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 .
Detailed Action
This action is in response to the claims filed 12/6/2023:
Claims 1 – 24 are pending.
Claims 1, 12, and 23 are independent.
Specification
The disclosure is objected to because of the following informalities:
The title of the invention is not descriptive. A new title is required that is clearly indicative of the invention to which the claims are directed.
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.
Claims 6 and 13 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 4 and 14, "high dimensional" is indefinite. This is a relative term with no relative basis for comparison. In the interest of further examination high dimensional is interpreted as having two or more dimensions.
Regarding claims 7 and 17, "linear-polation" appears to be a coined term with no associated definition. This does not appear to be a misspelling of "linear-interpolation" as the instant specification explicitly differentiates between "interpolation", "extrapolation", and "linear-polation".
Regarding claims 5-7 and 15-17, claims 5-7 and 15-17 are rejected with respect to their dependence on rejected claims 4 and 14.
Claim Rejections - 35 USC § 101
101 Rejection
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-4, 6-14, and 16-22 are rejected under 35 USC § 101 because the claimed invention is directed to non-statutory subject matter.
Regarding Claim 1: Claim 1 is rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1 Analysis: Claim 1 is directed to a method, which is directed to a process, one of the statutory categories.
Step 2A Prong One Analysis: Claim 1 under its broadest reasonable interpretation is a series of mental processes. For example, but for the generic computer components language, the above limitations in the context of this claim encompass machine learning processing, including the following:
analyzing the first output of the neural network to determine first correct predictions and first incorrect predictions using a classifier (observation, evaluation, and judgement),
mutating seeds of the validation dataset corresponding to the first correct predictions (observation, evaluation, and judgement)
analyzing the second output of the neural network to determine second correct predictions and second incorrect predictions using the classifier (observation, evaluation, and judgement)
determining whether there is an increase in neural network coverage for mutated seeds yielding the second correct predictions (observation, evaluation, and judgement)
performing steps of mutating the seeds […] and analyzing the second output of the neural network for the mutated seeds yielding the second correct predictions
Therefore, claim 1 recites an abstract idea which is a judicial exception.
Step 2A Prong Two Analysis: Claim 1 recites additional elements “running the mutated seeds through the neural network”. However, these additional features are computer components recited at a high-level of generality, such that they amount to no more than mere instructions to apply the judicial exception using a generic computer component. An additional element that merely recites the words “apply it” (or an equivalent) with the judicial exception, or merely includes instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea, does not integrate the judicial exception into a practical application (See MPEP 2106.05(f)). Claim 1 also recites additional elements “running a validation dataset through the neural network to provide a first output”, “running the mutated seeds through the neural network to provide a second output;” which amounts to gathering and outputting data which is insignificant extra-solution activity (See MPEP 2106.05(g)). Therefore, claim 1 is directed to a judicial exception.
Step 2B Analysis: Claim 1 does not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to the lack of integration of the abstract idea into a practical application, the additional elements recited in claim 1 amount to no more than mere instructions to apply the judicial exception using a generic computer component and insignificant extra-solution activity. The gathering and outputting of data is considered well-understood, routine, and conventional in the art (See MPEP 2106.05(d)(II)(i)).
For the reasons above, claim 1 is rejected as being directed to non-patentable subject matter under §101. This rejection applies equally to independent claim 12, which recites a device, respectively, as well as to dependent claims 2-11 and 13-22.
Independent claim 12 recites additional instructions to apply the judicial exceptions to generic computer components “A device for augmenting training data for a neural network, the device comprising: a processor; and a memory with program instructions stored thereon coupled to the processor, wherein the program instructions, when executed by the processor, enable the device to.”
The additional limitations of the dependent claims are addressed briefly below:
Dependent claims 2 and 13 recite additional instructions to apply the judicial exception using generic computer components “using the mutated seeds yielding the second correct predictions as training data to further train the neural network”
Dependent claim 3 recites additional observation, evaluation, and judgement “determining a neural network coverage metric for the validation dataset combined with the mutated seeds”
Dependent claims 4 and 14 recite additional observation, evaluation, and judgement “mutating the seeds comprises performing high dimensional perturbations or latent space mutations”
Dependent claims 6 and 16 recite additional observation, evaluation, and judgement “the high dimensional perturbations include: for infrared applications: at least one of flipping and rotating, brightness and contrast alterations, gaussian noise, blur, scaling and cropping, resolution alteration, or data mixup; for time of flight (TOF) applications: at least one of time shift, scaling, noise addition, temporal jitter, depth cropping and flipping, data interpolation, resolution alteration, or outlier injection; for radar applications: at least one of range compression, time shift, doppler shift, range scaling, noise addition, clutter addition, azimuth and elevation variation, or resolution change; for audio applications: at least one of time change, pitch shift, background noise addition, volume variation, time and frequency domain variations, clipping and distortion, audio concatenation, speed perturbation, or echo generation; for ultrasound applications: at least one of: flipping and rotation, zooming and scaling, noise addition, contrast and brightness alteration, shadow and artifact simulation, texture variation, or resolution alteration; or for WiFi applications: at least one of RSSI scaling, signal dropout, signal interpolation, noise addition, temporal jitter, location perturbation, access point (AP) dropout and rotation, data splitting, or AP density variation”
Dependent claims 7 and 17 recite additional observation, evaluation, and judgement “the latent space mutations include interpolation, extrapolation, linear-polation and resampling”
Dependent claims 8 and 18 recite additional observation, evaluation, and judgement “determining whether there is an increase in neural network coverage includes determining a k-multisectional neuron coverage (KMNC)”
Dependent claims 9 and 19 recite additional observation, evaluation, and judgement “determining the KMNC includes determining an output range of each neuron based on the validation dataset, dividing the output range into K bins, determining a coverage of each bin with respect to the validation dataset, and determining whether there is an increased number of bins covered when using a respective mutated seed”
Dependent claims 10 and 20 recite additional observation, evaluation, and judgement “determining whether there is an increase in neural network coverage includes determining a neuron coverage (NC), a neuron boundary coverage (NBC), a strong neuron activation coverage (SNAC), or a neural coverage (NLC)”
Dependent claims 11 and 21 recite additional instructions to apply the judicial exception using generic computer components “wherein the neural network is a deep neural network”
Dependent claim 22 recites additional instructions to apply the judicial exception using generic computer components “The device of claim 12, further comprising the neural network.”
Therefore, when considering the elements separately and in combination, they do not add significantly more to the inventive concept. Accordingly, claims 1-4, 6-14, and 16-22 are rejected under 35 U.S.C. § 101.
Claim Rejections - 35 USC § 102
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 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 the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claims 1-4, 10-14, and 20-22 are rejected under U.S.C. §102(a)(1) as being anticipated by Guo (“Coverage Guided Differential Adversarial Testing of Deep Learning Systems”, 2021).
PNG
media_image1.png
458
1476
media_image1.png
Greyscale
FIG. 1 of Guo
Regarding claim 1, Guo teaches A method for augmenting training data for a neural network, ([p. 934] "DLFuzz exhibits its practical use for steadily improving the deep learning systems by augmenting the training dataset and retraining the model" [p. 8] "we incorporated 114 adversarial images into the training set of three CNNs on MNIST and retrained them to see if their accuracy is able to increase" Guo's adversarial/mutated (augmented) DLfuzz generated inputs are explicitly incorporated into a training set and used for retraining.)
the method comprising: running a validation dataset through the neural network to provide a first output;([Abstract] "we propose DLFuzz, the coverage guided differential adversarial testing framework to guide deep learning systems exposing incorrect behaviors. DLFuzz keeps minutely mutating the input to maximize the neuron coverage and the prediction difference between the original input and the mutated input [...] The incorrect behaviors obtained by DLFuzz are then exploited for retraining and improving the dependability of the models [...] the test inputs" [p. 936] "input_list <- unlabeled inputs for testing dnn <- DNN under test" Guo's input_list is the dataset of test/validation inputs to obtain first predictions. For each seed xs the DNN classifier is run and produces the first output: original class c and top-k competing labels c_topk.)
analyzing the first output of the neural network to determine first correct predictions and first incorrect predictions using a classifier;([p. 935] "the whole test input set X is composed of images to be classified and each input x is the one during testing. The DNN is a particular convolutional neural network (CNN) under test, such as VGG-16 [35]. The mutation algorithm applies tiny perturbation to x and gets x0, which is visibly indistinguishable from x. If the mutated input x0 and the original input x are both fed to the CNN but classified to be of different class labels, we treat this as an incorrect behavior and x0 to be one of the adversarial inputs. The inconsistent classification results before and after mutation indicate that at least one of them is wrong so that manually labeling effort is not required here. In contrast, if the two are predicted of the same class label by the CNN, x0 will continue to be mutated by the mutation algorithm to test the CNN’s robustness" The classifier is Guo's DNN. Guo analyzes the first classifier output into original class c and competing labels c_topk. The original label is the baseline expected classification for identity-preserving mutations; the top-k different labels are candidate incorrect classifications.)
mutating seeds of the validation dataset corresponding to the first correct predictions;([p. 935] "The mutation algorithm applies tiny perturbation to x and gets x0 […] if the two are predicted of the same class label by the CNN, x0 will continue to be mutated by the mutation algorithm to test the CNN’s robustnes [...] a seed list is well maintained for each given input x, keeping those intermediate mutated inputs which could increase the neuron coverage and satisfy the limit for the perturbation" See FIG. 1 Seed List ("seeds for subsequent fuzzing") to test inputs to mutation algorithm)
running the mutated seeds through the neural network to provide a second output;([p. 936 FIG. 2] "13: x'=xs+perturbation //mutated input obtained 14: c' =dnn.predict(x') //label after mutation See FIG. 1)
analyzing the second output of the neural network to determine second correct predictions and second incorrect predictions using the classifier;([p. 935] "the whole test input set X is composed of images to be classified and each input x is the one during testing. The DNN is a particular convolutional neural network (CNN) under test, such as VGG-16 [35]. The mutation algorithm applies tiny perturbation to x and gets x0, which is visibly indistinguishable from x. If the mutated input x0 and the original input x are both fed to the CNN but classified to be of different class labels, we treat this as an incorrect behavior and x0 to be one of the adversarial inputs. The inconsistent classification results before and after mutation indicate that at least one of them is wrong so that manually labeling effort is not required here. In contrast, if the two are predicted of the same class label by the CNN, x0 will continue to be mutated by the mutation algorithm to test the CNN’s robustness" See FIG. 1)
determining whether there is an increase in neural network coverage for mutated seeds yielding the second correct predictions; and([p. 936 FIG. 2] "15. update conv_tracker […] 17: if the coverage improved by x' > " [p. 936] "These neurons are selected considering many strategies to improve the neuron coverage" [p. 934] "DLFuzz keeps the mutated inputs which contribute to a certain increase of the neuron coverage for the subsequent fuzzing.")
performing steps of mutating the seeds, running the mutated seeds through the neural network, and analyzing the second output of the neural network for the mutated seeds yielding the second correct predictions.([p. 934] "DLFuzz keeps the mutated inputs which contribute to a certain increase of the neuron coverage for the subsequent fuzzing." [p. 936 FIG. 2] "2: for i=0 to len(input_list) do" [p. 936] "In the mutation process, DLFuzz iteratively applies the processed gradient as the perturbation to xs and obtains the intermediate input x0").
Regarding claim 2, Guo teaches The method of claim 1, further comprising using the mutated seeds yielding the second correct predictions as training data to further train the neural network.(Guo [p. 934] "DLFuzz exhibits its practical use for steadily improving the deep learning systems by augmenting the training dataset and retraining the model" [p. 8] "we incorporated 114 adversarial images into the training set of three CNNs on MNIST and retrained them to see if their accuracy is able to increase" See also FIG. 1 and algorithm in FIG. 2. After the first loop additional adversarial inputs may still be obtained to further retrain the neural network, the adversarial examples explicitly determined from mutated seeds).
Regarding claim 3, Guo teaches The method of claim 1, further comprising determining a neural network coverage metric for the validation dataset combined with the mutated seeds.(Guo [p. 936] "After each mutation, the intermediate class label c0, coverage information, l2 distance of x and x0 are acquired").
Regarding claim 4, Guo teaches The method of claim 1, wherein mutating the seeds comprises performing high dimensional perturbations or latent space mutations.(Guo [p. 940] "DLFuzz currently aims at image classification tasks" [p. 935] "The mutation algorithm applies tiny perturbation to x and gets x0" Applying mutation to images interpreted as performing high dimensional perturbations).
Regarding claim 10, Guo teaches The method of claim 1, wherein determining whether there is an increase in neural network coverage includes determining a neuron coverage (NC), a neuron boundary coverage (NBC), a strong neuron activation coverage (SNAC), or a neural coverage (NLC).(Guo [p. 2] "maximize the neuron coverage").
Regarding claim 11, Guo teaches The method of claim 1, wherein the neural network is a deep neural network.(Guo [p. 935] "the whole test input set X is composed of images to be classified and each input x is the one during testing. The DNN is a particular convolutional neural network (CNN)").
Regarding claim 12, claim 12 is directed towards a device for performing the method of claim 1. Therefore, the rejection applied to claim 1 also applies to claim 12. Claim 12 recites additional elements A device for augmenting training data for a neural network, the device comprising: a processor; and a memory with program instructions stored thereon coupled to the processor, wherein the program instructions, when executed by the processor, enable the device to: (Guo [p. 937] "We implemented DLFuzz based on various widespread frameworks of deep learning systems, Tensorflow 1.2.1, Keras 2.1.3 and Caffe 1.0.0. DLFuzz exhibits high portability on these general frameworks. We developed and evaluated DLFuzz on a computer with 4 cores (Intel i7 7700HQ @3.6 GHz), 16 GB of memory, an NVIDIA GTX 1070 GPU and Ubuntu 16.04.4 as the host OS.").
Similarly, regarding claims 13-14 and 20-21, claims 13-14 and 20-21 are directed towards a device for performing the methods of claims 2, 4, 10, and 11, respectively. Therefore, the rejection applied to claims 2, 4, 10, and 11 also apply to claims 13-14 and 20-21.
Regarding claim 22, Guo teaches The device of claim 12, further comprising the neural network.(Guo [p. 937] "We implemented DLFuzz based on various widespread frameworks of deep learning systems, Tensorflow 1.2.1, Keras 2.1.3 and Caffe 1.0.0. DLFuzz exhibits high portability on these general frameworks. We developed and evaluated DLFuzz on a computer with 4 cores (Intel i7 7700HQ @3.6 GHz), 16 GB of memory, an NVIDIA GTX 1070 GPU and Ubuntu 16.04.4 as the host OS." See FIG. 1 which shows the DNN is part of DLFuzz which is explicitly comprised by the device running Ubuntu).
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 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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 5, 6, 7, 15, 16, and 17 are rejected under U.S.C. §103 as being unpatentable over the combination of Guo and Wong (“LEARNING PERTURBATION SETS FOR ROBUST MACHINE LEARNING”, 2020).
Regarding claim 5, Guo teaches The method of claim 4.
However, Guo doesn't explicitly teach, wherein the latent space mutations are performed using a conditional variational autoencoder (CVAE).
Wong, in the same field of endeavor, teaches The method of claim 4, wherein the latent space mutations are performed using a conditional variational autoencoder (CVAE). ([Abstract] "we use a conditional generator that defines the perturbation set over a constrained region of the latent space. We formulate desirable properties that measure the quality of a learned perturbation set, and theoretically prove that a conditional variational autoencoder naturally satisfies these criteria").
Guo as well as Wong are directed towards training data augmentation for machine learning. Therefore, Guo as well as Wong are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of Guo with the teachings of Wong by using Wong’s CVAE as the DNN in Guo. Wong provides as additional motivation for combination ([p. 7] “We highlight some empirical results in Table 2, where we first find that training with the CVAE perturbation set can improve generalization. Specifically, using the CVAE perturbation set during training achieves 3-5% improved accuracy”).
Regarding claim 6, Guo teaches The method of claim 4.
However, Guo doesn't explicitly teach, wherein the high dimensional perturbations include: for infrared applications: at least one of flipping and rotating, brightness and contrast alterations, gaussian noise, blur, scaling and cropping, resolution alteration, or data mixup; for time of flight (TOF) applications: at least one of time shift, scaling, noise addition, temporal jitter, depth cropping and flipping, data interpolation, resolution alteration, or outlier injection; for radar applications: at least one of range compression, time shift, doppler shift, range scaling, noise addition, clutter addition, azimuth and elevation variation, or resolution change; for audio applications: at least one of time change, pitch shift, background noise addition, volume variation, time and frequency domain variations, clipping and distortion, audio concatenation, speed perturbation, or echo generation; for ultrasound applications: at least one of: flipping and rotation, zooming and scaling, noise addition, contrast and brightness alteration, shadow and artifact simulation, texture variation, or resolution alteration; or for WiFi applications: at least one of RSSI scaling, signal dropout, signal interpolation, noise addition, temporal jitter, location perturbation, access point (AP) dropout and rotation, data splitting, or AP density variation..
Wong, in the same field of endeavor, teaches The method of claim 4, wherein the high dimensional perturbations include: for infrared applications: at least one of flipping and rotating, brightness and contrast alterations, gaussian noise, blur, scaling and cropping, resolution alteration, or data mixup; for time of flight (TOF) applications: at least one of time shift, scaling, noise addition, temporal jitter, depth cropping and flipping, data interpolation, resolution alteration, or outlier injection; for radar applications: at least one of range compression, time shift, doppler shift, range scaling, noise addition, clutter addition, azimuth and elevation variation, or resolution change; for audio applications: at least one of time change, pitch shift, background noise addition, volume variation, time and frequency domain variations, clipping and distortion, audio concatenation, speed perturbation, or echo generation; for ultrasound applications: at least one of: flipping and rotation, zooming and scaling, noise addition, contrast and brightness alteration, shadow and artifact simulation, texture variation, or resolution alteration; or for WiFi applications: at least one of RSSI scaling, signal dropout, signal interpolation, noise addition, temporal jitter, location perturbation, access point (AP) dropout and rotation, data splitting, or AP density variation. ([p. 27] "Randomized smoothing is done in its most basic form, with normal Gaussian data augmentation at the specification noise level").
Guo as well as Wong are directed towards training data augmentation for machine learning. Therefore, Guo as well as Wong are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of Guo with the teachings of Wong by using Wong’s CVAE as the DNN in Guo. Wong provides as additional motivation for combination ([p. 7] “We highlight some empirical results in Table 2, where we first find that training with the CVAE perturbation set can improve generalization. Specifically, using the CVAE perturbation set during training achieves 3-5% improved accuracy”).
Regarding claim 7, Guo teaches The method of claim 4.
However, Guo doesn't explicitly teach wherein the latent space mutations include interpolation, extrapolation, linear-polation and resampling.
Wong, in the same field of endeavor, teaches The method of claim 4 wherein the latent space mutations include interpolation, extrapolation, linear-polation and resampling.([p. 2] "The resulting perturbation set can interpolate between common corruptions, produce diverse samples, and be used in adversarial training and randomized smoothing frameworks. The adversarially trained models have improved generalization performance to both in- and out-of-distribution corruptions and better robustness to adversarial corruptions." producing diverse samples interpreted as resampling. Generalization interpreted as extrapolation.).
Guo as well as Wong are directed towards training data augmentation for machine learning. Therefore, Guo as well as Wong are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of Guo with the teachings of Wong by using Wong’s CVAE as the DNN in Guo. Wong provides as additional motivation for combination ([p. 7] “We highlight some empirical results in Table 2, where we first find that training with the CVAE perturbation set can improve generalization. Specifically, using the CVAE perturbation set during training achieves 3-5% improved accuracy”).
Regarding claims 15-17, claims 15-17 are directed towards a device for performing the methods of claims 5-7, respectively. Therefore, the rejections applied to claims 5-7 also apply to claims 15-17.
Claims 8, 9, 18, and 19 are rejected under U.S.C. §103 as being unpatentable over the combination of Guo and Xie (“DeepHunter: A Coverage-Guided Fuzz Testing Framework for Deep Neural Networks”, 2019).
Regarding claim 8, Guo teaches The method of claim 1.
However, Guo doesn't explicitly teach wherein determining whether there is an increase in neural network coverage includes determining a k-multisectional neuron coverage (KMNC).
Xie, in the same field of endeavor, teaches The method of claim 1, wherein determining whether there is an increase in neural network coverage includes determining a k-multisectional neuron coverage (KMNC).([p. 147] "we propose DeepHunter, a general-purpose coverage-guided fuzz testing framework for DNNs […] for example, for KMNC in LeNet-5, DeepHunter has achieved 69.14% coverage").
Guo as well as Xie are directed towards data augmentation for machine learning. Therefore, Guo as well as Xie are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of Guo with the teachings of Xie by using KMNC for coverage determination. Xie provides as additional motivation for combination ([p. 148] “KMNC measures the ratio of all covered sections of all neurons of a DNN”). This motivation for combination also applies to the remaining claims which depend on this combination.
Regarding claim 9, the combination of Guo, and Xie teaches The method of claim 8, wherein determining the KMNC includes determining an output range of each neuron based on the validation dataset, dividing the output range into K bins, determining a coverage of each bin with respect to the validation dataset, (Xie [p. 3] "We propose a metamorphic mutation strategy towards generating valid test cases [...] We perform a systematically large-scale study to evaluate the effectiveness of different testing strategies and criteria on increasing coverage and detecting defects of DNNs" [p. 3] "(KMNC). For each neuron, the range of its values (obtained from training data) are partitioned into k sections. An input covers a section of a neuron if the output value falls into the corresponding value section range. KMNC measures the ratio of all covered sections of all neurons of a DNN")
and determining whether there is an increased number of bins covered when using a respective mutated seed.(Xie [p. 3] "We propose a metamorphic mutation strategy towards generating valid test cases [...] We perform a systematically large-scale study to evaluate the effectiveness of different testing strategies and criteria on increasing coverage and detecting defects of DNNs").
Regarding claims 18 and 19, claims 18 and 19 are directed towards a device for performing the method of claims 8 and 9, respectively. Therefore, the rejection applied to claims 8 and 9 also applies to claim 18 and 19.
Claims 23 and 24 are rejected under U.S.C. §103 as being unpatentable over the combination of Guo and Sekhon (“White-box Testing of NLP models with Mask Neuron Coverage”, 2022).
Regarding claim 23, Guo teaches A method for retraining a neural network, the method comprising:([p. 934] "DLFuzz exhibits its practical use for steadily improving the deep learning systems by augmenting the training dataset and retraining the model" [p. 8] "we incorporated 114 adversarial images into the training set of three CNNs on MNIST and retrained them to see if their accuracy is able to increase" Guo's adversarial/mutated (augmented) DLfuzz generated inputs are explicitly incorporated into a training set and used for retraining.)
providing a first set of seeds to the neural network to provide a first output;([p. 934] "it keeps the mutated inputs contributing to the increase of coverage in a seed list" See FIG. 1 Seed list which is used to augment the original input which is provided directly to the DNN as input to generate an output)
applying a classifier to the first output to determine first seeds of the first set of seeds corresponding to first correct predictions by the classifier;([p. 935] "the whole test input set X is composed of images to be classified and each input x is the one during testing. The DNN is a particular convolutional neural network (CNN) under test, such as VGG-16 [35]. The mutation algorithm applies tiny perturbation to x and gets x0, which is visibly indistinguishable from x. If the mutated input x0 and the original input x are both fed to the CNN but classified to be of different class labels, we treat this as an incorrect behavior and x0 to be one of the adversarial inputs. The inconsistent classification results before and after mutation indicate that at least one of them is wrong so that manually labeling effort is not required here. In contrast, if the two are predicted of the same class label by the CNN, x0 will continue to be mutated by the mutation algorithm to test the CNN’s robustness" The classifier is Guo's DNN. Guo analyzes the first classifier output into original class c and competing labels c_topk. The original label is the baseline expected classification for identity-preserving mutations; the top-k different labels are candidate incorrect classifications.)
mutating the first seeds to provide first mutated seeds;([p. 935] "The mutation algorithm applies tiny perturbation to x and gets x0 […] if the two are predicted of the same class label by the CNN, x0 will continue to be mutated by the mutation algorithm to test the CNN’s robustnes [...] a seed list is well maintained for each given input x, keeping those intermediate mutated inputs which could increase the neuron coverage and satisfy the limit for the perturbation" See FIG. 1 Seed List ("seeds for subsequent fuzzing") to test inputs to mutation algorithm)
running the first mutated seeds through the neural network to provide a second output;([p. 936 FIG. 2] "13: x'=xs+perturbation //mutated input obtained 14: c' =dnn.predict(x') //label after mutation See FIG. 1)
applying the classifier to the second output to determine second seeds of the first set of seeds corresponding to second correct predictions by the classifier;([p. 935] "the whole test input set X is composed of images to be classified and each input x is the one during testing. The DNN is a particular convolutional neural network (CNN) under test, such as VGG-16 [35]. The mutation algorithm applies tiny perturbation to x and gets x0, which is visibly indistinguishable from x. If the mutated input x0 and the original input x are both fed to the CNN but classified to be of different class labels, we treat this as an incorrect behavior and x0 to be one of the adversarial inputs. The inconsistent classification results before and after mutation indicate that at least one of them is wrong so that manually labeling effort is not required here. In contrast, if the two are predicted of the same class label by the CNN, x0 will continue to be mutated by the mutation algorithm to test the CNN’s robustness" The classifier is Guo's DNN. Guo analyzes the first classifier output into original class c and competing labels c_topk. The original label is the baseline expected classification for identity-preserving mutations; the top-k different labels are candidate incorrect classifications.)
determining whether there is an increase in neural network coverage for determined second seeds; and([p. 936 FIG. 2] "15. update conv_tracker […] 17: if the coverage improved by x' > " [p. 936] "These neurons are selected considering many strategies to improve the neuron coverage" [p. 934] "DLFuzz keeps the mutated inputs which contribute to a certain increase of the neuron coverage for the subsequent fuzzing.").
However, Guo does not explicitly teach using at least one of the second seeds to retrain the neural network in response to a determination that the second seed of the second seeds causes an increase in neural network coverage..
Sekhon, in the same field of endeavor, teaches using at least one of the second seeds to retrain the neural network in response to a determination that the second seed of the second seeds causes an increase in neural network coverage.([p. 7] "We focus on using coverage to guide generation of augmented samples. We propose a greedy search algorithm to coverage as guide to generate a new training set with selected augmentations […] We then add the coverage selected samples into the training set and retrain a target model").
Guo as well as Sekhon are directed towards data augmentation for machine learning. Therefore, Guo as well as Sekhon are analogous art in the same field of endeavor. It would have been obvious before the effective filing date of the claimed invention to combine the teachings of Guo with the teachings of Sekhon by using Sekhon's coverage guided retraining philosophy on the outputs of Guo’s coverage-focused mutation loop. Sekhon provides as motivation for combination ([p. 7] “Our results show that using MNCOVER to guide data augmentation can improve test accuracy”).
Regarding claim 24, the combination of Guo and Sekhon teaches The method of claim 23, wherein: applying the classifier to the second output further comprising applying the classifier to the second output to determine third seeds of the first set of seeds corresponding to incorrect predictions by the classifier; and using at least one of the third seeds to retrain the neural network.(Guo [p. 934] "DLFuzz exhibits its practical use for steadily improving the deep learning systems by augmenting the training dataset and retraining the model" [p. 935] "If the mutated input x0 and the original input x are both fed to the CNN but classified to be of different class labels, we treat this as an incorrect behavior and x0 to be one of the adversarial inputs" [p. 940] "we incorporated 114 adversarial images into the training set of three CNNs on MNIST and retrained them to see if their accuracy is able to increase" Guo's adversarial/mutated (augmented) DLFuzz generated inputs are explicitly incorporated into a training set and used for retraining. See FIG. 1 where mutated/augmented samples are classified as incorrect/adversarial inputs).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Eisele (US20230205677A1) is directed towards a neural network system for iteratively mutating inputs and augmenting the training data set to improve coverage.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SIDNEY VINCENT BOSTWICK whose telephone number is (571)272-4720. The examiner can normally be reached M-F 7:30am-5:00pm EST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Miranda Huang can be reached on (571)270-7092. 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.
/SIDNEY VINCENT BOSTWICK/Examiner, Art Unit 2124
Prosecution Insights