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 .
Remarks - 35 USC 101
The claims appear to be statutory under 101. With respect to the judicial exception –
The claimed invention is directed to specific training objective that blends Bayesian learning with robustness and uncertainty awareness. The claims require loss optimization based on Bayesian fidelity and robustness awareness and enforce both simultaneously during training including on perturbed inputs. The training data is intentionally generated to be modified (perturb) to add noise, distortions, adversarial. The uncertainly is computed on the perturbed training data as an explicit training term combined with a variational posterior objectives that simultaneously enforces posterior correctness and encourages high uncertainty on perturbed inputs by adding an uncertainty term computed from perturbed data into the training objective.
Therefore, the claims are determined to be significantly more than the abstract idea and thus, statutory under 35 USC 101.
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.
Claims 22, 29, 32-33, 36, 38-39 is/are rejected under 35 U.S.C. 103 as being unpatentable over Liu et al. “ADV-BNN: IMPROVED ADVERSARIAL DEFENSE THROUGH ROBUST BAYESIAN NEURAL NETWORK” in view of Kendall et al. “What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?”.
Regarding claim 22, Liu teaches a method for training a deep neural network (DNN) capable of adversarial detection, wherein the DNN is configured with a plurality of sets of weights candidates (Abstract – see adversarial-trained Bayesian neural network, p.2 Notations), the method comprising the following steps:
inputting training data selected from a training data set to the DNN (p.2 Notations);
calculating, based on the training data, a first term for indicating a difference between a variational posterior probability distribution and a true posterior probability distribution of the DNN (p.4 ¶ 3. METHOD see Eq.(6)(7), where q(w) variational posterior, p(w) true posterior, where
PNG
media_image1.png
19
120
media_image1.png
Greyscale
is a first term and preset a difference between posteriors);
perturbing the training data to generate perturbed training data (p.2 see Notations - adversarial example is generated by
PNG
media_image2.png
15
83
media_image2.png
Greyscale
and specifically Eq.(4));
calculating a second term for indicating a quantification of predictive robustness NOTE); and
updating the plurality of sets of weights candidates of the DNN based on augmenting a summation of the first term and the second term (p.4-5 ¶3. METHOD see Eq.(7)-(10)).
NOTE Liu teaches computing prediction over sampled weight on perturbed inputs Eq.(4), wherein the uncertainty derives from . Thus, it is reasonable to conclude that the predictive uncertainty is represented through the posterior predictive expectation over the perturbed sample
PNG
media_image2.png
15
83
media_image2.png
Greyscale
()i.e. variability across samples implies uncertainly
Still, Liu does not explicitly teach, however Kendall discloses calculating a second term for indicating a quantification of predictive uncertainty (p.5 Eq(5) )on the perturbed training data and updating the plurality of sets of weights candidates of the DNN based on augmenting a summation of the first term and the second term (i.e. predictive uncertainty) (p.4 Eq (3)-(6), (10-11)).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Liu to include calculating a second term for indicating a quantification of predictive uncertainty on the perturbed training data as disclosed by Kendall. Doing so leads to new loss functions for deep learning models and makes the loss more robust to noisy data (Kendall Abstract).
Regarding claim 29, Liu as modified teaches the method of claim 21, wherein the summation of the first term and the second term is augmented by performing stochastic gradient ascent (Liu p.4-5 ¶3. METHOD).
Regarding claim 32, Liu as modified teaches the method of claim 21, wherein the first term is an evidence lower bound (ELBO) of the variational posterior probability distribution (Liu p.4-5 ¶3. METHOD).
Regarding claim 33, Liu as modified teaches the method of claim 21, wherein the first term is a log-likelihood function on the training data, wherein each instance of the training data corresponds to one stochastically assigned set of weights candidate of the plurality of sets of weights candidates (Liu p.4-5 ¶3. METHOD).
Regarding claim 36, Liu as modified teaches the method of claim 21, wherein the second term is added to the first term by a tradeoff coefficient (Kendall p.4. Eq(5), Liu p.4-5 ¶3. METHOD).
Regarding claim 38, Liu teaches a computer system, comprising: one or more processors; and one or more non-transitory storage devices storing computer-executable instructions for training a deep neural network (DNN) capable of adversarial detection, wherein the DNN is configured with a plurality of sets of weights candidates, the instructions, when executed by the one or more processors, causing the one or more processors to perform the following steps: inputting training data selected from a training data set to the DNN; calculating, based on the training data, a first term for indicating a difference between a variational posterior probability distribution and a true posterior probability distribution of the DNN; perturbing the training data to generate perturbed training data; calculating a second term for indicating a quantification of predictive uncertainty on the perturbed training data; and updating the plurality of sets of weights candidates of the DNN based on augmenting a summation of the first term and the second term.
Claim 38 recites substantially the same limitations as claim 1, and is rejected for substantially the same reasons.
Regarding claim 39, Liu teaches one or more non-transitory computer readable storage media on which are stored computer-executable instructions for training a deep neural network (DNN) capable of adversarial detection, wherein the DNN is configured with a plurality of sets of weights candidates, the instructions, when executed by one or more processors, causing the one or more processors to perform the following steps: inputting training data selected from a training data set to the DNN; calculating, based on the training data, a first term for indicating a difference between a variational posterior probability distribution and a true posterior probability distribution of the DNN; perturbing the training data to generate perturbed training data; calculating a second term for indicating a quantification of predictive uncertainty on the perturbed training data; and updating the plurality of sets of weights candidates of the DNN based on augmenting a summation of the first term and the second term.
Claim 39 recites substantially the same limitations as claim 1, and is rejected for substantially the same reasons.
Claims 22-28, 30, 34 is/are rejected under 35 U.S.C. 103 as being unpatentable over Liu as modified and in further view of Gal et al. “Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning”, hereafter Gal which refers to Gal et al “Dropout as a Bayesian Approximation: Appendix”, hereafter Appendix.
Regarding claim 22, Liu as modified does not explicitly teach, however Gal/ Appendix discloses the method of claim 21, wherein the plurality of sets of weights candidates includes a first subset of weights and a plurality of second subsets of weights candidates, each set of the plurality of sets of weights candidates includes the first subset of weights and one second subset of the plurality of second subsets of weights candidates (Appendix ¶2.1).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Liu to include plurality of first, second subsets of weights candidates as disclosed by Gal/ Appendix. Doing so would represent and mitigate uncertainty in deep learning without sacrificing either computational complexity or test accuracy (Gal/ Appendix Abstract).
Regarding claim 23, Liu as modified teaches the method of claim 22, wherein the first subset of weights is updated with the training data set (Appendix p. 3 2.2, Liu p.4-5 ¶3. METHOD, Kendall Eq.5).
Regarding claim 24, Liu as modified teaches the method of claim 22, wherein each second subset of the plurality of second subsets of weights is updated with training data stochastically selected from the training data set (Liu p.4-5 ¶3. METHOD, Appendix p.19 ¶E.1-E.2).
Regarding claim 25, Liu as modified teaches the method of claim 22, wherein each second subset of the plurality of second subset of weights candidates corresponds to at least one layer of the DNN (Gal ¶3, Appendix p.11 ¶4.5, p.13 ¶5.2).
Regarding claim 26, Liu as modified teaches the method of claim 25, wherein the at least one layer of the DNN includes a last layer of the DNN (Gal ¶3, Appendix p.11 ¶4.5, p.13 ¶5.2).
Regarding claim 27, Liu as modified teaches the method of claim 25, wherein the at least one layer of the DNN includes a plurality of successive layers of the DNN (Gal ¶3, Appendix p.11 ¶4.5).
Regarding claim 28, Liu as modified teaches the method of claim 22, wherein the DNN is pre-trained, and wherein the first subset of weights and the plurality of second subsets of weights candidates are initialized with the pre-trained weights of the DNN (Gal ¶5.1, Appendix p.3 ¶2.1, p.19 ¶E1).
Regarding claim 30, Liu as modified teaches the method of claim 29, wherein the updating of the plurality of sets of weights candidates of the DNN based on augmenting the summation of the first term and the second term, further includes: updating the first subset of weights by a first optimizer with a first weight decay coefficient; and updating the plurality of second subsets of weights candidates with a second optimizer with a second weight decay coefficient (Kendall p.4 ¶2.2, p.5 Eq.(7), Liu p.4-5 ¶3. METHOD Algorithm 1, Gal ¶4. Eq (7), Appendix p.8 ¶3.4).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Liu to include plurality of first, second subsets of weights candidates as disclosed by Gal/ Appendix. Doing so would represent and mitigate uncertainty in deep learning without sacrificing either computational complexity or test accuracy (Gal/ Appendix Abstract).
Regarding claim 34, Liu as modified teaches the method of claim 22, wherein the predictive uncertainty on an instance is calculated as a scalar indicating variance of hidden features of the instance in at least one layer of the DNN corresponding to the second subsets of weights candidates (Liu p.4-5 ¶3. METHOD, Gal ¶5.1, Appendix p.2 ¶2.1 p.18 ¶D, p.19 ¶E2).
Claim 23 is/are rejected under 35 U.S.C. 103 as being unpatentable over Liu as modified and in further view of Finn et al. “Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks”.
Regarding claim 23, if Liu as modified does not explicitly teach, however Finn discloses the method of claim 22, wherein the first subset of weights is updated with the training data set (2.2 Eq(1)(2)).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Liu to include the first subset of weights is updated with the training data set as disclosed by Finn. Doing so provides good weight initialization of deep networks (Finn ¶4).
Claim 31 is/are rejected under 35 U.S.C. 103 as being unpatentable over Liu as modified and in further view of ZHANG et al. (US 20210012188).
Regarding claim 31, Liu as modified does not explicitly teach, however ZHANG discloses the method of claim 21, wherein the training data are perturbed uniformly, and the perturbation is within a training perturbation budget ([0057], [0064]-[0065], [0080]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Liu to include training perturbation budget as disclosed by ZHANG. Doing so would trains image model against adversarial attacks and improving model robustness (ZHANG [0004]).
Claims 35, 37 is/are rejected under 35 U.S.C. 103 as being unpatentable over Liu et al. “ADV-BNN: IMPROVED ADVERSARIAL DEFENSE THROUGH ROBUST BAYESIAN NEURAL NETWORK” in view of Kendall et al. “What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?” and in further view of the applicant’s admitted prior art Lakshminarayanan et al. “Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles”, hereafter D2.
Regarding claim 35, Liu as modified does not explicitly teach, however D2 discloses the method of claim 21, wherein the second term is a regularization term of the predictive uncertainty on the perturbed training data compared with a tunable threshold (p.3 ¶2.1, p.7-8 ¶3.5, where “we observe that MC-dropout produces over-confident predictions on unseen examples, whereas our method produces higher uncertainty on unseen classes” is a tunable threshold).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Liu to include regularization term of the predictive uncertainty on the perturbed training data compared with a tunable threshold as disclosed by D2. Doing so produces well-calibrated uncertainty estimates which are as good or better than approximate Bayesian NNs (D2 Abstract).
Regarding claim 37, Liu teaches a method for using a deep neural network trained for adversarial detection (Abstract – see adversarial-trained Bayesian neural network, p.2 Notations), the comprising:
feeding an input to the DNN (p.6 Algorithm 1);
generating one or more task-dependent predictions of the input (p.9 Eq (18), F4);
wherein the DNN is configured with a plurality of sets of weights candidates, and the DNN is trained by:
inputting training data selected from a training data set to the DNN (p.2 Notations);
calculating, based on the training data, a first term for indicating a difference between a variational posterior probability distribution and a true posterior probability distribution of the DNN (p.4 ¶ 3. METHOD see Eq.(6)(7), where q(w) variational posterior, p(w) true posterior, where
PNG
media_image1.png
19
120
media_image1.png
Greyscale
is a first term and preset a difference between posteriors);
perturbing the training data to generate perturbed training data (p.2 see Notations - adversarial example is generated by
PNG
media_image2.png
15
83
media_image2.png
Greyscale
and specifically Eq.(4));
calculating a second term for indicating a quantification of predictive robustness NOTE); and
updating the plurality of sets of weights candidates of the DNN based on augmenting a summation of the first term and the second term (p.4-5 ¶3. METHOD see Eq.(7)-(10)).
Liu does not explicitly teach, however D2 discloses estimating a predictive uncertainty of the one or more task-dependent predictions concurrently; and determining whether to accept the one or more task-dependent predictions based on the predictive uncertainty (p.3 2.2.1, p.4 2.4 Algorithm 1).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Liu to include estimating a predictive uncertainty of the one or more task-dependent predictions concurrently and determining whether to accept the one or more task-dependent predictions based on the predictive uncertainty as disclosed by D2. Doing so produces well-calibrated uncertainty estimates which are as good or better than approximate Bayesian NNs (D2 Abstract).
NOTE Liu teaches computing prediction over sampled weight on perturbed inputs Eq.(4), wherein the uncertainty derives from . Thus, it is reasonable to conclude that the predictive uncertainty is represented through the posterior predictive expectation over the perturbed sample
PNG
media_image2.png
15
83
media_image2.png
Greyscale
()i.e. variability across samples implies uncertainly
Still, Liu does not explicitly teach, however Kendall discloses calculating a second term for indicating a quantification of predictive uncertainty (p.5 Eq(5) )on the perturbed training data and updating the plurality of sets of weights candidates of the DNN based on augmenting a summation of the first term and the second term (i.e. predictive uncertainty) (p.4 Eq (3)-(6), (10-11)).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Liu to include calculating a second term for indicating a quantification of predictive uncertainty on the perturbed training data as disclosed by Kendall. Doing so leads to new loss functions for deep learning models and makes the loss more robust to noisy data (Kendall Abstract).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure is indicated on PTO-892.
Specifically note -
Szegedy et a. (US 10521718) which teaches training neural networks by generating perturbed training data.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to POLINA G PEACH whose telephone number is (571)270-7646. The examiner can normally be reached Monday-Friday, 9:30 - 5:30.
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, Aleksandr Kerzhner can be reached at 571-270-1760. 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.
/POLINA G PEACH/Primary Examiner, Art Unit 2165 April 22, 2026
Prosecution Insights