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 .
This action is in response to communications filed on 07/07/2026.
Claims 1-8 are canceled.
Claims 18-29 are added.
Claims 9-29 are pending and have been examined.
Priority
Receipt is acknowledged of certified copies of papers required by 37 CFR 1.55.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claim 27 is 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.
As per claim 27, it is unclear whether the features are limiting, or merely refer to intended use (all limitations recite “used for” which appears to indicate intended use, but intended use does not have patentable weight).
Response to Arguments
Previous objections to the claims have been withdrawn in view of amendments.
Previous rejections under 35 USC 112(a) have been withdrawn in view of amendments.
Applicant’s arguments with respect to the amended claims of claims 9 and 12 have been considered but are moot in view of new grounds of rejection. See rejections including Halder (US 20190310636 A1) below.
With respect to claim 15, in response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies (i.e., distance signals generated by respective modules collectively used for evaluating, separate discriminator outputs would be fused) are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993). The claim does not recite that the distance signals are generated by respective modules. It only recites “wherein the first monitoring modules generates a first distance signal”. The other signals are recited generally as limitations of “generating a second distance signal” and “generating a third distance signal”. There is no recitation that requires the second and third signal to be generated by any module. There is also no recitation that the signals relate to fused discriminator outputs. It is also noted that the newly amended limitation recites “used for” which appears to refer to intended use, which would not be accorded patentable weight. However, assuming that it is, Notz, for example, teaches “calibration steps may be carried out iteratively a plurality of times and the discriminator, generator and encoder networks preferably updated during each iteration… anomaly score may include…reconstruction loss… LGx=‖x−G(E(x))‖1… signal indicating anomalous vehicle scenario indicates that vehicle controller may have to be amended in order to the vehicle to autonomous handle the indicated scenario in a safe manner in the future… the discriminator network, the generator network, and/or the encoder network may be updated according to the calibrations steps so that the previously anomalous scenario is subsequently considered a normal scenario… anomaly score generated by the evaluation machine 150 may indicate how well the current scenario has been covered in previous drives, which have been used for the training of models for autonomous control of the vehicle” (e.g. in paragraphs 12, 26-28, 42, 46, and 58) and figure 1. In other words, how well the AI module of Notz covers scenarios is collectively used (via many loss signals) when driving (i.e. during inference) so that the AI module learns more scenarios and becomes better, which reads on the amended features. As such, the combination teaches the claimed features.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 9-14, 18-24 and 26 are rejected under 35 U.S.C. 103 as being unpatentable over Notz et al (EP 3654246) in view of Cricri et al (US 20190122072 A1) and Halder (US 20190310636 A1).
As per independent claim 9, Notz teaches a method for monitoring a real input data stream of an artificial intelligence (Al) module that forms part of a processing chain of a partially automated or automated driving function of a vehicle (e.g. in paragraphs 5, 7, 42-43, 51, and 59, “generating a signal indicating anomalous vehicle scenario using an evaluation machine comprising a discriminator network, a generator network, and an encoder network… wherein an input signal to the encoder network includes data from vehicle sensors… distinguish between driving scenarios that the vehicle is able to handle autonomous, from driving scenarios that the vehicle potentially will not be able to handle autonomous in a safe manner… sensors 120 generating signals 140 indicating current vehicle scenario… evaluation machine 150 generating signal 160 indicating anomalous scenario… evaluation machine 150 may be part of the vehicle… signal may indicate to the driver that additional precaution is needed, such as that driver's hands are put on the steering wheel and the driver has readiness to brake… may enter a safe mode, such as including stopping the vehicle in a controlled manner”), comprising:
in a training phase, training a monitoring module using real training data and generated false training data generated by a generator (e.g. in paragraphs 17, 50 and 53, “x comes from real data and z is generated from the encoder network with input x… Calibration (or training) is depicted in Fig. 2… Fig. 2 depicts a structure of the evaluation machine during calibration of the discriminator D, generator G, and encoder E” and figure 2);
feeding the real training data and the generated false training data to a discriminator (e.g. in paragraphs 17, 50 and 53, “x comes from real data and z is generated from the encoder network with input x…during calibration of the discriminator D, generator G, and encoder E” and figure 2 showing real training data and false (generated) training data fed to discriminator network D);
generating a calculation between the real training data and the generated false training data (e.g. in paragraphs 7, 26-27 and 53-54, “The reconstruction loss may be given by the function: LGx=‖x−G(E(x))‖1” showing difference between real and false; note: x is real, while G(E(x)) is false/generated, and figure 2);
in an inference phase, evaluating the Al module via the monitoring module (e.g. in paragraphs 7 and 51, “generating an anomaly score from an output signal of the encoder network, an output signal of the generator network, and/or an output signal of the discriminator network, wherein an input signal to the encoder network includes data from vehicle sensors… evaluation machine 150 generating signal 160 indicating anomalous scenario” and paragraphs 26-29, “For anomalous sensor signals x, the reconstruction loss may be large, since neither G nor E have been trained on anomalous sensor signals. Using the reconstruction loss…further improves the accuracy of indicating anomalous vehicle scenario”);
generating, via the Al module, output signals configured for a driving function based on the real input data stream (e.g. in paragraphs 51 and 59, “sensors 120 generating signals 140 indicating current vehicle scenario… evaluation machine 150 generating signal 160 indicating anomalous scenario… signal may indicate to the driver that additional precaution is needed, such as that driver's hands are put on the steering wheel and the driver has readiness to brake… may enter a safe mode, such as including stopping the vehicle in a controlled manner”); and
feeding the real input data stream to the discriminator to determine one or more calculations that are used for evaluating the real input data stream and for monitoring the output signals of the AI module for controlling driving functions (e.g. in paragraphs 5, 7, 15, 26-29, 34-35, 42-43, 51, and 59, “generating a signal indicating anomalous vehicle scenario using an evaluation machine comprising a discriminator network, a generator network, and an encoder network… Generate the anomaly score may include generating a reconstruction loss value… wherein an input signal to the encoder network includes data from vehicle sensors… distinguish between driving scenarios that the vehicle is able to handle autonomous, from driving scenarios that the vehicle potentially will not be able to handle autonomous in a safe manner… sensors 120 generating signals 140 indicating current vehicle scenario… evaluation machine 150 generating signal 160 indicating anomalous scenario… signal may indicate to the driver that additional precaution is needed, such as that driver's hands are put on the steering wheel and the driver has readiness to brake… may enter a safe mode, such as including stopping the vehicle in a controlled manner” and figures 1-2), the one or more calculations being indicative of whether the real input data stream corresponds to a trained situation or an untrained situation (e.g. in paragraph 58, “anomaly score generated by the evaluation machine 150 may indicate how well the current scenario has been covered in previous drives, which have been used for the training of models for autonomous control of the vehicle… the classification into normal [i.e. trained] and anomalous [i.e. untrained] scenarios may indicate how well driving models (and the vehicle) handle the given situation”),
but does not specifically teach generating, via the discriminator, the one or more calculation/calculations including one or more distance/distances, wherein the real input data stream is fed to the discriminator in parallel with generation of the output signals by the Al module, and the one or more distances being indicative of whether the real input data stream corresponds to a trained situation or an untrained situation.
However, Cricri teaches generating, via a discriminator, one or more calculation/calculations including one or more distance/distances (e.g. in paragraphs 22, 31, and 47, “discriminator distinguishes between generated and real content… the discriminator neural network is considered a critic, which…provides an estimate of the distance between the probability distributions of the real training data and the generated data… online…where models continuously receive new streams of data and keep on learning, the whole system comprising the generator G 220, discriminator D1 240, discriminator D2 250 and the model which extracts key features may be part of the deployed system”). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Notz to include the teachings of Cricri because one of ordinary skill in the art would have recognized the benefit of distinguishing between generated and real content and/or facilitating continuous learning.
but does not specifically teach wherein the real input data stream is fed to the discriminator in parallel with generation of the output signals by the Al module, and the one or more distances being indicative of whether the real input data stream corresponds to a trained situation or an untrained situation.
However, Halder teaches a real input data stream is fed to a discriminator in parallel with generation of output signals by an Al module (e.g. in paragraphs 193-195, “real time sensor data from sensors 902 may be communicated to trained AI model 906 and also to epistemic uncertainty checker (EUC) subsystem 904 [i.e. in parallel]. The trained AI model 906 may make a prediction 912…based upon the received real time sensory data input”, and in figure 9 showing in parallel; note the epistemic uncertainty checker discriminates “the distribution of real time sensory data inputs received from sensors 902 to the distribution of training data 903”; note also paragraph 265, “processes may be executed…in parallel”) and one or more distances being indicative of whether the real input data stream corresponds to a trained situation or an untrained situation (e.g. in paragraphs 48, 153, 160, 170, 185, and 195, “improve the overall safety of using AI models… generates a score (referred to as a model confidence score) that is indicative of how similar [i.e. trained] or dissimilar [i.e. untrained] the inferring data point is to the training dataset… degree of similarity… compares the distribution of the inferring data points (e.g., real time sensor data) received in 606 with the distribution of the training sensor data obtained in 602 and used to train the model in 604… measure the distance between the vectorized input inferring image and the training data”). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of the combination to include the teachings of Halder because one of ordinary skill in the art would have recognized the benefit of facilitating improved safety of using AI models.
As per claim 10, the rejection of claim 9 is incorporated and the combination further teaches wherein the generated false training data is generated by a background data source that are fed to the discriminator (e.g. Notz, in paragraph 17, “x comes from real data and z is generated from the encoder network with input x” and figure 2 showing feeding to discriminator; Cricri, in paragraph 38, “feature-removed images x_rem(t1) 215 may be created from the at least one image x(t1) 200 by applying feature extracting and removal 210 to the at least one image frame x(t1) 200. Feature extracting may for example comprise determining, manually by a user or by a pre-trained neural network or other model, one or more objects in image x(t1) and removing these objects from the image… generator G 220 creates generated images x_gen (t2) 225 comprising target objects having at least one replaced key feature from feature-removed images”).
As per claim 11, the rejection of claim 9 is incorporated and the combination further teaches wherein a training loss with respect to the real training data is determined, based on the distance, and wherein the training loss is used for training the generator and the discriminator of the monitoring module (e.g. Cricri, in paragraphs 22, 43, and 46, “discriminator distinguishes between generated and real content… the discriminator neural network is considered a critic, which…provides an estimate of the distance between the probability distributions of the real training data and the generated data… loss computed from an output of the discriminator D1 240… loss computed through the discriminator D1 240 and applied to train the discriminator… loss computed through the discriminator D1 240 and applied to train the generator G 220”).
As per claim 18, the rejection of claim 9 is incorporated and the combination further teaches wherein the one or more distances represent a degree of similarity between the real input data stream and the real training data used during the training phase (e.g. Notz, in paragraph 58, “anomaly score generated by the evaluation machine 150 may indicate how well the current scenario has been covered in previous drives, which have been used for the training of models for autonomous control of the vehicle… the classification into normal [i.e. trained] and anomalous [i.e. untrained] scenarios may indicate how well driving models (and the vehicle) handle the given situation… If the similar scenario coverage is very low, or not covered at all, the scenario should be classified as anomalous”; Halder, in paragraph 48, “generates a score (referred to as a model confidence score) that is indicative of how similar or dissimilar the inferring data point is to the training dataset… degree of similarity”),
As per claim 19, the rejection of claim 9 is incorporated and the combination further teaches comparing the one or more distances with at least one threshold to determine whether the real input data stream corresponds to a trained situation or an untrained situation (e.g. Notz, in paragraph 58, “anomaly score generated by the evaluation machine 150 may indicate how well the current scenario has been covered in previous drives, which have been used for the training of models for autonomous control of the vehicle… the classification into normal [i.e. trained] and anomalous [i.e. untrained] scenarios may indicate how well driving models (and the vehicle) handle the given situation… If the similar scenario coverage is very low, or not covered at all, the scenario should be classified as anomalous… if the estimated ability is too low (i.e. the anomaly score is above a pre-set threshold) the signal might indicate that the vehicle may not be confident in its ability to handle the given situation in a safe manner”; Halder, in paragraphs 48, 170, and 185, “generates a score (referred to as a model confidence score) that is indicative of how similar or dissimilar the inferring data point is to the training dataset… degree of similarity… measure the distance between the vectorized input inferring image and the training data”).
As per claim 20, the rejection of claim 19 is incorporated and the combination further teaches wherein the monitoring module determines that the real input data stream corresponds to the untrained situation when the one or more distances exceed the at least one threshold (e.g. Notz, in paragraph 58, “anomaly score generated by the evaluation machine 150 may indicate how well the current scenario has been covered in previous drives, which have been used for the training of models for autonomous control of the vehicle… the classification into normal [i.e. trained] and anomalous [i.e. untrained] scenarios may indicate how well driving models (and the vehicle) handle the given situation… If the similar scenario coverage is very low, or not covered at all, the scenario should be classified as anomalous… if the estimated ability is too low (i.e. the anomaly score is above a pre-set threshold) the signal might indicate that the vehicle may not be confident in its ability to handle the given situation in a safe manner”, i.e. untrained; Halder, in paragraphs 48, 170, and 185, “generates a score (referred to as a model confidence score) that is indicative of how similar or dissimilar the inferring data point is to the training dataset… degree of similarity… measure the distance between the vectorized input inferring image and the training data”).
Claims 12-14 and 21-23 are the apparatus method claims corresponding to method claims 9-11 and 18-20, and are rejected under the same reasons set forth, and the combination further teaches one or more processors and a memory, operatively coupled to the one or more processors, the memory storing instructions executable by the one or more processors (e.g. Notz, in paragraphs 40-41 and 51, and claim 11, “storage and processing requirements… storage for vehicle training… Controller 130… Computer readable storage medium containing instructions which, when executed by at least one processor, cause the processor to carry out the steps”; Cricri, in paragraph 57, “computer program code in a memory, and a processor that, when running the computer program code, causes the device to carry out the features of an embodiment”); and a monitoring module comprising a generator and a discriminator, wherein the one or more processors are configured to implement the generator and the discriminator (e.g. Notz, in paragraphs 5, 40-41 and 51, and claim 11, “an evaluation machine comprising a discriminator network, a generator network, and an encoder network… storage and processing requirements… storage for vehicle training… Controller 130… Computer readable storage medium containing instructions which, when executed by at least one processor, cause the processor to carry out the steps”; Cricri, in paragraphs 22, 47, and 57, “GAN includes at least two neural networks: a generator and a discriminator… continuously receive new streams of data and keep on learning, the whole system comprising the generator G 220, discriminator D1 240, discriminator D2 250 and the model which extracts key features… computer program code in a memory, and a processor that, when running the computer program code, causes the device to carry out the features of an embodiment”), the generator being configured to generate false training data for training the monitoring module using real training data and false training data generated by the generator (e.g. Notz, in paragraphs 17, 28-29, 50 and 53, “x comes from real data and z is generated from the encoder network with input x… Calibration (or training) is depicted in Fig. 2… Fig. 2 depicts a structure of the evaluation machine during calibration of the discriminator D, generator G, and encoder E”), wherein the discriminator is configured to output one or more distance values representing a distance between the real training data and the false training data generated by the generator (e.g. Cricri, in paragraphs 22, 31, and 47, “discriminator distinguishes between generated and real content… the discriminator neural network is considered a critic, which…provides an estimate of the distance between the probability distributions of the real training data and the generated data”).
As per claim 24, the rejection of claim 12 is incorporated and the combination further teaches wherein, during the inference phase, the monitoring module evaluates the real input data stream using the discriminator without using the generator (e.g. Halder, in paragraphs 193-195, “real time sensor data from sensors 902 may be communicated to trained AI model 906 and also to epistemic uncertainty checker (EUC) subsystem 904. The trained AI model 906 may make a prediction 912…based upon the received real time sensory data input”, and in figure 9 showing no use of a generator).
As per claim 26, the rejection of claim 12 is incorporated and the combination further teaches wherein the generator comprises an artificial intelligence module (e.g. Cricri, in paragraphs 33, 38 and 44, “generator needs to learn… training the generator G”, i.e. AI module).
Claims 15-17 and 27-29 are rejected under 35 U.S.C. 103 as being unpatentable over Notz et al (EP 3654246) in view of Cricri et al (US 20190122072 A1), Schiegg et al. (US 20200331473 A1), and Lim et al (US 20200065635 A1).
As per independent claim 15, Notz teaches a method for monitoring a real input data stream and an output data stream of an artificial intelligence (Al) module that forms part of a processing chain of a partially automated or automated driving function of a vehicle (e.g. in paragraphs 5, 7, 42-43, 51, and 59, “generating a signal indicating anomalous vehicle scenario using an evaluation machine comprising a discriminator network, a generator network, and an encoder network… wherein an input signal to the encoder network includes data from vehicle sensors… distinguish between driving scenarios that the vehicle is able to handle autonomous, from driving scenarios that the vehicle potentially will not be able to handle autonomous in a safe manner… sensors 120 generating signals 140 indicating current vehicle scenario… evaluation machine 150 generating signal 160 indicating anomalous scenario… evaluation machine 150 may be part of the vehicle… signal may indicate to the driver that additional precaution is needed, such as that driver's hands are put on the steering wheel and the driver has readiness to brake… may enter a safe mode, such as including stopping the vehicle in a controlled manner”), comprising:
training the Al module using real training data, and generating output data (e.g. in paragraphs 13-14, “output signal of the discriminator network… Updating the discriminator network may include affecting (increasing or decreasing) a networks loss value in a first direction (such as increasing and/or maximizing) by calibrating the discriminator network… x comes from real data”);
training a first monitoring module for monitoring the real input data stream of the Al module using the real training data and false training data, wherein the false training data are generated by a generator of the first monitoring module (e.g. in paragraphs 17, 50 and 53, “x comes from real data and z is generated from the encoder network with input x… Calibration (or training) is depicted in Fig. 2… Fig. 2 depicts a structure of the evaluation machine during calibration of the discriminator D, generator G, and encoder E” and figure 2 showing real training data and false (generated) training data fed to discriminator network D), and wherein the first monitoring module generates a first calculation signal (e.g. in paragraphs 27 and 53-54, “The reconstruction loss may be given by the function: LGx=‖x−G(E(x))‖1” and figures 1-2); and generating a second calculation signal and generating a third calculation signal, wherein the first calculation signal, the second calculation signal, and the third calculation signal are collectively used for evaluating operation of the Al module during the inference phase (e.g. in paragraphs 12, 26-28, 42, 46, and 58, “calibration steps may be carried out iteratively a plurality of times and the discriminator, generator and encoder networks preferably updated during each iteration… anomaly score may include…reconstruction loss… LGx=‖x−G(E(x))‖1… signal indicating anomalous vehicle scenario indicates that vehicle controller may have to be amended in order to the vehicle to autonomous handle the indicated scenario in a safe manner in the future… the discriminator network, the generator network, and/or the encoder network may be updated according to the calibrations steps so that the previously anomalous scenario is subsequently considered a normal scenario… anomaly score generated by the evaluation machine 150 may indicate how well the current scenario has been covered in previous drives, which have been used for the training of models for autonomous control of the vehicle”, and figure 1); and
monitoring, in an inference phase, both the real input data stream and the output data stream from the Al module (e.g. in paragraphs 28-29, “For anomalous sensor signals x, the reconstruction loss may be large, since neither G nor E have been trained on anomalous sensor signals. Using the reconstruction loss…further improves the accuracy of indicating anomalous vehicle scenario”), wherein the monitoring comprises: (i) feeding the real input data stream to a discriminator of the trained first monitoring module and determining one or more calculations therefrom that are used for evaluating the real input data stream of the Al module (e.g. in paragraphs 5, 7, 15, 26-27, 42-43, 51, and 59, “generating a signal indicating anomalous vehicle scenario using an evaluation machine comprising a discriminator network, a generator network, and an encoder network… Generate the anomaly score may include generating a reconstruction loss value… The reconstruction loss may be given by the function: LGx=‖x−G(E(x))‖1 [showing calculations for signals 140] … wherein an input signal to the encoder network includes data from vehicle sensors… distinguish between driving scenarios that the vehicle is able to handle autonomous, from driving scenarios that the vehicle potentially will not be able to handle autonomous in a safe manner… sensors 120 generating signals 140 indicating current vehicle scenario… evaluation machine 150 generating signal 160 indicating anomalous scenario… signal may indicate to the driver that additional precaution is needed, such as that driver's hands are put on the steering wheel and the driver has readiness to brake… may enter a safe mode, such as including stopping the vehicle in a controlled manner” and figures 1-2),
but does not specifically teach training the AI module using real training data and ground truth data, the calculation/calculations include distance/distances and training a second monitoring module using the output data of the Al module and false output data, wherein the false output data are generated by a generator of the second monitoring module; training a third monitoring module using ground truth data and false ground truth data, wherein the false ground truth data are generated by a generator of the third monitoring module, determining, via the discriminator, one more distances, and (ii) feeding the output data stream to a discriminator of the trained second monitoring module, and determining, via the discriminator, one or more distances therefrom that are used for evaluating the output data stream of the Al module, and/or (iii) feeding the output data stream to a discriminator of the trained third monitoring module, and determining, via the discriminator, one or more distances therefrom that are used for evaluating the output data stream of the Al module.
However, Cricri teaches determining, via a discriminator, one or more calculation/calculations including distance/distances (e.g. in paragraphs 22, 31, and 47, “discriminator distinguishes between generated and real content… the discriminator neural network is considered a critic, which…provides an estimate of the distance between the probability distributions of the real training data and the generated data… online…where models continuously receive new streams of data and keep on learning, the whole system comprising the generator G 220, discriminator D1 240, discriminator D2 250 and the model which extracts key features may be part of the deployed system”). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Notz to include the teachings of Cricri because one of ordinary skill in the art would have recognized the benefit of distinguishing between generated and real content and/or facilitating continuous learning.
but does not specifically teach training the AI module using real training data and ground truth data, training a second monitoring module using the output data of the Al module and false output data, wherein the false output data are generated by a generator of the second monitoring module; training a third monitoring module using ground truth data and false ground truth data, wherein the false ground truth data are generated by a generator of the third monitoring module, and (ii) feeding the output data stream to a discriminator of the trained second monitoring module, and determining, via the discriminator, one or more distances therefrom that are used for evaluating the output data stream of the Al module, and/or (iii) feeding the output data stream to a discriminator of the trained third monitoring module, and determining, via the discriminator, one or more distances therefrom that are used for evaluating the output data stream of the Al module.
However, the combination teaches determining, via a discriminator, one or more distances therefrom that are used for evaluating the output data stream of the Al module (e.g. Cricri, in paragraphs 20, 22, 31, and 47, “Artificial neural networks may be used for first extracting features… discriminator distinguishes between generated and real content… the discriminator neural network is considered a critic, which…provides an estimate of the distance between the probability distributions of the real training data and the generated data… online…where models continuously receive new streams of data and keep on learning, the whole system comprising the generator G 220, discriminator D1 240, discriminator D2 250 and the model which extracts key features may be part of the deployed system”) and Schiegg teaches training a second monitoring module using output data of an Al module and false output data, wherein the false output data are generated by a generator of the second monitoring module and feeding an output data stream to a discriminator of the trained second monitoring module and/or feeding the output data stream to a discriminator of a trained third monitoring module (e.g. in paragraphs 47, 92, and 94, “optimization of an objective function 5, in dependence upon which, parameters of generator 41 and of discriminator 42 are adjusted… a loss function is selected as objective function, which is implemented as…Wasserstein distance between the distributions” and figure 1 showing generated output signals 52 fed with false output data 43 generated by generator 41). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of the combination to include the teachings of Schiegg because one of ordinary skill in the art would have recognized the benefit of allowing components to be appropriately adjusted and/or optimized,
but the combination does not specifically teach training the AI module using real training data and ground truth data and training a third monitoring module using ground truth data and false ground truth data, wherein the false ground truth data are generated by a generator of the third monitoring module.
However, Lim teaches training an AI module using real training data and ground truth data (e.g. in paragraph 46, “learning apparatus may train the detector 130 by a loss function 140 based on a similarity level between the real image and ground truth 2 of the real image”) and training a third monitoring module using ground truth data and false ground truth data, wherein the false ground truth data are generated by a generator of the third monitoring module (e.g. in paragraphs 16, 45, 50, and 105, “generating a fake image of the first domain by decoding the latent variable in the first domain, generating a fake image of the second domain by decoding the latent variable in the second domain… ground truth information (GT) of the synthetic image [i.e. false ground truth] is also referred to as ground truth 1, and ground truth information of the real image is also referred to as ground truth 2… learning apparatus trains the detector 130 with the loss function 140 based on … the ground truth 1 … the ground truth 2… loss 423 may be an L2 distance between the real image (image 2) 420 and the fake image (image 2′) 421”, and figure 1). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of the combination to include the teachings of Lim because one of ordinary skill in the art would have recognized the benefit of facilitating updating of parameters for learning.
As per claim 16, the rejection of claim 15 is incorporated and the combination further teaches wherein generating false training data comprises using a respective background data source and wherein the generated false training data are fed to any of the discriminators of the respective monitoring modules (e.g. Notz, in paragraph 17, “x comes from real data and z is generated from the encoder network with input x” and figure 2 showing feeding to discriminator; Cricri, in paragraph 38, “feature-removed images x_rem(t1) 215 may be created from the at least one image x(t1) 200 by applying feature extracting and removal 210 to the at least one image frame x(t1) 200. Feature extracting may for example comprise determining, manually by a user or by a pre-trained neural network or other model, one or more objects in image x(t1) and removing these objects from the image… generator G 220 creates generated images x_gen (t2) 225 comprising target objects having at least one replaced key feature from feature-removed images”; Schiegg, in paragraphs 60-61, “a random vector is ascertained in block 3 and transmitted to generator 4l”).
As per claim 17, the rejection of claim 15 is incorporated and the combination further teaches wherein during the training phase, training losses with respect to the respective training data are determined from each of the distances, and the training losses are used for training the generators and discriminators of the respective monitoring modules (e.g. Cricri, in paragraphs 22, 43, and 46, “discriminator distinguishes between generated and real content… the discriminator neural network is considered a critic, which…provides an estimate of the distance between the probability distributions of the real training data and the generated data… loss computed from an output of the discriminator D1 240… loss computed through the discriminator D1 240 and applied to train the discriminator… loss computed through the discriminator D1 240 and applied to train the generator G 220”; Schiegg, in paragraphs 47 and 92, “optimization of an objective function 5, in dependence upon which, parameters of generator 41 and of discriminator 42 are adjusted” and figure 1; Lim, in paragraphs 100-101 and 105, “learning apparatus 400 acquires the first loss 422… the first loss 422 may be an L1 distance between the synthetic image (image 1) 418 and the fake image (image 1′) 419 of the first domain… gradient of the first loss 422 to the fake image (image 1′) 419 of the first domain is back-propagated relative to a path through which the synthetic image (image 1) is propagated in FIG. 4, and used to update parameters”).
As per claim 27, the rejection of claim 15 is incorporated and the combination further teaches wherein the first distance signal is used for evaluating a typicality of the real input data stream relative to training input data, the second distance signal is used for evaluating a typicality of the output data stream relative to training output data, and the third distance signal is used for evaluating a typicality of the output data stream relative to ground truth data from the training phase (e.g. Notz, in paragraphs 12, 26-28, 42, and 46, “calibration steps may be carried out iteratively a plurality of times and the discriminator, generator and encoder networks preferably updated during each iteration… anomaly score may include…loss… For typical (normal) sensor signals x,…have been trained on typical sensor signals. For anomalous sensor signals x,…have [not] been trained on anomalous sensor signals”, and figure 1, i.e. determining loss/anomaly means evaluating typicality; Schiegg, in paragraphs 47, 92, and 94, “optimization of an objective function 5, in dependence upon which, parameters of generator 41 and of discriminator 42 are adjusted… a loss function is selected as objective function, which is implemented as…Wasserstein distance between the distributions” and figure 1 showing generated output signals 52 fed to discriminator; Lim, in paragraphs 16, 45, 50, and 105, “generating a fake image of the first domain by decoding the latent variable in the first domain, generating a fake image of the second domain by decoding the latent variable in the second domain… ground truth information (GT) of the synthetic image [i.e. false ground truth] is also referred to as ground truth 1, and ground truth information of the real image is also referred to as ground truth 2… learning apparatus trains the detector 130 with the loss function 140 based on … the ground truth 1 … the ground truth 2… loss 423 may be an L2 distance between the real image (image 2) 420 and the fake image (image 2′) 421”).
As per claim 28, the rejection of claim 15 is incorporated and the combination further teaches wherein monitoring further comprises evaluating a relationship between a typicality of a current module output with respect to reference ground truth data and a typicality of the current module output with respect to reference training outputs (e.g. Notz, in paragraphs 12, 26-28, 41-42, 46, and 58, “calibration steps may be carried out iteratively a plurality of times and the discriminator, generator and encoder networks preferably updated during each iteration… anomaly score may include…loss… For typical (normal) sensor signals x,…have been trained on typical sensor signals. For anomalous [i.e. atypical] sensor signals x,…have [not] been trained on anomalous sensor signals… from vehicle training scenario data, scenario data associated with the signal indicating anomalous scenario acquired during open-loop operation, where anomalous scenario indicate irregular driver behaviour, wherein scenario data not associated with the signal indicating anomalous scenario is stored in a storage for vehicle training from the data storage, scenario data associated with the signal indicating anomalous scenario [i.e. reference truth data and reference training outputs]… anomaly score generated by the evaluation machine 150 may indicate how well the current scenario has been covered in previous drives [i.e. typicality including typical and atypical], which have been used for the training of models for autonomous control of the vehicle… the classification into normal and anomalous scenarios may indicate how well driving models (and the vehicle) handle the given situation… If the similar scenario coverage is very low, or not covered at all, the scenario should be classified as anomalous… if the estimated ability is too low (i.e. the anomaly score is above a pre-set threshold) the signal might indicate that the vehicle may not be confident in its ability to handle the given situation in a safe manner [current module output]”, and figure 1; Lim, in paragraphs 16, 45, 50, and 105, “ground truth information of the real image is also referred to as ground truth 2… learning apparatus trains the detector 130 with the loss function 140 based on … the ground truth 2”).
As per claim 29, the rejection of claim 28 is incorporated and the combination further teaches wherein evaluating the relationship provides information regarding a generalization capability of the Al module (e.g. Notz, in paragraph 58, “anomaly score generated by the evaluation machine 150 may indicate how well the current scenario has been covered in previous drives, which have been used for the training of models for autonomous control of the vehicle… the classification into normal and anomalous scenarios may indicate how well driving models (and the vehicle) handle the given situation [i.e. generalized capability] … If the similar scenario coverage is very low, or not covered at all, the scenario should be classified as anomalous… if the estimated ability is too low (i.e. the anomaly score is above a pre-set threshold) the signal might indicate that the vehicle may not be confident in its ability to handle the given situation in a safe manner”).
Claim 25 is rejected under 35 U.S.C. 103 as being unpatentable over Notz et al (EP 3654246) in in view of Cricri et al (US 20190122072 A1) and Halder (US 20190310636 A1) and further in view of Ceccaldi et al. (US 20190046068 A1).
As per claim 25, the rejection of claim 12 is incorporated and the combination further teaches wherein the background data source comprises an entity from which the generator generates the false training data (e.g. Notz, in paragraph 17, “x comes from real data and z is generated from the encoder network with input x” and figure 2 showing feeding to discriminator; Cricri, in paragraph 38, “feature-removed images x_rem(t1) 215 may be created from the at least one image x(t1) 200 by applying feature extracting and removal 210 to the at least one image frame x(t1) 200. Feature extracting may for example comprise determining, manually by a user or by a pre-trained neural network or other model, one or more objects in image x(t1) and removing these objects from the image… generator G 220 creates generated images x_gen (t2) 225 comprising target objects having at least one replaced key feature from feature-removed images” and figure 2a),
but does not specifically teach the entity including a latent space.
However, Ceccaldi teaches an entity including a latent space (e.g. in paragraph 46, “encoder network is configured to generate a latent space that represents a compact version of the first data. The latent space is input into a decoder network of the generator neural network”). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of the combination to include the teachings of Ceccaldi because one of ordinary skill in the art would have recognized the benefit of representing data (also amounts a simple substitution that yields predictable results [e.g. see KSR Int'l Co v. Teleflex Inc., 550 US 398,82 USPQ2d 1385,1396 (U.S. 2007) and MPEP 2143(B)]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
For example,
Cheng et al (US 20190130212 A1) teaches “inference is encapsulated in a generative adversarial training process… discriminator aims to distinguish the real data distribution custom-character.sub.data(x) from the fake sample distribution custom-character.sub.g(z), where N(0, custom-character). An Earth-Mover distance metric can be used, such that the problem is expressed as” a distance equation between real data and false data (e.g. in paragraphs 18 and 25).
Rezagholizadeh et al. (US 20180336471 A1) teaches “outputting, from a generator network, generated samples derived from a random noise vector (block 302); inputting, to a discriminator network, the generated samples, a plurality of labelled training samples, and a plurality of unlabelled training samples (block 304); outputting, from the discriminator network, for at least the generated samples, a predicted continuous label for a generated sample and a predicted probability that the sample is a real sample or a fake sample” (e.g. in paragraph 84).
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 WILLIAM WONG whose telephone number is (571)270-1399. The examiner can normally be reached Monday-Friday 9am-5pm.
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, TAMARA KYLE can be reached at (571)272-4241. 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.
/W.W/Examiner, Art Unit 2144 09/19/2026
/TAMARA T KYLE/Supervisory Patent Examiner, Art Unit 2144
