System and Method for Explainable Anomaly Detection

Application No. 18/068,537

• Filed: Dec 20, 2022
• Examiner: VU, TAYLOR P
• Art Unit: 2437
• Tech Center: 2400 — Computer Networks
• Assignee: Mitsubishi Electricresearch Laboratories Inc.
• OA Round: 4 (Non-Final)

Office 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 .

Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 02/05/2025 has been entered.

Response to Arguments
The present office action is responsive to communication filed on 02/05/2026. Claims 1-20 are currently pending.
Applicant arguments filed on 02/05/2026 have been fully considered but they are not fully persuasive.
On pages 1-2 of the Remarks, Applicant contends:
“Step 2A, Prong I assesses whether the claim recites a judicial exception, i.e., a law of nature, natural phenomenon, or abstract idea. In the Office Action, the Examiner lists certain limitations of Claim 1 and then summarily concludes that “the steps performing amount to an abstract idea which falls under a judicial exception.”
Applicant respectfully submits that this conclusion is incomplete and unsupported. The Office Action does not clearly identify which specific claim limitations are alleged to constitute an abstract idea, nor does it explain why any such limitations are abstract. Merely labeling the recited steps as “abstract” without articulating the purported abstract concept or providing a reasoned analysis fails to satisfy the Examiner’s burden under Step 2A, Prong One.
The Office Action further states that “abstract ideas fall in the category the claim are directed to abstract ideas because they merely apply generic AI-based models to a new data environment without reciting a technological improvement to the AI-based methods themselves.”
Applicant respectfully notes that this statement does not clearly articulate the Examiner’s rationale, making the Examiner’s rationale difficult to discern. To the extent the Examiner intends to assert that Claim 1 is abstract because it applies “generic AI-based models to a new data environment without reciting a technological improvement to the AI-based methods themselves,” such reasoning is misplaced under Step 2A, Prong I. Whether a claim recites a technological improvement or integrates an exception into a practical application is properly addressed under Step 2A, Prong II. The Office Action therefore fails to explain why the identified claim limitations are abstract in the first instance, relying instead on considerations that are legally relevant only after a judicial exception has been properly identified.
The Examiner next asserts that courts have found claims using artificial intelligence to generate network maps and schedules for television broadcasts to be attempts to patent an abstract idea, and that such claims lack an inventive concept sufficient to transform the abstract idea into patent-eligible subject matter. The Office Action further concludes that merely applying generic AI-based methods to a new data environment does not constitute an inventive concept, citing Recentive Analytics, Inc. v. Fox Corp.
Applicant respectfully submits that whether a claim includes an inventive concept sufficient to transform a judicial exception into patent-eligible subject matter is an inquiry properly conducted under Step 2B of the Alice/Mayo framework. As such, this analysis does not address the threshold question under Step 2A, Prong I: whether Claim 1 recites an abstract idea in the first instance. The Office Action therefore fails to provide a reasoned explanation as to why the identified claim limitations are abstract, instead relying on considerations that are legally relevant only after a judicial exception has been properly identified.
The determination of whether a claim recites an abstract idea is based on: (i) identifying the specific claim limitations that are alleged to recite an abstract idea, and (ii) determining whether those identified limitations fall within at least one of the abstract idea groupings defined in MPEP § 2106.04(a)(2), namely mathematical concepts, certain methods of organizing human activity, or mental processes.
The Office Action provides no rationale explaining why any identified limitations are abstract. Moreover, the Examiner does not identify which, if any, of the abstract idea groupings set forth in MPEP § 2106.04(a)(2) the limitations are alleged to fall within, nor does the Office Action explain how the limitations correspond to any such grouping. As a result, the rejection does not reflect a complete abstract idea determination under Step 2A, Prong I.
Accordingly, Applicant respectfully submits that the Examiner’s characterization of the limitations recited in Claim 1 as abstract is unsupported by the record and does not satisfy the requirements of Step 2A, Prong I of the subject matter eligibility analysis.”
Examiner takes note of the applicant’s comments and has clarified the rejection under 35 U.S.C. § 101 as set forth below. In Step 2A, Prong One, claim 1 recites an abstract idea falling within the mental process and mathematical concept groupings of MPEP § 2106.04(a)(2). In particular, claim 1 recites collecting input data, analyzing the data by comparing portions of the input data and reconstructed input data to determine a sequence of losses and classifying the sequence of losses to detect an anomaly type and/or severity, and rendering the result of anomaly detection. This is similar to the abstract idea identified in Electric Power Group, LLC v. Alstom S.A., 830 F.3d 1350, 119 USPQ2d 1739 (Fed. Cir. 2016), namely "collecting information, analyzing it, and displaying certain results of the collection and analysis." To the extent claim 1 recites generating reconstruction-losss values through mathematical comparison of portions of the input data and reconstructed input data, and classifying those values, the claim also recites mathematical concepts.
On pages 2-3 of the Remarks, Applicant contends:
“Step 2A, Prong II determines whether the claim as a whole integrates a judicial exception into a practical application such that the claim is not directed to a judicial exception.
The Office Action asserts that the alleged judicial exception in Claim 1 is not integrated into a practical application because the claim recites “generically recited computer elements,” such as a processor and a memory. Applicant respectfully submits that this analysis fails to consider the claim as a whole, as required under Step 2A, Prong Two, and instead focuses narrowly on isolated elements taken out of context.
The Office Action further states that “the implementation of neural networks results in enabling generic AI-based models without using an anomaly detector in any meaningful way to improve the functioning of a computer or another technology without reference to what is well-understood, routine, and conventional.”
Applicant respectfully submits that this statement is unclear and does not provide a coherent rationale under Step 2A, Prong II. The Examiner appears to assert that the implementation of neural networks merely enables generic AI-based models and does not employ an anomaly detector in a meaningful way to improve the functioning of a computer or another technology. However, this assertion is conclusory and inconsistent with the Examiner’s own analysis under 35 U.S.C. § 103, which acknowledges that the claimed architecture yields a technical benefit in the form of reduced computational burden. As it is further discussed in the analysis for Step 2A, Prong II under “35 U.S.C. § 101 – Substantive,” such reduction in computation constitutes an improvement to computer functionality and supports integration of any alleged abstract idea into a practical application.
The rationale presented in the Office Action then states “The claim do not include additional elements that are sufficient to amount to significantly more than the judicial exception because simply appending generic AI-based models, specified at a high level of generality, to the judicial exception, e.g., a claim to an abstract idea requiring no more than a generic computer to perform generic computer function that are well-understood, routine and conventional activities previously known industry”, citing Alice Corp., 573 U.S. at 225, 110 USPQ2d at 1984.
Whether the additional elements amount to “significantly more” than the alleged judicial exception pertains to Step 2B of the § 101 analysis, as opposed to Step 2A, Prong II.
Accordingly, it is in the opinion of the Applicant that the Examiner’s analysis under Step 2A, Prong II is incomplete and legally insufficient and therefore does not provide a proper basis for concluding that Claim 1 fails to integrate the alleged judicial exception into a practical application.”
Examiner respectfully disagrees.
Applicant’s Prong Two argument has been fully considered but is not persuasive. The proper inquiry is whether the claim as a whole integrates the recited abstract idea into a practical application by reciting a particular technological solution that improves computer functionality or another technology. Here, even when considered as an ordered combination, the claim does not do so. Applicant argues that classifying a sequence of reconstruction-loss values rather than raw high-dimensional input reduces computational burden and improves scalability. However, the claim does not recite the specific technical mechanism that achieves the asserted benefit. The claim recites, at a functional level, an autoencoder that reconstructs input data, a loss estimator that determines a sequence of losses, and a second neural network trained using labeled sequences of losses. It does not recite any specific data structure, loss formulation, dimensionality-reduction rule, network topology, parameter constraint, or other processing technique that changes how the computer or machine-learning model itself operates. Thus, the alleged reduction in computation is an asserted result, not a claim-recited technological improvement. Unlike Ex Parte Desjardins, where the claims reflected the specific improvement described in the specification as to how the machine-learning model operated, the presented claims do not recite a comparable operative mechanism that improves model or computer functionality. Instead, they use generic machine-learning components to analyze information and produce an anomaly result. Accordingly, the claim does not integrate the recited judicial exception into a practical application under Step 2A, Prong Two.

On pages 4-5 of the Remarks, Applicant contends:
“Applicant respectfully notes that the Office Action does not include a separate or substantive analysis under Step 2B, in contrast to the sections presented for Step 2A, Prong I and Step 2A, Prong II. The Examiner merely states, in the paragraph addressing Step 2A, Prong I, that courts have found claims using artificial intelligence to generate network maps and schedules for television broadcasts to be attempts to patent an abstract idea, and that such claims lack an inventive concept sufficient to transform the abstract idea into patent-eligible subject matter. The Office Action then concludes that merely applying generic AI-based methods to a new data environment does not constitute an inventive concept, citing Recentive Analytics, Inc. v. Fox Corp. Applicant respectfully submits that this discussion is presented at a high level of generality and fails to establish any meaningful connection between the cited case law and the specific limitations of the present claims. Merely invoking a judicial decision, without identifying corresponding claim limitations or explaining how the reasoning of that decision applies to Claim 1, does not constitute a sufficient or reasoned Step 2B analysis.Additionally, with respect to Step 2B, the Examiner states, within the paragraph addressing Step 2A, Prong II, that “the claim do not include additional elements that are sufficient to amount to significantly more than the judicial exception because simply appending generic AI-based models, specified at a high level of generality, to the judicial exception” results in a claim requiring no more than a generic computer to perform well-understood, routine, and conventional activities, citing Alice Corp, 573 U.S. at 225, 110 USPQ2d at 1984. The Examiner alleges that the Applicant merely appends generic AI-based models, at a high level of generality to an abstract idea. However, this conclusion is unsupported. The Office Action does not identify which specific aspects of the claimed neural-network models are allegedly generic. Furthermore, alleging that such models approach a high level of generality is a mischaracterization, as Claim 1 defines how the models are configured, trained, and used on a technical level. Eligibility under Step 2B turns on whether this claimed arrangement and operation of the models, considered as a whole, was well-understood, routine, and conventional, not on a generalized assessment of whether artificial-intelligence models are known in isolation.Furthermore, it is in the opinion of the Applicant that the Examiner references, in a conclusory manner, a characterization of the claimed elements as “well-understood, routine, and conventional,” the Office Action provides no rationale or evidentiary support establishing that the claimed neural-network architectures, loss estimation operations, or anomaly-classification functionality were conventional at the time of the invention, as required under Step 2B. The USPTO bears the burden of establishing that the claimed additional elements are well-understood, routine, and conventional, as required under Berkheimer v. HP Inc., 881 F.3d 1360 (Fed. Cir. 2018), and MPEP § 2106.05(d). Moreover, Aatrix Software, Inc. v. Green Shades Software, Inc., 882 F.3d 1121 (Fed. Cir. 2018), confirms that such a determination must be supported by evidence and may not be made in a conclusory manner. Accordingly, the Office Action fails to provide a legally sufficient Step 2B analysis demonstrating that the additional claim elements are well-understood, routine, and conventional, and therefore does not support a conclusion of ineligibility under 35 U.S.C. § 101.
”
Examiner respectfully disagrees.
Applicant’s Step 2B arguments have been considered but are not persuasive. Step 2B asks whether the claim includes additional elements that amount to significantly more than the recited abstract idea. Here, the additional elements include the recited processor, memory/instructions, input interface, first neural network having an autoencoder architecture, loss estimator, second neural network trained in a supervised manner, and output interface. Considered individually, these elements are well-understood, routine, and conventional (see Berkheimer v. HP Inc., 881 F.3d 1360, 125 USPQ2d 1649 (Fed. Cir. 2018)).
The specification itself acknowledges that anomaly detectors generally use autoencoders that reconstruct input data and use reconstruction loss between input data and reconstructed data to detect anomalies, evidencing that these features where conventional in the art (see Specification, ¶4). The specification further describes standard ML building blocks and ordinary training/loss mechanisms, rather than any unconventional computer component or architecture. Considered as an ordered combination, the claim likewise does not add significantly more, because it merely arranges known ML components to perform their expected functions in sequence, reconstructing input data, computing loss values, classifying those values, and outputting a result. The claim does not recite a non-conventional interaction among these components that changes how the computer or machine-learning system itself operates. Accordingly, the additional elements, individually and as an ordered combination, do not amount to significantly more than the abstract idea.
Applicant arguments and amendments filed on 07/30/2025 with regards to 35 USC 103, as seen in pages 7-10, with respect to under Hu et al. (US PGPub No. 20220301689-A1 ) in view of Kobayashi et al. (US PGPub No. 20230419169-A1) and Park et al (US PGPub No. 20230065385-A1), with respect to the independent claims, have been fully considered and persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new grounds of rejection is made in view of newly found prior art.

Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.

Claim 1-14 rejected under 35 U.S.C. 101 because the claimed invention is directed to non-statutory subject matter because claim invention is directed to an abstract idea without significantly more.
Applying the SMET as found in MPEP § 2106, to claim 1:

Step 1: Does the claim recite a statutory category of invention?
Claim 1 recites an anomaly detector comprising a processor and memory which is interpreted to be a ‘machine’ and one of the four statutory categories of invention (Step 1 of the Subject Matter Eligibility Test).
However, the claim as a whole appear to not qualify for a streamlined analysis thus a full eligibility and thus a full eligibility analysis is necessary (Step 2A and Step 2B of the Subject Matter Eligibility Test).

Step 2A, Prong One: Does the claim recite a judicial exception (e.g. abstract idea, law of nature, etc.)?
Under Step 2A, Prong One, claim 1 recites a judicial exception. In particular, claim 1 recites collecting input data, analyzing the data by comparing portions of the input data and reconstructed input data to determine a sequence of losses and by classifying the sequence of losses to detect an anomaly and produce a result indicating anomaly type and/or severity, and rendering the result of anomaly detection. These limitations recite an abstract idea falling within the mental process and mathematical concept groupings. The claim recites information collection, analysis, and presentation of results, similar to the abstract idea identified in Electric Power Group, LLC v. Alstom S.A., 830 F.3d 1350, 119 USPQ2d 1739 (Fed. Cir. 2016). In addition, to the extent the claim recites generating reconstructed input data, and classifying those values, the claim also recites mathematical concepts.
Step 2A, Prong Two: Does the claim integrate the judicial exception into a practical application?
Under Step 2A, Prong Two, claim 1 does not integrate the judicial exception into a practical application. Applicant argues that the claimed architecture improves computer functionality by reducing computation burden because classification is performed on reconstruction-loss sequences rather than on raw high-dimensional input data. That argument has been considered but is not persuasive. As the courts explained in Recentive Analytics, Inc. v. Fox Corp., No. 2023-2437 (Fed. Cir. Apr. 18, 2025), in the context of computer-assisted methods, claims are not made patent eligible under § 101 simply because they speed up human activity. Even when evaluated as a whole and as an ordered combination, claim 1 does not recite the specific technological mechanism that achieves the asserted improvement. The claim recites, at a functional level, an input interface, an autoencoder, a loss estimator that determines a sequence of losses, a second neural network trained using labeled sequences of losses, and an output interface. However, the claim does not recite any specific data structure for the loss sequence, any particular dimensionality-reduction rule, any specific loss formulation beyond a generic sequence of losses, any network topology, parameter constraint, processing rule, or other concrete technical mechanism that changes how the computer or machine-learning model itself operates. Thus, the alleged reduction in computation is an asserted result rather than a claim-recited technological improvement Unlike claims found eligible because they reflected a disclosed improvement in how the machine-learning model itself operated, the present claim recites a result-oriented analytic pipeline using known machine-learning components to analyze information and produce an anomaly result. Accordingly, claim 1 does not integrate the judicial exception into a practical application.
Step 2B: Does the claim recite additional elements that amount to significantly more than the judicial exception?
Under Step 2B, claim 1 does not recite additional elements that amount to significantly more than the judicial exception. The additional elements include the recited processor, memory/instructions, input interface, first neural network having an autoencoder architecture, loss estimator, second neural network trained in a supervised manner, and output interface. Considered individually, these additional elements are well-understood, routine, and conventional. The specification in the background (see Spec., ¶4) describes anomaly detectors as generally using autoencoders that reconstruct input data and using reconstruction loss between input data and reconstructed data to detect anomalies, indicating that such features were known in the art. Later, in the summary, the specification further describes the use of cross-entropy and mean square error loss function for different feature types, and describes the classification model as being constructed using logistic regression or a multilayer perceptron network, rather than any unconventional computer component. Considered as an ordered combination, the claim likewise does not add significantly more because it merely arranges known machine-learning components to perform their expected functions in sequence: receiving data, reconstructing data, computing loss values, classifying those values, and outputting a result. The claim does not recite a non-conventional interaction among these components that changes how the computer or machine-learning system itself operates. Accordingly, the additional elements, individually and as an ordered combination, do not amount to significantly more than the abstract idea.
Therefore, claim 1 is ineligible under 35 U.S.C. § 101.

Claims 15 and 20:
Claims 15 and 20 recite substantially the same judicial exception and substantially the same additional elements as claim 1, in method and computer-readable-medium form (both statutory categories under Step 1), respectively. For the same reasons discussed with respect to claim 1, claims 15 and 20 recite an abstract idea under Step 2A, Prong One, do not integrate the exception into a practical application under Step 2A, Prong Two, and do not recite significantly more under Step 2B. Accordingly, claims 15 and 20 are ineligible under 35 U.S.C. § 101.

With regards to claim 2-14 and 16-19:
Claims 2-15 and 16-19 do not cure the deficiency of their respective base claims. The additional limitations, such as specifying a deep neural network, joint training, end-to-end training, staged training, domain-specific training, explainability, attribution scores, and particular data types, either futher define how the abstract anlaysis is performed or add field-of-use/output-related details, but do not recite a specific technological mechanism that integrates the judicial exception into a practical application or adds significantly more than the judicial exception. Accordingly, claims 2-14 and 16-19 are also ineligible under 35 U.S.C. § 101.

Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.

The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.

Claims 1-20 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.

Regarding claims 1, 11, and 20:
“…a severity of the anomaly…” and “…determine a sequence of losses…” is recited in claim 1, 11, and 20, but the specification does not explicitly disclose how the severity of an anomaly is calculated by the neural network and how the inputs, sequence of losses, are used to determine the severity, or how the sequence of losses is determined. In ¶0043 of the specification states “…Additionally or alternatively, the result of anomaly detection 121 includes a severity of the anomaly. Examples of the severity of anomaly may include low, medium, high, and critical. For instance, the second neural network 115 may determine an anomaly of class2 with severity as high…” Nowhere in the specification discloses how the severity of the anomaly of how the sequences of losses are determined.
As discussed in MPEP § 2161.01, “[t]he level of detail required to satisfy the written description requirement varies depending on the nature and scope of the claims and on the complexity and predictability of the relevant technology. Ariad, 598 F.3d at 1351, 94 USPQ2d at 1172; Capon v. Eshhar, 418 F.3d 1349, 1357-58, 76 USPQ2d 1078, 1083-84 (Fed. Cir. 2005). Computer-implemented inventions are often disclosed and claimed in terms of their functionality. For computer-implemented inventions, the determination of the sufficiency of disclosure will require an inquiry into the sufficiency of both the disclosed hardware and the disclosed software due to the interrelationship and interdependence of computer hardware and software. The critical inquiry is whether the disclosure of the application relied upon reasonably conveys to those skilled in the art that the inventor had possession of the claimed subject matter as of the filing date. Vasudevan Software, Inc. v. MicroStrategy, Inc., 782 F.3d 671, 682. 114 USPQ2d 1349, 1356 (citing Ariad Pharm., Inc. V. Eli Lilly & Co, 598 F.3d 1336, 1351, 94 USPQ2d 1161, 1172 (Fed. Cir. 2010) in the context of determining possession of a claimed means of accessing disparate databases).” In particular, the courts cautioned that “[i]t is not enough that one skilled in the art could write a program to achieve the claimed function because the specification must explain how the inventor intends to achieve the claimed function to satisfy the written description requirement. See, e.g., Vasudevan Software, Inc. v. MicroStrategy, Inc., 782 F.3d 671, 681-683, 114 USPQ2d 1349, 1356, 1357 (Fed. Cir. 2015).” A skilled artisan reading Appellant's generic disclosures would not reasonably understand that the inventors had possession of all possible implementations of determining a sequence of losses and producing a type of the anomaly and a severity of the anomaly.
Claims 2-10 and 12-19 do not overcome the rejections of their respective base claims that have been rejected above, and therefore rejected under the same grounds provided to claims 1 and 11.

The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.

The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.

Claims 1-20 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 claims 1, 11, and 20:
“…a severity of the anomaly…” is recited in claim 1, 11, and 20 is indefinite because the claims does not specify how the severity of an anomaly are calculated using the sequence of losses. Further one in the ordinary skill in the art would not know how severe of an anomaly is just by using the sequences of losses.

Claims 2-10 and 12-19 do not overcome the rejections of their respective base claims that have been rejected above, and therefore rejected under the same grounds provided to claims 1 and 11.

Claim Rejections - 35 USC § 103
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.
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.

Claims 1-4,14, 15, and 17-20 are rejected under 35 U.S.C. 103 as being unpatentable over Hu et al. (US PGPub No. 20220301689-A1 ) in view of Zhou et al. (US PGPub No.20200076842), Cosatto et al. (US Pat No. 10964011-B2), and Park et al (US PGPub No. 20230065385-A1) .

With respect to claim 1, Hu teaches an anomaly detector, comprising: at least one processor; and a memory having instructions stored thereon that form modules of the anomaly detector, (¶0007: The embodiments disclosed herein build an anomaly detection model by training on medical images depicting normal tissue. The anomaly detection model may incorporate aspects of a Generative Adversarial Network (GAN) model and an autoencoder; ¶0079-0082: … processor 702 may retrieve (or fetch) the instructions from an internal register, an internal cache, memory 704….)
wherein the at least one processor is configured to execute the instructions of the modules of the anomaly detector, (¶0079-0082: Illustrated in Figure 7, an example and not by way of limitation, to execute instructions, processor 702 may retrieve (or fetch) the instructions from an internal register, an internal cache, memory 704, or storage 706; decode and execute them; and then write one or more results to an internal register, an internal cache, memory 704, or storage 706.).
the modules comprising: an input interface configured to accept input data; (¶0081: In particular embodiments, computer system 700 includes a processor 702, memory 704, storage 706, an input/output (I/O) interface 708, a communication interface 710, and a bus 712. ¶0085: An input device coupled to the I/O interface 708 may include devices for converting different forms of volitional user input (accepting input data) into digital signals that can be processed by computer system 700) .
a first neural network having an autoencoder architecture including an encoder trained to encode the input data and a decoder trained to decode the encoded input data to reconstruct the input data; (¶0024: The generative model may be implemented using an autoencoder, which may learn to encode an unlabeled input image to a compressed latent space and then decode the representation from latent space back to an image. ¶0026-0027: Both the input image x and the output of the generative model G—reconstructed image G(x) (reconstruct the input data))

an output interface to render the result of anomaly detection. (¶0052: At step 390, a heatmap of the image of tissue is reconstructed, based on the anomaly scores for each tile, to provide a visualization (outputting) of any identified anomalies in the tissue. ¶0077: For example, the user can access a user device 630 that is in communication with the whole slide image processing system 630 that is in communication with the whole slide image processing system 610 and provide a query image for analysis).
Hu does not disclose:
a loss estimator configured to compare a plurality of parts of the input data with corresponding plurality of parts of the reconstructed input data to determine a sequence of losses each loss in the sequence corresponding to a different feature or component of the input data;
However, Zhou teaches a loss estimator configured to compare a plurality of parts of the input data with corresponding plurality of parts of the reconstructed input data to determine a sequence of losses each loss in the sequence corresponding to a different feature or component of the input data; (¶0220-221: Figure 14, depicts a training harness to improve dense encoding. Reconstruction loss 1470 may be calculated based on comparing original and reconstructed raw vectors for dense sequence 1420. For example, reconstructed raw vector 1465 may be compared to original raw vector 1452. In an embodiment, values within reconstructed raw vector 1465 are probabilistic, such that one original value of ‘1’ is approximated by a probability approaching ‘1’, such as ‘0.81’ as shown in circles. Likewise, prediction loss 1440 may be calculated based on comparing original and predicted vectors for dense sequence 1420. Both or either of losses 1440 and 1470 may entail calculating and integrating (e.g. summing) individual losses of each vector in a sequence. For example, individual losses for raw/sparse or dense vectors may be measured by comparison between corresponding vectors of an original sequence versus a predicted sequence (not shown), such as discussed for Figure 7);
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Zhou with regards to loss estimator to the method of Hu in order to better manage traffic and activity analysis and improve accuracy. (Zhou : ¶0003 & ¶0220).
Hu in view of Zhou does not disclose:
a second neural network trained in a supervised manner ,using labeled sequences of losses, to classify the sequence of losses jointly to detect an anomaly to produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly; and
However, Cosatto teaches a second neural network trained in a supervised manner ,using labeled sequences of losses, to classify the sequence of losses jointly to detect an anomaly to [produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly;] and ( ¶0103-0105: Figure 12 is a flow diagram illustrating a method 1200 for using a trained anomaly detector with predictive normalization. In an embodiment, the classifying block 1230 is performed using a Deep Convolutional Neural Network (DCNN) to output a classification score, and wherein deemed defective responsive (detecting an anomaly) to the particular one of the patch portions having the classification score below a threshold. In an embodiment, the classifying block (1230) aggregates classifications of all of the patch portions sampled from the input test image, classifies the input test image as including a defect responsive to a number of defective ones of the patch portions sampled from the input test image exceeding a threshold, and displays locations of the patch portions including the detect by overlaying identifying information on the input test image. At block 1240, detect a defect in the product based on the classified samples.);
Although, Hu teaches second neural network is disclosed with regards to detecting an anomaly to produce a result anomaly detection including one or a combination of a type of the anomaly and the severity of the anomaly as shown in Figure 3B and ¶0025: The anomaly detection model may use a Markovian discriminator (e.g., PatchGAN) to improve perceptual detection of localized anomalies ¶0010: The loss of the discriminator model may be determined as between an output of the discriminator model for the tile and an output of the discriminator model for the corresponding reconstruction of the tile. The anomaly scores may be normalized to fall within a range from 0 to 1 which i. Anomaly scores may be computed for each of a plurality of overlapping patches of the tiles. & ¶0052: In Figure 3B further elaborates the discriminator at step 375-380 wherein, step 375, the tile of interest (at each of the multiple scales) may be input into the discriminative model (second neural network trained). At step 380, an anomaly score may be computed for each patch of a set of overlapping patches of the tile of interest, based on the concatenated channels of input. Hu does not disclose the classification of the sequence of losses rather they compare the results to a baseline determines the anomaly as discussed in ¶0049, reconstruction error, and does not explicitly disclose the second neural network being trained in a supervised manner. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Cosatto with regards to the neural network detecting an anomaly to the method of Hu in view of Zhou in order to increase accuracy for detecting anomalies while maintaining quality. (Cosatto : ¶0004).
Hu in view of Zhou and Cosatto does not disclose:
produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly
However, Park teaches produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly; and (¶0109-0110: As such, when an abnormality occurs, the detection map is outputted together. Because the detection map indicates an attention region AT having an attention value equal to or greater than a threshold value in the input data as shown in part (C) of Figure 7 or 8 , it is possible to know not only whether an abnormality has occurred but also a region causing the abnormality. Moreover, according to the present disclosure, it is possible to perform object recognition it is possible to perform object recognition by using an anomaly detection and anomaly type classification model as it is without building a separate algorithm. ).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Park with regards to producing a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly to the method of Hu in view of Zhou and Cosatto in order to produce the cost, including effort to build an object recognition algorithm and reduce the amount of computation (Park: ¶0110).

With respect to claim 2, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 1 (see rejection of claim 1 above) wherein the second neural network is a deep neural network. (Hu: ¶0031: The discriminator D may be a convolutional neural network with adversarial objective of distinguishing between real or generated images. In addition to the adversarial objective, the discriminator minimizes the difference between its real and fake features, f(D(x)) and f(D(G(x))), which are typically vectors z∈R.sup.n×1.).

With respect to claim 3, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 1 (see rejection of claim 1 above) wherein the first neural network and the second neural network are jointly trained. (Hu: ¶0011: The anomaly detection model may involve jointly training the generator model (first neural network) and the discriminator model (second neural network) based on a combined loss comprising an adversarial loss, a reconstruction loss, and a latent loss.).

With respect to claim 4, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 1 (see rejection of claim 1 above) wherein the first neural network and the second neural network are jointly trained end-to-end. (Hu: ¶0024: A GAN architecture (used to be jointly trained end-to-end) pairs a generative model with a discriminative model in order to learn a loss function on reconstructed images by generating plausible fake images and comparing them to real, in-distribution images).

With respect to claim 14, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 1 (see rejection of claim 1 above) wherein the input data corresponds one of internet proxy log data, image data, video data, or audio data. (Hu : ¶0007: An image processing system may receive an image of a tissue sample and generate a set of tiles from the image of the tissue sample (image data) . The set of tiles may be input into an anomaly detection model comprising a generator model comprising functional skip-connections and a Markovian discriminator model. The anomaly detection model may be trained to isolate a feature space of normal tissue samples.).

With respect to claim 15, Hu teaches a method for anomaly detection, ( Abstract: An image processing system may receive an image of a tissue sample. A set of tiles may be generated from the image of the tissue sample. The set of tiles may be input into an anomaly detection model (anomaly detection) comprising a generator model comprising functional skip-connections and a Markovian discriminator model.)
wherein the method uses a processor coupled with stored instructions implementing the method, wherein the instructions, when executed by the processor carry out steps of the method, comprising: receiving input data; ( ¶0079-0082: As seen in Figure 7, wherein a processor 702 includes hardware for executing instructions, such as those making up a computer program. As an example, and not by way of limitation, to execute instructions, processor 702 may retrieve (or fetch) the instructions from an internal register, an internal cache, memory 704, or storage 706; decode and execute them; and then write one or more results to an internal register, an internal cache, memory 704, or storage 706. In particular embodiments, processor 702 may include one or more internal caches for data, instructions, or addresses.)
encoding the input data and decoding the encoded data to reconstruct the input data, based on a first neural network having an autoencoder architecture; (¶0024: The generative model may be implemented using an autoencoder, which may learn to encode an unlabeled input image to a compressed latent space and then decode the representation from latent space back to an image. ¶0026-0027: Both the input image x and the output of the generative model G—reconstructed image G(x) (reconstruct the input data))
rendering the result of anomaly detection. (¶0052: At step 390, a heatmap of the image of tissue is reconstructed, based on the anomaly scores for each tile, to provide a visualization (outputting) of any identified anomalies in the tissue. ¶0077: For example, the user can access a user device 630 that is in communication with the whole slide image processing system 630 that is in communication with the whole slide image processing system 610 and provide a query image for analysis).
Hu does not disclose:
determining a sequence of losses by comparing a plurality of parts of the input data with corresponding plurality of parts of the reconstructed input data, each loss in the sequence corresponding to a difference feature or component of the input data;
However, Zhou teaches determining a sequence of losses by comparing a plurality of parts of the input data with corresponding plurality of parts of the reconstructed input data, each loss in the sequence corresponding to a difference feature or component of the input data; (¶0220-221: Figure 14, depicts a training harness to improve dense encoding. Reconstruction loss 1470 may be calculated based on comparing original and reconstructed raw vectors for dense sequence 1420. For example, reconstructed raw vector 1465 may be compared to original raw vector 1452. In an embodiment, values within reconstructed raw vector 1465 are probabilistic, such that one original value of ‘1’ is approximated by a probability approaching ‘1’, such as ‘0.81’ as shown in circles. Likewise, prediction loss 1440 may be calculated based on comparing original and predicted vectors for dense sequence 1420. Both or either of losses 1440 and 1470 may entail calculating and integrating (e.g. summing) individual losses of each vector in a sequence. For example, individual losses for raw/sparse or dense vectors may be measured by comparison between corresponding vectors of an original sequence versus a predicted sequence (not shown), such as discussed for Figure 7);
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Zhou with regards to loss estimator to the method of Hu in order to better manage traffic and activity analysis and improve accuracy. (Zhou : ¶0003 & ¶0220).
Hu in view of Zhou does not disclose:
classifying, using a second neural network trained in a supervised manner using labeled sequences of losses, the sequence of losses to detect an anomaly to produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly; and
However, Cosatto teaches classifying, using a second neural network trained in a supervised manner using labeled sequences of losses, the sequence of losses to detect an anomaly to produce a result of anomaly detection [including one or a combination of a type of the anomaly and a severity of the anomaly; and] ( ¶0103-0105: Figure 12 is a flow diagram illustrating a method 1200 for using a trained anomaly detector with predictive normalization. In an embodiment, the classifying block 1230 is performed using a Deep Convolutional Neural Network (DCNN) to output a classification score, and wherein deemed defective responsive (detecting an anomaly) to the particular one of the patch portions having the classification score below a threshold. In an embodiment, the classifying block (1230) aggregates classifications of all of the patch portions sampled from the input test image, classifies the input test image as including a defect responsive to a number of defective ones of the patch portions sampled from the input test image exceeding a threshold, and displays locations of the patch portions including the detect by overlaying identifying information on the input test image. At block 1240, detect a defect in the product based on the classified samples.);
Although, Hu teaches second neural network is disclosed with regards to detecting an anomaly to produce a result anomaly detection including one or a combination of a type of the anomaly and the severity of the anomaly as shown in Figure 3B and ¶0025: The anomaly detection model may use a Markovian discriminator (e.g., PatchGAN) to improve perceptual detection of localized anomalies ¶0010: The loss of the discriminator model may be determined as between an output of the discriminator model for the tile and an output of the discriminator model for the corresponding reconstruction of the tile. The anomaly scores may be normalized to fall within a range from 0 to 1 which i. Anomaly scores may be computed for each of a plurality of overlapping patches of the tiles. & ¶0052: In Figure 3B further elaborates the discriminator at step 375-380 wherein, step 375, the tile of interest (at each of the multiple scales) may be input into the discriminative model (second neural network trained). At step 380, an anomaly score may be computed for each patch of a set of overlapping patches of the tile of interest, based on the concatenated channels of input. Hu does not disclose the classification of the sequence of losses rather they compare the results to a baseline determines the anomaly as discussed in ¶0049, reconstruction error, and does not explicitly disclose the second neural network being trained in a supervised manner. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Cosatto with regards to the neural network detecting an anomaly to the method of Hu in view of Zhou in order to increase accuracy for detecting anomalies while maintaining quality. (Cosatto : ¶0004).
Hu in view of Zhou and Cosatto does not disclose:
produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly
However, Park teaches produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly (¶0109-0110: As such, when an abnormality occurs, the detection map is outputted together. Because the detection map indicates an attention region AT having an attention value equal to or greater than a threshold value in the input data as shown in part (C) of Figure 7 or 8 , it is possible to know not only whether an abnormality has occurred but also a region causing the abnormality. Moreover, according to the present disclosure, it is possible to perform object recognition it is possible to perform object recognition by using an anomaly detection and anomaly type classification model as it is without building a separate algorithm. ).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Park with regards to producing a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly to the method of Hu in view of Zhou and Cosatto in order to produce the cost, including effort to build an object recognition algorithm and reduce the amount of computation (Park: ¶0110).

With respect to claim 17, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 15 (see rejection of claim 15 above) wherein the second neural network is a deep neural network. (Hu: ¶0031: The discriminator D may be a convolutional neural network with adversarial objective of distinguishing between real or generated images. In addition to the adversarial objective, the discriminator minimizes the difference between its real and fake features, f(D(x)) and f(D(G(x))), which are typically vectors z∈R.sup.n×1.).

With respect to claim 18, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 15 (see rejection of claim 15 above) wherein the first neural network and the second neural network are jointly trained. (Hu: ¶0011: The anomaly detection model may involve jointly training the generator model (first neural network) and the discriminator model (second neural network) based on a combined loss comprising an adversarial loss, a reconstruction loss, and a latent loss.).

With respect to claim 19, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 15 (see rejection of claim 15 above), wherein the first neural network and the second neural network are jointly trained end-to-end. (Hu: ¶0024: A GAN architecture (used to be jointly trained end-to-end) pairs a generative model with a discriminative model in order to learn a loss function on reconstructed images by generating plausible fake images and comparing them to real, in-distribution images.).

With respect to claim 20, Hu teaches a non-transitory computer readable storage medium embodied thereon a program executable by a processor for performing a method, the method comprising: (¶0083: As seen in Figure 7, in particular embodiments, memory 704 includes main memory for storing instructions for processor 702 to execute or data for processor 702 to operate on. As an example, and not by way of limitation, computer system 700 may load instructions from storage 706 or another source (such as, for example, another computer system 700) to memory 704.).
receiving input data; (¶0081: In particular embodiments, computer system 700 includes a processor 702, memory 704, storage 706, an input/output (I/O) interface 708, a communication interface 710, and a bus 712. ¶0085: An input device coupled to the I/O interface 708 may include devices for converting different forms of volitional user input (accepting input data) into digital signals that can be processed by computer system 700) .
encoding the input data and decoding the encoded data to reconstruct the input data, based on a first neural network having an autoencoder architecture; (¶0024: The generative model may be implemented using an autoencoder, which may learn to encode an unlabeled input image to a compressed latent space and then decode the representation from latent space back to an image. ¶0026-0027: Both the input image x and the output of the generative model G—reconstructed image G(x) (reconstruct the input data)).
and rendering the result of anomaly detection. (¶0052: At step 390, a heatmap of the image of tissue is reconstructed, based on the anomaly scores for each tile, to provide a visualization (outputting) of any identified anomalies in the tissue. ¶0077: For example, the user can access a user device 630 that is in communication with the whole slide image processing system 630 that is in communication with the whole slide image processing system 610 and provide a query image for analysis).
Hu does not disclose:
determining a sequence of losses by comparing a plurality of parts of the input data with corresponding plurality of parts of the reconstructed input data, each loss in the sequence corresponding to a different feature or component of the input data;
However, Zhou teaches determining a sequence of losses by comparing a plurality of parts of the input data with corresponding plurality of parts of the reconstructed input data; (¶0220-221: Figure 14, depicts a training harness to improve dense encoding. Reconstruction loss 1470 may be calculated based on comparing original and reconstructed raw vectors for dense sequence 1420. For example, reconstructed raw vector 1465 may be compared to original raw vector 1452. In an embodiment, values within reconstructed raw vector 1465 are probabilistic, such that one original value of ‘1’ is approximated by a probability approaching ‘1’, such as ‘0.81’ as shown in circles. Likewise, prediction loss 1440 may be calculated based on comparing original and predicted vectors for dense sequence 1420. Both or either of losses 1440 and 1470 may entail calculating and integrating (e.g. summing) individual losses of each vector in a sequence. For example, individual losses for raw/sparse or dense vectors may be measured by comparison between corresponding vectors of an original sequence versus a predicted sequence (not shown), such as discussed for Figure 7);
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Zhou with regards to loss estimator to the method of Hu in order to better manage traffic and activity analysis and improve accuracy. (Zhou : ¶0003 & ¶0220).
Hu in view of Zhou does not disclose:
classifying, using a second neural network trained in a supervised manner using labeled sequences of losses, the sequence of losses to detect an anomaly to produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly;
However, Cosatto teaches classifying, using a second neural network trained in a supervised manner using labeled sequences of losses, the sequence of losses to detect an anomaly to produce a result of anomaly detection [including one or a combination of a type of the anomaly and a severity of the anomaly;] ( ¶0103-0105: Figure 12 is a flow diagram illustrating a method 1200 for using a trained anomaly detector with predictive normalization. In an embodiment, the classifying block 1230 is performed using a Deep Convolutional Neural Network (DCNN) to output a classification score, and wherein deemed defective responsive (detecting an anomaly) to the particular one of the patch portions having the classification score below a threshold. In an embodiment, the classifying block (1230) aggregates classifications of all of the patch portions sampled from the input test image, classifies the input test image as including a defect responsive to a number of defective ones of the patch portions sampled from the input test image exceeding a threshold, and displays locations of the patch portions including the detect by overlaying identifying information on the input test image. At block 1240, detect a defect in the product based on the classified samples.);
Although, Hu teaches second neural network is disclosed with regards to detecting an anomaly to produce a result anomaly detection including one or a combination of a type of the anomaly and the severity of the anomaly as shown in Figure 3B and ¶0025: The anomaly detection model may use a Markovian discriminator (e.g., PatchGAN) to improve perceptual detection of localized anomalies ¶0010: The loss of the discriminator model may be determined as between an output of the discriminator model for the tile and an output of the discriminator model for the corresponding reconstruction of the tile. The anomaly scores may be normalized to fall within a range from 0 to 1 which i. Anomaly scores may be computed for each of a plurality of overlapping patches of the tiles. & ¶0052: In Figure 3B further elaborates the discriminator at step 375-380 wherein, step 375, the tile of interest (at each of the multiple scales) may be input into the discriminative model (second neural network trained). At step 380, an anomaly score may be computed for each patch of a set of overlapping patches of the tile of interest, based on the concatenated channels of input. Hu does not disclose the classification of the sequence of losses rather they compare the results to a baseline determines the anomaly as discussed in ¶0049, reconstruction error, and does not explicitly disclose the second neural network being trained in a supervised manner. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Cosatto with regards to the neural network detecting an anomaly to the method of Hu in view of Zhou in order to increase accuracy for detecting anomalies while maintaining quality. (Cosatto : ¶0004).
Hu in view of Zhou and Cosatto does not disclose:
produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly;
However, Park teaches produce a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly; (¶0109-0110: As such, when an abnormality occurs, the detection map is outputted together. Because the detection map indicates an attention region AT having an attention value equal to or greater than a threshold value in the input data as shown in part (C) of Figure 7 or 8 , it is possible to know not only whether an abnormality has occurred but also a region causing the abnormality. Moreover, according to the present disclosure, it is possible to perform object recognition it is possible to perform object recognition by using an anomaly detection and anomaly type classification model as it is without building a separate algorithm. ).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Park with regards to producing a result of anomaly detection including one or a combination of a type of the anomaly and a severity of the anomaly to the method of Hu in view of Zhou and Cosatto in order to produce the cost, including effort to build an object recognition algorithm and reduce the amount of computation (Park: ¶0110).

Claims 5-9 are rejected under 35 U.S.C. 103 as being unpatentable over Hu et al. (US PGPub No. 20220301689-A1 ) in view of Zhou (US PGPub No.20200076842), Cosatto et al. (US Pat No. 10964011-B2), Kallur Palli Kumar et al. (US PGPub No. 20190147333-A1 ), and Park et al (US PGPub No. 20230065385-A1) .

With respect to claim 5, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 1 (see rejection of claim 1 above) but does not disclose wherein, during a first training stage, the first neural network is trained with unlabeled data samples in an unsupervised learning manner.
However, Kallur Palli Kumar teaches wherein, during a first training stage, the first neural network is trained with unlabeled data samples in an unsupervised learning manner. (¶0065: During operation, generator 162 takes as input a noise 168 z and an attribute 170 y, and produces data x.sub.G based on z and y, i.e., G(z,y)=x.sub.G. Unsupervised discriminator 164 takes as input x.sub.G from generator 162 (communication 176) and the unlabeled data x.sub.U and the labeled x.sub.T (of labeled data (x.sub.T, y.sub.T) from training data 166 (communication 178). Unsupervised discriminator 164 then calculates D.sub.u(x) to determine output 172 (e.g., whether the x is real or fake ¶0065: During operation, generator 162 takes as input a noise 168 z (unlabeled data) and an attribute 170 y, and produces data x.sub.G based on z and y, i.e., G(z,y)=x.sub.G. Unsupervised discriminator 164 takes as input x.sub.G from generator 162 (communication 176) and the unlabeled data x.sub.U and the labeled x.sub.T (of labeled data (x.sub.T, y.sub.T) from training data 166 (communication 178).).
Although, Hu does disclose the first neural network is trained in an unsupervised learning manner but Hut do not disclose the first neural network being trained with unlabeled samples as seen in Figure 3 (rather the first neural network is trained on labeled data to distinguish normal tissue/images/tiles ). However, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Kallur Palli Kumar with regards to the first neural network is trained with unlabeled samples the method of Hu in view of Zhou, Cosatto, and Parkin order to enable the generative model/first neural network to learn whether a given data is real or fake (Kallur Palli Kumar: ¶0043-0044)

With respect to claim 6, the combination of Hu in view of Zhou, Cosatto, Park, and Kallur Palli Kumar teaches the method of claim 5 (see rejection of claim 5 above) wherein, during a second training stage, the first neural network trained with the unlabeled data samples (Kallur Palli Kumar: ¶0044: In the unsupervised setting where the data is comprised solely of unlabeled data (i.e., n=0), the goal is to learn a generative model G.sub.u(z; θ.sub.u) that samples from the marginal image distribution p(x) by transforming vectors of noise z as x=G.sub.u(z; θ.sub.u) ¶0065: During operation, generator 162 takes as input a noise 168 z (unlabeled data) and an attribute 170 y, and produces data x.sub.G based on z and y, i.e., G(z,y)=x.sub.G. Unsupervised discriminator 164 takes as input x.sub.G from generator 162 (communication 176) and the unlabeled data x.sub.U and the labeled x.sub.T (of labeled data (x.sub.T, y.sub.T) from training data 166 (communication 178).).
and the second neural network are trained with labeled data samples in a supervised learning manner. (Kallur Palli Kumar: ¶0010: learns a second probability that pairs comprised of a data object and a corresponding attribute label are real based on data objects which only have a corresponding attribute label. ¶0068: As seen in Figure 2A, Subsequently, supervised discriminator 165 takes as input pairs of values to determine an output 186. That is, supervised discriminator takes as input (h(x),y) pairs from generator 162 (e.g., h(x.sub.G) (labeled data samples) from communication 180, and y from communication 182), and also takes as input (h(x),y) pairs from training data 166 (e.g., h(x.sub.T) via communications 178 and 180, and y.sub.T from communication 184).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Kallur Palli Kumar with regards to the first neural network being trained with unlabeled data samples and second neural network trained in a supervise manner the method of Hu in view of Zhou, Cosatto, and Park in order to efficiently addresses and reduces the cost of resources for the system by adopting a semi supervised learning method (Kallur Palli Kumar: ¶0038).

With respect to claim 7, the combination of Hu in view of Zhou, Cosatto, Park, and Kallur Palli Kumar teaches the method of claim 6 (see rejection of claim 6 above) wherein, during the second training stage, the second neural network is trained with the labeled data samples in the supervised learning manner. (Kallur Palli Kumar: ¶0068: Subsequently, supervised discriminator 165 takes as input pairs of values to determine an output 186. That is, supervised discriminator takes as input (h(x),y) pairs from generator 162 (e.g., h(x.sub.G) (labeled data) from communication 180, and y from communication 182), and also takes as input (h(x),y) pairs from training data 166 (e.g., h(x.sub.T) via communications 178 and 180, and y.sub.T from communication 184).)).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Kallur Palli Kumar with regards to the third stage the method of Hu in view of Zhou, Cosatto, and Park in order to efficiently addresses and reduces the cost of resources.

With respect to claim 8, the combination of Hu in view of Zhou, Cosatto, Park, and Kallur Palli Kumar teaches the method of claim 7 (see rejection of claim 7 above) wherein, during a third training stage, the first neural network and the second neural network trained in the second training stage are trained with labeled samples. (Kallur Palli Kumar: ¶ 0068: That is, supervised discriminator 165 determines which of the (data, label) pairs are produced by generator 162 (i.e., the (x.sub.G, y) fake pairs) and which of the (data, label) pairs are part of the training data (i.e., the (x.sub.T, y.sub.T) real pairs) (first neural network trained with labeled samples as seen in Figure 2A) Architecture 260 of FIG. 2D is similar to architecture 160 of FIG. 1D. ).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Kallur Palli Kumar with regards to the third stage the method of Hu in view of Zhou, Cosatto, and Park in order to reduces the cost of resources (from the semi supervise setting) and distinguish the attribute as real.

With respect to claim 9, the combination of Hu in view of Zhou, Cosatto, Park, and Kallur Palli Kumar teaches the method of claim 8 (see rejection of claim 8 above) wherein, during the third training stage, the second neural network trained in the second training stage is trained with the labeled samples. (Kallur Palli Kumar: ¶0066: As discriminator 164 learns which of the data or images is real or fake, generator 162 continues (continually trained with labeled samples) to produce data or images to confuse discriminator 164, such that generator 162 improves in producing data or images x.sub.G which look more and more realistic. At the same time, discriminator 164 continues to iterate through the training process, improving its own ability to distinguish between a real image and a fake image,).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Kallur Palli Kumar with regards to the second neural network trained in a supervise manner the method of Hu in view of Zhou, Cosatto, and Park in order to efficiently addresses and reduces the cost of resources for the system by adopting a semi supervised learning method (Kallur Palli Kumar: ¶0038).

Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Hu et al. (US PGPub No. 20220301689-A1 ) in view of Zhou (US PGPub No.20200076842), Cosatto et al. (US Pat No. 10964011-B2) , Kallur Palli Kumar et al. (US PGPub No. 20190147333-A1 ), Park et al. (US PGPub No. 20230065385-A1), and Cortes et al. (US PGPub No. 20240144635-A1) .

With respect to claim 10, the combination of Hu in view of Zhou, and Cosatto, Park, and Kallur Palli Kumar teaches the method of claim 9 (see rejection of claim 9 above) but does not disclose wherein a domain of the labeled samples is different from a domain of the unlabeled data samples and the unlabeled data samples.
However, Cortes teaches wherein a domain of the labeled samples is different from a domain of the unlabeled data samples and the unlabeled data samples. (¶0025-0026: As seen in Figure 1, the labeled data 101 may, for example, include one or more labeled images. Labeled data 101 is training data in the same domain of the use case. For the training operation 102, the labels around the region of interest must be available. As seen in Figure 2, unlabeled data 201 can for example, include one or more unlabeled images. Unlabeled data 201 is training data in a neighbor or general domain of the use case, where data should be available).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Cortes with regards to the domains the method of Hu in view of Zhou, Cosatto, Park, and Kallur Palli Kumar in order to ensure robust model training.

Claims 11-12 are rejected under 35 U.S.C. 103 as being unpatentable over Hu et al. (US PGPub No. 20220301689-A1 ) in view of Zhou (US PGPub No.20200076842), Cosatto et al. (US Pat No. 10964011-B2), Park et al. (US PGPub No. 20230065385-A1), and Seyfi et al. (US PGPub No. 20230376366-A1) .

With respect to claim 11, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 1 (see rejection of claim 1 above) but does not teach , the modules further comprising an explainability module configured to predict a class of anomaly.
However, Seyfi teaches the modules further comprising an explainability module configured to predict a class of anomaly. (¶0033-0034: As seen in Figure 1, in operation, anomaly detector 160 is applied to a tuple 150 to generate an inference such a numeric anomaly score 170 that is value of regression or prediction. In an embodiment, anomaly score 170 is compared to a threshold to detect whether or not tuple 150 is anomalous. ¶0035-0041: As discussed below 110 and/or anomaly detector 160 participate in a sequence of phases that include: training of anomaly detector 160 and MLX innovation that generates neighborhood 140 based on anomalous tuple to explain 151 before generating a local explanation. In various scenarios, anomalous tuple to explain 151 and its anomaly score 12, and or anomaly detector are reviewed for various reasons. MLX herein can provide combinations of any of the following functionalities: Explainability: The ability to explain the local reasons why anomaly score 12 occurred for anomalous tuple to explain 151 Interpretability: The level at which a human can understand the explanation, What-If Explanations: Understand how changes in anomalous tuple to explain 151 may or may not cause same anomaly score 12 , and Model-Agnostic Explanations: Explanations treat anomaly detector 160 as a black box, instead of using internal properties from anomaly detector 160 to guide the explanation).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Seyfi with regards to the explainability module the method of Hu in view of Zhou, Cosatto, and Park in order to aid the interpretation of complex ML and deep learning model and (Seyfi ¶0006-0007 & ¶0030-0035) .

With respect to claim 12, the combination of Hu in view of Zhou, Cosatto, and Seyfi teaches the method of claim 11 (see rejection of claim 11 above), wherein the explainability module corresponds to an attribution based classifier (Kobayashi: ¶0058:As seen in Figure 1, explanation 180 is an attribute-based explanation (ABX). To generate explanation 180, the explainer may analyze the three reconstruction errors of respective feature F1-F3 as shown in reconstruction error 110 for anomalous tuple 160.) configured to determine a class of anomaly based on attribution scores corresponding to the input data. ( Kobayashi: ¶0063-0069: As seen in Figure 2, Step 204 automatically generates explanation 180 that correctly or incorrectly identifies an identified feature as a cause of anomalous tuple 140 being anomalous. During steps 202-204, tuples 120 and 140 are identical except for their respective values of perturbed feature F2. For example during steps 202-204, tuples 120 and 140 have a same first value for feature F1 and a same second value for feature F3.).
Although Seyfi teaches an explainability module configured to predict a class of anomaly as disclosed in ¶0035-0041 but does not explicitly disclose the explainability module corresponding to an attribute-based classifier, but it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of Kobayashi with regards to the attribution-based classifier the method of Hu in view of Zhou, Cosatto, Park, and Seyfi in order to identify corrupted features of the system (Kobayashi : ¶0005-0011 & ¶0021-0023).

Claim 13 is rejected under 35 U.S.C. 103 as being unpatentable over Hu et al. (US PGPub No. 20220301689-A1 ) in view of Zhou (US PGPub No.20200076842), Cosatto et al. (US Pat No. 10964011-B2), Park et al. (US PGPub No. 20230065385-A1), Seyfi et al. (US PGPub No. 20230376366-A1), and Tate et al. (US PGPub No. 20220138490-A1 ).

With respect to claim 13, the combination of Hu in view of Zhou, Cosatto, Park, and Seyfi teaches the method of claim 12 (see rejection of claim 12 above) , but does not disclose wherein the second neural network is configured to determine the attribution scores corresponding to the input data.
However, Tate teaches wherein the second neural network is configured to determine the attribution scores corresponding to the input data. ( ¶0057 & 0125: As seen in Figure 2, the attribute determination unit 205 includes the likelihood determination units 205a to 205g serving as discriminators each of which is learned in advance so as to determine the likelihood of each attribute. As each discriminator, for example, a linear support vector machine (SVM) can be used. In step S111, the attribute determination unit 205 generates the attribute score map 208 (determining attribution scores)).
Although Hu in view of Zhou, Cosatto, Park, and Seyfi teaches a second neural network but does not explicitly disclose the second neural network is configured to determine rather the explainability module configured to the second neural network is used to determine the attribution scores. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the teachings of with regards the second neural network the method of Hu in view of Zhou, Cosatto, and Park in order to reduce the cost in calculation and by offering a relatively low calculation cost, a detailed result from the attribute score maps with low resolutions. This point is particularly emphasized since this is one of features of this embodiment in which the plurality of attributes are determined and integrated to recognize a target. (Tate: ¶0004-0005 & 0090-0095).

Claim 16 is rejected under 35 U.S.C. 103 as being unpatentable over Hu et al. (US PGPub No. 20220301689-A1 ) in view of Zhou (US PGPub No.20200076842), Cosatto et al. (US Pat No. 10964011-B2), Park et al. (US PGPub No. 20230065385-A1), and Blake et al. (US PGPub No. 20110274352-A1) .

With respect to claim 16, the combination of Hu in view of Zhou, Cosatto, and Park teaches the method of claim 15 (see rejection of claim 15 above) but does not disclose wherein the input data corresponds to internet proxy log data, image data, video data, and audio data.
However, Blake teaches wherein the input data corresponds to internet proxy log data, image data, video data, and audio data. (¶0078: The computing-based device 1700 comprises one or more inputs 1706 which are of any suitable type for receiving media content, Internet Protocol (IP) input, images, videos, user input, audio input, or other types of input.).
Although, Hu teaches input data corresponding image data as seen in ¶0031 wherein the discriminator D may be a convolutional neural network with adversarial objective of distinguishing between real or generated images, but Hu does explicitly not disclose that the input data correspond to internet proxy log data, video data, and audio data. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize the known teachings of Blake with regard to input data to the method of Hu in view of Zhou, Cosatto, and Park in order to provide flexibility of inputs for the method to evaluate.

Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Longari et al. (CANnolo: An Anomaly Detection System Based on LSTM Autoencoders for Controller Area Network," in IEEE Transactions on Network and Service Management, vol. 18, no. 2, pp. 1913-1924) teaches the comparison of raw input and reconstructed input to detect an anomaly.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to TAYLOR P VU whose telephone number is [phone]. The examiner can normally be reached MON - FRI (7:30 - 5:00).
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, Alexander Lagor can be reached at [phone]. 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.

/T.P.V./ Examiner, Art Unit 2437

/ALEXANDER LAGOR/ Supervisory Patent Examiner, Art Unit 2437

Prosecution Timeline

• Dec 20, 2022: Application Filed
• Nov 07, 2024: Non-Final Rejection mailed — §101, §103, §112
• Feb 13, 2025: Response Filed
• May 09, 2025: Non-Final Rejection mailed — §101, §103, §112
• Jul 30, 2025: Response Filed
• Nov 12, 2025: Final Rejection mailed — §101, §103, §112
• Feb 05, 2026: Request for Continued Examination
• Feb 20, 2026: Response after Non-Final Action
• Apr 13, 2026: Non-Final Rejection mailed — §101, §103, §112
• Jul 02, 2026: Interview Requested
• Jul 08, 2026: Examiner Interview Summary