DETAILED ACTION
This action is responsive to the amendment filed on 05/27/2026 for application 18/186,101. Claims 1, 5-9 and 13-17 are pending in the case. Claims 1 and 9 are independent claims. Claims 1, 5, 7, 9, 11, 13 15-17 are amended with claims 2-4 and 10-12 canceled.
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 .
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.
Claim Rejections - 35 USC § 112
Claims 1, 5-9 and 13-17 is/are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Regarding claim 1 and 9, the phrase upon determining that the restoration error is greater than or equal to the reference value renders the claim indefinite because it is unclear whether the limitation(s) following the phrase are part of the claimed invention. See MPEP § 2173.05(d).
Claim Rejections - 35 USC § 103
Claim(s) 1-15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Sokhandan et al.(US 20220108163A1, henceforth known as Sokhandan) and further in view of Clark et al.(“ Adaptive Threshold for Outlier Detection on Data Streams” henceforth known as Clark) in further view of Cheng et al(US20190286506A1, henceforth known as Cheng) with Ikeda et al(“Human-Assisted Online Anomaly Detection with Normal Outlier Retraining” henceforth known as Ikeda)
Regarding Claim 1:
Sokhandan discloses (a) after learning of a detection network and a classification network is completed, acquiring, by a detection and classification apparatus, information about a medium from an inspection target to detect whether there is an abnormality in the medium, the abnormality of which is unknown(Sokhandan, [0109], “The sequence 500 may begin at operation 510 by supplying, to the image encoder 105, an input image 150 of the object, the input image 150 of the object containing zero or more anomalies” where the input image of an object is considered information about a medium of anomaly detection from an inspection target.)
Sokhandan discloses (b) generating, by the detection and classification apparatus, an input value, which is a feature vector matrix including a plurality of feature vectors, from the medium information(Sokhandan, [0082], “…The image models are placed in a latent space 115. In more details, a neural network is used to extract a compact set of image features, smaller than the size of the original images, to form the image models placed in the latent space 115. In a non-limiting embodiment, the neural network may be based on a normalizing flow structure” where extracting a set of features corresponds to generating an input value which is a feature vector matrix from the medium information(See Also Sokhandan, [0019], “supplying, to the latent space, a statistically sufficient sample of information contained in the vectors containing the means and standard deviations”))
Sokhandan discloses (c) deriving, by the detection and classification apparatus, a restored value imitating the input value through the detection network learned to generate the restored value for the input value(Sokhandan, [0078], “Having learned a rich representation of the non-anomalous object, the system is able to receive an input image of a particular object that may contain anomalies, generate an image model and regenerate a substitute non-anomalous image of the object” where generating an image model and regenerate a substitute image of the object corresponds to deriving a restored value imitating the input value to generate a restored value for the input value)
Sokhandan discloses (d) calculating, by the detection and classification apparatus, a classification value(Sokhandan, [0121], “The training engine 400 may further use classification values obtained from the classifier 260 in training the system 200 or 300”) indicating a probability that the input value belongs to a category of an anomaly state(Sokhandan, [0096], “Additionally, the system 200 may comprise a classifier 260 that is supplied with the labels from the supplier 240 and with at least some of the content of the latent space 115. The classifier 260 may use the content of the latent space 115 to generate classification information for each anomaly type. The latent space 115 contains a set of extracted features at the output of the encoder 105” and Figures 11-14 that show heat-map displaying anomaly probabilities on objects) through the classification network learned to calculate the probability for the input value(Sokhandan, [0097] The classifier 260 may further use the labels identifying the one or more anomaly types to calculate a classification loss 265 for each of the anomaly types. The system 200 may further be trained using the one or more classification losses 265 calculated for the one or more anomaly types. The classification loss 265 may, for example and without limitation, be calculated as expressed in https://en.wikipedia.org/wiki/Cross entropy, the disclosure of which is incorporated by reference herein” where cross-entropy loss being defined over predicted class probabilities corresponds to a classification network learned to calculate the probability or the input value)
Sokhandan discloses (h) detecting, by the detection and classification apparatus, occurrence of an event requiring a model update(Sokhandan, [0099], “[73] In the industrial context where the object is produced or tested, new anomaly types may be detected after a few weeks or a few months of production.” where new anomaly types being detected triggering retraining is corresponds to an event requiting a model update),
Sokhandan discloses upon detecting the occurrence of the event…learning the classification network using the stored category data when category data of a second predetermined number or more are stored(Sokhandan, [0099], “When new anomaly types are identified for the object, one or more new sets of anomalous images of the object are supplied to the image encoder. These images contain anomalies corresponding to one or more new anomaly types for the object…The supplier 240 provides new labels to the anomaly encoder 245, each new label corresponding to a given one of the new anomalous images of the object and identifying a related new anomaly type” and [0033] “…the training engine being also adapted to retrain the system by: supplying, to the image encoder, one or more second sets of images corresponding to one or more second anomaly types for the object” where supplying a new set of anomalous images to train the network and the supply providing new labels to the new anomalous images and identifying the new anomaly type corresponds to learning the classification network using the stored category data and where sending images corresponds to one or more anomalous type corresponds to when category data of a second predetermined number or more are stored as the new images are used in retraining all the time, including when a predetermined number or more are stored)
Sokhandan does not explicitly teach the following limitations:
(e) determining, by the detection and classification apparatus, whether a restoration error indicating a difference between the input value and the restored value is greater than or equal to a previously calculated reference value
determining whether the classification value is greater than or equal to a predetermined threshold, upon determining that the restoration error is greater than or equal to the reference value
(g) storing, by the detection and classification apparatus, the input value as category data upon determining that the classification value is greater than or equal to the predetermined threshold
(f) storing, by the detection and classification apparatus, the input value as normal data upon determining that the restoration error is less than the reference value
upon detecting the occurrence of the event, by the detection and classification apparatus learning the detection network using the stored normal data when normal data of a first predetermined number or more are stored
Clark discloses (e) determining, by the detection and classification apparatus, whether a restoration error indicating a difference between the input value and the restored value is greater than or equal to a previously calculated reference value(Clark, Algorithm 1 lines 6-10, where the scoret > threshold if else statement and scoret <- f(xt) corresponds to determining restoration error indicating a difference between the input value and the restored value is greater than or equal to a previously calculated reference value)
References Sokhandan and Clark are analogous art because they are from the same field of endeavor of using machine-learning methods to determine anomaly/outlier in detection systems.
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Sokhandan and Clark before him or her, to modify the anomaly scoring and thresholding of Sokhandan to include the inference step (computing anomaly score(scoret)), threshold decision classifying as outlier or normal and selective memory updating of Clark to reduce false alarms and ensuring normal-identified data has influence in the network. The suggestion/motivation for doing so would have been “we describe the method we have designed which outfits a one-class learning anomaly detection system with an adaptive threshold setting method whose aim is to decrease the false alarm rate” (Clark, Page 3, Col. 1, Paragraph 2)
Sokhandan-Clark does not explicitly teach the following limitations:
determining whether the classification value is greater than or equal to a predetermined threshold, upon determining that the restoration error is greater than or equal to the reference value
(g) storing, by the detection and classification apparatus, the input value as category data upon determining that the classification value is greater than or equal to the predetermined threshold
(f) storing, by the detection and classification apparatus, the input value as normal data upon determining that the restoration error is less than the reference value
upon detecting the occurrence of the event, by the detection and classification apparatus learning the detection network using the stored normal data when normal data of a first predetermined number or more are stored
Cheng discloses determining whether the classification value is greater than or equal to a predetermined threshold(Cheng, Figure 7 and [0050]-[0051], “…Where the original sensor data and reconstructed sensor data deviate from each other by greater than a threshold amount, the data can be determined to indicate anomalous behavior.… Thus, the anomaly evaluator 406 quantifies and scores reconstructed data. Where the score rises above a threshold level, the data is considered anomalous…” where comparing the reconstructed data to the sensor/original data and with an anomaly evaluator to determine anomalies with the anomaly evaluator checking whether a score rises above a threshold corresponds to determining whether the classification value is greater than or equal to a predetermined threshold ), upon determining that the restoration error is greater than or equal to the reference value (Cheng, [0046], “…Accordingly, the reconstructed data and the original set of data can be compared to determine a difference. Where the difference is high, for example, above a threshold error level, the set of data is deemed anomalous, and thus corresponding to a suspected fault.” and [0064]- [0065] “As a result, the autoencoder 430 is trained to reconstruct sensor data 502 according to normal operating behavior. Therefore, if the sensor data 502 includes anomalous operating behavior, the generated vectors from the decoder 520 will differ from the sensor data 502. However, the sensor data 502 includes normal operating behavior, then the generated vectors and the sensor data 502 will match within a degree of acceptable error…The degree of acceptable error can include a threshold that is, e.g., learned, predetermine, or user selectable” where acceptable error or difference corresponds to a reference value and, when exceeded, the data being determined to be anomalous corresponds to determining that the restoration error is greater than or equal to the reference value)
References Sokhandan-Clark and Cheng are analogous art because they are from the same field of endeavor of continuously retraining machine learning models as an anomaly detector on new anomaly types.
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Sokhandan-Clark and Cheng before him or her, to modify the classification of anomalies of Sokhandan-Clark to include the threshold-based determination of Cheng to allow for an acceptable degree of error. The suggestion/motivation for doing so would have been Sokhandan, [0064]-[0065] “…However, the sensor data 502 includes normal operating behavior, then the generated vectors and the sensor data 502 will match within a degree of acceptable error… The degree of acceptable error can include a threshold that is, e.g., learned, predetermine, or user selectable”
Cheng discloses (g) storing, by the detection and classification apparatus, the input value as category data upon determining that the classification value is greater than or equal to the predetermined threshold(Cheng, Figure 7 and [0076], “Where the similarity score exceeds a threshold value, the ranked anomalies 601 and the corresponding historical anomaly are deemed similar” where 714 and 715 show determining anomalies from sensor data, ranking the anomalies with 703 producing an anomaly classification based on historical anomalies with an anomaly classifier corresponds to storing, by the detection and classification apparatus, the input value as category data upon determining that the classification value is greater than or equal to the predetermined threshold as to produce an anomaly classification the data must be stored in the model/machine to be able to be accessed and the classification of the sensors data as anomalous/not anomalous as well as categorizing the anomalies is determining category data for input values)
Sokhandan-Clark-Cheng does not explicitly teach the following limitations:
(f) storing, by the detection and classification apparatus, the input value as normal data upon determining that the restoration error is less than the reference value
upon detecting the occurrence of the event, by the detection and classification apparatus learning the detection network using the stored normal data when normal data of a first predetermined number or more are stored
Ikeda discloses (f) storing, by the detection and classification apparatus, the input value as normal data upon determining that the restoration error is less than the reference value, (Ikeda, Algorithm 1, Lines 26-28, “If MSE(xt) < (γαano) or operator judged xt as normal then…Update preprocessing parameters…Add xt to Xnormal” where adding xt to Xnormal based on the comparison of MSE(xt) < (γαano) corresponds to storing the input value as normal data upon determining that the restoration error is less than the reference value)
References Sokhandan-Clark-Cheng and Ikeda are analogous art because they are from the same field of endeavor of machine learning anomaly detection with continuing/online model adaptation.
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Sokhandan-Clark-Cheng and Ikeda before him or her, to modify the training routine of Sokhandan-Clark-Cheng to include the online learning and retraining procedure of Ikeda to allow for a more robust scalability in large systems. The suggestion/motivation for doing so would have been Ikeda, Page 3, Col. 2, Paragraph 3, “In addition, the batch learning approach requires past training data, which is not promising for large ICT systems.”
Ikeda discloses (upon detecting the occurrence of the event, by the detection and classification apparatus learning the detection network using the stored normal data when normal data of a first predetermined number or more are stored(Ikeda, Page 3, Col. 1, Paragraph 6,“If online test data is determined as normal with the algorithm or, the operator confirmed that the detection is a FP even though the data was determined as anomaly, the data is used for retraining”)
Regarding Claim 5:
The rejection of claim 1 is incorporated and further:
Sokhandan discloses initializing, by the detection and classification apparatus, the detection network(Sokhandan, [0082], “FIG. 1 is a block diagram of an anomaly detection system”) inputting, by the detection and classification apparatus, the stored normal data as a training input value to the initialized detection network(Sokhandan, [0082], FIG. 1 is a block diagram of an anomaly detection system 100 adapted to be trained in unsupervised mode in accordance with an embodiment of the present technology. The system 100 includes an image encoder 105 that receives input images 110 and forms an image model for each input image 110) calculating, by the detection and classification apparatus, an uncompressed latent value from the training input value(Sokhandan, [0082], “The image models are placed in a latent space 115. In more details, a neural network is used to extract a compact set of image features, smaller than the size of the original images, to form the image models placed in the latent space”) calculating, by the detection and classification apparatus, the restored value from the latent value(Sokhandan, [0082], “An image decoder 120 produces regenerated images 125 based on the image models placed in the latent space 115.”) calculating, by the detection and classification apparatus, a loss that is a difference between the restored value and the training input value(Sokhandan, [0086], “The system 100 calculates a reconstruction loss 130 using equation (1)”) and performing, by the detection and classification apparatus, optimization of updating a parameter of the detection network to minimize the loss(Sokhandan, [0091], “The system 100 is trained using the reconstruction loss 130 and may further be trained using the log-likelihood loss 135 and the regularization loss, following which the system 100 is ready to identify anomalies in a particular object similar to the anomaly-free object. This training process is sometimes called “optimization through backpropagation”, a technique that has been used for training various types of neural networks. In this process, the gradient of the loss with respect to each layer in the neural network is computed and is used to update the corresponding weights in that layer.”)
Regarding claim 6:
The rejection of claim 5 is incorporated and further:
Clark discloses after the learning, calculating, by the detection and classification apparatus, the reference value in accordance with Equation 0=p+(kxa), wherein p denotes an average of a mean squared error (MSE) between a plurality of training input values and a plurality of restored values corresponding to the plurality of training input values used for learning on the detection network, wherein a denotes a standard deviation of the MSE between the plurality of training input values and the plurality of restored values corresponding to the plurality of training input values, and wherein k is a weight for the standard deviation. (Clark, Page 4, Paragraph 1, “The first part of the algorithm consists of obtaining the initial threshold (lines1-3). S is divided into a training data set and a validation data set. The training set is used to train a model (in this work, either an autoencoder or LOF), and the validation set is used to obtain the initial threshold. This threshold is set to the mean + 2 standard deviations of the outlier scores output by the model on the validation data” where the threshold being determined by mean + 2 * standard deviates corresponds to having a reference value in accordance with the Equation θ= μ+(k × σ),
Regarding Claim 7:
The rejection of claim 1 is incorporated and further:
Sokhandan discloses initializing, by the detection and classification apparatus, the classification network; (Sokhandan, [0096], “Additionally, the system 200 may comprise a classifier”) preparing, by the detection and classification apparatus, a training input value(Sokhandan, [0096], “…that is supplied with the labels from the supplier 240 and with at least some of the content of the latent space 115.”) by setting a label corresponding to a category of the stored category data(Sokhandan, [0097], “The classifier 260 may further use the labels identifying the one or more anomaly types to calculate a classification loss 265 for each of the anomaly types.”) inputting, by the detection and classification apparatus, the training input value to the initialized classification network(Sokhandan, [0096], “…The latent space 115 contains a set of extracted features at the output of the encoder 105. The classifier 260 may take these features as input”) calculating, by the det
ection and classification apparatus, a classification value from the training input value by performing an operation in which a plurality of inter-layer weights are applied(Sokhandan, [0097], “…The system 200 may further be trained using the one or more classification losses 265 calculated for the one or more anomaly types.”) calculating, by the detection and classification apparatus, a classification loss that is a difference between the classification value and the label(Sokhandan, [0097], “…The classification loss 265 may, for example and without limitation, be calculated as expressed in https://en.wikipedia.org/wiki/Cross entropy, the disclosure of which is incorporated by reference herein” where the cross entropy quantifying the difference between the classification value produced by the classification network and the ground-truth label corresponds to the calculating a log loss that is the difference between the classification value and the label) and performing, by the detection and classification apparatus, optimization of updating a parameter of the classification network(Sokhandan, [0104], “The training engine 400 may also obtain values for the classification loss 265 from the systems 200 or 300” where the training engine explicitly consuming the classification loss to input into the training when updating the parameters corresponds optimization of the classification network ) to minimize the classification loss(Sokhandan, [0091], “The system 100 is trained using the reconstruction loss 130 and may further be trained…This training process is sometimes called “optimization through backpropagation”…the gradient of the loss with respect to each layer in the neural network is computed and is used to update the corresponding weights in that layer” where the use of backpropagation is specifically designed to minimize the loss function by and the classification loss being used in the training corresponds to updating parameters to of the classification network to minimize the classification loss.)
Regarding Claim 8:
The rejection of claim 1 is incorporated and further:
Sokhandan discloses a non-transitory computer-readable recording medium that records a program for executing the method for continuous learning according to claim 1(Sokhandan, [0075], “Similarly, it will be appreciated that any flowcharts, flow diagrams, state transition diagrams, pseudo-code, and the like represent various processes that may be substantially represented in non-transitory computer-readable media and so executed by a computer or processor, whether or not such computer or processor is explicitly shown.”)
Regarding Claim 9:
Sokhandan discloses data processing unit configured to(Sokhandan, [0076] “…In some embodiments of the present technology, the processor may be a general-purpose processor, such as a central processing unit (CPU)” where a CPU corresponds to a data processing unit) generate an input value, which is a feature vector matrix including a plurality of feature vectors, from information about a medium, wherein after learning oSokhandan, [0082], “…The image models are placed in a latent space 115. In more details, a neural network is used to extract a compact set of image features, smaller than the size of the original images, to form the image models placed in the latent space 115. In a non-limiting embodiment, the neural network may be based on a normalizing flow structure” where extracting a set of features corresponds to generating an input value which is a feature vector matrix from the medium information(See Also Sokhandan, [0019], “supplying, to the latent space, a statistically sufficient sample of information contained in the vectors containing the means and standard deviations”)) to detect whether there is an abnormality in the medium, the abnormality of which is unknown; (Sokhandan,[0099], “In the industrial context where the object is produced or tested, new anomaly types may be detected after a few weeks or a few months of production” where detecting new anomalies corresponds to detecting unknown abnormalities as the anomalies are new and therefore ukknown)
Sokhandan discloses (a) after learning of a detection network and a classification network is completed, acquiring, by a detection and classification apparatus, information about a medium from an inspection target to detect whether there is an abnormality in the medium, the abnormality of which is unknown(Sokhandan, [0109], “The sequence 500 may begin at operation 510 by supplying, to the image encoder 105, an input image 150 of the object, the input image 150 of the object containing zero or more anomalies” where the input image of an object is considered information about a medium of anomaly detection from an inspection target.)
Sokhandan discloses (b) generating, by the detection and classification apparatus, an input value, which is a feature vector matrix including a plurality of feature vectors, from the medium information(Sokhandan, [0082], “…The image models are placed in a latent space 115. In more details, a neural network is used to extract a compact set of image features, smaller than the size of the original images, to form the image models placed in the latent space 115. In a non-limiting embodiment, the neural network may be based on a normalizing flow structure” where extracting a set of features corresponds to generating an input value which is a feature vector matrix from the medium information(See Also Sokhandan, [0019], “supplying, to the latent space, a statistically sufficient sample of information contained in the vectors containing the means and standard deviations”))
Sokhandan discloses (c) deriving, by the detection and classification apparatus, a restored value imitating the input value through the detection network learned to generate the restored value for the input value(Sokhandan, [0078], “Having learned a rich representation of the non-anomalous object, the system is able to receive an input image of a particular object that may contain anomalies, generate an image model and regenerate a substitute non-anomalous image of the object” where generating an image model and regenerate a substitute image of the object corresponds to deriving a restored value imitating the input value to generate a restored value for the input value)
Sokhandan discloses (d) calculating, by the detection and classification apparatus, a classification value(Sokhandan, [0121], “The training engine 400 may further use classification values obtained from the classifier 260 in training the system 200 or 300”) indicating a probability that the input value belongs to a category of an anomaly state(Sokhandan, [0096], “Additionally, the system 200 may comprise a classifier 260 that is supplied with the labels from the supplier 240 and with at least some of the content of the latent space 115. The classifier 260 may use the content of the latent space 115 to generate classification information for each anomaly type. The latent space 115 contains a set of extracted features at the output of the encoder 105” and Figures 11-14 that show heat-map displaying anomaly probabilities on objects) through the classification network learned to calculate the probability for the input value(Sokhandan, [0097] The classifier 260 may further use the labels identifying the one or more anomaly types to calculate a classification loss 265 for each of the anomaly types. The system 200 may further be trained using the one or more classification losses 265 calculated for the one or more anomaly types. The classification loss 265 may, for example and without limitation, be calculated as expressed in https://en.wikipedia.org/wiki/Cross entropy, the disclosure of which is incorporated by reference herein” where cross-entropy loss being defined over predicted class probabilities corresponds to a classification network learned to calculate the probability or the input value)
Sokhandan discloses (h) detecting, by the detection and classification apparatus, occurrence of an event requiring a model update(Sokhandan, [0099], “[73] In the industrial context where the object is produced or tested, new anomaly types may be detected after a few weeks or a few months of production.” where new anomaly types being detected triggering retraining is corresponds to an event requiting a model update),
Sokhandan discloses upon detecting the occurrence of the event…learning the classification network using the stored category data when category data of a second predetermined number or more are stored(Sokhandan, [0099], “When new anomaly types are identified for the object, one or more new sets of anomalous images of the object are supplied to the image encoder. These images contain anomalies corresponding to one or more new anomaly types for the object…The supplier 240 provides new labels to the anomaly encoder 245, each new label corresponding to a given one of the new anomalous images of the object and identifying a related new anomaly type” and [0033] “…the training engine being also adapted to retrain the system by: supplying, to the image encoder, one or more second sets of images corresponding to one or more second anomaly types for the object” where supplying a new set of anomalous images to train the network and the supplier providing new labels to the new anomalous images and identifying the new anomaly type corresponds to learning the classification network using the stored category data and where sending images corresponds to one or more anomalous type corresponds to when category data of a second predetermined number or more are stored as the new images are used in retraining all the time, including when a predetermined number or more are stored)
Sokhandan does not explicitly teach the following limitations:
determining, by the detection and classification apparatus, whether a restoration error indicating a difference between the input value and the restored value is greater than or equal to a previously calculated reference value
determining whether the classification value is greater than or equal to a predetermined threshold, upon determining that the restoration error is greater than or equal to the reference value
storing, by the detection and classification apparatus, the input value as category data upon determining that the classification value is greater than or equal to the predetermined threshold
storing, by the detection and classification apparatus, the input value as normal data upon determining that the restoration error is less than the reference value
upon detecting the occurrence of the event, by the detection and classification apparatus learning the detection network using the stored normal data when normal data of a first predetermined number or more are stored
Clark discloses determining, by the detection and classification apparatus, whether a restoration error indicating a difference between the input value and the restored value is greater than or equal to a previously calculated reference value(Clark, Algorithm 1 lines 6-10, where the scoret > threshold if else statement and scoret <- f(xt) corresponds to determining restoration error indicating a difference between the input value and the restored value is greater than or equal to a previously calculated reference value)
References Sokhandan and Clark are analogous art because they are from the same field of endeavor of using machine-learning methods to determine anomaly/outlier in detection systems.
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Sokhandan and Clark before him or her, to modify the anomaly scoring and thresholding of Sokhandan to include the inference step (computing anomaly score(scoret)), threshold decision classifying as outlier or normal and selective memory updating of Clark to reduce false alarms and ensuring normal-identified data has influence in the network. The suggestion/motivation for doing so would have been “we describe the method we have designed which outfits a one-class learning anomaly detection system with an adaptive threshold setting method whose aim is to decrease the false alarm rate” (Clark, Page 3, Col. 1, Paragraph 2)
Sokhandan-Clark does not explicitly teach the following limitations:
determining whether the classification value is greater than or equal to a predetermined threshold, upon determining that the restoration error is greater than or equal to the reference value
storing, by the detection and classification apparatus, the input value as category data upon determining that the classification value is greater than or equal to the predetermined threshold
storing, by the detection and classification apparatus, the input value as normal data upon determining that the restoration error is less than the reference value
upon detecting the occurrence of the event, by the detection and classification apparatus learning the detection network using the stored normal data when normal data of a first predetermined number or more are stored
Cheng discloses determining whether the classification value is greater than or equal to a predetermined threshold(Cheng, Figure 7 and [0050]-[0051], “…Where the original sensor data and reconstructed sensor data deviate from each other by greater than a threshold amount, the data can be determined to indicate anomalous behavior.… Thus, the anomaly evaluator 406 quantifies and scores reconstructed data. Where the score rises above a threshold level, the data is considered anomalous…” where comparing the reconstructed data to the sensor/original data and with an anomaly evaluator to determine anomalies with the anomaly evaluator checking whether a score rises above a threshold corresponds to determining whether the classification value is greater than or equal to a predetermined threshold ), upon determining that the restoration error is greater than or equal to the reference value (Cheng, [0046], “…Accordingly, the reconstructed data and the original set of data can be compared to determine a difference. Where the difference is high, for example, above a threshold error level, the set of data is deemed anomalous, and thus corresponding to a suspected fault.” and [0064]- [0065] “As a result, the autoencoder 430 is trained to reconstruct sensor data 502 according to normal operating behavior. Therefore, if the sensor data 502 includes anomalous operating behavior, the generated vectors from the decoder 520 will differ from the sensor data 502. However, the sensor data 502 includes normal operating behavior, then the generated vectors and the sensor data 502 will match within a degree of acceptable error…The degree of acceptable error can include a threshold that is, e.g., learned, predetermine, or user selectable” where acceptable error or difference corresponds to a reference value and, when exceeded, the data being determined to be anomalous corresponds to determining that the restoration error is greater than or equal to the reference value)
References Sokhandan-Clark and Cheng are analogous art because they are from the same field of endeavor of continuously retraining machine learning models as an anomaly detector on new anomaly types.
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Sokhandan-Clark and Cheng before him or her, to modify the classification of anomalies of Sokhandan-Clark to include the threshold-based determination of Cheng to allow for an acceptable degree of error. The suggestion/motivation for doing so would have been Sokhandan, [0064]-[0065] “…However, the sensor data 502 includes normal operating behavior, then the generated vectors and the sensor data 502 will match within a degree of acceptable error… The degree of acceptable error can include a threshold that is, e.g., learned, predetermine, or user selectable”
Cheng discloses (g) storing, by the detection and classification apparatus, the input value as category data upon determining that the classification value is greater than or equal to the predetermined threshold(Cheng, Figure 7 and [0076], “Where the similarity score exceeds a threshold value, the ranked anomalies 601 and the corresponding historical anomaly are deemed similar” where 714 and 715 show determining anomalies from sensor data, ranking the anomalies with 703 producing an anomaly classification based on historical anomalies with an anomaly classifier corresponds to storing, by the detection and classification apparatus, the input value as category data upon determining that the classification value is greater than or equal to the predetermined threshold as to produce an anomaly classification the data must be stored in the model/machine to be able to be accessed and the classification of the sensors data as anomalous/not anomalous as well as categorizing the anomalies is determining category data for input values)
Sokhandan-Clark-Cheng does not explicitly teach the following limitations:
storing, by the detection and classification apparatus, the input value as normal data upon determining that the restoration error is less than the reference value
upon detecting the occurrence of the event, by the detection and classification apparatus learning the detection network using the stored normal data when normal data of a first predetermined number or more are stored
Ikeda discloses storing, by the detection and classification apparatus, the input value as normal data upon determining that the restoration error is less than the reference value, (Ikeda, Algorithm 1, Lines 26-28, “If MSE(xt) < (γαano) or operator judged xt as normal then…Update preprocessing parameters…Add xt to Xnormal” where adding xt to Xnormal based on the comparison of MSE(xt) < (γαano) corresponds to storing the input value as normal data upon determining that the restoration error is less than the reference value)
References Sokhandan-Clark-Cheng and Ikeda are analogous art because they are from the same field of endeavor of machine learning anomaly detection with continuing/online model adaptation.
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Sokhandan-Clark-Cheng and Ikeda before him or her, to modify the training routine of Sokhandan-Clark-Cheng to include the online learning and retraining procedure of Ikeda to allow for a more robust scalability in large systems. The suggestion/motivation for doing so would have been Ikeda, Page 3, Col. 2, Paragraph 3, “In addition, the batch learning approach requires past training data, which is not promising for large ICT systems.”
Ikeda discloses (upon detecting the occurrence of the event, by the detection and classification apparatus learning the detection network using the stored normal data when normal data of a first predetermined number or more are stored(Ikeda, Page 3, Col. 1, Paragraph 6,“If online test data is determined as normal with the algorithm or, the operator confirmed that the detection is a FP even though the data was determined as anomaly, the data is used for retraining”)
Regarding Claim 13:
The rejection of claim 9 is incorporated and further:
Sokhandan discloses initialize the detection network(Sokhandan, [0082], “FIG. 1 is a block diagram of an anomaly detection system”), and input the stored normal data as a training input value to the initialized detection network(Sokhandan, [0082], FIG. 1 is a block diagram of an anomaly detection system 100 adapted to be trained in unsupervised mode in accordance with an embodiment of the present technology. The system 100 includes an image encoder 105 that receives input images 110 and forms an image model for each input image 110), when an encoder of the detection network calculates an uncompressed latent value from the training input value, and calculates the restored value from the latent value(Sokhandan, [0082], “An image decoder 120 produces regenerated images 125 based on the image models placed in the latent space 115.”), calculate a loss that is a difference between the restored value and the training input value(Sokhandan, [0086], “The system 100 calculates a reconstruction loss 130 using equation (1)”), and perform optimization of updating a parameter of the detection network to minimize the loss(Sokhandan, [0091], “The system 100 is trained using the reconstruction loss 130 and may further be trained using the log-likelihood loss 135 and the regularization loss, following which the system 100 is ready to identify anomalies in a particular object similar to the anomaly-free object. This training process is sometimes called “optimization through backpropagation”, a technique that has been used for training various types of neural networks. In this process, the gradient of the loss with respect to each layer in the neural network is computed and is used to update the corresponding weights in that layer.”)
Regarding Claim 14:
The rejection of claim 13 is incorporated and further:
Sokhandan discloses wherein the learning unit is configured to:calculate the reference value in accordance with Equation θ= μ+(k × σ), wherein denotes an average of a mean squared error (MSE) between a plurality of training input values and a plurality of restored values corresponding to the plurality of training input values used for learning on the detection network, wherein a denotes a standard deviation of the MSE between the plurality of training input values and the plurality of restored values corresponding to the plurality of training input values, and wherein k is a weight for the standard deviation(Clark, Page 4, Paragraph 1, “The first part of the algorithm consists of obtaining the initial threshold (lines1-3). S is divided into a training data set and a validation data set. The training set is used to train a model (in this work, either an autoencoder or LOF), and the validation set is used to obtain the initial threshold. This threshold is set to the mean + 2 standard deviations of the outlier scores output by the model on the validation data” where the threshold being determined by mean + 2 * standard deviates corresponds to having a reference value in accordance with the Equation θ= μ+(k × σ),
References Sokhandan and Clark are analogous art because they are from the same field of endeavor of using machine-learning methods to determine anomaly/outlier in detection systems.
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Sokhandan and Clark before him or her, to modify the thresholding of Sokhandan to include threshold rule of Clark to reduce false alarms and ensuring normal-identified data has influence in the network. The suggestion/motivation for doing so would have been “we describe the method we have designed which outfits a one-class learning anomaly detection system with an adaptive threshold setting method whose aim is to decrease the false alarm rate” (Clark, Page 3, Col. 1, Paragraph 2)
Regarding Claim 15:
The rejection of claim 9 is incorporated and further:
Sokhandan discloses initialize the classification network(Sokhandan, [0096], “Additionally, the system 200 may comprise a classifier”), prepare a training input value(Sokhandan, [0096], “…that is supplied with the labels from the supplier 240 and with at least some of the content of the latent space 115.”) by setting a label corresponding to a category of the stored category data(Sokhandan, [0097], “The classifier 260 may further use the labels identifying the one or more anomaly types to calculate a classification loss 265 for each of the anomaly types.”), and input the training input value to the initialized classification network(Sokhandan, [0096], “…The latent space 115 contains a set of extracted features at the output of the encoder 105. The classifier 260 may take these features as input”), when the classification network calculates a classification value from the training input value by performing an operation in which a plurality of inter-layer weights are applied(Sokhandan, [0097], “…The classification loss 265 may, for example and without limitation, be calculated as expressed in https://en.wikipedia.org/wiki/Cross entropy, the disclosure of which is incorporated by reference herein” where the cross entropy quantifying the difference between the classification value produced by the classification network and the ground-truth label corresponds to the calculating a log loss that is the difference between the classification value and the label), calculate a classification loss that is a difference between the classification value and the label, and perform optimization of updating a parameter of the classification network(Sokhandan, [0104], “The training engine 400 may also obtain values for the classification loss 265 from the systems 200 or 300” where the training engine explicitly consuming the classification loss to input into the training when updating the parameters corresponds optimization of the classification network ) to minimize the classification loss(Sokhandan, [0091], “The system 100 is trained using the reconstruction loss 130 and may further be trained…This training process is sometimes called “optimization through backpropagation”…the gradient of the loss with respect to each layer in the neural network is computed and is used to update the corresponding weights in that layer” where the use of backpropagation is specifically designed to minimize the loss function by and the classification loss being used in the training corresponds to updating parameters to of the classification network to minimize the classification loss.)
Claim(s) 16-17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Sokhandan et al.(US 20220108163A1, henceforth known as Sokhandan) and further in view of Clark et al.(“ Adaptive Threshold for Outlier Detection on Data Streams” henceforth known as Clark) in further view of Cheng et al(US20190286506A1, henceforth known as Cheng) with Ikeda et al(“Human-Assisted Online Anomaly Detection with Normal Outlier Retraining” henceforth known as Ikeda) and Chang et al. (US20190051310A1, henceforth known as Chang)
Regarding Claim 16:
The rejection of claim 13 is incorporated and further:
Sokhandan discloses wherein the learning unit(Sokhandan, [0083], “In an embodiment, the image encoder 105 implements an encoding function ge and the image decoder”)
Sokhandan-Clark-Cheng-Ikeda does not disclose, however, Chang discloses is configured to generate a mean square loss of an input value(Chang, [0050], “In Equation 1 and Equation 2, the LSGAN may indicate the least square GAN” where least square corresponds to a mean square loss of an input value ) and a restored value or use a restoration error(Chang, [0053], “In this case, a waveform in which the final lost packet is concealed may be restored at the top of the decoder” where the restored packet corresponds to a restored value)
References Sokhandan-Clark-Cheng-Ikeda and Clark are analogous art because they are both concerned about withed training generative/ Chang neural networks to infer or reconstruct useful information5.
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Sokhandan-Clark-Cheng-Ikeda and Clark before him or her, to modify the model of Sokhandan-Clark-Cheng-Ikeda to include the mean square loss and a restored value of Chang to train the generative model to generate realistic data. The suggestion/motivation for doing so would have been “…The generative model G 330 may perform training so that the classification model D 320 generates data…that makes the fake data…classified as 1 like real"(Chang, [0046])
Regarding Claim 17:
The rejection of claim 13 is incorporated and further:
Chang discloses wherein the learning unit(Chang, [0045], “Referring to FIG. 3, the GAN may include a classification model D and a generative model G” and Chang, [0050], “In Equation 1 and Equation 2, the LSGAN may indicate the least square GAN” where least square corresponds to a mean square loss of an input value with classification model D corresponding to a discriminator), and, is configured to: use a mean square loss of a discriminator output(Chang, [0050], “In Equation 1 and Equation 2, the LSGAN may indicate the least square GAN” where least square corresponds to a mean square loss of a discriminator output) for an actual input and a generated input(Chang, [0048], Equation 1, where X is real data, G(z) is generated data, D(X) is the discriminator output of real input, D(G(z)) is the discriminator output for the generated input) as a discrimination error(Chang, [0048], “For example, the classification model training unit 210 may perform training so that a classification model classifies the input X as a 1 and classifies the input ~X as 0 based on a loss function for the classification model D defined in Equation 1” where the equation 1 being defined as the loss function for the classification model D corresponds to a discrimination error)
References Sokhandan-Clark-Cheng-Ikeda and Clark are analogous art because they are both concerned about withed training generative/ Chang neural networks to infer or reconstruct useful information5.
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Sokhandan-Clark-Cheng-Ikeda and Clark before him or her, to modify the model of Sokhandan-Clark-Cheng-Ikeda to include the mean square loss and a restored value of Chang to train the generative model to generate realistic data. The suggestion/motivation for doing so would have been “…The generative model G 330 may perform training so that the classification model D 320 generates data…that makes the fake data…classified as 1 like real"(Chang, [0046])
Responds to Arguments
Applicant's arguments filed 5/27/2026 have been fully considered and are partially persuasive. A breakdown of arguments can be found below.
101:
Applicants arguments are persuasive and 101 has been withdrawn, specifically, applicant agrees with Page 14 that specifies additional elements that integrate the judicial exception(s) into a practical application
103:
Applicant appears to argue on pages 18-20 that the cited art do not teach the amended language of
Applicant’s arguments with respect to claim(s) have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument. Specifically, new prior art Ikeda is mapped to new/amended limitation (i) upon detecting the occurrence of the event, by the detection and classification apparatus learning the detection network using the stored normal data when normal data of a first predetermined number or more are stored and (f) storing, by the detection and classification apparatus, the input value as normal data upon determining that the restoration error is less than the reference value with the new prior art Cheng mapped to (g) storing, by the detection and classification apparatus, the input value as category data upon determining that the classification value is greater than or equal to the predetermined threshold as well as new prior art Chang mapped to the new/amended claims 16 and 17.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to CHARLES JEFFREY JONES JR whose telephone number is (703)756-1414. The examiner can normally be reached Monday - Friday 8:00 - 5:00 EST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Kakali Chaki can be reached at 571-272-3719. 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.
/C.J.J./Examiner, Art Unit 2122
/KAKALI CHAKI/Supervisory Patent Examiner, Art Unit 2122
Prosecution Insights