DETAILED ACTION
This action is responsive to Applicant’s reply filed September 01, 2026. This action is made final.
Status of the Claims
Claims 1 and 9-11 are amended.
Claim 6 is canceled.
Claim status is currently pending and under examination for claims 1-5 and 7-11 of which independent claims are 1, 10 and 11.
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 Amendment
Applicant’s amendments to the Claims have overcome each objection and 35 USC § 101 rejections previously set forth in the Non-Final Office Action mailed July 14, 2026.
Applicant’s arguments regarding the art rejections are moot in view of the new grounds of rejection necessitated by Applicant’s amendment.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The following are the references relied upon in the rejections below:
Li (US 20210264106 A1)
Wang, Shaolei, et al. "Combining self-training and self-supervised learning for unsupervised disfluency detection." Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020.
Claims 1, 3-5, 7 and 9-11 are rejected under 35 U.S.C. 103 as being unpatentable over Li / Wang.
Regarding Claim 1, Li teaches:
An information processing system comprising ([0003] “a data processing system”):
at least one memory that is configured to store instructions ([0003] “memory stores executable instructions”);
and at least one processor that is configured to execute the instructions to input labeled data and unlabeled data ([0003] processor, executable instructions
[0037] “The teacher models 330 may include multiple teachers which may be trained using different features, datasets (labeled or unlabeled)”);
add a pseudo-label to the unlabeled data, by using a teacher model learned by using the labeled data (The Examiner interprets “pseudo-label” according to its broadest reasonable interpretation (BRI) in view of the Applicant’s specification as encompassing “soft label”. This interpretation is consistent with the illustrative descriptions in the Applicant’s specification at [0024], (see excerpt below).
Applicant’s written description at [0024] “The "pseudo-label" is a pseudo-correct answer label, and is generated by a model that is learned by using the labeled data.”
[0037] “multiple teachers which may be trained using different features, datasets (labeled or unlabeled)”
[0042] “the teacher models may be executed with the training data set 455 to generate the resulting training data set 460. … The resulting training data set 460 may be referred to as a transfer data set which contains soft labeled training data. The soft labeled training data may provide the probability of a class for each label. The probability may be calculated by utilizing the logit function which may also be regularized by adding noise to the function. In another example, the probability distribution may be obtained by using a softmax function”
[0039] “during the training process small probabilities are assigned to incorrect answers. Even when these probabilities are very small, some of them are much larger than others. The relative probabilities of incorrect answers contain important information about the differences among incorrect answers”
A teacher model is trained by using a labeled dataset. The teacher model produces a soft label which provides the probability of a class for each sample in a training data set. Small probabilities of a soft label are used to represent incorrect answers (classes), therefore a soft label contains probabilities that are pseudo-correct answer labels and are generated by a teacher model trained on labeled data (and therefore soft labels are pseudo-labels).);
evaluate the pseudo-label added to the unlabeled data, by using an evaluation model learned by using at least one of the labeled data and the unlabeled data ([0039] “in deep neural networks, the normal training objective is to maximize the average log probability of the correct answers. However, during the training process small probabilities are assigned to incorrect answers. Even when these probabilities are very small, some of them are much larger than others. The relative probabilities of incorrect answers contain important information about the differences among incorrect answers. Accordingly, a set of accurate teacher models may be trained to learn the small probabilities among incorrect labels of a training data set.”
[0037] “the teacher models 330 are deep neural network (DNN) teachers such as traditional semi-supervised neural network (NN) teachers”
A set of accurate teacher models (‘evaluation model’) are trained to learn a soft label (‘pseudo label’) that is comprised of small probabilities of incorrect labels and maximized log probabilities of correct labels (answers). By learning small probabilities of incorrect labels, the models are able to distinguish (learn the differences) between incorrect and correct answers (and therefore evaluating the soft label (pseudo-label) by using an evaluation model).),
and output the pseudo-label that reaches a predetermined evaluation criterion, as an evaluation pseudo-label ([0039] “This learned information may then be used to train a student model that learns from both the original labels and the soft labels generated by the teacher models to improve its accuracy without adding significant model parameters”
In a soft-label (‘pseudo label’), small probabilities are assigned to incorrect answers and probabilities are maximized for correct answers (see [0039]). Therefore, small probabilities are a predetermined evaluation criterion since small probabilities represent wrong answers and maximized probabilities (which are larger than small probabilities) are treated as correct answers. Correct answers reach a predetermined evaluation criterion (they surpass the small probabilities / wrong answers), and correct answers therefore are ‘evaluation pseudo-labels’ used to train a student model (and therefore outputting evaluation pseudo-labels).);
learn a student model, by using the labeled data, and pseudo-labeled data obtained by adding the evaluation pseudo-label to the labeled data ([0043] “Once the soft labeled training data is generated from the teacher models, the resulting training data set 460 can be provided to a student training mechanism 480 to train the student model 485.”
[0039] “train a student model that learns from both the original labels and the soft labels generated by the teacher models to improve its accuracy without adding significant model parameters”
Soft labels (‘pseudo-labels’) are comprised of correct answers (‘evaluation pseudo-labels’), see [0039] and are used along with original labels (‘labeled data’) to train a student model (therefore adding evaluation pseudo-labels to labeled data).);
and output the learned student model ([0018] “knowledge is distilled from the more complex training models to train a smaller and simpler student model that is not only smaller in size and easier to deploy and operate but can also provide more accurate results.”),
wherein the evaluation model is trained using a difference between an output of the teacher model and a label added to the labeled data ([0039] “in deep neural networks, the normal training objective is to maximize the average log probability of the correct answers. However, during the training process small probabilities are assigned to incorrect answers. Even when these probabilities are very small, some of them are much larger than others. The relative probabilities of incorrect answers contain important information about the differences among incorrect answers. Accordingly, a set of accurate teacher models may be trained to learn the small probabilities among incorrect labels of a training data set”
A set of accurate teacher models (‘evaluation model’) are trained to learn small probabilities of incorrect labels and maximize the log probabilities of correct answers. By learning small probabilities of incorrect labels, the models are able to learn the differences between incorrect and correct answers (labels added to labeled data), and therefore the evaluation model is learned by using a difference between an output of a teacher model and a label added to the labeled data.);
However, Li does not teach estimating a difference between a pseudo-label and a true label and evaluating the quality of a pseudo-label, which is taught by Wang:
the data to which the pseudo-label has been added is input to the trained evaluation model to estimate a difference between the pseudo-label and a true label ((P. 1815, Sec. 2.2, ¶4) “we train a sentence grammaticality judgment model to help select sentences with high-quality pseudo labels.”
(P. 1815, Sec. 2.2, ¶6) “The training goal is to detect the type of an input sentence, where the labels right and error means that the sentence is grammatically-correct and grammatically incorrect, respectively.”
(P. 1816, Sec. 2.2, ¶1) “Given a sentence
S
=
{
w
1
,
w
2
,
…
,
w
n
}
and its pseudo labeles
T
=
{
t
1
,
t
2
.
.
.
,
t
n
}
, we get a sub-sentence
S
s
u
b
… by deleting the words with the label D. If the sentence grammaticality judgment model generates right label on
S
s
u
b
, we assume that the pseudo labels T is the same as gold labels and keep (S,D) for student model training.”
A sentence grammaticality judgment model is trained to select sentences (‘data’) with high-quality pseudo-labels. Pseudo labels T are added to a sentence. A sentence is decomposed into a sub-sentence for evaluation by the model. The model uses the sub-sentence as input to generate a label (either “right” or “error”). If the model generates a “right” label then the pseudo labels T are assumed (estimated) to be the same as gold labels (‘true labels’), therefore estimating a difference between pseudo-labels and a true label.);
the quality of the pseudo-label is evaluated based on the estimated difference (Sentences with pseudo labels T determined to be the same as gold labels are considered high-quality pseudo labels, therefore the quality of pseudo labels is evaluated based on the estimated difference (whether the gold labels and pseudo labels T are the same).);
and only a pseudo-label that satisfies the predetermined evaluation criterion is output as an evaluation pseudo-label ((P. 1818, Sec. 4.1) “We conjecture that grammaticality judgment model can help filter out the sentence with false pseudo labels.”
(P. 1815, Sec. 2.1, Last Paragraph) “we use the sentence grammaticality judgment model to help select sentences with high-quality pseudo labels. We then train a student model on the selected pseudo labeled sentences.”
Sentences with pseudo labels that are high-quality are selected to be used to train a student model. Pseudo labels that are false pseudo labels are filtered out. Therefore, a predetermined evaluation criterion is quality, and sentences with high-quality pseudo labels satisfy the evaluation criterion and are used to train a student model (therefore outputting pseudo labels as evaluation pseudo labels, the evaluation pseudo labels being the selected high-quality pseudo labels).).
Before the effective filing date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify the training method of Li with the pseudo label filtering technique disclosed by Wang to filter out data samples with low-quality pseudo labels. By filtering out data samples with low-quality pseudo labels, only data samples with high-quality pseudo labels are retained and used for downstream training tasks, thereby increasing model accuracy and mitigating model bias.
Regarding Claim 3, the combined training method of Li / Wang teaches:
The information processing system according to claim 1, wherein the evaluation model is learned by using only the unlabeled data (A set of accurate teacher models (‘evaluation model’) are trained to learn small probabilities of incorrect labels and maximize the log probabilities of correct labels (see Li [0039]).
Li [0037] “The teacher models 330 may include multiple teachers which may be trained using different features, datasets (labeled or unlabeled) or hyperparameters.”).
Regarding Claim 4, the combined training method of Li / Wang teaches:
The information processing system according to claim 1, wherein the evaluation model is learned by using a part of the labeled data and is then learned by using the unlabeled data (A set of accurate teacher models (‘evaluation model’) are trained to learn small probabilities of incorrect labels and maximize the log probabilities of correct labels (see Li [0039]).
Li [0041] “a training mechanism 430 for training each of the teacher models which may then utilize the information to train each teacher model. In one example, this may involve training several semi-supervised DNN models”).
Regarding Claim 5, the combined training method of Li / Wang teaches:
The information processing system according to claim 1, wherein the evaluation model is learned by using only the labeled data (A set of accurate teacher models (‘evaluation model’) are trained to learn small probabilities of incorrect labels and maximize the log probabilities of correct labels (see Li [0039]).
Li [0037] “The teacher models 330 may include multiple teachers which may be trained using different features, datasets (labeled or unlabeled) or hyperparameters.”).
Regarding Claim 7, the combined training method of Li / Wang teaches:
The information processing system according to claim 1, wherein the at least one processor is configured to execute the instructions to evaluate the pseudo-label by using a plurality of evaluation models that are separately learned (Li [0039] “a set of accurate teacher models may be trained to learn the small probabilities among incorrect labels of a training data set”
Li [0037] “The teacher models 330 may include multiple teachers which may be trained using different features, datasets (labeled or unlabeled) or hyperparameters.”
A set of accurate teacher models (plurality of evaluation models) are trained using different features, datasets, and hyperparameters, therefore the set of accurate teacher models are separately learned and are used to evaluate pseudo-labels (soft labels).).
Regarding Claim 9, the combined training method of Li / Wang teaches:
The information processing system according to claim 1, wherein the at least one processor is configured to execute the instructions to learn a part of layers of the learned student model, by using the labeled data (Li [0038] “train a light-weight student model 350. In one implementation, the student model 350 may be a shallow neural network model (e.g., having one or two layers). Thus, the student model is configured to distill knowledge from the deep neural network models and the pretrained NLP model to provide a shallow neural network student model at minimal loss of accuracy.”
Li [0043] “the student model may be a semi-supervised DNN model with a modified training objective. The modified training objective may be achieved by using the weighted sum of the original label (training data set 455) and the soft target label (the resulting training data set 460). … By utilizing the hyper parameters 465, model type 470, target weight 475, original training data set 455 and the resulting training data set 460, the student model 485 may be trained to provide more accurate results while having a smaller structure.”).
Regarding Claim 10, the rejection of Claim 1 is incorporated. The difference in scope being:
An information processing method comprising (Li [0012] “an example method for training a student model”).
Regarding Claim 11, the rejection of Claim 1 is incorporated. The difference in scope being:
A non-transitory recording medium on which a computer program that allows a computer to execute an information processing method is recorded, the information processing method including (Li [0005] “a non-transitory computer readable medium on which are stored instructions that when executed cause a programmable device to train a first ML model”).
The following are the references relied upon in the rejections below:
Bhatt (US 20160253597 A1)
Claim 2 is rejected under 35 U.S.C. 103 as being unpatentable over Li / Wang / Bhatt.
Regarding Claim 2, the combined training method of Li / Wang teaches: The information processing system according to claim 1, wherein the at least one processor is configured to execute the instructions to, however the combination does not teach converting unlabeled and labeled data into domains that are common to each other, which is taught by Bhatt:
convert the unlabeled data and the labeled data, into domains that are common to each other ([Abstract] “An adaptation method includes using a first classifier trained on projected representations of labeled objects from a first domain to predict pseudo-labels for unlabeled objects in a second domain, based on their projected representations.”
[0015] “an adaptation method includes learning a transformation based on features extracted from objects in first and second domains. A similarity is computed between the first and second domains. Original representations of labeled objects in the first domain and unlabeled objects in the second domain are projected with the learned projection. A first classifier is trained on the projected representations of the objects from the first domain and respective labels. Pseudo-labels for the projected representations of the unlabeled objects are predicted with the first classifier.”
[0029] “the exemplary method facilitates this adaptation in a content-aware manner by seamlessly unifying the similarity between the two domains in the adaptation setting. This is also useful in practical scenarios where there are multiple candidate source domains to learn from and method is able to identify the best source domain from which to learn.”
[0030] “efficiently adapt classifier models trained on one domain to perform well for classification on different domains, without requiring any labeled data from the target domain”).
Before the effective filing date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify the combined training method of Li / Wang with the domain adaptation technique disclosed by Bhatt to adapt labeled and unlabeled data into a common domain. By adapting labeled and unlabeled data into a common domain, a classifier trained on a source domain can be used to generate pseudo-labels for the unlabeled data, thereby saving time and computer resources since a separate model does not need to be trained to predict labels for unlabeled data from a different domain.
The following are the references relied upon in the rejections below:
Sharifi (US 20230036764 A1)
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over Li / Wang / Sharifi.
Regarding Claim 8, the combined training method of Li / Wang teaches:
The information processing system according to claim 1, wherein the at least one processor is configured to execute the instructions to learn the teacher model by using the labeled data (Li [0037] “multiple teachers which may be trained using different features, datasets (labeled or unlabeled)”),
However, the combination does not teach re-learning a teacher model by using an evaluation result of an evaluation model, which is taught by Sharifi:
and re-learn the teacher model by using an evaluation result of the evaluation model ([0067] “Further, the teacher output data 208, the student output data 212, and/or the ground truth training data 210 can include data extracted from hidden layers of the teacher machine-learned model 202 and/or student machine-learned model 204. Parameters of the teacher machine-learned model 202 and/or student machine-learned model 204 can be adjusted based on comparisons of the teacher output data 208, student output data, and/or ground truth training data 210”
[0054] “train a teacher model 140 based on a set of training data 162. The training data 162 can include, for example, labeled and/or unlabeled training examples.”).
Before the effective filing date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify the combined training method of Li / Wang with the teacher training technique disclosed by Sharifi to re-train a teacher model based on comparing ground truths and model-generated outputs. By re-training a teacher model based on comparing ground truths and model-generated outputs, a teacher model can be updated to generate outputs that match ground truths, thereby improving the accuracy of both a teacher and student model.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Velizhev et al. (US 20210326749 A1) teaches a data evaluator that predicts the quality of a predicted annotation for a sample by comparing the predicted annotation with ground truth. Predicted quality is used to determine which samples are added to a set of training data.
Shan et al. (US 20220172269 A1) teaches calculating the confidence of a pseudo label and comparing the calculated confidence to a confidence threshold to determine if the pseudo label is a true tag.
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 PEDRO J MORALES whose telephone number is (571)272-6106. The examiner can normally be reached 8:30 AM - 6:00 PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, MIRANDA M HUANG can be reached at (571)270-7092. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/PEDRO J MORALES/Examiner, Art Unit 2124
/MIRANDA M HUANG/Supervisory Patent Examiner, Art Unit 2124