Notice of Pre-AIA or AIA Status
1. 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
Continued Examination Under 37 CFR 1.114
2. 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and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114.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 9/25/2025 has been entered. Claims 1, 14, and 16 have been amended. Claims 1-14, 16-18, and 20-21 remain pending in the application.
Information Disclosure Statement
3. The information disclosure statement (IDS(s)) submitted on 08/11/2025 is/are in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Response to Arguments
4. Applicant’s arguments with respect to claims have been considered but are moot in view of new ground of rejection. See rejections below for details.
Claim Rejections – 35 USC § 103
5. 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 of this title, 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.
6. Claims 1-3, 5-14, 16-17, and 20-21 are rejected under 35 U.S.C. 103 as being unpatentable over Cooper et al. (U.S. Patent Application Pub. No. US 20200117953 A1) in view of Karam et al. (U.S. Patent Application Pub. No. US 20190303720 A1).
Claim 1: Cooper teaches a computer-implemented method for improving robustness of a neural network in a function-specific manner (i.e. By maintaining, saving, storing or using the camera properties of the images to train data models while still adding to the training data with augmented images, the resulting model may be useful across many different devices having the same camera properties. Moreover, the augmentation may provide generalization and greater robustness to the model prediction, particularly when images are clouded, occluded, or otherwise do not provide clear views of the detectable objects; para. [0017]), comprising:
accessing the neural network (i.e. The model training module 340 trains a computer model based on the images captured by the image collection system 140 and the augmented images generated by the image augmentation module 330. These images may be used as an image training set for the model training. In one embodiment, the machine-learned models are neural network models such as feed-forward networks, convolutional neural networks (CNN), deep neural networks (DNN), recurrent neural networks (RNN), self-organizing maps (SOM), and the like, that are trained by the model training module 340 based on training data; para. [0039]), wherein the neural network is trained on the basis of a training data set comprising training data having original image data (i.e. The training images 410 may include a training image 410A that is not augmented, having associated training output 420A that corresponds with the labeled data in the labeled training image 400; para. [0028, 0031]), the neural network being trained with a function for processing raw image data from a sensor (i.e. A model receives the sensor data (e.g., an image) as an input and outputs an output prediction according to the training of the model; para. [0024, 0039]);
generating at least one changed training data set having manipulated image data by manipulating the original image data (i.e. The image augmentation module 330 generates augmented images by applying an image manipulation function to the labeled training image 400. The image manipulation function generates a modified version of the labeled training image 400 to vary the characteristics of the image for training the model. The image manipulation function used to generate the training images maintains the camera properties of the labeled training image 400; para. [0030, 0032]), wherein the training data is changed while maintaining semantically meaningful content, wherein the semantically meaningful content is image data content (i.e. the manipulation function may maintain the scale, perspective, orientation, and other characteristics of the view of the environment that may be affected by the physical capture characteristics of the camera or the position of the camera when capturing the environment that may be consistent across various devices. Accordingly, the image manipulation functions may affect how viewable objects or other features of the environment are or how clearly these are seen in a scene, but may not affect the location or size of objects in the image; para. [0018, 0032]), which image data content the function of the neural network is intended to recognize as part of a semantic segmentation or classification (i.e. The model training system 130 in this example typically trains models based on image inputs to generate an output predicting information about the image. For example, in various embodiments these outputs may identify objects in the image (identify objects, either by bounding box or by segmentation; para. [0022]), including one or more objects in an environment of the automated vehicle, a behavior of the one or more objects, a road geometry or topology (i.e. These approaches may be particularly useful for object detection and in autonomous vehicles; para. [0017, 0024]);
changing one or more of parameters and an architecture of the neural network (i.e. That prediction for the training inputs can be compared with the designated output (e.g., the label) to assess a loss (e.g., with a loss function) and the parameters may be revised via an optimization algorithm to optimize the set of parameters to reduce the loss function; para. [0039]) to obtain a changed neural network (i.e. After training, the computer model may be stored in the trained computer model store 370; para. [0039]) an application of the original training data set and the at least one changed training data set on the trained neural network (i.e. That is, during training the training data is assessed according to a current set of parameters to generate a prediction. That prediction for the training inputs can be compared with the designated output (e.g., the label) to assess a loss (e.g., with a loss function) and the parameters may be revised via an optimization algorithm to optimize the set of parameters to reduce the loss function; para. [0039]);
training the changed neural network on the basis of the original image data of the training data set and at least part of the manipulated image data of the at least one changed training data set (i.e. The model training module 340 trains a computer model based on the images captured by the image collection system 140 and the augmented images generated by the image augmentation module 330; para. [0039]); and
using the changed neural network in one or more of a driver assistance system and an automated motor vehicle to provide a self-driving function in the vehicle (i.e. The image collection system 140 may be manually operated or may be operated be a partially-or fully-automated vehicle … when the model application system 110 is a vehicle, the model may predict the presence of objects in the image, which may be used by the model application system 110 as part of a safety system or as a part of an autonomous (or semi-autonomous) control system; para. [0024, 0025]).
Cooper does not explicitly teach according to a comparison result of a comparison between an application of the original training data set and the at least one changed training data set on the trained neural network.
However, Karam teaches according to a comparison result of a comparison (i.e. Computing the output predictions by swapping a distortion affected filter output with its corresponding clean output for each of the ranked filters would improve classification performance; para. [0043-0049]) between an application of the original training data set (i.e. The terms “pre-trained” or “baseline” network are used herein to refer to any network that is trained on undistorted images; para. [0040, 0043]) and the at least one changed training data set on the trained neural network (i.e. it is clear that the networks trained on clean images perform poorly when presented with input images that are distorted even at low distortion levels; para. [0043, 0044, 0046]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 2: Cooper and Karam teach the method of claim 1. Cooper does not explicitly teach wherein changing one or more of the parameters and the architecture of the neural network and training the changed neural network are repeated until at least one termination criterion is met.
However, Karam further teaches wherein changing one or more of the parameters and the architecture of the neural network and training the changed neural network are repeated until at least one termination criterion is met (i.e. the features can be corrected by appending correction units taking the form of small blocks of stacked convolutional layers such as a residual block with a single skip connection or other CNN-based block or other blocks implementing a transformation, at the output of select filters and train them to correct the worst distortion-affected filter activations using a target-oriented loss or other desired loss function … Starting with an initial learning rate that is ten times lower than the initial learning rate used to generate the pre-trained model, i.e., 0.01, and momentum equal to 0.9, network parameters in various layers are updated for a fixed number of iterations (40 epochs), with the learning rate reduced by a factor of 10 after every 10 epochs; para. [0031, 0059, 0061]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 3: Cooper and Karam teach the method of claim 1. Cooper does not explicitly teach wherein a robustness measure is determined for the neural network on the basis of the comparison, wherein the changing is conducted on the basis of the determined robustness measure.
However, Karam further teaches wherein a robustness measure is determined for the neural network on the basis of the comparison (i.e. The correction priority of a convolutional filter ϕi,j is defined as the improvement in DNN performance on a validation set, generated by replacing i,j(xi) with ϕi,j(xi) for a pre-trained network. Let the baseline performance (computed over distorted images) for a network pb, which can be obtained by computing the average top-1 accuracy of the network over a set of images or another task-specific performance measure. Let pswp(i,j) denote the new improved performance of the network after swapping; para. [0047]), wherein the changing is conducted on the basis of the determined robustness measure (i.e. Metrics are disclosed herein to rank the convolutional filters in order of the highest gain in classification accuracy upon correction. In one embodiment, the features can be corrected by appending correction units taking the form of small blocks of stacked convolutional layers such as a residual block with a single skip connection or other CNN-based block or other blocks implementing a transformation, at the output of select filters and train them to correct the worst distortion-affected filter activations using a target-oriented loss or other desired loss function, whilst leaving the rest of the pre-trained filter outputs in the network unchanged; para. [0031, 0065]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 5: Cooper and Karam teach the method of claim 2. Cooper does not explicitly teach wherein the termination criterion is one or more of reaching a target value for a functional quality of the trained changed neural network and reaching a target value for the determined robustness measure.
However, Karam further teaches wherein the termination criterion is one or more of reaching a target value for a functional quality of the trained changed neural network (i.e. The proposed generative sensing framework aims at transforming low-end, low-quality sensor data into higher quality sensor data in terms of achieved classification accuracy; para. [0033, 0047]) and reaching a target value for the determined robustness measure (i.e. computing the average top-1 accuracy of the network over a set of images or another task-specific performance measure. Let pswp(i,j) denote the new improved performance of the network after swapping (xi) with ϕi,j(xi). As the implementation focuses on classification tasks, the average top-1 accuracy over a set of distorted images is used to measure pb and pswp(i,j). The correction priority for filter ϕi,j is then given by: τ(i,j)=p swp(i,j)−p b ; para. [0047-0050, 0093]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 6: Cooper and Karam teach the method of claim 1. Cooper does not explicitly teach wherein changing one or more of the parameters and of the architecture is made separately by one or more of neurons and regions.
However, Karam further teaches wherein changing one or more of the parameters and of the architecture is made separately by one or more of neurons and regions (i.e. In the DeepCorrect framework, a task-driven corrective transform acts as a distortion masker for convolutional filters or features of those filters that are most susceptible to input distortion, while leaving all the other pre-trained filter outputs in the layer unchanged; para. [0050, 0055, 0057, 0065]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 7: Cooper and Karam teach the method of claim 1. Cooper does not explicitly teach wherein upon comparison at least one activation differential between an activation of the neural network via the training data of the original training data set and an activation via the respective corresponding training data of the at least one changed training data set is determined, wherein the change is made on the basis of the determined at least one activation differential.
However, Karam further teaches wherein upon comparison at least one activation differential between an activation of the neural network via the training data of the original training data set and an activation via the respective corresponding training data of the at least one changed training data set is determined, wherein the change is made on the basis of the determined at least one activation differential (i.e. For a given layer, let ϕm(u) be the output (activation map) of the mth convolutional filter with kernel weights represented by the tensor (3-dimensional array) Wm for an input u. Let em=ϕm(u+r)−ϕm(u) be the additive noise (perturbation) that is caused in the output activation map ϕm(u) as a result of applying an additive perturbation r to the input u; para. [0031, 0046, 0053]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 8: Cooper and Karam teach the method of claim 7. Cooper does not explicitly teach wherein activation differentials are determined and taken into account by neurons and/or regions.
However, Karam further teaches wherein activation differentials are determined and taken into account by neurons and/or regions (i.e. the aforementioned questions are addressed by first evaluating the effect of image distortions like blur, noise, and adversarial perturbations on the outputs of pre-trained convolutional filters. The present disclosure observes that for every layer of convolutional filters in the DNN, certain filters are far more susceptible to input distortions than others and that correcting the activations of these filters can help recover lost performance; para. [0031, 0046, 0057]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 9: Cooper and Karam teach the method of claim 8. Cooper does not explicitly teach wherein determined activation differentials are in each case taken into account averaged over multiple neurons and/or over a region.
However, Karam further teaches wherein determined activation differentials (i.e. For a given layer, let ϕm(u) be the output (activation map) of the mth convolutional filter with kernel weights represented by the tensor (3-dimensional array) Wm for an input u. Let em=ϕm(u+r)−ϕm(u) be the additive noise (perturbation) that is caused in the output activation map ϕm(u) as a result of applying an additive perturbation r to the input u; para. [0031, 0046, 0053]) are in each case taken into account averaged over multiple neurons and/or over a region (i.e. Let the baseline performance (computed over distorted images) for a network pb, which can be obtained by computing the average top-1 accuracy of the network over a set of images or another task-specific performance measure; para. [0047-0050]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 10: Cooper and Karam teach the method of claim 7. Cooper does not explicitly teach wherein determined activation differentials are taken into account in a weighted manner according to a position of an associated neuron layer within the neural network.
However, Karam further teaches wherein determined activation differentials (i.e. For a given layer, let ϕm(u) be the output (activation map) of the mth convolutional filter with kernel weights represented by the tensor (3-dimensional array) Wm for an input u. Let em=ϕm(u+r)−ϕm(u) be the additive noise (perturbation) that is caused in the output activation map ϕm(u) as a result of applying an additive perturbation r to the input u; para. [0031, 0046, 0053]) are taken into account in a weighted manner according to a position of an associated neuron layer within the neural network (i.e. The main DeepCorrect model for ImageNet shown in FIG. 8b is generated by correcting βiNi ranked distortion susceptible filters in the first 5 layers out of the 8 layered pre-trained AlexNet DNN shown in FIG. 3b and referred to as Deepcorr-5 in Tables VII-VIII, using the notation defined in Section 5.1. Deepcorr-5 is generated by correcting 75% ranked filter outputs in the first two layers (βi, β2=0.75) and 50% ranked filter outputs in the next three layers (β3, β4, β5=0.5). The correction priority computed using Algorithm 1 is used to independently rank the distortion susceptible filters within each layer (local ranking) and correct the activations of the top βiNi filters in each layer; para. [0068]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 11: Cooper and Karam teach the method of claim 7. Cooper does not explicitly teach wherein determined activation differentials are in each case taken into account averaged over multiple inference runs.
However, Karam further teaches wherein determined activation differentials (i.e. For a given layer, let ϕm(u) be the output (activation map) of the mth convolutional filter with kernel weights represented by the tensor (3-dimensional array) Wm for an input u. Let em=ϕm(u+r)−ϕm(u) be the additive noise (perturbation) that is caused in the output activation map ϕm(u) as a result of applying an additive perturbation r to the input u; para. [0031, 0046, 0053]) are in each case taken into account averaged over multiple inference runs (i.e. Let the baseline performance (computed over distorted images) for a network pb, which can be obtained by computing the average top-1 accuracy of the network over a set of images or another task-specific performance measure; para. [0047-0050]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 12: Cooper and Karam teach the method of claim 7. Cooper does not explicitly teach wherein determined activation differentials are in each case taken into account in a weighted manner according to an associated manipulation method.
However, Karam further teaches wherein determined activation differentials (i.e. For a given layer, let ϕm(u) be the output (activation map) of the mth convolutional filter with kernel weights represented by the tensor (3-dimensional array) Wm for an input u. Let em=ϕm(u+r)−ϕm(u) be the additive noise (perturbation) that is caused in the output activation map ϕm(u) as a result of applying an additive perturbation r to the input u; para. [0031, 0046, 0053]) are in each case taken into account in a weighted manner according to an associated manipulation method (i.e. The identified filters most susceptible to Gaussian blur are mainly frequency- and orientation-selective filters, most of which are color agnostic, while filters most susceptible to AWGN were a mix of both color specific blobs and frequency- and orientation-selective filters; para. [0047-0051]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 13: Cooper and Karam teach the method of claim 7. Cooper does not explicitly teach wherein neurons and/or regions of the neural network are sorted according to the activation differentials determined in each case for these, wherein changing is made on the basis of an associated ranking.
However, Karam further teaches wherein neurons and/or regions of the neural network are sorted according to the activation differentials (i.e. For a given layer, let ϕm(u) be the output (activation map) of the mth convolutional filter with kernel weights represented by the tensor (3-dimensional array) Wm for an input u. Let em=ϕm(u+r)−ϕm(u) be the additive noise (perturbation) that is caused in the output activation map ϕm(u) as a result of applying an additive perturbation r to the input u; para. [0031, 0046, 0053]) determined in each case for these, wherein changing is made on the basis of an associated ranking (i.e. Using the ranking measure in Equation (1), correction priorities were computed for every convolutional filter in the network and the filters were ranked in descending order of correction priority. The detailed overview and pseudo-code for computing correction priorities is summarized in Algorithm 1. The correction priority for convolutional filters can be used to rank filters locally (i.e., rank only the filters within a layer) as well as globally (i.e., rank the filters across layers). The impact of both a local ranking scheme and a global ranking scheme on DNN robustness to distortions were evaluated and are described herein; para. [0049, 0068]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 14 is similar in scope to Claim 1 and is rejected under a similar rationale.
Cooper teaches a device for improving robustness of a neural network in a function-specific manner (i.e. By maintaining, saving, storing or using the camera properties of the images to train data models while still adding to the training data with augmented images, the resulting model may be useful across many different devices having the same camera properties. Moreover, the augmentation may provide generalization and greater robustness to the model prediction, particularly when images are clouded, occluded, or otherwise do not provide clear views of the detectable objects; para. [0017]), the device comprising a processor and memory having instructions that are configured to (i.e. a system having one or more processors and non-transitory computer storage media storing instructions that when executed by the one or more processors, cause the processors to perform operations; para. [0008]).
Claim 16 is similar in scope to Claim 1 and is rejected under a similar rationale.
Cooper teaches a non-transitory computer-readable storage medium comprising commands which, when run by a computer, prompt the latter to (i.e. a non-transitory computer-readable medium having instructions for execution by a processor, the instructions when executed by the processor causing the processor to; para. [0009]).
Claims 17, 20, and 21 are similar in scope to Claims 3, 5, 6 and are rejected under a similar rationale.
7. Claims 4 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Cooper in view of Karam, and further in view of Mummadi et al. (U.S. Patent Application Pub. No. US 20180308012 A1).
Claim 4: Cooper and Karam teach the method of claim 2. Cooper does not explicitly teach wherein the termination criterion is a convergence of the determined robustness measure.
However, Karam further teaches wherein the termination criterion is a of the determined robustness measure (i.e. the features can be corrected by appending correction units taking the form of small blocks of stacked convolutional layers such as a residual block with a single skip connection or other CNN-based block or other blocks implementing a transformation, at the output of select filters and train them to correct the worst distortion-affected filter activations using a target-oriented loss or other desired loss function … Starting with an initial learning rate that is ten times lower than the initial learning rate used to generate the pre-trained model, i.e., 0.01, and momentum equal to 0.9, network parameters in various layers are updated for a fixed number of iterations (40 epochs), with the learning rate reduced by a factor of 10 after every 10 epochs; para. [0031, 0059, 0061]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Cooper to include the feature of Karam. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
However, Mummadi teaches wherein the termination criterion is a convergence of the determined robustness measure (i.e. Such a numerical minimization may take place, for example, using an iterative method, the iterative method being terminated when a predefinable abort criterion (for example, after a predefinable number of iteration steps) is reached … Then 2070, it is checked whether a convergence criterion is me; para. [0024, 0027, 0086]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the combination of Cooper and Karam to include the feature of Mummadi. One would have been motivated to make this modification because it improves robustness of DNN-based computer vision models against degraded or changed image inputs.
Claim 18 is similar in scope to Claim 4 and is rejected under a similar rationale.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure.
Yang et al. (Pub. No. US 10083171 B1), the modified convolutional neural networks model is retrained until a desired convergence criterion is met or achieved.
It is noted that any citation to specific pages, columns, lines, or figures in the prior art references and any interpretation of the references should not be considered to be limiting in any way. A reference is relevant for all it contains and may be relied upon for all that it would have reasonably suggested to one having ordinary skill in the art. In re Heck, 699 F.2d 1331, 1332-33, 216 U.S.P.Q. 1038, 1039 (Fed. Cir. 1983) (quoting In re Lemelson, 397 F.2d 1006, 1009, 158 U.S.P.Q. 275, 277 (C.C.P.A. 1968)).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to TAN TRAN whose telephone number is (303)297-4266. The examiner can normally be reached on Monday - Thursday - 8:00 am - 5:00 pm MT.
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, Matt Ell can be reached on 571-270-3264. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/TAN H TRAN/Primary Examiner, Art Unit 2141
Prosecution Insights