in the directed subgraph that are close to a spam page will have classification function values that are likely to indicate that it will be a spam page; i.e., different closeness between nodes (i.e., different partitions/segments/groups/clusters) will have different weights to penalize non-smoothness; ¶¶ [0039]-[0080]: a path is a tuple of vertices with the property that (vi, vi+1) [Symbol font/0xCE] E for 1 [Symbol font/0xA3] i [Symbol font/0xA3] p-1; for a strongly connected graph, there is an integer k [Symbol font/0xB3] 1 and a unique partition V=V0 U V1 ... U Vk-1 such that for all 0 [Symbol font/0xB3] r [Symbol font/0xB3] k-1 each edge (u, v) [Symbol font/0xCE] E with u [Symbol font/0xCE] V, has v[Symbol font/0xCE]Vr+1, where Vk= V0; and k is maximal, that is, there is no other such partition with k > k; a graph G is weighted if it is associated with a function w; Let G=(V, E, w) denote a weighted directed graph; the function w is called the weight function of G; in the objective function, the first term forces the classification function to be relatively smooth, and perhaps as smooth as possible and the second term forces the classification function to fit the given labels as well as possible).
Zhou further discloses a system (Zhou, ¶¶ [0081]-[0082] and [0084] with 300/310 in FIG. 6: a suitable computing system environment 300; multiprocessor systems, micro
processor-based systems; a computer 310) comprising: at least one hardware processor (Zhou, ¶¶ [0082] and [0084] with 300/310 in FIG. 6: multiprocessor; a processing unit 320); and software (Zhou, ¶ [0086] with 334-337 in FIG. 6: operating system 334, application programs 335, other program modules 336, and program data 337) configured to, when executed by the at least one hardware processor, perform the method described above (Zhou, ¶¶ [0083] and [0086] with FIG. 6: computer-executable instructions, such as program modules, being executed by a computer; RAM 332 typically contains data and/or program modules that are immediately accessible to and/or presently being operated on by processing unit 320.).
Zhou fails to explicitly disclose clustering the training dataset into a plurality of clusters.
HEWAGE teaches a system and a method relating to classification (HEWAGE, ¶ [0001]), wherein clustering the training dataset into a plurality of clusters (HEWAGE, ¶¶ [0022] and [0027]-[0029]: regularising the style vector z further comprises training an encoder network of the autoencoder with input training data to enforce a selected probability distribution on at least a portion of the style vector z; regularising the style vector z during training causes the label vectors y associated with input data to form multiple or two or more clusters of label vectors y, wherein each contains a subgroup of label vectors y that are substantially the same or similar, and the set of label vectors y are substantially time invariant; each cluster is defined by a region or boundary and the subgroup of label vectors y for each cluster are contained within the defined region or boundary, and label vectors y are substantially the same or similar when they are contained within the region or boundary of the same cluster, wherein the cluster relates to a true state or class label; clustering the set of label vectors y to form multiple clusters of label vectors y in which each cluster contains a subgroup of label vectors y that are substantially the same or similar; and mapping each of the clusters of label vectors y to a class or state label from a set of class or state labels S associated with the input data for use by the trained autoencoder in classifying input data; ¶¶ [0033]-[0034]: outputting the generator loss function value associated with label vector y for use by the autoencoder in training an encoder network to enforce the categorical distribution on the label vector y; the autoencoder further comprising a decoding network coupled to the latent representation layer, wherein the training set of input data comprises a training set of neurological sample vector sequences in which Lk is the length of the k-th neurological sample vector sequence and T is the number of training neurological sample vector sequences, for each k-th neurological sample vector sequence corresponding to a k-th neural activity that is passed through the autoencoder; generating a loss or cost function based on the output of the one or more regularising networks and/or the adversarial network, an estimate of k-th neurological sample vector sequence output from the decoding network, the original k-th neurological sample vector sequence; and updating the weights of the hidden layer(s) of the encoding network and/or decoding network based on the generated loss of cost function; ¶¶ [0123] and [0176] with FIGS. 2a and 3: the training set may be generated from a collected set of unlabelled neurological sample vector sequences using autoencoder 200 as a classifier that outputs, from encoder network 202a, the label vector y 206 (e.g. this may be a soft vector) for each of the input neurological sample vector sequences 201a; this may produce a set of label vectors y which can be mapped to a plurality of true states or classes associated with the bodily variables encoded in the neural activity; e.g., the set of label vectors y 206 may be used to determine the bodily variable labels (e.g. true state or classes) by observing whether the set of label vectors y 206 form cluster regions, in which each cluster region may be labelled with a bodily variable label; the bodily variable label for each cluster region may be identified by, firstly, comparing each of the neural activities of (e.g. automatically analysed) that generate the label vectors y 206 within the cluster region with corresponding sensor data (e.g. video, audio, motion tracking, blood, heart rate etc.) recorded/stored/collected at the same time the multichannel neurological signal sample vector sequences were recorded/stored/sampled and collected. This is used to analyse the neural activity and corresponding sensor data associated with said cluster region and determining a bodily variable label based on the analysed neural activity and corresponding sensor data; thus a mapping from the cluster region to the bodily variable label or true state/classes may be generated and used for classifying label vectors y in accordance with the bodily variable labels or true states/classes etc.; ¶ [0175]: the label vector y 206 may include a vector, a tensor or otherwise, where the vector, tensor or otherwise includes at least one or more from the group of: a one hot vector; a measure of entropy; be regularized to L1 L2 or both, or other norms; a discrete boltzmann distributed vector; a representation of a prior class state; a known feature or configuration set; L1 and L2 are the well-known "distance" measures or measure of total value of all the elements in a vector; to penalise or regularise using this measure is to minimise this measure; ¶¶ [0177]-[0192] with FIG. 4a: a multiple of clusters (or two or more clusters) may be determined based on the output set of label vectors y 206; this may involve detecting whether each of the clusters contains a subgroup of label vectors y 206 that are substantially the same or similar; each cluster may be defined by a region or boundary and the subgroup of label vectors y 206 for each cluster are contained within the defined region or boundary, and label vectors y 206 are substantially the same or similar when they are contained within the region or boundary of the same cluster; clustering the set of label vectors y to form multiple clusters of label vectors y in which each cluster contains a subgroup of label vectors y that are substantially the same or similar, and mapping each of the clusters of label vectors y to a class or state label from a set of class or state labels S associated with the input data for use by an autoencoder defined by the set of hyperparameters in classifying input data; the latent vector comprising a label vector y 206 and a style vector z 208 where the autoencoder is configured based on data representative of the weights and/or parameters of one or more neural network(s) and/or hidden layer(s) associated with the trained autoencoder of the set of autoencoder configuration data, wherein the trained autoencoder regularized the style vector z 208 and outputs substantially time invariant label vector(s) y 206).
Zhou and HEWAGE are analogous art because they are from the same field of endeavor, a system and a method relating to classification. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of HEWAGE to Zhou. Motivation for doing so would improve the robustness of ML classification techniques (HEWAGE, ¶¶ [0013]- [0017]).
Claim 2
Zhou in view of HEWAGE discloses all the elements as stated in Claim 1 and further discloses wherein the non-smoothness penalization function comprises a Laplacian norm of the output values (Zhou, ¶¶ [0039]-[0080]: the discrete Laplacian Δ: H(V)->H(V) is defined by: -½ div·[Symbol font/0xD1] (Eq. 10); Eq. 14 has been widely used to define the Laplacian for an undirected graph; matrix in Eq. 16 has been used to define Laplacian for directed graphs; given a directed graph G=(V, E, w), and a discrete label set L={-1, 1}, the vertices in a subset S [Symbol font/0xCD] V have labels in L; the task is to predict the labels of those unclassified vertices in Sc, the complement of S; define a function y with y(v)=1 or -1 if v[Symbol font/0xCE]S, and 0 if v[Symbol font/0xCE]Sc; for classifying those unclassified vertices in Sc, define a discrete regularization in Eq. 17, where C>0 is the regularization parameter; when choosing the basis, Eq. 17 can be written as Eq. 18; in the objective function, the first term (e.g.,
∇
φ
H
(
E
)
2
) forces the classification function to be relatively smooth, and perhaps as smooth as possible).
Claim 3
Zhou in view of HEWAGE discloses all the elements as stated in Claim 1 and further discloses wherein the non-smoothness penalization function approximates a gradient of the machine-learning model on a data manifold of the machine-learning model (Zhou, ¶¶ [0039]-[0080]: the discrete gradient [Symbol font/0xD1]: H(V)->H(E) is defined as an operator in Eq. 7; the discrete Laplacian Δ: H(V)->H(V) is defined by: -½ div·[Symbol font/0xD1] (Eq. 10); Eq. 14 has been widely used to define the Laplacian for an undirected graph; matrix in Eq. 16 has been used to define Laplacian for directed graphs; given a directed graph G=(V, E, w), and a discrete label set L={-1, 1}, the vertices in a subset S [Symbol font/0xCD] V have labels in L; the task is to predict the labels of those unclassified vertices in Sc, the complement of S; define a function y with y(v)=1 or -1 if v[Symbol font/0xCE]S, and 0 if v[Symbol font/0xCE]Sc; for classifying those unclassified vertices in Sc, define a discrete regularization in Eq. 17, where C>0 is the regularization parameter; when choosing the basis, Eq. 17 can be written as Eq. 18; in the objective function, the first term (e.g.,
∇
φ
H
(
E
)
2
) forces the classification function to be relatively smooth, and perhaps as smooth as possible).
Claim 5
Zhou in view of HEWAGE discloses all the elements as stated in Claim 1 and further discloses wherein adaptively learning the weights for the plurality of clusters comprises, for each of a plurality of outer iterations: determining the weights for the plurality of clusters; and for each of a plurality of inner iterations, adjusting the machine-learning model to minimize the estimated error, produced by the training dataset, in the optimization problem, using the determined weights for the plurality of clusters in the non-smoothness penalization function (HEWAGE, ¶¶ [0177]-[0192] with FIGS. 4a and 2d: controlling the regularisation of style vector z and thus selecting suitable hyperparameters for use with the example ML technique(s); one or more probability distribution(s) 214a-214n and corresponding one or more vector(s) 208a-208n partitioning the style vector z 208 may be selected in which the regularisation of style vector z 208 improves or increases the time invariance of label vector(s) y 206 and/or ensures that the label vector(s) y 206 are substantially time invariant or even time invariant; if it is considered that the set of label vectors y 206 are clustered (e.g. 'Y'), then the method 400 proceeds to step 414, otherwise (e.g. 'N') it proceeds to step 404 to select another set of hyperparameters; in step 414, in response to detecting that each cluster contains a subgroup of label vectors y 206 that are substantially the same or similar, the method proceeds to detect whether the set of label vectors y 206 are substantially time invariant; in step 416, in response to detecting that each cluster contains a subgroup of label vectors y 206 that are substantially the same or similar and detecting that the set of label vectors y 206 are substantially time invariant, then the selected set of hyperparameters are considered to be a set of hyperparameters that may be stored in an optimised hyperparameter dataset; the set of hyperparameters may further include one or more from the group of: autoencoder size, wherein the autoencoder size comprises a length of the encoder state; initial learning rate or decay; batch size, wherein batch size comprises the number of samples and defines the update of weights or parameters of the autoencoder neural network or hidden layer(s); size of the label vector y; number of classes or states associated with the label vector y; number of hidden layer(s), neural network cells, and/or long short term memory cells; feed size, wherein feed size comprises the number of time steps per data point or batch; loss weighting coefficient, wherein the loss weighting coefficient comprises a relative weighting to give to generative and discriminative losses when the autoencoder uses a discriminator and/or a generator neural network components; optimisation function for optimising the weights of the autoencoder neural network structure(s); type of weight update algorithm or procedure of the weights of the autoencoder neural network structure(s ); learning rate decay factor, wherein learning rate decay factor is used to adjust learning rate when the loss associated with a loss cost function of the autoencoder plateaus or stagnates; and one or more performance checkpoint(s) for determining how often learning rate is to be adjusted) (Zhou, ¶¶ [0029]-[0032] and [0039]-[0080] in generating the classification function, the values of the classification function are forced to be close to known values for examples 116 and 118; in other words, the values of the classification function are forced to be close to the values that indicate spam and normal pages at the nodes in the graph that are actually known to be spam and normal pages as identified by the trusted entity 102; the closeness between the classification function and the known values can be measured in a variety of different ways; e.g. the closeness can be measured using least square loss, hinge loss, precision/recall measure, the F1-score, the ROC score, or the AUC score; the classification function is not only close to known values at known nodes, but it is relatively smooth in that it changes relatively slowly on densely connected subgraphs; in other words, the nodes that reside close to one another on the subgraph may likely have values which are relatively close to one another; however, if they are known to be one spam node and one normal node, respectively, then the classification function changes by a large amount between those nodes, but this lack of smoothness is penalized in the chosen cost function that is optimized; i.e., the classification function value can change abruptly, if necessary; again, however, this is penalized; given a directed graph G=(V, E, w), and a discrete label set L={-1, 1}, the vertices in a subset S [Symbol font/0xCD] V have labels in L; the task is to predict the labels of those unclassified vertices in Sc, the complement of S; define a function y with y(v)=1 or -1 if v[Symbol font/0xCE]S, and 0 if v[Symbol font/0xCE]Sc; for classifying those unclassified vertices in Sc, define a discrete regularization in Eq. 17, where C>0 is the regularization parameter; when choosing the basis, Eq. 17 can be written as Eq. 18; in the objective function, the first term forces the classification function to be relatively smooth, and perhaps as smooth as possible and the second term forces the classification function to fit the given labels as well as possible; the first term makes the function relatively smooth over all nodes while the second term forces the function to fit the labeled nodes to a desired closeness; a random walk over a given directed graph can be defined in many different ways; three exemplary types of random walk used in spam detection are: (a) following outlinks uniformly at random; (b) following links uniformly at random regardless of directionality; (c) following inlinks uniformly at random; assigning values to the nodes in directed graph 104 basically requires selection of a random walk definition (transition probabilities) and solving Eq. 18 above for each of the nodes; to solve the optimization problem in Eq. 18, differentiate the objective function with respect to s and then obtain Eq. 23; the classification is based on the sign of the function value on each vertex, which is equivalent to setting the classification threshold to 0; Claims 2-3, 12-13, and 15: optimizing a cost function that assigns a penalty based on a size of a difference in classification function values between nodes in the directed graph; optimizing the cost function that assigns a penalty based on a difference between classification function values assigned to nodes representing the training examples and the target function values for those nodes).
Claim 6
Zhou in view of HEWAGE discloses all the elements as stated in Claim 5 and further discloses wherein the weights for the plurality of clusters are determined in each of the plurality of outer iterations using a multi-class classification algorithm (HEWAGE, ¶¶ [0194]-[0195] with FIG. 4b: the 5 true states were "stand", "walk", "shuffle", "reverse" and "tum"; FIG. 4b is the t-SNE clustering graph 430 of the set of label vectors y output from the autoencoder 200 after training; as can be seen several clusters 432a, 432b and 434 formed; the clusters 432a and 432b were identified to be associated mainly with the true state "stand" and the cluster 434 was identified to be associated mainly with the true state "shuffle").
Claim 7
Zhou in view of HEWAGE discloses all the elements as stated in Claim 6 and further discloses wherein the machine-learning model is adjusted in each of the plurality of inner iterations using a regression algorithm (HEWAGE, ¶ [0175]: the vector, tensor or otherwise is penalised or regularised by at least one or more from the group of, by way of example only but is not limited to: a probability distribution; L1 L2 or both, or other norms; nuisance variables; and error variables; L1 and L2 are the well-known "distance" measures or measure of total value of all the elements in a vector; to penalise or regularise using this measure is to minimise this measure; ¶¶ [0177]-[0189] with FIG. 4: each set of hyperparameters in the optimized hyperparameter dataset defines an autoencoder structure that can be trained to output substantially time invariant label vector(s) y 206, or label vector(s) y 206 with increased time invariance, by regularising style vector z 208 during training; the autoencoder is configured based on data representative of the weights and/or parameters of one or more neural network(s) and/or hidden layer(s) associated with the trained autoencoder of the set of autoencoder configuration data, wherein the trained autoencoder regularized the style vector z 208 and outputs substantially time invariant label vector(s) y 206).
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over Zhou in view of HEWAGE as applied to Claim 1 above, and further in view of Gambella et al. ("Optimization Problems for Machine Learning: A Survey", ARXIV ID: 1901.05331, 01/16/2019, pp. 1-65), hereinafter Gambella.
Claim 8
Zhou in view of HEWAGE discloses all the elements as stated in Claim 1 and except failing to explicitly disclose prior to acquiring the training dataset, acquire data having a first dimensionality, and embed the data into a lower dimensional space having a second dimensionality that is lower than the first dimensionality, wherein the training dataset is acquired from the data in the lower dimensional space.
Gambella teaches a system and a method relating to present machine learning as optimization models (Gambella, Abstract of Page 1), wherein prior to acquiring the training dataset, acquire data having a first dimensionality, and embed the data into a lower dimensional space having a second dimensionality that is lower than the first dimensionality, wherein the training dataset is acquired from the data in the lower dimensional space (Gambella, Section 2 of Pages 4-6 and Section 2.2 of Pages 7-8: linear regression aims to find a regression function f that expresses the linear relation between input vector and real-valued output Y via the regression coefficients; dimension reduction methods search for M < p linear combinations of the predictors (also called projections); PCA can be used for several data analysis problems which benefit from reducing the problem dimension; PCR has the advantage of including less predictors than the original set and of retaining the variability of the dataset in the derived features; Section 3.3 of Pages 12-13: linear discriminant analysis(LDA) is an approach for classification and dimensionality reduction; it is often applied to data that contains a large number of features (such as image data) where reducing the number of features is necessary to obtain robust classification; the input data is partitioned into K groups where πk denotes the sample set of the k-th class which contains nk data points; LDA maps the features space to a lower dimensional space through a linear transformation).
Zhou in view of HEWAGE, and Gambella are analogous art because they are from the same field of endeavor, a system and a method relating to detecting the occurrence of a fault. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of Gambella to Zhou in view of HEWAGE. Motivation for doing so would obtain robust classification (Gambella, Section 3.3 of Pages 12-13).
Claims 9-13 are rejected under 35 U.S.C. 103 as being unpatentable over Zhou in view of HEWAGE as applied to Claim 1 above, and further in view of Cardoso et al. ("Application of neural-network modules to electric power system fault section estimation", IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 19, NO. 3, JULY 2004, pp. 1034-1041), hereinafter Cardoso.
Claim 9
Zhou in view of HEWAGE discloses all the elements as stated in Claim 1 and except failing to explicitly disclose wherein each of the plurality of feature sets represents a set of measurement signals for a power line, and wherein each target value represents a fault location on the power line.
Cardoso teaches a system and a method relating to detecting the occurrence of a fault (Cardoso, 2nd paragraph of Section I in Page 1034), wherein each of the plurality of feature sets represents a set of measurement signals for a power line, and wherein each target value represents a fault location on the power line (Cardoso, Abstract of Page 1034: a neural system intended to aid the control center operator in the task of fault section estimation; in order to allow the diagnosis task, the protection system philosophy of busbars, transmission lines, and transformers are modeled with the use of two types of neural networks: the general regression neural network and the multilayer perceptron neural network; Section I of Pages 1034-1035: dealing with topological changes and real power systems dimension, besides excessive effort in training procedures in case of neural-network applications; instead of dividing the power system into smaller parts and training a neural network for each set of a few buses, neural networks are employed here to model the protection system philosophy of the main transmission system components (buses, lines, and transformers); the alarms corresponding to the operation of protection relays and CBs are the inputs to MLP neural networks trained with the backpropagation algorithm; the output of MLP neural networks feed general regression neural networks (GRNNs), which conclude whether the equipment is faulted, not faulted, or there is a lack of information; the NN models developed for transformers and lines have additional information, which indicate the direction of external faults; this information can be combined and helps to conclude about which equipment started the occurrence; Section II of Pages 1035 with FIGS.1-3: the MLP nets with backpropagation learning algorithm (Fig. 1) constitute the NNs architecture most commonly used since it is capable to approximate any nonlinear function; GRNNs are feedforward networks that can be used to estimate an output vector Y from a measured vector X; an overview on GRNN architecture is shown in Fig. 2; the input units distribute the variables (input variables) to all neurons in the pattern units; each neuron belonging to the pattern layer corresponds to a sample or a cluster center; when a new vector X is presented to the network, the distance, usually Euclidian, between this vector and each cluster center previously defined and stored is computed; Section III of Pages 1035-1036: after a contingency, only the NNs corresponding to de-energized equipment should be activated; thus, first the fault scenario should be determined; an expert system has been designed for this task; the expert system is a topology tracker, which determines the active elements before and after the fault, using information about the breaker status; the fault scenario set for total number of elements involved in the fault is the difference between active elements before and after the fault; finally, for each element that composes the fault scenario set S, a neural-network module is activated according to its respective alarm inputs; each net would try to classify the element Si ( i-th element of S) into one of its outputs previously determined; Section VI with FIGS. 4-6 and TABLES 2-4 of Pages 1036-1038: the main reason for using MLP together with GRNN is the reduction in the dimension of the neural network, without oversimplifying the protection system; NN_TR determines whether the component is faulty or not, whether the fault is external toward the high-voltage or medium-voltage side, or even whether the information provided to the network is insufficient for the diagnosis; the output of this network determines whether the main protection of the line on the sending or receiving side (MP.S/MP.R) operated, or if the protection indicates the possibility of an external fault to the S (send) side or R (receive); NN_TL determines whether the component is faulty or not, if the fault is external toward the S or R side, or if the information is not sufficient to produce a diagnosis; the output of the network NN_BUS indicates whether the component is faulty or not, or whether there is not enough information for the diagnosis).
Zhou in view of HEWAGE, and Cardoso are analogous art because they are from the same field of endeavor, a system and a method relating to detecting the occurrence of a fault. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of Cardoso to Zhou in view of HEWAGE. Motivation for doing so would open up the possibility for extension to .
Claim 10
Zhou in view of HEWAGE and Cardoso discloses all the elements as stated in Claim 9 and further disclose by a protection device: receiving measurement signals from a plurality of sensors connected to the power line; applying the trained machine-learning model to the measurement signals to estimate a fault location on the power line; comparing the estimated fault location to a threshold; and when the estimated fault location satisfies the threshold, tripping at least one circuit breaker to electrically isolate the fault location on the power line (Cardoso, Abstract of Page 1034: a neural system intended to aid the control center operator in the task of fault section estimation; in order to allow the diagnosis task, the protection system philosophy of busbars, transmission lines, and transformers are modeled with the use of two types of neural networks: the general regression neural network and the multilayer perceptron neural network; Section I of Pages 1034-1035: the protection devices are responsible for detecting the occurrence of a fault and, when necessary, they send trip signals to circuit breakers (CBs) in order to isolate the defective part of the system (selectivity); dealing with topological changes and real power systems dimension, besides excessive effort in training procedures in case of neural-network applications; instead of dividing the power system into smaller parts and training a neural network for each set of a few buses, neural networks are employed here to model the protection system philosophy of the main transmission system components (buses, lines, and transformers); the alarms corresponding to the operation of protection relays and CBs are the inputs to MLP neural networks trained with the backpropagation algorithm; the output of MLP neural networks feed general regression neural networks (GRNNs), which conclude whether the equipment is faulted, not faulted, or there is a lack of information; the NN models developed for transformers and lines have additional information, which indicate the direction of external faults; this information can be combined and helps to conclude about which equipment started the occurrence; Section II of Pages 1035 with FIGS.1-3: the MLP nets with backpropagation learning algorithm (Fig. 1) constitute the NNs architecture most commonly used since it is capable to approximate any nonlinear function; GRNNs are feedforward networks that can be used to estimate an output vector Y from a measured vector X; an overview on GRNN architecture is shown in Fig. 2; the input units distribute the variables (input variables) to all neurons in the pattern units; each neuron belonging to the pattern layer corresponds to a sample or a cluster center; when a new vector X is presented to the network, the distance, usually Euclidian, between this vector and each cluster center previously defined and stored is computed; Section III of Pages 1035-1036: after a contingency, only the NNs corresponding to de-energized equipment should be activated; thus, first the fault scenario should be determined; an expert system has been designed for this task; the expert system is a topology tracker, which determines the active elements before and after the fault, using information about the breaker status; the fault scenario set for total number of elements involved in the fault is the difference between active elements before and after the fault; finally, for each element that composes the fault scenario set S, a neural-network module is activated according to its respective alarm inputs; each net would try to classify the element Si ( i-th element of S) into one of its outputs previously determined; Section VI with FIGS. 4-6 and TABLES 2-4 of Pages 1036-1038: the main reason for using MLP together with GRNN is the reduction in the dimension of the neural network, without oversimplifying the protection system; NN_TR determines whether the component is faulty or not, whether the fault is external toward the high-voltage or medium-voltage side, or even whether the information provided to the network is insufficient for the diagnosis; the output of this network determines whether the main protection of the line on the sending or receiving side (MP.S/MP.R) operated, or if the protection indicates the possibility of an external fault to the S (send) side or R (receive); NN_TL determines whether the component is faulty or not, if the fault is external toward the S or R side, or if the information is not sufficient to produce a diagnosis; the output of the network NN_BUS indicates whether the component is faulty or not, or whether there is not enough information for the diagnosis).
Claim 11
Zhou in view of HEWAGE disclose all the elements as stated in Claim 1 and except failing to explicitly disclose wherein each of the plurality of feature sets represents a set of measurement signals for a power line, and wherein each target value represents a decision to either trip or not trip a circuit breaker on the power line.
Cardoso teaches a system and a method relating to detecting the occurrence of a fault (Cardoso, 2nd paragraph of Section I in Page 1034), wherein each of the plurality of feature sets represents a set of measurement signals for a power line, and wherein each target value represents a decision to either trip or not trip a circuit breaker on the power line (Cardoso, Abstract of Page 1034: a neural system intended to aid the control center operator in the task of fault section estimation; in order to allow the diagnosis task, the protection system philosophy of busbars, transmission lines, and transformers are modeled with the use of two types of neural networks: the general regression neural network and the multilayer perceptron neural network; Section I of Pages 1034-1035: the protection devices are responsible for detecting the occurrence of a fault and, when necessary, they send trip signals to circuit breakers (CBs) in order to isolate the defective part of the system (selectivity); dealing with topological changes and real power systems dimension, besides excessive effort in training procedures in case of neural-network applications; instead of dividing the power system into smaller parts and training a neural network for each set of a few buses, neural networks are employed here to model the protection system philosophy of the main transmission system components (buses, lines, and transformers); the alarms corresponding to the operation of protection relays and CBs are the inputs to MLP neural networks trained with the backpropagation algorithm; the output of MLP neural networks feed general regression neural networks (GRNNs), which conclude whether the equipment is faulted, not faulted, or there is a lack of information; the NN models developed for transformers and lines have additional information, which indicate the direction of external faults; this information can be combined and helps to conclude about which equipment started the occurrence; Section II of Pages 1035 with FIGS.1-3: the MLP nets with backpropagation learning algorithm (Fig. 1) constitute the NNs architecture most commonly used since it is capable to approximate any nonlinear function; GRNNs are feedforward networks that can be used to estimate an output vector Y from a measured vector X; an overview on GRNN architecture is shown in Fig. 2; the input units distribute the variables (input variables) to all neurons in the pattern units; each neuron belonging to the pattern layer corresponds to a sample or a cluster center; when a new vector X is presented to the network, the distance, usually Euclidian, between this vector and each cluster center previously defined and stored is computed; Section III of Pages 1035-1036: after a contingency, only the NNs corresponding to de-energized equipment should be activated; thus, first the fault scenario should be determined; an expert system has been designed for this task; the expert system is a topology tracker, which determines the active elements before and after the fault, using information about the breaker status; the fault scenario set for total number of elements involved in the fault is the difference between active elements before and after the fault; finally, for each element that composes the fault scenario set S, a neural-network module is activated according to its respective alarm inputs; each net would try to classify the element Si ( i-th element of S) into one of its outputs previously determined; Section VI with FIGS. 4-6 and TABLES 2-4 of Pages 1036-1038: the main reason for using MLP together with GRNN is the reduction in the dimension of the neural network, without oversimplifying the protection system; NN_TR determines whether the component is faulty or not, whether the fault is external toward the high-voltage or medium-voltage side, or even whether the information provided to the network is insufficient for the diagnosis; the output of this network determines whether the main protection of the line on the sending or receiving side (MP.S/MP.R) operated, or if the protection indicates the possibility of an external fault to the S (send) side or R (receive); NN_TL determines whether the component is faulty or not, if the fault is external toward the S or R side, or if the information is not sufficient to produce a diagnosis; the output of the network NN_BUS indicates whether the component is faulty or not, or whether there is not enough information for the diagnosis).
Zhou in view of HEWAGE, and Cardoso are analogous art because they are from the same field of endeavor, a system and a method relating to detecting the occurrence of a fault. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of Cardoso to Zhou in view of HEWAGE. Motivation for doing so would open up the possibility for extension to .
Claim 12
Zhou in view of HEWAGE discloses all the elements as stated in Claim 1 and except failing to explicitly disclose wherein each of the plurality of feature sets represents a state of a power system, wherein each target value represents a value of a continuous variable of the power system, and wherein the method further comprises using the trained machine-learning model to estimate the value of the continuous variable.
Cardoso teaches a system and a method relating to detecting the occurrence of a fault (Cardoso, 2nd paragraph of Section I in Page 1034), wherein each of the plurality of feature sets represents a state of a power system, wherein each target value represents a value of a continuous variable of the power system, and wherein the method further comprises using the trained machine-learning model to estimate the value of the continuous variable (Cardoso, Abstract of Page 1034: a neural system intended to aid the control center operator in the task of fault section estimation; in order to allow the diagnosis task, the protection system philosophy of busbars, transmission lines, and transformers are modeled with the use of two types of neural networks: the general regression neural network and the multilayer perceptron neural network; Section I of Pages 1034-1035: the protection devices are responsible for detecting the occurrence of a fault and, when necessary, they send trip signals to circuit breakers (CBs) in order to isolate the defective part of the system (selectivity); dealing with topological changes and real power systems dimension, besides excessive effort in training procedures in case of neural-network applications; instead of dividing the power system into smaller parts and training a neural network for each set of a few buses, neural networks are employed here to model the protection system philosophy of the main transmission system components (buses, lines, and transformers); the alarms corresponding to the operation of protection relays and CBs are the inputs to MLP neural networks trained with the backpropagation algorithm; the output of MLP neural networks feed general regression neural networks (GRNNs), which conclude whether the equipment is faulted, not faulted, or there is a lack of information; the NN models developed for transformers and lines have additional information, which indicate the direction of external faults; this information can be combined and helps to conclude about which equipment started the occurrence; Section II of Pages 1035 with FIGS.1-3: the MLP nets with backpropagation learning algorithm (Fig. 1) constitute the NNs architecture most commonly used since it is capable to approximate any nonlinear function; GRNNs are feedforward networks that can be used to estimate an output vector Y from a measured vector X; an overview on GRNN architecture is shown in Fig. 2; the input units distribute the variables (input variables) to all neurons in the pattern units; each neuron belonging to the pattern layer corresponds to a sample or a cluster center; when a new vector X is presented to the network, the distance, usually Euclidian, between this vector and each cluster center previously defined and stored is computed; Section III of Pages 1035-1036: after a contingency, only the NNs corresponding to de-energized equipment should be activated; thus, first the fault scenario should be determined; an expert system has been designed for this task; the expert system is a topology tracker, which determines the active elements before and after the fault, using information about the breaker status; the fault scenario set for total number of elements involved in the fault is the difference between active elements before and after the fault; finally, for each element that composes the fault scenario set S, a neural-network module is activated according to its respective alarm inputs; each net would try to classify the element Si ( i-th element of S) into one of its outputs previously determined; Section VI with FIGS. 4-6 and TABLES 2-4 of Pages 1036-1038: the main reason for using MLP together with GRNN is the reduction in the dimension of the neural network, without oversimplifying the protection system; NN_TR determines whether the component is faulty or not, whether the fault is external toward the high-voltage or medium-voltage side, or even whether the information provided to the network is insufficient for the diagnosis; the output of this network determines whether the main protection of the line on the sending or receiving side (MP.S/MP.R) operated, or if the protection indicates the possibility of an external fault to the S (send) side or R (receive); NN_TL determines whether the component is faulty or not, if the fault is external toward the S or R side, or if the information is not sufficient to produce a diagnosis; the output of the network NN_BUS indicates whether the component is faulty or not, or whether there is not enough information for the diagnosis).
Zhou in view of HEWAGE, and Cardoso are analogous art because they are from the same field of endeavor, a system and a method relating to detecting the occurrence of a fault. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of Cardoso to Zhou in view of HEWAGE. Motivation for doing so would open up the possibility for extension to .
Claim 13
Zhou in view of HEWAGE discloses all the elements as stated in Claim 1 and except failing to explicitly disclose using the trained machine-learning model for one of topology estimation, parameter estimation, power flow estimation, or load forecasting.
Cardoso teaches a system and a method relating to detecting the occurrence of a fault (Cardoso, 2nd paragraph of Section I in Page 1034), wherein using the trained machine-learning model for one of topology estimation, parameter estimation, power flow estimation, or load forecasting (Cardoso, Abstract of Page 1034: a neural system intended to aid the control center operator in the task of fault section estimation; in order to allow the diagnosis task, the protection system philosophy of busbars, transmission lines, and transformers are modeled with the use of two types of neural networks: the general regression neural network and the multilayer perceptron neural network; Section I of Pages 1034-1035: the protection devices are responsible for detecting the occurrence of a fault and, when necessary, they send trip signals to circuit breakers (CBs) in order to isolate the defective part of the system (selectivity); dealing with topological changes and real power systems dimension, besides excessive effort in training procedures in case of neural-network applications; instead of dividing the power system into smaller parts and training a neural network for each set of a few buses, neural networks are employed here to model the protection system philosophy of the main transmission system components (buses, lines, and transformers); the alarms corresponding to the operation of protection relays and CBs are the inputs to MLP neural networks trained with the backpropagation algorithm; the output of MLP neural networks feed general regression neural networks (GRNNs), which conclude whether the equipment is faulted, not faulted, or there is a lack of information; the NN models developed for transformers and lines have additional information, which indicate the direction of external faults; this information can be combined and helps to conclude about which equipment started the occurrence; Section II of Pages 1035 with FIGS.1-3: the MLP nets with backpropagation learning algorithm (Fig. 1) constitute the NNs architecture most commonly used since it is capable to approximate any nonlinear function; GRNNs are feedforward networks that can be used to estimate an output vector Y from a measured vector X; an overview on GRNN architecture is shown in Fig. 2; the input units distribute the variables (input variables) to all neurons in the pattern units; each neuron belonging to the pattern layer corresponds to a sample or a cluster center; when a new vector X is presented to the network, the distance, usually Euclidian, between this vector and each cluster center previously defined and stored is computed; Section III of Pages 1035-1036: after a contingency, only the NNs corresponding to de-energized equipment should be activated; thus, first the fault scenario should be determined; an expert system has been designed for this task; the expert system is a topology tracker, which determines the active elements before and after the fault, using information about the breaker status; the fault scenario set for total number of elements involved in the fault is the difference between active elements before and after the fault; finally, for each element that composes the fault scenario set S, a neural-network module is activated according to its respective alarm inputs; each net would try to classify the element Si ( i-th element of S) into one of its outputs previously determined; Section VI with FIGS. 4-6 and TABLES 2-4 of Pages 1036-1038: the main reason for using MLP together with GRNN is the reduction in the dimension of the neural network, without oversimplifying the protection system; NN_TR determines whether the component is faulty or not, whether the fault is external toward the high-voltage or medium-voltage side, or even whether the information provided to the network is insufficient for the diagnosis; the output of this network determines whether the main protection of the line on the sending or receiving side (MP.S/MP.R) operated, or if the protection indicates the possibility of an external fault to the S (send) side or R (receive); NN_TL determines whether the component is faulty or not, if the fault is external toward the S or R side, or if the information is not sufficient to produce a diagnosis; the output of the network NN_BUS indicates whether the component is faulty or not, or whether there is not enough information for the diagnosis).
Zhou in view of HEWAGE, and Cardoso are analogous art because they are from the same field of endeavor, a system and a method relating to detecting the occurrence of a fault. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of Cardoso to Zhou in view of HEWAGE. Motivation for doing so would open up the possibility for extension to various applications.
Claims 14-16 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Zhou in view of HEWAGE as applied to Claim 1 above, and further in view of Guo et al. ("Spatially Penalised Registration of Multivariate Functional Data", ARXIV ID: 2207.10914, 07/22/2022, pp. 1-33), hereinafter Guo.
Claim 14
Zhou in view of HEWAGE discloses all the elements as stated in Claim 1 and further discloses wherein each of at least a subset of the plurality of feature sets in the training dataset is labeled with a ground-truth (Zhou, ¶ [0013]: features must be selected that are useful in detecting spam, and each web page is then represented as a vector having each element described by one type of spam feature; the features can be the number of inlinks, the number of outlinks, scores under any of the above-mentioned algorithms, etc.; ¶¶ [0023]-[0024], [0027]-[0028], and [0034]-[0036] with 114, 102, 116, 118, 104, 120, and 122 in FIG. 1, 152-154 in FIG.2, and FIG. 3: the collection of web pages 114 is considered a directed graph, in that the web pages themselves in collection 114 are nodes in the graph while hyperlinks between those web pages are directed edges in the directed graph; for applying at the domain/host level, the domains/hosts are nodes in the graph and hyperlinks among the web pages in the domains/hosts are the directed edges; the web page collection 114 is also provided to trusted entity 102, such as a human expert, in identifying link spam; trusted entity 102 then identifies some examples of spam web pages in web page collection 114 as spam training examples 116. Trusted entity 102 also identifies good web pages (or normal web pages) in web page collection 114 as normal training examples 118; random walk component 106 then receives a definition of a random walk on directed graph 104 (or each strongly connected component in directed graph 104), the random walk being defined by transition probabilities; based on the defined random walk, component 106 obtains stationary probabilities associated with each node in directed graph 104; the “transition probabilities' are the probabilities of transitioning from any given node on graph 104 to another node; the stationary probability distributions are the probabilities of being in any given node on directed graph 104),
wherein the optimization problem further comprises a (Zhou, ¶ [0013]: a classifier is chosen, such as a neural network; ¶¶ [0029]-[0032] and [0039]-[0080] in generating the classification function, the values of the classification function are forced to be close to known values for examples 116 and 118; in other words, the values of the classification function are forced to be close to the values that indicate spam and normal pages at the nodes in the graph that are actually known to be spam and normal pages as identified by the trusted entity 102; the closeness between the classification function and the known values can be measured in a variety of different ways; e.g. the closeness can be measured using least square loss, hinge loss, precision/recall measure, the F1-score, the ROC score, or the AUC score; the classification function is not only close to known values at known nodes, but it is relatively smooth in that it changes relatively slowly on densely connected subgraphs; in other words, the nodes that reside close to one another on the subgraph may likely have values which are relatively close to one another; however, if they are known to be one spam node and one normal node, respectively, then the classification function changes by a large amount between those nodes, but this lack of smoothness is penalized in the chosen cost function that is optimized; i.e., the classification function value can change abruptly, if necessary; again, however, this is penalized; given a directed graph G=(V, E, w), and a discrete label set L={-1, 1}, the vertices in a subset S [Symbol font/0xCD] V have labels in L; the task is to predict the labels of those unclassified vertices in Sc, the complement of S; define a function y with y(v)=1 or -1 if v[Symbol font/0xCE]S, and 0 if v[Symbol font/0xCE]Sc; for classifying those unclassified vertices in Sc, define a discrete regularization in Eq. 17, where C>0 is the regularization parameter; when choosing the basis, Eq. 17 can be written as Eq. 18; in the objective function, the first term forces the classification function to be relatively smooth, and perhaps as smooth as possible and the second term forces the classification function to fit the given labels as well as possible; the first term makes the function relatively smooth over all nodes while the second term forces the function to fit the labeled nodes to a desired closeness; a random walk over a given directed graph can be defined in many different ways; three exemplary types of random walk used in spam detection are: (a) following outlinks uniformly at random; (b) following links uniformly at random regardless of directionality; (c) following inlinks uniformly at random; assigning values to the nodes in directed graph 104 basically requires selection of a random walk definition (transition probabilities) and solving Eq. 18 above for each of the nodes; to solve the optimization problem in Eq. 18, differentiate the objective function with respect to s and then obtain Eq. 23; the classification is based on the sign of the function value on each vertex, which is equivalent to setting the classification threshold to 0; Claims 2-3, 12-13, and 15: optimizing a cost function that assigns a penalty based on a size of a difference in classification function values between nodes in the directed graph; optimizing the cost function that assigns a penalty based on a difference between classification function values assigned to nodes representing the training examples and the target function values for those nodes) (HEWAGE, ¶¶ [0020]-[0024], [0033], [0035], and [0086]: a machine learning technique that uses a latent vector from a latent vector space in classifying input data, where the latent vector includes a label vector y and a style vector z in which at least a part of the style vector z is regularised causing the machine learning technique to output label vectors y that are substantially time invariant, time invariant, or more time invariant that compared with when the machine learning technique does not regularise the style vector z; training an autoencoder, the autoencoder outputting a latent vector of an N-dimensional latent space for classifying input data, the latent vector comprising a label vector y and a style vector z; regularising the style vector z during training for effecting time invariance in the set of label vectors y associated with the input data; training an encoder network of the autoencoder with input training data to enforce a selected probability distribution on at least a portion of the style vector z; regularising the style vector z increases the time invariance of the set of label vectors y during training; the latent representation layer outputting the label vector, y, of the latent space; and an adversarial network coupled comprising an input layer, one or more hidden layer(s), and an output layer for outputting and/or evaluating an generator loss function, LGY, associated with label vector y, wherein the input layer of the adversarial network is connected to the label vector, y; training the adversarial network to distinguish between label vectors, y, generated by the latent representation layer and sample vectors from a categorical distribution of a set of one hot vectors of the same dimension as the label vector, y; and outputting the generator loss function value, LGy, associated with label vector y for use by the autoencoder in training an encoder network to enforce the categorical distribution on the label vector y; optimizing an autoencoder, the autoencoder outputting a latent vector of an N-dimensional latent space for classifying input data, the latent vector comprising a label vector y and a style vector z; controlling the regularization of style vector z to increase or decrease the time invariance of the label vectors y; the style vector z comprises a vector, a tensor or otherwise, wherein the vector, tensor or otherwise is penalised or regularised by at least one or more from the group of: a probability distribution; L1, L2 or both, or other norms; nuisance variables; and error variables).
Zhou in view of HEWAGE fails to explicitly disclose wherein (1) a deviation penalization function is a phasor- deviation penalization function; and (2) a ground-truth value/and estimated value is a ground-truth phasor value/an estimated phasor value.
Guo teaches a system and a method relating to optimize an objective function (Guo, Abstract of Page 1), wherein (1) a deviation penalization function is a phasor- deviation penalization function; and (2) a ground-truth value/and estimated value is a ground-truth phasor value/an estimated phasor value (Guo, Abstract of Page 1: Registration of multivariate functional data involves handling of both cross-component and cross-observation phase variations; allowing for the two phase variations to be modelled as general diffeomorphic time warpings, in this work, focus on the hitherto unconsidered setting where phase variation of the component functions are spatially correlated; propose an algorithm to optimize a metric-based objective function for registration with a novel penalty term that incorporates the spatial correlation between the component phase variations through a kriging estimate of an appropriate phase random field; the penalty term encourages the overall phase at a particular location to be similar to the spatially weighted average phase in its neighborhood, and thus engenders a regularization that prevents over-alignment; Section 1 with FIG. 1 of Pages 1-5: the notion of phase variation can be decomposed into two types: (i) cross-observation phase, which is common across all K component univariate functions fij within an observation Fi, and (ii) cross-component phase within each observed unit; motivated by the type of registration task associated with the EEG data, propose a registration procedure that exploits correlated phase variation between component functions { fij } of multivariate functional data { Fi } and offers a compromise between the two extreme cases of independent component-wise and universal registration methods alluded to above; propose a metric-based penalized registration method for multivariate functional data with a penalty term defined using the spatial correlations between the cross-component phases { ξj }, using which the overall phases { γij } are estimated by eliminating cross-observation phase variation { αi }; the penalty term is defined using a spatially weighted combination of the (estimated) cross-component phases { ξj }, and is, owing to the invariant property of the metric to (simultaneous) time warping, impervious to the cross-observation phases { αi }; this allows us to entirely circumvent having to estimate the { αi }; Section 4 with Algorithm 1 of Pages 11-17: propose a penalized multiple registration method for multivariate functional data wherein cross-component phase in each observation is spatially correlated; propose a spatially penalized registration approach that takes spatial cross-component phase correlation into account, but allows each component to have its own phase; within each Fi the dependence between the component functions { fij } arises through spatial correlation in the cross-component phases { ξj }; consider the optimization problem for each component j [Symbol font/0xCE] [K] in Eq. 6; the first term in the objective function in (12) provides a measure of synchronization for component j, across observations i, with respect to the template [Symbol font/0x6D]j; the second term, a penalty on phase, measures the distance between the estimated phase γij and a target
γ
~
i
j
determined by the correlated phase in the other components; the regularization penalty attempts to preserve the phase-induced correlation structure in the aligned components).
Zhou in view of HEWAGE, and Guo are analogous art because they are from the same field of endeavor, a system and a method relating to optimize an objective . Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of Guo to Zhou in view of HEWAGE. Motivation for doing so would open up the possibility for extension to settings involving di.
Claim 15
Zhou in view of HEWAGE and Guo discloses all the elements as stated in Claim 14 and further discloses wherein the at least a subset of the plurality of feature sets in the training dataset consists of only a portion of the plurality of feature sets in the training dataset that comprises, for each of the plurality of features, a time series that encompasses at least a predefined length of time (HEWAGE, ¶¶ [0013]-[0016] with FIGS. 1a-1c: FIG. 1b illustrates the example classification 100 of FIG. la in which two random time windows 106a and 106b (e.g. TWN and TWN+1); FIG. 1c is a schematic diagram illustrating an example cluster diagram 110 showing clustering of the vector labels in cluster regions 112, 114, and 116 for time windows 106a and 106b; ¶¶ [0141]-[0143]: the sensor data was video footage of the subject that was collected for a period of 24 minutes; 1,000,000 samples were generated with unit amplitudes and random periods between 4 and 30 time steps).
Claim 16
Zhou in view of HEWAGE and Guo discloses all the elements as stated in Claim 14 and further discloses wherein the encoder network is an artificial neural network (Zhou, ¶ [0013]: a classifier is chosen, such as a neural network) (HEWAGE, ¶¶ [0046]: the autoencoder is configured based on data representative of the weights and/or parameters of one or more neural network(s) and/or hidden layer(s) associated with the trained autoencoder of the set of autoencoder configuration data).
Claim 18
Zhou in view of HEWAGE discloses all the elements as stated in Claim 17 and further discloses wherein each of at least a subset of the plurality of feature sets in the training dataset is labeled with a ground-truth (Zhou, ¶ [0013]: features must be selected that are useful in detecting spam, and each web page is then represented as a vector having each element described by one type of spam feature; the features can be the number of inlinks, the number of outlinks, scores under any of the above-mentioned algorithms, etc.; ¶¶ [0023]-[0024], [0027]-[0028], and [0034]-[0036] with 114, 102, 116, 118, 104, 120, and 122 in FIG. 1, 152-154 in FIG.2, and FIG. 3: the collection of web pages 114 is considered a directed graph, in that the web pages themselves in collection 114 are nodes in the graph while hyperlinks between those web pages are directed edges in the directed graph; for applying at the domain/host level, the domains/hosts are nodes in the graph and hyperlinks among the web pages in the domains/hosts are the directed edges; the web page collection 114 is also provided to trusted entity 102, such as a human expert, in identifying link spam; trusted entity 102 then identifies some examples of spam web pages in web page collection 114 as spam training examples 116. Trusted entity 102 also identifies good web pages (or normal web pages) in web page collection 114 as normal training examples 118; random walk component 106 then receives a definition of a random walk on directed graph 104 (or each strongly connected component in directed graph 104), the random walk being defined by transition probabilities; based on the defined random walk, component 106 obtains stationary probabilities associated with each node in directed graph 104; the “transition probabilities' are the probabilities of transitioning from any given node on graph 104 to another node; the stationary probability distributions are the probabilities of being in any given node on directed graph 104),
wherein the optimization problem further comprises a (Zhou, ¶ [0013]: a classifier is chosen, such as a neural network; ¶¶ [0029]-[0032] and [0039]-[0080] in generating the classification function, the values of the classification function are forced to be close to known values for examples 116 and 118; in other words, the values of the classification function are forced to be close to the values that indicate spam and normal pages at the nodes in the graph that are actually known to be spam and normal pages as identified by the trusted entity 102; the closeness between the classification function and the known values can be measured in a variety of different ways; e.g. the closeness can be measured using least square loss, hinge loss, precision/recall measure, the F1-score, the ROC score, or the AUC score; the classification function is not only close to known values at known nodes, but it is relatively smooth in that it changes relatively slowly on densely connected subgraphs; in other words, the nodes that reside close to one another on the subgraph may likely have values which are relatively close to one another; however, if they are known to be one spam node and one normal node, respectively, then the classification function changes by a large amount between those nodes, but this lack of smoothness is penalized in the chosen cost function that is optimized; i.e., the classification function value can change abruptly, if necessary; again, however, this is penalized; given a directed graph G=(V, E, w), and a discrete label set L={-1, 1}, the vertices in a subset S [Symbol font/0xCD] V have labels in L; the task is to predict the labels of those unclassified vertices in Sc, the complement of S; define a function y with y(v)=1 or -1 if v[Symbol font/0xCE]S, and 0 if v[Symbol font/0xCE]Sc; for classifying those unclassified vertices in Sc, define a discrete regularization in Eq. 17, where C>0 is the regularization parameter; when choosing the basis, Eq. 17 can be written as Eq. 18; in the objective function, the first term forces the classification function to be relatively smooth, and perhaps as smooth as possible and the second term forces the classification function to fit the given labels as well as possible; the first term makes the function relatively smooth over all nodes while the second term forces the function to fit the labeled nodes to a desired closeness; a random walk over a given directed graph can be defined in many different ways; three exemplary types of random walk used in spam detection are: (a) following outlinks uniformly at random; (b) following links uniformly at random regardless of directionality; (c) following inlinks uniformly at random; assigning values to the nodes in directed graph 104 basically requires selection of a random walk definition (transition probabilities) and solving Eq. 18 above for each of the nodes; to solve the optimization problem in Eq. 18, differentiate the objective function with respect to s and then obtain Eq. 23; the classification is based on the sign of the function value on each vertex, which is equivalent to setting the classification threshold to 0; Claims 2-3, 12-13, and 15: optimizing a cost function that assigns a penalty based on a size of a difference in classification function values between nodes in the directed graph; optimizing the cost function that assigns a penalty based on a difference between classification function values assigned to nodes representing the training examples and the target function values for those nodes) (HEWAGE, ¶¶ [0020]-[0024], [0033], [0035], and [0086]: a machine learning technique that uses a latent vector from a latent vector space in classifying input data, where the latent vector includes a label vector y and a style vector z in which at least a part of the style vector z is regularised causing the machine learning technique to output label vectors y that are substantially time invariant, time invariant, or more time invariant that compared with when the machine learning technique does not regularise the style vector z; training an autoencoder, the autoencoder outputting a latent vector of an N-dimensional latent space for classifying input data, the latent vector comprising a label vector y and a style vector z; regularising the style vector z during training for effecting time invariance in the set of label vectors y associated with the input data; training an encoder network of the autoencoder with input training data to enforce a selected probability distribution on at least a portion of the style vector z; regularising the style vector z increases the time invariance of the set of label vectors y during training; the latent representation layer outputting the label vector, y, of the latent space; and an adversarial network coupled comprising an input layer, one or more hidden layer(s), and an output layer for outputting and/or evaluating an generator loss function, LGY, associated with label vector y, wherein the input layer of the adversarial network is connected to the label vector, y; training the adversarial network to distinguish between label vectors, y, generated by the latent representation layer and sample vectors from a categorical distribution of a set of one hot vectors of the same dimension as the label vector, y; and outputting the generator loss function value, LGy, associated with label vector y for use by the autoencoder in training an encoder network to enforce the categorical distribution on the label vector y; optimizing an autoencoder, the autoencoder outputting a latent vector of an N-dimensional latent space for classifying input data, the latent vector comprising a label vector y and a style vector z; controlling the regularization of style vector z to increase or decrease the time invariance of the label vectors y; the style vector z comprises a vector, a tensor or otherwise, wherein the vector, tensor or otherwise is penalised or regularised by at least one or more from the group of: a probability distribution; L1, L2 or both, or other norms; nuisance variables; and error variables).
Zhou in view of HEWAGE fails to explicitly disclose wherein (1) a deviation penalization function is a phasor- deviation penalization function; and (2) a ground-truth value/and estimated value is a ground-truth phasor value/an estimated phasor value.
Guo teaches a system and a method relating to optimize an objective function (Guo, Abstract of Page 1), wherein (1) a deviation penalization function is a phasor- deviation penalization function; (2) a ground-truth value/and estimated value is a ground-truth phasor value/an estimated phasor value (Guo, Abstract of Page 1: Registration of multivariate functional data involves handling of both cross-component and cross-observation phase variations; allowing for the two phase variations to be modelled as general diffeomorphic time warpings, in this work, focus on the hitherto unconsidered setting where phase variation of the component functions are spatially correlated; propose an algorithm to optimize a metric-based objective function for registration with a novel penalty term that incorporates the spatial correlation between the component phase variations through a kriging estimate of an appropriate phase random field; the penalty term encourages the overall phase at a particular location to be similar to the spatially weighted average phase in its neighborhood, and thus engenders a regularization that prevents over-alignment; Section 1 with FIG. 1 of Pages 1-5: the notion of phase variation can be decomposed into two types: (i) cross-observation phase, which is common across all K component univariate functions fij within an observation Fi, and (ii) cross-component phase within each observed unit; motivated by the type of registration task associated with the EEG data, propose a registration procedure that exploits correlated phase variation between component functions { fij } of multivariate functional data { Fi } and offers a compromise between the two extreme cases of independent component-wise and universal registration methods alluded to above; propose a metric-based penalized registration method for multivariate functional data with a penalty term defined using the spatial correlations between the cross-component phases { ξj }, using which the overall phases { γij } are estimated by eliminating cross-observation phase variation { αi }; the penalty term is defined using a spatially weighted combination of the (estimated) cross-component phases { ξj }, and is, owing to the invariant property of the metric to (simultaneous) time warping, impervious to the cross-observation phases { αi }; this allows us to entirely circumvent having to estimate the { αi }; Section 4 with Algorithm 1 of Pages 11-17: propose a penalized multiple registration method for multivariate functional data wherein cross-component phase in each observation is spatially correlated; propose a spatially penalized registration approach that takes spatial cross-component phase correlation into account, but allows each component to have its own phase; within each Fi the dependence between the component functions { fij } arises through spatial correlation in the cross-component phases { ξj }; consider the optimization problem for each component j [Symbol font/0xCE] [K] in Eq. 6; the first term in the objective function in (12) provides a measure of synchronization for component j, across observations i, with respect to the template [Symbol font/0x6D]j; the second term, a penalty on phase, measures the distance between the estimated phase γij and a target
γ
~
i
j
determined by the correlated phase in the other components; the regularization penalty attempts to preserve the phase-induced correlation structure in the aligned components).
Zhou in view of HEWAGE, and Guo are analogous art because they are from the same field of endeavor, a system and a method relating to optimize an objective function. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of Guo to Zhou in view of HEWAGE. Motivation for doing so would open up the possibility for extension to settings involving di.
Claims 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Zhou in view of Guo and HEWAGE.
Independent Claim 19
Zhou discloses a method of manifold regularization while training a machine-learning model (Zhou, ¶¶ [0029], [0045], and [0064]-[0065] with FIG. 1: training a classifier is performed by generating a smooth classification function based on the probabilities 120 and 122 over the detected graph 104; a number of discrete operators on directed graphs are discrete analogs of the corresponding differential operators on Riemannian manifolds; the classification algorithm for directed graphs is derived from the discrete regularization; for classifying those unclassified vertices in Sc, define a discrete regularization in Eq. 17, where C>0 is the regularization parameter), the method comprising using at least one hardware processor (Zhou, ¶¶ [0082] and [0084] with 320 in FIG. 6: multiprocessor; processing unit 320) to:
acquire a training dataset that comprises a plurality of feature sets, wherein each of the plurality of feature sets comprises a feature value for each of a plurality of features and is labeled with a target value for each of one or more targets, and wherein each of at least a subset of the plurality of feature sets is labeled with a ground-truth (Zhou, ¶ [0013]: features must be selected that are useful in detecting spam, and each web page is then represented as a vector having each element described by one type of spam feature; the features can be the number of inlinks, the number of outlinks, scores under any of the above-mentioned algorithms, etc.; ¶¶ [0023]-[0024], [0027]-[0028], and [0034]-[0036] with 114, 102, 116, 118, 104, 120, and 122 in FIG. 1, 152-154 in FIG.2, and FIG. 3: the collection of web pages 114 is considered a directed graph, in that the web pages themselves in collection 114 are nodes in the graph while hyperlinks between those web pages are directed edges in the directed graph; for applying at the domain/host level, the domains/hosts are nodes in the graph and hyperlinks among the web pages in the domains/hosts are the directed edges; the web page collection 114 is also provided to trusted entity 102, such as a human expert, in identifying link spam; trusted entity 102 then identifies some examples of spam web pages in web page collection 114 as spam training examples 116. Trusted entity 102 also identifies good web pages (or normal web pages) in web page collection 114 as normal training examples 118; random walk component 106 then receives a definition of a random walk on directed graph 104 (or each strongly connected component in directed graph 104), the random walk being defined by transition probabilities; based on the defined random walk, component 106 obtains stationary probabilities associated with each node in directed graph 104; the “transition probabilities' are the probabilities of transitioning from any given node on graph 104 to another node; the stationary probability distributions are the probabilities of being in any given node on directed graph 104);
generate an optimization problem comprising an objective function, a non-smoothness penalization function. and a a neural network, and a ground-truth (Zhou, ¶ [0013]: a classifier is chosen, such as a neural network; ¶¶ [0029]-[0032] and [0039]-[0080] in generating the classification function, the values of the classification function are forced to be close to known values for examples 116 and 118; in other words, the values of the classification function are forced to be close to the values that indicate spam and normal pages at the nodes in the graph that are actually known to be spam and normal pages as identified by the trusted entity 102; the closeness between the classification function and the known values can be measured in a variety of different ways; e.g. the closeness can be measured using least square loss, hinge loss, precision/recall measure, the F1-score, the ROC score, or the AUC score; the classification function is not only close to known values at known nodes, but it is relatively smooth in that it changes relatively slowly on densely connected subgraphs; in other words, the nodes that reside close to one another on the subgraph may likely have values which are relatively close to one another; however, if they are known to be one spam node and one normal node, respectively, then the classification function changes by a large amount between those nodes, but this lack of smoothness is penalized in the chosen cost function that is optimized; i.e., the classification function value can change abruptly, if necessary; again, however, this is penalized; given a directed graph G=(V, E, w), and a discrete label set L={-1, 1}, the vertices in a subset S [Symbol font/0xCD] V have labels in L; the task is to predict the labels of those unclassified vertices in Sc, the complement of S; define a function y with y(v)=1 or -1 if v[Symbol font/0xCE]S, and 0 if v[Symbol font/0xCE]Sc; for classifying those unclassified vertices in Sc, define a discrete regularization in Eq. 17, where C>0 is the regularization parameter; when choosing the basis, Eq. 17 can be written as Eq. 18; in the objective function, the first term forces the classification function to be relatively smooth, and perhaps as smooth as possible and the second term forces the classification function to fit the given labels as well as possible; the first term makes the function relatively smooth over all nodes while the second term forces the function to fit the labeled nodes to a desired closeness; a random walk over a given directed graph can be defined in many different ways; three exemplary types of random walk used in spam detection are: (a) following outlinks uniformly at random; (b) following links uniformly at random regardless of directionality; (c) following inlinks uniformly at random; assigning values to the nodes in directed graph 104 basically requires selection of a random walk definition (transition probabilities) and solving Eq. 18 above for each of the nodes; to solve the optimization problem in Eq. 18, differentiate the objective function with respect to s and then obtain Eq. 23; the classification is based on the sign of the function value on each vertex, which is equivalent to setting the classification threshold to 0; Claims 2-3, 12-13, and 15: optimizing a cost function that assigns a penalty based on a size of a difference in classification function values between nodes in the directed graph; optimizing the cost function that assigns a penalty based on a difference between classification function values assigned to nodes representing the training examples and the target function values for those nodes); and
train the machine-learning model by adjusting the machine-learning model to minimize the estimated error, produced by the training dataset, in the optimization problem (Zhou, ¶ [0013]: a classifier is chosen, such as a neural network, a decision tree, a support vector machine (SVM), etc., and it is trained with a set of examples of normal and spam web pages which have been judged by human experts; ¶¶[0029]-[0030] with 108 in FIG. 1 and 156 in FIG. 2: once examples 116 and 118 and probabilities 120 and 122 are obtained, classifier training component 108 trains a classifier that can be used in link spam detection; training a classifier is performed by generating a smooth classification function based on the probabilities 120 and 122 over the detected graph 104; in generating the classification function, the values of the classification function are forced to be close to known values for examples 116 and 118; in other words, the values of the classification function are forced to be close to the values that indicate spam and normal pages at the nodes in the graph that are actually known to be spam and normal pages as identified by the trusted entity 102; the closeness between the classification function and the known values can be measured in a variety of different ways; e.g. the closeness can be measured using least square loss, hinge loss, precision/recall measure, the F1-score, the ROC score, or the AUC score; Claim 9: a classifier training component configured to receive a first set of training pages labeled as normal pages and a second set of training pages labeled as spam pages and to train a web page classifier based on both the first set of training pages and the second set of training pages), wherein training the machine-learning model comprises (Zhou, ¶ [0012]: consider a general ranking function that depends on incoming paths of various lengths weighted by some chosen damping function that decreases with distance; i.e., links from pages that are a greater distance from the subject web page are weighted by weight that is damped less than links from closer web pages; ¶¶ [0029]-[0033]: training a classifier is performed by generating a smooth classification function based on the probabilities 120 and 122 over the detected graph 104; the closeness between the classification function and the known values can be measured in a variety of different ways; e.g., the closeness can be measured using least square loss, hinge loss, precision/recall measure, the F1-score, the ROC score, or the AUC score, as examples; the classification function is not only close to known values at known nodes, but it is relatively smooth in that it changes relatively slowly on densely connected subgraphs; i.e., the nodes that reside close to one another on the subgraph may likely have values which are relatively close to one another; if they are known to be one spam node and one normal node, respectively, then the classification function changes by a large amount between those nodes, but this lack of smoothness is penalized in the chosen cost function that is optimized; because pages that are relatively close to one another on the directed graph are assumed to be the same type (pages close to a known spam page are likely to be spam pages, while pages close to a known normal page are likely to be normal pages) by making the function smooth and relatively slow changing pages in the directed subgraph that are close to a known normal content page will have classification function values that more likely indicate it to be a normal content page; similarly, those pages in the directed subgraph that are close to a spam page will have classification function values that are likely to indicate that it will be a spam page; i.e., different closeness between nodes (i.e., different partitions/segments/groups) will have different weights to penalize non-smoothness; ¶¶ [0039]-[0080]: a path is a tuple of vertices with the property that (vi, vi+1) [Symbol font/0xCE] E for 1 [Symbol font/0xA3] i [Symbol font/0xA3] p-1; for a strongly connected graph, there is an integer k [Symbol font/0xB3] 1 and a unique partition V=V0 U V1 ... U Vk-1 such that for all 0 [Symbol font/0xB3] r [Symbol font/0xB3] k-1 each edge (u, v) [Symbol font/0xCE] E with u [Symbol font/0xCE] V, has v[Symbol font/0xCE]Vr+1, where Vk= V0; and k is maximal, that is, there is no other such partition with k > k; a graph G is weighted if it is associated with a function w; Let G=(V, E, w) denote a weighted directed graph; the function w is called the weight function of G; in the objective function, the first term forces the classification function to be relatively smooth, and perhaps as smooth as possible and the second term forces the classification function to fit the given labels as well as possible).
Zhou fails to explicitly disclose wherein (1) a deviation penalization function is a phasor- deviation penalization function; (2) a ground-truth value/and estimated value is a ground-truth phasor value/an estimated phasor value; (3) a neural network includes an encoder network; and (4) clustering the training dataset into a plurality of clusters.
Guo teaches a system and a method relating to optimize an objective function (Guo, Abstract of Page 1), wherein (1) a deviation penalization function is a phasor- deviation penalization function; and (2) a ground-truth value/and estimated value is a ground-truth phasor value/an estimated phasor value (Guo, Abstract of Page 1: Registration of multivariate functional data involves handling of both cross-component and cross-observation phase variations; allowing for the two phase variations to be modelled as general diffeomorphic time warpings, in this work, focus on the hitherto unconsidered setting where phase variation of the component functions are spatially correlated; propose an algorithm to optimize a metric-based objective function for registration with a novel penalty term that incorporates the spatial correlation between the component phase variations through a kriging estimate of an appropriate phase random field; the penalty term encourages the overall phase at a particular location to be similar to the spatially weighted average phase in its neighborhood, and thus engenders a regularization that prevents over-alignment; Section 1 with FIG. 1 of Pages 1-5: the notion of phase variation can be decomposed into two types: (i) cross-observation phase, which is common across all K component univariate functions fij within an observation Fi, and (ii) cross-component phase within each observed unit; motivated by the type of registration task associated with the EEG data, propose a registration procedure that exploits correlated phase variation between component functions { fij } of multivariate functional data { Fi } and offers a compromise between the two extreme cases of independent component-wise and universal registration methods alluded to above; propose a metric-based penalized registration method for multivariate functional data with a penalty term defined using the spatial correlations between the cross-component phases { ξj }, using which the overall phases { γij } are estimated by eliminating cross-observation phase variation { αi }; the penalty term is defined using a spatially weighted combination of the (estimated) cross-component phases { ξj }, and is, owing to the invariant property of the metric to (simultaneous) time warping, impervious to the cross-observation phases { αi }; this allows us to entirely circumvent having to estimate the { αi }; Section 4 with Algorithm 1 of Pages 11-17: propose a penalized multiple registration method for multivariate functional data wherein cross-component phase in each observation is spatially correlated; propose a spatially penalized registration approach that takes spatial cross-component phase correlation into account, but allows each component to have its own phase; within each Fi the dependence between the component functions { fij } arises through spatial correlation in the cross-component phases { ξj }; consider the optimization problem for each component j [Symbol font/0xCE] [K] in Eq. 6; the first term in the objective function in (12) provides a measure of synchronization for component j, across observations i, with respect to the template [Symbol font/0x6D]j; the second term, a penalty on phase, measures the distance between the estimated phase γij and a target
γ
~
i
j
determined by the correlated phase in the other components; the regularization penalty attempts to preserve the phase-induced correlation structure in the aligned components).
Zhou and Guo are analogous art because they are from the same field of endeavor, a system and a method relating to optimize an objective function. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of Guo to Zhou. Motivation for doing so would open up the possibility for extension to settings involving different forms of correlation between the component functions and their phases (Guo, Abstract).
Zhou in view of Guo fails to explicitly disclose wherein (1) a neural network includes an encoder network; and (2) clustering the training dataset into a plurality of clusters.
HEWAGE teaches a system and a method relating to classification (HEWAGE, ¶ [0001]), wherein (1) a neural network includes an encoder network (HEWAGE, ¶¶ [0020]-[0024], [0033], [0035], and [0086]: a machine learning technique that uses a latent vector from a latent vector space in classifying input data, where the latent vector includes a label vector y and a style vector z in which at least a part of the style vector z is regularised causing the machine learning technique to output label vectors y that are substantially time invariant, time invariant, or more time invariant that compared with when the machine learning technique does not regularise the style vector z; training an autoencoder, the autoencoder outputting a latent vector of an N-dimensional latent space for classifying input data, the latent vector comprising a label vector y and a style vector z; regularising the style vector z during training for effecting time invariance in the set of label vectors y associated with the input data; training an encoder network of the autoencoder with input training data to enforce a selected probability distribution on at least a portion of the style vector z; regularising the style vector z increases the time invariance of the set of label vectors y during training; the latent representation layer outputting the label vector, y, of the latent space; and an adversarial network coupled comprising an input layer, one or more hidden layer(s), and an output layer for outputting and/or evaluating an generator loss function, LGY, associated with label vector y, wherein the input layer of the adversarial network is connected to the label vector, y; training the adversarial network to distinguish between label vectors, y, generated by the latent representation layer and sample vectors from a categorical distribution of a set of one hot vectors of the same dimension as the label vector, y; and outputting the generator loss function value, LGy, associated with label vector y for use by the autoencoder in training an encoder network to enforce the categorical distribution on the label vector y; optimizing an autoencoder, the autoencoder outputting a latent vector of an N-dimensional latent space for classifying input data, the latent vector comprising a label vector y and a style vector z; controlling the regularization of style vector z to increase or decrease the time invariance of the label vectors y; the style vector z comprises a vector, a tensor or otherwise, wherein the vector, tensor or otherwise is penalised or regularised by at least one or more from the group of: a probability distribution; L1, L2 or both, or other norms; nuisance variables; and error variables); and (2) clustering the training dataset into a plurality of clusters (HEWAGE, ¶¶ [0022] and [0027]-[0029]: regularising the style vector z further comprises training an encoder network of the autoencoder with input training data to enforce a selected probability distribution on at least a portion of the style vector z; regularising the style vector z during training causes the label vectors y associated with input data to form multiple or two or more clusters of label vectors y, wherein each contains a subgroup of label vectors y that are substantially the same or similar, and the set of label vectors y are substantially time invariant; each cluster is defined by a region or boundary and the subgroup of label vectors y for each cluster are contained within the defined region or boundary, and label vectors y are substantially the same or similar when they are contained within the region or boundary of the same cluster, wherein the cluster relates to a true state or class label; clustering the set of label vectors y to form multiple clusters of label vectors y in which each cluster contains a subgroup of label vectors y that are substantially the same or similar; and mapping each of the clusters of label vectors y to a class or state label from a set of class or state labels S associated with the input data for use by the trained autoencoder in classifying input data; ¶¶ [0033]-[0034]: outputting the generator loss function value associated with label vector y for use by the autoencoder in training an encoder network to enforce the categorical distribution on the label vector y; the autoencoder further comprising a decoding network coupled to the latent representation layer, wherein the training set of input data comprises a training set of neurological sample vector sequences in which Lk is the length of the k-th neurological sample vector sequence and T is the number of training neurological sample vector sequences, for each k-th neurological sample vector sequence corresponding to a k-th neural activity that is passed through the autoencoder; generating a loss or cost function based on the output of the one or more regularising networks and/or the adversarial network, an estimate of k-th neurological sample vector sequence output from the decoding network, the original k-th neurological sample vector sequence; and updating the weights of the hidden layer(s) of the encoding network and/or decoding network based on the generated loss of cost function; ¶¶ [0123] and [0176] with FIGS. 2a and 3: the training set may be generated from a collected set of unlabelled neurological sample vector sequences using autoencoder 200 as a classifier that outputs, from encoder network 202a, the label vector y 206 (e.g. this may be a soft vector) for each of the input neurological sample vector sequences 201a; this may produce a set of label vectors y which can be mapped to a plurality of true states or classes associated with the bodily variables encoded in the neural activity; e.g., the set of label vectors y 206 may be used to determine the bodily variable labels (e.g. true state or classes) by observing whether the set of label vectors y 206 form cluster regions, in which each cluster region may be labelled with a bodily variable label; the bodily variable label for each cluster region may be identified by, firstly, comparing each of the neural activities of (e.g. automatically analysed) that generate the label vectors y 206 within the cluster region with corresponding sensor data (e.g. video, audio, motion tracking, blood, heart rate etc.) recorded/stored/collected at the same time the multichannel neurological signal sample vector sequences were recorded/stored/sampled and collected. This is used to analyse the neural activity and corresponding sensor data associated with said cluster region and determining a bodily variable label based on the analysed neural activity and corresponding sensor data; thus a mapping from the cluster region to the bodily variable label or true state/classes may be generated and used for classifying label vectors y in accordance with the bodily variable labels or true states/classes etc.; ¶ [0175]: the label vector y 206 may include a vector, a tensor or otherwise, where the vector, tensor or otherwise includes at least one or more from the group of: a one hot vector; a measure of entropy; be regularized to L1 L2 or both, or other norms; a discrete boltzmann distributed vector; a representation of a prior class state; a known feature or configuration set; L1 and L2 are the well-known "distance" measures or measure of total value of all the elements in a vector; to penalise or regularise using this measure is to minimise this measure; ¶¶ [0177]-[0192] with FIG. 4a: a multiple of clusters (or two or more clusters) may be determined based on the output set of label vectors y 206; this may involve detecting whether each of the clusters contains a subgroup of label vectors y 206 that are substantially the same or similar; each cluster may be defined by a region or boundary and the subgroup of label vectors y 206 for each cluster are contained within the defined region or boundary, and label vectors y 206 are substantially the same or similar when they are contained within the region or boundary of the same cluster; clustering the set of label vectors y to form multiple clusters of label vectors y in which each cluster contains a subgroup of label vectors y that are substantially the same or similar, and mapping each of the clusters of label vectors y to a class or state label from a set of class or state labels S associated with the input data for use by an autoencoder defined by the set of hyperparameters in classifying input data; the latent vector comprising a label vector y 206 and a style vector z 208 where the autoencoder is configured based on data representative of the weights and/or parameters of one or more neural network(s) and/or hidden layer(s) associated with the trained autoencoder of the set of autoencoder configuration data, wherein the trained autoencoder regularized the style vector z 208 and outputs substantially time invariant label vector(s) y 206).
Zho and HEWAGE are analogous art because they are from the same field of endeavor, a system and a method relating to classification. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to apply the teaching of HEWAGE to Zhou in view of Guo. Motivation for doing so would improv.
Claim 20
Zhou in view of Guo and HEWAGE discloses all the elements as stated in Claim 19 and further discloses wherein the at least a subset of the plurality of feature sets in the training dataset consists of only a portion of the plurality of feature sets in the training dataset that comprises, for each of the plurality of features, a time series that encompasses at least a predefined length of time (HEWAGE, ¶¶ [0013]-[0016] with FIGS. 1a-1c: FIG. 1b illustrates the example classification 100 of FIG. la in which two random time windows 106a and 106b (e.g. TWN and TWN+1); FIG. 1c is a schematic diagram illustrating an example cluster diagram 110 showing clustering of the vector labels in cluster regions 112, 114, and 116 for time windows 106a and 106b; ¶¶ [0141]-[0143]: the sensor data was video footage of the subject that was collected for a period of 24 minutes; 1,000,000 samples were generated with unit amplitudes and random periods between 4 and 30 time steps).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
SRKRISHNAN et al. (US 2011/0158474 A1, pub. date: 06/30/2011) discloses in ¶¶ [0013]-[0040] with FIG. 1 that (1) at block 110, an input image with at least one image object to be tracked is received; (2) an initial contour for tracking an image object is defined (block 120); (3) at block 130, the initial contour is partitioned into a plurality of contour segments; (4) a weighted length of each of the plurality of contour segments is estimated (block 140); (5) at block 150, an internal energy value for each of the plurality of contour segments is computed; (6) at block 160, a desired contour is generated by guiding the plurality of contour segments towards edges of the image object using the computed internal energy; and (7) at block 170, an output image with the desired contour is stored. SRKRISHNAN further discloses in ¶¶ [0045]-[0049] with FIG. 2 that (1) the image processing circuit utilizes gradient vector force (GVF) segmentation with a space varying smoothness term based on the continuity prior to generate the desired contour; (2) the space varying smoothness term applies a penalty to contour segments moving away from a long edge thereby raising the energy of the contour; (3) as a result, such contour segments are guided towards the object of interest; (4) in operation, points of the initial contour are grouped into the plurality of contour segments based upon pre-determined thresholds; (5) the internal energy is computed based upon a weighted length of each of the plurality of contour segments and a smoothness index of the contour segments; (6) the internal energy is a spatially adaptive quantity that is estimated using gradient data of points along the contour segments thereby facilitating converging of the contour segments to edges of the image object 230 rather than to any distracting edges in the vicinity of the image object 230; (7) relatively high weighted lengths are assigned to contour segments lying along edges of the image object 230; and (8) further, deviation from contour segments with high weighted lengths is penalized using the smoothness functional.
ISOLA et al. (US 2016/0082287 A1, pub. date: 03/24/2016) discloses in ¶¶ [0057]-[0061] with FIG. 3 that (1) assign regularization values Pr to the graph edges, wherein the regularization values may depend on deviations between the extensions br, br-1 of the illumination distributions represented by the nodes connected by the respective graph edges; (2) the regularization values, which can also be regarded as being regularization weights, can be added to the first weights, which are preferentially local gradient-based weights; (3) the regularization values can be used to enforce smooth opening contours and/or larger opening areas; e.g., following regularization term may be added to the first weights er to enforce smoother segment opening contours shown in Eq. (7); (4) a weighting parameter α can be used to tune for a best tradeoff between the first weight er and the smoothing penalization Pr; and (5) the weighting parameter α may be determined by trying different weighting parameters α and selecting the weighting parameter α, which leads to the best approximation of the ideal fluence map.
Siewerdsen et al. (US 2021/0256716 A1, pub. date: 08/19/2021) discloses in ¶¶ [0032]-[0037] that (1) the preoperative stage 102 uses a segmentation technique 108, such as a continuous max-flow min-cut segmentation (Yuan et al 2010a), to obtain 3D models of bone fragments 110 from a patient CT 1-6; (2) a semi-automatic segmentation method was implemented to segment bone fragments in preoperative CT images of fractured pelvis, requiring only an input "seed point" for each bone fragment of interest; (3) the segmentation is formulated as an N-label continuous max-flow problem such that each bone fragment and the background are segmented into different labels, according to the cost function shown in Eq. (1); (3) the last term in the integrand of Eq. (1) is a regularization to enforce the smoothness of the segmentation by penalizing large segmentation gradient; (4) β and γ are scalar parameters to control the relative strength of the three terms in the objective of Eq. (1).
Wang et al. ("Sparse and Structured Function-on-Function Quality Predictive Modeling by Hierarchical Variable Selection and Multitask Learning", IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 17, NO. 10, December 2, 2020, pp. 6720-6730) discloses in ABSTRACT of Page 6720 that (1) propose a novel sparse and structured function-on-function regression (SSF2R) model, where a hierarchical variable selection is developed to identify informative signals and further screen significant elements within the selected signals, and a multitask learning is devised to exploit the smoothness nature of surface response and the similarity structure among a series of sub-regression tasks; (2) SSF2R model is concisely formulated as a convex problem with an efficient iterative algorithm derived to obtain the global optimum; and (3) quality prediction can be performed dynamically during an ongoing manufacturing process when only partial observations of the signal predictors are available. Wang further discloses in Section I of Pages 6720-6721 that (1) in the proposed function-on-function regression model, follow the functional data analysis (FDA) reasoning where either the process signal or the product surface is regarded as a single complete functional datum of a continuity or smoothness nature, which captures maximum data information and circumvents the traditional feature extraction efforts; (2) the main idea is that first, at the function level, identify the informative signals globally that highly impact the quality response, and second, at the element level, further determine the dominant segments or parts locally within the selected signals; (3) when the response is also conveyed by a function with a smoothness functional integrity, the quality measurements at neighbouring locations in this functional response would exhibit certain similarities; (4) to exploit this explicitly in the predictive model, adopt the multitask learning (MTL) to solve these multiple interconnected sub-regression tasks jointly; (5) the devised mechanism is to intentionally penalize the pairwise coefficients’ differences of these sub-regression tasks and the penalties will become pronounced when two locations in the functional response get close to each other; (6) in this way, achieve the transfer of knowledge on model parameters across multiple related tasks, i.e., the information on the coefficient values in each task will be transferred into its adjacent tasks to make these tasks together tend to have similar coefficients; and (7) the proposed FR model which integrates the above two favorable properties, i.e., sparsity and structure, is formulated as a tractable convex problem. Wang also discloses in Section II of Pages 6721-6722 that (1) a second-order derivative roughness penalty is often imposed to control the degree of smoothness of the estimated coefficient functions; (2) besides the basis function representation and the roughness level governance, the sparsity regularization has also been widely used in the FR models to cope with the high-dimensional challenge, and more importantly, to make the regression model more interpretable via VS; and (3) the use of the MTL in the function-on-function regression scenario to leverage the inherent smoothness integrity of the functional response, that is, to explore the potential similarity structure among a range of sub-regression tasks. Wang further teaches in Section III with FIGS. 2-3 of Pages 6722-6725 that (1) provide the preliminaries of function-on-function regression model; (2) then the TLH-VS via a sparsity regularization and the MTL via a structure regularization are developed to formulize the proposed SSF2R model; (3) an efficient iterative optimization algorithm is derived for model parameter estimation; (4) the extension to dynamical prediction with partial observations is finally discussed; (5) the function-on-function regression model is shown in Eq. (1); (6) the discrete version of (1) is shown in Eq. (2); (7) the matrix form of (2) is shown in Eq. (3); (8) the quality predictive model can be established by minimizing the composite objective function shown in Eq. (4); (9) h(Θ) is a regularization term which is usually used to suppress overly model complexity and facilitate proper parameter structure, and λ ≥ 0 is the tuning parameter; (10) proposed sparsity regularization for the TLH-VS is a combination of h1(Θ) and h2(Θ) in (5) and (6) shown in Eq. (7), where λ1 ≥ 0 and λ2 ≥ 0 are the tuning parameters; (11) the function-on-function regression model in (2) is actually comprised of a series of S sub-regression tasks, each of which predicts the surface response evaluated at one particular spatial index s; (12) for any two spatial indexes s and s', we first calculate a similarity measure c(s, s') based on the distance of their respective 2-D coordinates, where c(s, s') can be numerical in [0, 1] or binary in {0, 1}; (13) MTL is devised in light of the intuition that when two spatial indexes s and s' are highly adjacent in the surface response with a large c(s, s'), their associated sub-regression tasks would have approximate coefficients; (14) to explicitly utilize such a similarity structure among these sub-regression tasks, a weighed fused LASSO penalty is defined as Eq. (8), where c(s, s') acts as a weight to impose more severe penalty when s and s' tend to be closer, and C is used to calculate the weighted pairwise differences of coefficients in each row of Θ; (15) the potential similarity structure of our bivariate coefficient function θj(t, s) over s has been exploited in (8), but it also possesses a smoothness nature over t when s is fixed; (16) finally admit another penalty term to the MTL which controls the coefficient function’s roughness longitudinally in the time domain as shown in Eq. (9); (17) the structure regularization is a summation of h3(Θ) and h4(Θ) in (8) and (9) shown in Eq. (10), where λ3 ≥ 0 and λ4 ≥ 0 are the tuning parameters; (18) SSF2R model is finally built by combining (4), (7), and (10) as shown in Eq. (11), where the tuning parameters are re-parameterized as λ1 = λγα, λ2 = λγ(1 − α), λ3 = λ(1 − γ)β and λ4 = λ(1 − γ)(1 − β) with λ ≥ 0 and γ,α, β ∈ [0, 1] to simplify their selections; (19) to solve non-differentiable for L1-norm when evaluated at the zero point, apply the ADMM to derive an efficient algorithm shown in Eq. (12), where ρ > 0 is called the penalty parameter in ADMM; (20) the dominant computing burden of our algorithm lies in the updating of Θ which involves the inversion of A; (21) fortunately, this difficulty can be skipped based on the property of the Kronecker sum operator; (22) the overall computation cost of the algorithm is modest, and the updating of Z can be performed parallelly to further enhance the computational efficiency; (23) based on the convergency property of the ADMM for convex optimization, the algorithm is guaranteed to get the global optimal solution; (24) the SSF2R model built in the above sections takes the complete signatures of signal predictors as input and outputs a predicted surface after the manufacturing is finished, which is useful in the soft sensing applications to save measurement costs and detect quality anomalies, but it can also be extended to a dynamical situation where the data points of signal predictors arrive progressively during an ongoing manufacturing process; i.e., at any time before the manufacturing process is completely over, one can predict the final product surface based only on the partial observations of the signal predictors; (25) the quality response can be foreseen as Eq. (15), where the prediction is based on a supervised functional predictor completion (SFPC) as the regression coefficients which are estimated from the training dataset are utilized as weights in (13); and (26) such an SFPC is also adaptive since the contribution of each training sample, i.e., wi(s), could be different when predicting quality at different spatial indexes.
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 HWEI-MIN LU whose telephone number is (313)446-4913. The examiner can normally be reached Mon - Fri: 9:00 AM - 6:00 PM 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, Mariela D. Reyes can be reached at (571) 270-1006. 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.
/HWEI-MIN LU/Primary Examiner, Art Unit 2142