DETAILED ACTION
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 2, 10-13, and 21 are rejected under 35 U.S.C. 103 as being unpatentable over Olabiyi et al. (Pub. No. US 2020/0265296) in view of Yao et al. (Pub. No. US 2021/0133518).
Regarding claims 2, 12, and 13, Olabiyi teaches a computer-implemented method comprising: receiving (selecting) training data (training set data) for training a machine learning model to map model inputs to model outputs in order to perform a particular machine learning task [Para. 10 “More particularly, some aspects described herein may provide a computer-implemented method for training a model having a deep neural network architecture and a plurality of model parameters”; Para. 52 “the system may select an initial mini-batch D.sub.0 from the training set data D”; Para. 32 “Minimizing the objective of equation (1) using SGD in equation (2) without regularization may increase the conditional probability of the target outputs, p.sub.θ(y|x), while decreasing the conditional probability of alternative incorrect outputs.”; Para. 27 “Artificial neural networks may have many applications, including object classification, image recognition, speech recognition, natural language processing, text recognition, regression analysis, behavior modeling, and others.”], the training data comprising: a plurality of labeled training inputs and, for each labeled training input, a ground truth output (ground truth) that should be generated by the machine learning model by performing the particular machine learning task on the labeled training input [Para. 52 “the system may select an initial mini-batch D.sub.0 from the training set data D”; Para. 10 “A training set comprising a plurality of examples may be determined and used to train the model.”; Para. 31 “ground truth”]; and training the machine learning model on the training data (training set data), comprising: training the machine learning model on the labeled training inputs to optimize a supervised objective (DeepBoost Objective) that measures a difference between (i) a model output generated by the machine learning model for a given labeled training input and (ii) the ground truth output (ground truth) for the given labeled training input, [Para. 52 “the system may select an initial mini-batch D.sub.0 from the training set data D”; Para. 53 “The system may train the model, based on the training set, to refine the model parameters through a plurality of first iterations using a first loss function and a plurality of second iterations using a second, weighted loss function, as described further herein”], wherein the model output comprises a probability for the ground truth output (ground truth) [Para. 32 “Minimizing the objective of equation (1) using SGD in equation (2) without regularization may increase the conditional probability of the target outputs, p.sub.θ(y|x), while decreasing the conditional probability of alternative incorrect outputs”. Para. 31 “ground truth”] and wherein the supervised objective (DeepBoost objective) is: based on a negative log likelihood of the given ground truth output when the probability assigned to the given ground truth output by the model output is less than a confidence threshold [Para. 54 “The second loss function/objective may be the DeepBoost objective set forth above in equation (4), and/or any of the functions derived therefrom above.”; Para. 33 and 39]; and
However, Olabiyi doesn’t explicitly teach equal to zero when the probability assigned to the given ground truth output by the model output is equal to or greater than the confidence threshold.
Yao teaches equal to zero when the probability assigned to the given ground truth output by the model output is equal to or greater than the confidence threshold [Para. 21 and 22].
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Olabiyi’s DeepBoost loss function by applying Yao’s loss threshold sample selection and setting the loss threshold. This modification improves Olabiyi by concentrating back propagation on high-loss labeled inputs, and excluding low loss, high confidence inputs, thereby reducing unnecessary gradient computation while preserving training signal for harder examples.
Regarding claims 10 and 21, Olabiyi teaches wherein the machine learning model (deep neural network model) is trained to map model inputs comprising visual data (image recognition) to model outputs to perform a computer vision task (object classification) [Para. 27].
Regarding claim 11, Olabiye teaches wherein the machine learning model (deep neural network model) is trained to map inputs comprising audio data or text data to an output [Para. 27].
Claims 3-7, 9, 14-18, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Olabiyi et al. (Pub. No. US 2020/0265296) in view of Yao et al. (Pub. No. US 2021/0133518) further view Laine et al. (Pub. No. US 2018/0101768).
Regarding claims 3 and 14, Olabiyi teaches training the machine learning model comprises using a confidence threshold [Para. 39].
However, Olabiyi in view of Yao doesn’t explicitly teach the rest of claim limitations.
Laine teaches wherein training the machine learning model comprises: increasing the confidence threshold as training progresses [Para. 65 “The weighted sum block 650 may implement an unsupervised loss weighting function that ramps up, starting from zero, along a Gaussian curve during the first P training epochs (e.g., 80 training epochs)”].
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Olabiyi in view of Yao’s confidence threshold (confidence threshold Pth) by using Laine’s training progress ramp-up schedule to increase the confidence threshold as training progresses. This modification reducing early overfitting while allowing more selective training later.
Regarding claims 4 and 15, Olabiyi teaches training the machine learning model comprises using a confidence threshold [Para. 39].
However, Olabiyi in view of Yao doesn’t explicitly teach the rest of claim limitations.
Laine wherein increasing the confidence threshold comprises increasing the confidence threshold after each training step [Para. 72 and table].
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Olabiyi in view of Yao’s confidence threshold (confidence threshold Pth) by using Laine’s training progress ramp-up schedule to increase the confidence threshold as training progresses. This modification reducing early overfitting while allowing more selective training later.
Regarding claims 5 and 16, Olabiyi teaches training the machine learning model comprises using a confidence threshold [Para. 39].
However, Olabiyi in view of Yao doesn’t explicitly teach the rest of claim limitations.
Laine teaches wherein the training data further comprises a plurality of unlabeled training inputs (input vectors without a corresponding label vector), and wherein training the machine learning model on the training data comprises: training the machine learning model on the unlabeled training inputs to optimize an unsupervised objective (unsupervised component) [para. 18 “some input vectors in the plurality of input vectors may not be associated with a label vector in the plurality of label vectors.”, 23].
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Olabiyi in view of Yao’s confidence threshold (confidence threshold Pth) by using Laine’s training progress ramp-up schedule to increase the confidence threshold as training progresses. This modification reducing early overfitting while allowing more selective training later.
Regarding claims 6 and 17, Olabiyi teaches training the machine learning model comprises using a confidence threshold [Para. 39].
However, Olabiyi in view of Yao doesn’t explicitly teach the rest of claim limitations.
Laine teaches unsupervised objective measures a difference between (i) a model output generated by the machine learning model for a given unlabeled training input and (ii) a model output generated by the machine learning model for a respective augmented training input generated from the unlabeled training input [Para. 62 “The unsupervised component, evaluated for all inputs, penalizes different predictions for the same training input x.sub.i by taking the mean square difference between prediction vectors z.sub.i and {tilde over (z)}.sub.i. To combine the supervised (cross-entropy loss) and unsupervised (squared-difference loss) loss terms, the unsupervised loss term is scaled by a time-dependent weighting function w(t).”. Para. 64 “The first network evaluation generates the first prediction vector z.sub.i using the first noisy input, and the second network evaluation generates the second prediction vector {tilde over (z)}.sub.i using the second noisy input.”.
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Olabiyi’s, modified by Yao, confidence threshold (confidence threshold Pth) by using Laine’s training progress ramp-up schedule to increase the confidence threshold as training progresses. This modification reducing early overfitting while allowing more selective training later.
Regarding claims 7 and 18, Olabiyi teaches training the machine learning model comprises using a confidence threshold [Para. 39].
However, Olabiyi in view of Yao doesn’t explicitly teach the rest of claim limitations.
Laine teaches generating, for each of the plurality of unlabeled training inputs, the respective augmented training input (noisy input) by applying a data augmentation technique (stochastic augmentation layer) to the unlabeled training input [Para. 63 “the Π-model structure processes the training input x.sub.i by a stochastic augmentation block 610 that introduces noise into the input vector for each of the two evaluations.” and “In the case of an image, for example, the stochastic augmentation layer 610 may change the color of a percentage of pixels in the image, may add random noise by adjusting color values up or down for each pixel within some limits, may perform affine transformations of the image (i.e., translations, rotations, etc.), or may warp the image, etc.”].
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Olabiyi’s, modified by Yao, confidence threshold (confidence threshold Pth) by using Laine’s training progress ramp-up schedule to increase the confidence threshold as training progresses. This modification reducing early overfitting while allowing more selective training later.
Regarding claims 9 and 20, Olabiyi teaches training the machine learning model comprises using a confidence threshold [Para. 39].
However, Olabiyi in view of Yao doesn’t explicitly teach the rest of claim limitations.
Laine teaches wherein the labeled and unlabeled training inputs have been augmented by applying a different data augmentation technique from the data augmentation technique used to generate the augmented unlabeled training inputs [Para. 39 and 40].
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Olabiyi’s, modified by Yao, training data pipeline by applying one Laine stochastic augmentation to labeled and unlabeled inputs. This medication improves Olabiyi by increasing diversity of perturbations and improving the model to different input transformations.
Claims 8 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Olabiyi et al. (Pub. No. US 2020/0265296) in view of Yao et al. (Pub. No. US 2021/0133518) further view Laine et al. (Pub. No. US 2018/0101768), and further in view of Aslan et al. (Pub. No. 2017/0132528).
Regarding claims 8 and 19, Olabiyi teaches training the machine learning model comprises using a confidence threshold [Para. 39].
However, Olabiyi in view of Yao further in view of Laine doesn’t explicitly teach the rest of claim limitations.
Aslan teaches wherein the model outputs are probability distributions and wherein the unsupervised objective (unsupervised component) is based on a K-L divergence (Kullback-leibler divergence (“KL divergence”)) between (i) the model output generated by the machine learning model for the given unlabeled training input and (ii) the model output generated by the machine learning model for the augmented training input generated from the unlabeled training input [Para. 37, fig. 2 and related description].
It would have been obvious to one of ordinary skill in the art before the effective filing date to modify Olabiyi’s, modified by Yao and Laine, unsupervised prediction consistency objective in Olabiyi’s training framework by using Aslan’s K-L divergence as the distance between probability distribution model output. This medication improves Olabiyi by using an established probability.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SOLOMON G BEZUAYEHU whose telephone number is (571)270-7452. The examiner can normally be reached on Monday-Friday 10 AM-7 PM.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, O’Neal Mistry can be reached on 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 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 800-786-0101 (IN USA OR CANADA) or 571-272-1000.
/SOLOMON G BEZUAYEHU/ Primary Examiner, Art Unit 2666
Prosecution Insights