DETAILED ACTION
This non-final rejection is response to the claims filed in the RCE submission filed on 4/27/2026. Claims 1-20 are pending. Claims 1, 8, and 15 are independent claims.
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 .
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 1, 2, 7, 8, 9, 14, 15, and 16 remain rejected under 35 U.S.C. 103 as being unpatentable over Nguyen et al., “SELF: Learning to Filter Noisy Labels with Self-Ensembling” (hereinafter Nguyen) in view of Moazzami et al., (US20200042903A1) (hereinafter Moazzami) further in view of Wang et al., “A Recursive Ensemble Learning Approach With Noisy Labels or Unlabeled Data” (herein after Wang).
Regarding claim 1 Nguyen teaches:
A computer-implemented method for training a machine learning model using noisily labeled data, the method comprising:
building a classification model that receives a classified dataset for which classes are pre-labeled as input, wherein said classified dataset comprises label noise, wherein said classification model generates a prediction of class probabilities; (Nguyen: “In the beginning, we assume that the labels of the training set are noisy. The model starts in iteration 0 with training from the noisy label set” – Section 2.1, Fig. 2 caption, p. 2-3; “the labels are only used for
supervised training if … the model predicts the respective label to be the correct class with the highest likelihood” – Section 2.2, p. 4) (Note: This discloses a single CNN (“student”) trained on noisy-labeled data and outputting per-class likelihoods.)
building a second model with a same architecture as said classification model, wherein said second model is a moving average of said classification model, wherein said second model generates a prediction of class probabilities; (Nguyen: “Maintains a running average of model snapshots (realized by the Mean-Teacher model …) to provide a stable learning signal.” – Section 2.1, Fig. 2 caption, p. 3; “A natural way to form a model ensemble is by using an exponential running average of model snapshots (Fig. 3a).” – Section 2.3, p. 4) (Note: This discloses Mean-Teacher is an exponential moving average copy of the student network, sharing the architecture and producing class-probability outputs.)
Nguyen does not teach but Moazzami teaches:
generating weight factors used to weight said predictions of class probabilities of said classification model and said second model by an artificial neural network; (Moazzami: “This process can be performed over all of the base models 308 a-308 n to identify the goodness factors for all of the base model 308 a-308 n. Later, when processing unlabeled real-use data or other unlabeled data, the meta learner 304 can use the goodness factors to weight the intermediate outputs generated using the base models 308 a-308 n.” – Paragraph [0061, 60], p.16; “Example types of learners that can be used here include… and multilayer neural network learners or other deep learning learners.” – Paragraph [0056], p.15) (Note: The goodness factors Gₖ are identified and applied by meta learner 304, which Moazzami discloses can be a multilayer neural network learner (i.e., an artificial neural network) acting on each base model’s probability vector Pₖ(y | x).)
obtaining a prediction of class probabilities using said weighted predictions by said artificial neural network; and (Moazzami: “the fusion model 310 can combine the intermediate outputs 404a-404n in the following manner: P (y | x) = Σk=1 n(Gₖ × Pₖ(y | x)) – Eq. (10) … Here, P (y | x) represents a final output value… the sum of the values can be normalized to produce a final output 406 that is normalized to some scale, such as a value between zero and one (inclusive).” – Paragraph [0078], Eq. (10), p.18) (Note: Equation (10) shows the ANN forming P(y | x) by summing the goodness-factor-weighted vectors Gₖ × Pₖ(y | x). Further, the reference states that the result is normalized to the 0-to-1 range. Thus the ANN obtains a new class-probability distribution from the weighted predictions.)
combining said predictions of class probabilities of said artificial neural network and said classification model to train said machine learning model. (Moazzami: “The final outputs 406 can then be used in any suitable manner, such as to retrain the artificial-intelligence system using the now-labeled real-use data.” – Paragraph [0064], p.16) (Note: The “final outputs 406” are the ANN-weighted probabilities fused with each base model’s probabilities. Moazzami teaches feeding this combined result back to retrain the target model.)
Nguyen and Moazzami are analogous art to the present invention because they both address improving classification reliability when model outputs or labels are noisy, focusing on generating and refining class probability vectors to guide retraining. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Nguyen (for employing a student network together with an exponential
moving average “Mean Teacher” copy to stabilize predictions on noisy label data) and Moazzami (for identifying numeric “goodness” weights for each base model and weighting their outputs before feeding the fused distribution back into training) in order to obtain a single, dynamically weighted probability vector that enhances Nguyen’s label filtering loop. One of ordinary skill in the art would have been motivated to make such a combination because Moazzami states that its meta learner “can use the goodness factors to weight the intermediate outputs” and that the resulting fused output “can then be used … to retrain the artificial intelligence system” (Moazzami, Paragraph [0061], [0064]).
Nguyen and Moazzami do not state that the intermediate outputs are class probabilities, but Wang teaches this limitation:
… predictions of class probabilities (Wang: “The outputs of the networks are the probability value that each sample belongs to each of the possible classes.” – Section III-B, p.5) (Note: Wang describes a network’s output as “the probability value that each sample belongs to each of the possible classes.” In light of that teaching, those skilled in the art would read Moazzami’s notation Pₖ(y | x) as a class-probability distribution.)
Wang is analogous art to the present invention because it addresses how each network’s output is a “probability value that each sample belongs to each of the possible classes,” thereby confirming that ensemble outputs are class-probability vectors. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Nguyen + Moazzami (which provide a student/Mean-Teacher architecture and a goodness-weighted fusion vector) with Wang (which defines network outputs
as per-class probabilities) in order to treat Moazzami’s Pₖ(y | x) and Nguyen’s student/teacher outputs as probability distributions and apply the goodness-factor weighting directly to those probabilities, yielding a properly normalized fused distribution. One of ordinary skill in the art would have been motivated to make such a combination because Wang states that “the outputs of the networks are the probability value that each sample belongs to each of the possible classes.” (Wang, Section III-B, p.5).
Regarding claim 2, Nguyen in view of Moazzami further in view of Wang teaches the method as recited in claim 1.
Nguyen further teaches: further comprising: updating parameters of said second model as said moving average of said classification model. (Nguyen: “Concretely, in each iteration, the model … maintains a running average of model snapshots (realized by the Mean-Teacher model …).” – Section 2.1, p.3; “A natural way to form a model ensemble is by using an exponential running average of model snapshots (Fig. 3a).” – § 2.3, p.4; “Update the weights of Mit as an exponential moving-average of the student weights.” – Section Appendix A.1, p. 11) (Note: The Mean-Teacher network Mit is the claimed “second model.” The Appendix line states that its parameters are updated each step by an exponential moving average (EMA) of the student model’s weights; Section 2.1 and 2.3 describe the same process conceptually. Together these passages show that the second model’s parameters are continuously updated as a moving average of the classification model’s parameters.)
Regarding claim 7, Nguyen in view of Moazzami further in view of Wang teaches the method as recited in claim 1.
Nguyen further teaches: wherein said machine learning model is a deep learning. (Nguyen: “Deep neural networks (DNNs) have been shown to over-fit a dataset when being trained with noisy labels for a long enough time.” – Abstract, p.1) (Note: Nguyen’s disclosure identifies the very model being trained as a deep neural network, which is typical form of a deep learning model.). Nguyen fails to teach model applied to one of the following fields: computer vision, speech recognition, natural language processing, machine translation, bioinformatics, drug design, medical image analysis, material inspection and board game programs. However, Moazzami teaches a machine learning system applied to electronic devices, such as smart glasses, medical devices, etc. (15). It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to combine the references, because Moazzami teaches saving time and money by labeling data from unlabeled data to train AI systems (31).
Claim 8 is a computer-program-product claim that corresponds directly to method claim 1 and is rejected for the same reasons given in the rejection of claim 1.
Claim 9 is a computer-program-product claim that corresponds directly to method claim 2 and is rejected for the same reasons given in the rejection of claim 2.
Claim 14 is a computer-program-product claim that corresponds directly to method claim 7 and is rejected for the same reasons given in the rejection of claim 7.
Claim 15 is a system claim that corresponds directly to method claim 1 and is rejected for the same reasons given in the rejection of claim 1.
Claim 16 is a system claim that corresponds directly to method claim 2 and is rejected for the same reasons given in the rejection of claim 2.
Claims 3, 10 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Nguyen in view of Moazzami further in view of Wang further in view of Mayer--US 20210287089.
Regarding claim 3, Nguyen in view of Moazzami further in view of Wang teaches the method as recited in claim 1.
Nguyen and Moazzami do not teach but Wang further teaches:
further comprising:
combining said predictions of class probabilities of said artificial neural network and said classification model as a log sum of corresponding class probabilities… (Wang: “The outputs of the networks are the probability value that each sample belongs to each of the possible classes.” – Section III-B, p. 5; “the discriminant function can be reduced to (see Eq. (5))” – Section III-C, p.6) (Note: Wang Eq. (5) the indicator hji = 1 selects the relevant class entries, making each term log (pi) up to the constant offset –log (1 – pi). Because the constant does not affect optimization or relative weighting, the expression is still a sum of log-probability contributions from the participating models. Thus, read under the broadest reasonable interpretation, Wang teach “log sum of corresponding class probabilities”.)
Nguyen, Moazzami, and Wang do not teach but Mayer teaches using a residual connection 460 to bypass non-linear layers, such as a prelu layer, which can then be added to the nonlinear output of a hidden layer 420 (111-112, fig.4)-- using a residual connection configured to allow gradients to flow through said artificial neural network directly without passing through non-linear activation functions.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Nguyen + Moazzami + Wang (which already provide two class-probability vectors, student/Mean-Teacher and goodness-weighted fusion and show that ensemble decisions can be expressed as a log-sum of class probabilities) with Mayer. One of ordinary skill in the art would have been motivated to make such a combination because Mayer discloses “reduce or avoid information loss that can result from the use of hidden layers.” (111).
Claim 10 is a computer-program-product claim that corresponds directly to method claim 3 and is rejected for the same reasons given in the rejection of claim 3.
Claim 17 is a system claim that corresponds directly to method claim 3 and is rejected for the same reasons given in the rejection of claim 3.
Claims 4, 11 and 18 remain rejected under 35 U.S.C. 103 as being unpatentable over Nguyen in view of Moazzami further in view of Wang further in view of Hall et al., (US20220344049A1) (hereinafter Hall).
Regarding claim 4, Nguyen in view of Moazzami further in view of Wang teaches the method as recited in claim 1.
Nguyen, Moazzami, and Wang do not teach but Hall teaches:
wherein a categorical cross entropy loss corresponds to a distance between ground truth labels and said combined predictions of class probabilities of said artificial neural network and said classification model. (Moazzami: “P(y | x)=Σk=1n (Gₖ × Pₖ(y | x))” – Paragraph [0078] Eq. (10); “The final outputs 406 can then be used in any suitable manner, such as to retrain the artificial-intelligence system using the now-labeled real-use data.” – Paragraph [0064]; Hall: “We first select a standard cross entropy loss function used for training a machine learning model which we will refer to generically as “CrossEntropyLoss.” … If x represents a batch of training data to be minimized, y is the target (ground-truth values) associated with each element of the batch x, and S(x) and T(x) are the distributions obtained from the Student and Teacher models, respectively, then we define the distillation loss function as: Loss(x, y) = CrossEntropyLoss(S(x), y) + D(S(x), T(x)).” – Paragraph [0143] & Eq. 1) (Note: Moazzami supplies the single fused probability vector produced by the meta learner (Eq. 10) and states that this fused output is fed back to retrain the system. Hall then gives a definition of a CrossEntropyLoss (prediction, ground-truth ) that is applied to a probability distribution such as Moazzami’s fused vector. Taken together,
these passages teach the claimed use of categorical cross-entropy as the “distance” between ground-truth labels and the combined ANN + classification model probabilities.)
Hall is analogous art to the present invention because it teaches using a standard CrossEntropyLoss between a model’s predicted probability distribution and the ground-truth labels in a student/teacher training scheme. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Nguyen + Moazzami + Wang (which together yield a fused class-probability vector from the student, Mean-Teacher, and goodness-factor weighting and confirm that these outputs are probability distributions) with Hall, which applies CrossEntropyLoss(prediction, ground truth), in order to measure the distance between that fused distribution and the true labels and thereby drive more reliable retraining of the student network. One of ordinary skill in the art would have been motivated to make such a combination because Hall states: “We first select a standard cross entropy loss function … Loss(x, y) = CrossEntropyLoss(S(x), y) + D(S(x), T(x))” (Hall, Paragraph [0143]).
Claim 11 is a computer-program-product claim that corresponds directly to method claim 4 and is rejected for the same reasons given in the rejection of claim 4.
Claim 18 is a system claim that corresponds directly to method claim 4 and is rejected for the same reasons given in the rejection of claim 4.
Claims 5, 12 and 19 remain rejected under 35 U.S.C. 103 as being unpatentable over Nguyen in view of Moazzami further in view of Wang further in view of Masataki et al., (US11081105B2) (hereinafter Masataki).
Regarding claim 5, Nguyen in view of Moazzami further in view of Wang teaches the method as recited in claim 1.
Nguyen, Moazzami, and Wang do not teach but Masataki teaches:
wherein a categorical cross entropy loss corresponds to a distance between said prediction of class probabilities of said second model and said combined predictions of class probabilities of said artificial neural network and said classification model. (Nguyen: “Maintaining a running average of model snapshots (realized by the Mean Teacher) …” – Section 2.1; (Note: The Mean Teacher is the claimed second model producing its own probability distribution.)) (Moazzami: “P (y | x) = Σk=1 n(Gₖ × Pₖ(y | x))” – Eq. (10) (Note: Defines the combined probability vector output 406.))
(Masataki: “calculates a cross entropy between the first output probability distribution and the second output probability distribution” – col. 4, lines 65-67) (Note: Nguyen teach the teacher/EMA probability distribution. Moazzami teach the single fused (ANN + base model) probability vector. Masataki teach computing a cross entropy between two such probability distributions to update model weights.)
Masataki is analogous art to the present invention because it teaches computing a cross-entropy loss between two probability distributions output by a teacher model and a student model. It would have been obvious to one of ordinary skill in the art before the effective filing
date of the claimed invention to combine the teachings of Nguyen + Moazzami + Wang (which together provide two probability vectors, the Mean-Teacher distribution and the goodness-weighted fusion vector, and confirm that these vectors are class-probability distributions) with Masataki (which applies cross-entropy directly between a first and a second output distribution) so that the distance between the Mean-Teacher distribution and the fused distribution is measured via categorical cross-entropy, thereby guiding further training of the student network. One of ordinary skill would have been motivated to make such a combination because Masataki teaches “calculat[ing] a cross entropy between the first output probability distribution and the second output probability distribution.” (Masataki, col. 4, lines 65-67).
Claim 12 is a computer-program-product claim that corresponds directly to method claim 5 and is rejected for the same reasons given in the rejection of claim 5.
Claim 19 is a system claim that corresponds directly to method claim 5 and is rejected for the same reasons given in the rejection of claim 5.
Claims 6, 13 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Nguyen in view of Moazzami further in view of Wang further in view of Sawada et al., “Trade-offs in Top-k Classification Accuracies on Losses for Deep Learning” (hereinafter Sawada).
Regarding claim 6, Nguyen in view of Moazzami further in view of Wang teaches the method as recited in claim 1. Nguyen, Moazzami, and Wang do not teach but Sawada teaches:
wherein a categorical cross entropy loss corresponds to a sum of a top-k probability weighted log probabilities, where k is a positive integer number. (Sawada: “The extended loss, which we call top-k grouping loss Lₖ, is defined as CE loss calculated for the top-k union and the remaining classes as below, see Eq. (5).” – Section III-C, p.3) (Note: Sawada’s top-k grouping loss is presented as an extension of categorical cross-entropy that operates only on the k highest-probability classes. In Eq. (5) the first term groups those k classes and applies the log to their aggregated probability mass. The loss therefore behaves as if one summed the log-probabilities of the k winners weighted by the target label, so under the broadest reasonable interpretation, Sawada teaches a categorical cross-entropy that “corresponds to a sum of top-k probability weighted log probabilities”.)
Sawada is analogous art to the present invention because it introduces a top-k grouping loss that rewrites categorical cross-entropy so the loss is computed only over the k highest-probability classes, thereby tailoring training to top-k performance. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Nguyen + Moazzami + Wang (which already generate and confirm a fused class-probability vector for retraining) with Sawada (which shows how to modify cross-entropy into a top-k form) so that the retraining step uses a top-k cross-entropy objective that sums the log-probabilities of the k strongest classes, matching the performance metric and further mitigating noise. One of ordinary skill in the art would have been motivated to make such a combination because Sawada states: “The extended loss, which we call top-k grouping loss Lₖ, is defined as CE loss calculated for the top-k union and the remaining classes as below, see Eq. (5)” (Sawada, Section III-C, p.3).
Claim 13 is a computer-program-product claim that corresponds directly to method claim 6 and is rejected for the same reasons given in the rejection of claim 6.
Claim 20 is a system claim that corresponds directly to method claim 6 and is rejected for the same reasons given in the rejection of claim 6.
Response to Arguments
Applicant’s arguments, see in pages 7-22, filed 4/27/2026, with respect to claims 1-20 have been fully considered but are not persuasive.
Regarding claims 1-2, 7-9 and 14-16 , the Applicant states that “Nguyen, Moazzami and Wang, taken singly or in combination, do not teach or suggest "generating weight factors used to weight said predictions of class probabilities of said classification model and said second model by an artificial neural network" as recited in claim 1 and similarly in claims 8 and 15.”. The examiner disagrees, since Moazzami discloses using goodness factors to weigh output or potential labels for the input of the various models (61, 60)-- weight said predictions of class probabilities of said classification model and said second model by an artificial neural network.
Additionally, the Applicant argues that “Nguyen, Moazzami and Wang, taken singly or in combination, do not teach or suggest "obtaining a prediction of class probabilities using said weighted predictions by said artificial neural network"”. The examiner disagrees, Moazzami teaches summing the goodness factors weigh vectors G to generate probabilities of the output for each of the models (78).
“Applicant further asserts that Nguyen, Moazzami and Wang, taken singly or in combination, do not teach or suggest "combining said predictions of class probabilities of said artificial neural network and said classification model to train said machine learning model"…”. The examiner disagrees with this conclusion, because Moazzami shows using fused probabilities to train the AI system (64).
Claims 2, 7, 9, 14, and 16 are not patentable at least for the reasons outlined above.
Claims 3, 10, and 17 are not patentable at least based on their dependency on claim 1, and the reasons shown above.
Claims 5, 12, and 19 are not patentable over the prior art at least based on their dependency on claims 1, and the reasons shown above.
Claims 6, 13, and 20 are not patentable at least based on their dependency on claim 1, and the reasons shown above.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to CESAR PAULA whose telephone number is (571)272-4128. The examiner can normally be reached Monday - Friday, 6.30am- 4:30 pm ET.
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, David Wiley can be reached at (571)272-3923. 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.
/CESAR B PAULA/Supervisory Patent Examiner, Art Unit 2145
Prosecution Insights