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 .
Response to Arguments
The rejection of claims 6 and 13 under 35 U.S.C. 112(b) is withdrawn in view of the cancellation of these claims.
The rejection of claims 1-14 under 35 U.S.C. 101 is withdrawn. Claims 1 and 8 have been amended to recite generating an input image based on data obtained from at least one sensor and controlling an actuator in response to the output. These amendments integrate the abstract idea into a practical application consistent with the analysis of claims 15-20 in the prior action.
Applicant's arguments with respect to the rejection under 35 U.S.C. 103 have been fully considered but moot in view of the newly cited prior art because of the amendments. Applicant argues that Wang discloses a "single backward pass" method that does not generate adversarial gradients from "respective starting points obtained by random initialization of multiple starting points around the input image." Applicant's characterization of Wang is noted. However, the amended limitations are addressed by the addition of Smilkov et al. , "SmoothGrad: removing noise by adding noise," (hereinafter "Smilkov"), which is applied to the rejection below. Smilkov teaches generating multiple noisy variants of an input by sampling from a Gaussian distribution centered on the input (i.e., random initialization of multiple starting points around the input image), computing gradients for each variant, and averaging the resulting gradient maps (i.e., aggregating the plurality of adversarial gradients). The rejection is maintained with the application of Smilkov as set forth below.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
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.
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 1-5, 7-12, and 14-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Perez et al. (Attribution of Predictive Uncertainties in Classification Models – hereinafter “Perez”) in view of Wang et al. (Gradient-based Uncertainty Attribution for Explainable Bayesian Deep Learning – hereinafter “Wang”) in view of Smilkov et al. (SmoothGrad: removing noise by adding noise – hereinafter “Smilkov”) in view of Cui et al. (US 2020/0086862 A1 – hereinafter “Cui”).
Claims 1 and 8.
The combination of Perez, Wang, Smilkov, and Cui discloses a computing device configured to obtain an uncertainty attribution of a prediction of objects in an input image (Perez: "Predictive uncertainties in classification tasks are often a consequence of model inadequacy ... In popular applications, such as image processing, we are often required to scrutinise these uncertainties by meaningfully attributing them to input features … we present a novel framework that combines path integrals, counterfactual explanations and generative models" (Abstract)), the computing device including a processing device configured to execute instructions stored in memory to:
(Cui: "The input may be image capture data from the real-world environment." (¶28); "the image capture sensor may detect or e capture images of the surrounding real-world environment" ¶41); Fig. 1 shows image capture device 140 within system 100.);
generate a prediction of objects in the input image (Perez defines the model f(x) generally as a classifier that outputs probabilities for classes. “fc(x) represents the predicted probability of class-c membership.” (Page 2, § 2));
estimate an uncertainty associated with the prediction of the objects in the input image (Perez: "We readily assign F(x) = H(x) [entropy] .. . and thus combine scores across all classes ... to identify pixels that confuse the model." (Page 3, top left column) Where, using Entropy H(x) is the standard estimation of uncertainty described in the paper.);
calculate an uncertainty attribution that represents regions of the input image that cause the estimated uncertainty (Perez: "regions in red are identified as contributors to predictive uncertainties. We readily comprehend why the model struggles to predict any single class ... " (Fig. 1 & Page 2, top right column). Wang: "Specifically, we introduce a gradient-based uncertainty attribution method to identify the most problematic regions of the input that contribute to the prediction uncertainty ... UA-Backprop." (Abstract)), wherein calculating the uncertainty attribution , (“State-of-the-art tools instead proceed by creating counterfactual or adversarial feature vectors, and assign attributions by direct comparison to original images." (Abstract) "These methods proceed by identifying counterfactual. .. or adversarial. .. explanations, i.e. small variations in the value of input features which output new model scores with minimal uncertainty." (P. 1, right column) This matches the claim requirement of modifications configured to change the uncertainty to a "minimal" level. “Popular resampling or gradient-based methods can easily be adapted in order to attribute measures of uncertainty such as H(x) to input features in an image. This includes tools such … integrated gradients (IG)” (page. 2, bottom left column). Perez calculates integrated gradients along the path from the adversarial counterfactual. This teaches a plurality of gradients (the integration steps) that correspond to the modifications (the path steps) derived from the adversarial example. Smilkov samples n noisy variants x + N(0, σ2)from a Gaussian distribution centered on the input image (random initialization of multiple starting points), Section 3.); and
generate an output indicative of the calculated uncertainty attribution (Smilkov defines the smoothed sensitivity map as
PNG
media_image1.png
82
455
media_image1.png
Greyscale
(Section 2.2); Perez generates "saliency masks, such as SHAP or integrated gradients," and visualizes them: "regions in red are identified as contributors to predictive uncertainties." (Abstract & p. 2, Fig. 1’s explanation)) by
.
Perez discloses all of the subject matter as described above except for specifically teaching “includes generating a plurality of adversarial gradients.” Instead, Perez relies on generative models (VAEs/GANs) to create “counterfactual” references. While these are adversarial in nature, Perez states that standard gradient method adapt poorly to uncertainty. and “an actuator configured to control an operation of the computer-controlled machine in response to the output signal.” However, Wang in the same field of endeavor teaches “generating a plurality of adversarial gradients” as a uncertainty attribution (UA) backdrop that computes the gradient of uncertainty with respect to the input pixels, without needing a complex generative model (“Predictions made by deep learning models are prone to data perturbations, adversarial attacks, and out-of-distribution inputs.” (Abstract); “Specifically, we introduce a gradient-based uncertainty attribution method to identify the most problematic regions of the input that contribute to the prediction uncertainty … UA-Backdrop.” (Abstract); “identify uncertainty sources and take actions to mitigate their effects on predictions.” (Abstract)).
Therefore, it would have been obvious to one of ordinary skill in the art to combine Perez and Wang before the effective filing date of the claimed invention. Perez teaches attributing uncertainty using ‘counterfactual’ reference generated by complex generative models (e.g. VAEs). Wang teaches a ‘gradient-based uncertainty attribution method’ that identifies uncertainty scores using gradients corresponding to adversarial perturbations. The motivation for this combination of references would have been to substitute Perez’s generative model approach with the gradient-based attribution of Wang to reduce computational complexity by eliminating the need to train a separate generative model while still effectively identifying the input regions causing uncertainty.
Perez and Wang disclose the prediction, uncertainty estimation, uncertainty attribution, and saliency-mask output described above, but do not explicitly disclose: (i) generating adversarial gradients "from respective starting points obtained by random initialization of multiple starting points around the input image" and "aggregating the plurality of adversarial gradients as calculated from the respective starting points"; or (ii) generating the input image based on data from a sensor and controlling an actuator in response to the output. Smilkov, in the same field of endeavor, teaches the multi-sample averaging limitation (Smilkov: "The core idea is to take an image of interest, sample similar images by adding noise to the image, then take the average of the resulting sensitivity maps for each sampled image." (Abstract); equation
PNG
media_image1.png
82
455
media_image1.png
Greyscale
(Section 2.2)). Cui, in the same field of endeavor (autonomous-vehicle Bayesian deep-learning policies), teaches the sensor-input and actuator-output limitations (Cui: image capture device 140 generating an input image (¶24, ¶41); controller 110 executing actions on vehicle actuators in response to the uncertainty array U[T] (Abstract; Fig. 1; Fig. 3; ¶¶29-30).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine Perez, Wang, Smilkov, and Cui. The motivation for the combination would have been: (a) to substitute Wang's gradient-based attribution for Perez's generative-model approach so as to reduce computational overhead while preserving attribution accuracy; (b) to apply Smilkov's multi-sample gradient averaging to mitigate gradient noise and produce sharper, more reliable uncertainty-attribution maps, per Smilkov's stated motivation that gradient-based sensitivity maps may "fluctuate sharply at small scales" (Section 2.2); and (c) to deploy the resulting uncertainty-attribution pipeline in Cui's autonomous-vehicle architecture so that the vehicle's actuators (e.g., steering, brake) execute safer and more targeted control actions informed by specific sources of prediction uncertainty.
Claims 2 and 9.
The combination of Perez, Wang, Smilkov, and Cui discloses the computing device of claim 8, wherein, to generate the plurality of adversarial gradients, the processing device is configured to execute instructions to at least one of (i) introduce noise into the input image and (ii) modify selected pixels of the input image (Perez discloses "creating counterfactual or adversarial feature vectors" (Abstract) which involves "small variations in the value of input features" (Page 1) i.e., pixel modification/noise. Wang discusses “data perturbations” and “adversarial attacks” (Abstract) which inherently involve modifying pixels.).
Claims 3 and 10.
The combination of Perez, Wang, Smilkov, and Cui discloses the computing device of claim 8, wherein, to generate the plurality of adversarial gradients, the processing device is configured to execute instructions to identify at least one adversarial gradient that causes the estimated uncertainty to decrease below a threshold (Perez explicitly searches for a reference point that "bears no predictive uncertainty" or has "minimal uncertainty." This effectively means decreasing the uncertainty below a threshold (to near zero).).
Claims 4 and 11.
The combination of Perez, Wang, Smilkov, and Cui discloses the computing device of claim 8, wherein, to generate the plurality of adversarial gradients, the processing device is configured to execute instructions to identify a one of the adversarial gradients that causes the estimated uncertainty to decrease below a threshold (Perez defines the goal of the adversarial search as finding a point with minimal uncertainty. “These methods proceed by identifying counterfactual (in-distribution) or adversarial (out-of-distribution) explanations, i.e. small variations in the value of input features which output new model scores with minimal uncertainty” (Page 1, right column). Minimal serves as the threshold (effectively meaning zero). ) with a least amount of modification of the input image (Perez discloses these as “small variations in the value of input features.” Where, “counterfactual” in the art is defined as the closest possible world “least modification” that change the outcome.).
Claims 5 and 12:
The combination of Perez, Wang, Smilkov, and Cui discloses the computing device of claim 8, wherein, to calculate the uncertainty attribution, the processing device is configured to execute instructions to identify an uncertainty attribution for a selected class of objects (Perez teaches attributing uncertainty to features that cause confusion regarding specific predictions (classes). “Predictive uncertainties in classification tasks are often a consequence of model inadequacy … In popular applications, such as image processing, we are often required to scrutinise these uncertainties by meaningfully attributing them to input features.” (Abstract)).
Claims 7 and 14:
The combination of Perez, Wang, Smilkov, and Cui discloses the computing device of claim 8, wherein the processing device is configured to execute instructions to modify the input image based on the calculated uncertainty attribution and generate a second prediction of the objects in the modified input image, and wherein, to modify the input image, the processing device is configured to execute instructions to mask portions of the image corresponding to the calculated uncertainty attribution (Perez Abstract discloses “saliency masks.” In the field of XAI (Explainable AI), the standard way to use or evaluate a “saliency mask” is the “deletion metric” or “masking test.” These evaluations mask the pixels highlighted in the saliency mask and generate a second prediction to see if the model’s uncertainty changes.).
Claim 15.
The combination of Perez, Wang, Smilkov, and Cui renders claim(s) 15 obvious for the reasons discussed above for claims 1 and 8, mutatis mutandis.
Claim 16.
The combination of Perez, Wang, Smilkov, and Cui renders claim(s) 16 obvious for the reasons discussed above for claims 2 and 9, mutatis mutandis.
Claim 17.
The combination of Perez, Wang, Smilkov, and Cui renders claim(s) 17 obvious for the reasons discussed above for claims 3 and 10, mutatis mutandis.
Claims 18.
The combination of Perez, Wang, Smilkov, and Cui renders claim(s) 18 obvious for the reasons discussed above for claims 4 and 11, mutatis mutandis.
Claim 19.
The combination of Perez, Wang, Smilkov, and Cui renders claim(s) 19 obvious for the reasons discussed above for claims 7 and 17, mutatis mutandis.
Claim 20.
The combination of Perez, Wang, Smilkov, and Cui discloses the computer-controlled machine of claim 15, wherein the computer-controlled machine includes an autonomous robot (Cui “end-to-end robotic systems” ¶2).
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Ross Varndell whose telephone number is (571)270-1922. The examiner can normally be reached M-F, 9-5 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, O’Neal Mistry can be reached at (313)446-4912. 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 https://ppair-my.uspto.gov/pair/PrivatePair. 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.
/Ross Varndell/Primary Examiner, Art Unit 2674
Prosecution Insights